linux/fs/xfs/xfs_trans_ail.c

936 lines
24 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
* Copyright (c) 2008 Dave Chinner
* All Rights Reserved.
*/
#include "xfs.h"
#include "xfs_fs.h"
#include "xfs_shared.h"
#include "xfs_format.h"
#include "xfs_log_format.h"
#include "xfs_trans_resv.h"
#include "xfs_mount.h"
#include "xfs_trans.h"
#include "xfs_trans_priv.h"
#include "xfs_trace.h"
#include "xfs_errortag.h"
#include "xfs_error.h"
#include "xfs_log.h"
xfs: AIL needs asynchronous CIL forcing The AIL pushing is stalling on log forces when it comes across pinned items. This is happening on removal workloads where the AIL is dominated by stale items that are removed from AIL when the checkpoint that marks the items stale is committed to the journal. This results is relatively few items in the AIL, but those that are are often pinned as directories items are being removed from are still being logged. As a result, many push cycles through the CIL will first issue a blocking log force to unpin the items. This can take some time to complete, with tracing regularly showing push delays of half a second and sometimes up into the range of several seconds. Sequences like this aren't uncommon: .... 399.829437: xfsaild: last lsn 0x11002dd000 count 101 stuck 101 flushing 0 tout 20 <wanted 20ms, got 270ms delay> 400.099622: xfsaild: target 0x11002f3600, prev 0x11002f3600, last lsn 0x0 400.099623: xfsaild: first lsn 0x11002f3600 400.099679: xfsaild: last lsn 0x1100305000 count 16 stuck 11 flushing 0 tout 50 <wanted 50ms, got 500ms delay> 400.589348: xfsaild: target 0x110032e600, prev 0x11002f3600, last lsn 0x0 400.589349: xfsaild: first lsn 0x1100305000 400.589595: xfsaild: last lsn 0x110032e600 count 156 stuck 101 flushing 30 tout 50 <wanted 50ms, got 460ms delay> 400.950341: xfsaild: target 0x1100353000, prev 0x110032e600, last lsn 0x0 400.950343: xfsaild: first lsn 0x1100317c00 400.950436: xfsaild: last lsn 0x110033d200 count 105 stuck 101 flushing 0 tout 20 <wanted 20ms, got 200ms delay> 401.142333: xfsaild: target 0x1100361600, prev 0x1100353000, last lsn 0x0 401.142334: xfsaild: first lsn 0x110032e600 401.142535: xfsaild: last lsn 0x1100353000 count 122 stuck 101 flushing 8 tout 10 <wanted 10ms, got 10ms delay> 401.154323: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x1100353000 401.154328: xfsaild: first lsn 0x1100353000 401.154389: xfsaild: last lsn 0x1100353000 count 101 stuck 101 flushing 0 tout 20 <wanted 20ms, got 300ms delay> 401.451525: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x0 401.451526: xfsaild: first lsn 0x1100353000 401.451804: xfsaild: last lsn 0x1100377200 count 170 stuck 22 flushing 122 tout 50 <wanted 50ms, got 500ms delay> 401.933581: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x0 .... In each of these cases, every AIL pass saw 101 log items stuck on the AIL (pinned) with very few other items being found. Each pass, a log force was issued, and delay between last/first is the sleep time + the sync log force time. Some of these 101 items pinned the tail of the log. The tail of the log does slowly creep forward (first lsn), but the problem is that the log is actually out of reservation space because it's been running so many transactions that stale items that never reach the AIL but consume log space. Hence we have a largely empty AIL, with long term pins on items that pin the tail of the log that don't get pushed frequently enough to keep log space available. The problem is the hundreds of milliseconds that we block in the log force pushing the CIL out to disk. The AIL should not be stalled like this - it needs to run and flush items that are at the tail of the log with minimal latency. What we really need to do is trigger a log flush, but then not wait for it at all - we've already done our waiting for stuff to complete when we backed off prior to the log force being issued. Even if we remove the XFS_LOG_SYNC from the xfs_log_force() call, we still do a blocking flush of the CIL and that is what is causing the issue. Hence we need a new interface for the CIL to trigger an immediate background push of the CIL to get it moving faster but not to wait on that to occur. While the CIL is pushing, the AIL can also be pushing. We already have an internal interface to do this - xlog_cil_push_now() - but we need a wrapper for it to be used externally. xlog_cil_force_seq() can easily be extended to do what we need as it already implements the synchronous CIL push via xlog_cil_push_now(). Add the necessary flags and "push current sequence" semantics to xlog_cil_force_seq() and convert the AIL pushing to use it. One of the complexities here is that the CIL push does not guarantee that the commit record for the CIL checkpoint is written to disk. The current log force ensures this by submitting the current ACTIVE iclog that the commit record was written to. We need the CIL to actually write this commit record to disk for an async push to ensure that the checkpoint actually makes it to disk and unpins the pinned items in the checkpoint on completion. Hence we need to pass down to the CIL push that we are doing an async flush so that it can switch out the commit_iclog if necessary to get written to disk when the commit iclog is finally released. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-by: Allison Henderson <allison.henderson@oracle.com> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
2021-08-11 09:00:44 +08:00
#include "xfs_log_priv.h"
#ifdef DEBUG
/*
* Check that the list is sorted as it should be.
*
* Called with the ail lock held, but we don't want to assert fail with it
* held otherwise we'll lock everything up and won't be able to debug the
* cause. Hence we sample and check the state under the AIL lock and return if
* everything is fine, otherwise we drop the lock and run the ASSERT checks.
* Asserts may not be fatal, so pick the lock back up and continue onwards.
*/
STATIC void
xfs_ail_check(
struct xfs_ail *ailp,
struct xfs_log_item *lip)
__must_hold(&ailp->ail_lock)
{
struct xfs_log_item *prev_lip;
struct xfs_log_item *next_lip;
xfs_lsn_t prev_lsn = NULLCOMMITLSN;
xfs_lsn_t next_lsn = NULLCOMMITLSN;
xfs_lsn_t lsn;
bool in_ail;
if (list_empty(&ailp->ail_head))
return;
/*
* Sample then check the next and previous entries are valid.
*/
in_ail = test_bit(XFS_LI_IN_AIL, &lip->li_flags);
prev_lip = list_entry(lip->li_ail.prev, struct xfs_log_item, li_ail);
if (&prev_lip->li_ail != &ailp->ail_head)
prev_lsn = prev_lip->li_lsn;
next_lip = list_entry(lip->li_ail.next, struct xfs_log_item, li_ail);
if (&next_lip->li_ail != &ailp->ail_head)
next_lsn = next_lip->li_lsn;
lsn = lip->li_lsn;
if (in_ail &&
(prev_lsn == NULLCOMMITLSN || XFS_LSN_CMP(prev_lsn, lsn) <= 0) &&
(next_lsn == NULLCOMMITLSN || XFS_LSN_CMP(next_lsn, lsn) >= 0))
return;
spin_unlock(&ailp->ail_lock);
ASSERT(in_ail);
ASSERT(prev_lsn == NULLCOMMITLSN || XFS_LSN_CMP(prev_lsn, lsn) <= 0);
ASSERT(next_lsn == NULLCOMMITLSN || XFS_LSN_CMP(next_lsn, lsn) >= 0);
spin_lock(&ailp->ail_lock);
}
#else /* !DEBUG */
#define xfs_ail_check(a,l)
#endif /* DEBUG */
/*
* Return a pointer to the last item in the AIL. If the AIL is empty, then
* return NULL.
*/
static struct xfs_log_item *
xfs_ail_max(
struct xfs_ail *ailp)
{
if (list_empty(&ailp->ail_head))
return NULL;
return list_entry(ailp->ail_head.prev, struct xfs_log_item, li_ail);
}
/*
* Return a pointer to the item which follows the given item in the AIL. If
* the given item is the last item in the list, then return NULL.
*/
static struct xfs_log_item *
xfs_ail_next(
struct xfs_ail *ailp,
struct xfs_log_item *lip)
{
if (lip->li_ail.next == &ailp->ail_head)
return NULL;
return list_first_entry(&lip->li_ail, struct xfs_log_item, li_ail);
}
/*
* This is called by the log manager code to determine the LSN of the tail of
* the log. This is exactly the LSN of the first item in the AIL. If the AIL
* is empty, then this function returns 0.
*
* We need the AIL lock in order to get a coherent read of the lsn of the last
* item in the AIL.
*/
static xfs_lsn_t
__xfs_ail_min_lsn(
struct xfs_ail *ailp)
{
struct xfs_log_item *lip = xfs_ail_min(ailp);
if (lip)
return lip->li_lsn;
return 0;
}
xfs_lsn_t
xfs_ail_min_lsn(
struct xfs_ail *ailp)
{
xfs_lsn_t lsn;
spin_lock(&ailp->ail_lock);
lsn = __xfs_ail_min_lsn(ailp);
spin_unlock(&ailp->ail_lock);
return lsn;
}
/*
* Return the maximum lsn held in the AIL, or zero if the AIL is empty.
*/
static xfs_lsn_t
xfs_ail_max_lsn(
struct xfs_ail *ailp)
{
xfs_lsn_t lsn = 0;
struct xfs_log_item *lip;
spin_lock(&ailp->ail_lock);
lip = xfs_ail_max(ailp);
if (lip)
lsn = lip->li_lsn;
spin_unlock(&ailp->ail_lock);
return lsn;
}
/*
* The cursor keeps track of where our current traversal is up to by tracking
* the next item in the list for us. However, for this to be safe, removing an
* object from the AIL needs to invalidate any cursor that points to it. hence
* the traversal cursor needs to be linked to the struct xfs_ail so that
* deletion can search all the active cursors for invalidation.
*/
STATIC void
xfs_trans_ail_cursor_init(
struct xfs_ail *ailp,
struct xfs_ail_cursor *cur)
{
cur->item = NULL;
list_add_tail(&cur->list, &ailp->ail_cursors);
}
/*
* Get the next item in the traversal and advance the cursor. If the cursor
* was invalidated (indicated by a lip of 1), restart the traversal.
*/
struct xfs_log_item *
xfs_trans_ail_cursor_next(
struct xfs_ail *ailp,
struct xfs_ail_cursor *cur)
{
struct xfs_log_item *lip = cur->item;
if ((uintptr_t)lip & 1)
lip = xfs_ail_min(ailp);
if (lip)
cur->item = xfs_ail_next(ailp, lip);
return lip;
}
/*
* When the traversal is complete, we need to remove the cursor from the list
* of traversing cursors.
*/
void
xfs_trans_ail_cursor_done(
struct xfs_ail_cursor *cur)
{
cur->item = NULL;
list_del_init(&cur->list);
}
/*
* Invalidate any cursor that is pointing to this item. This is called when an
* item is removed from the AIL. Any cursor pointing to this object is now
* invalid and the traversal needs to be terminated so it doesn't reference a
* freed object. We set the low bit of the cursor item pointer so we can
* distinguish between an invalidation and the end of the list when getting the
* next item from the cursor.
*/
STATIC void
xfs_trans_ail_cursor_clear(
struct xfs_ail *ailp,
struct xfs_log_item *lip)
{
struct xfs_ail_cursor *cur;
list_for_each_entry(cur, &ailp->ail_cursors, list) {
if (cur->item == lip)
cur->item = (struct xfs_log_item *)
((uintptr_t)cur->item | 1);
}
}
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
/*
* Find the first item in the AIL with the given @lsn by searching in ascending
* LSN order and initialise the cursor to point to the next item for a
* ascending traversal. Pass a @lsn of zero to initialise the cursor to the
* first item in the AIL. Returns NULL if the list is empty.
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
*/
struct xfs_log_item *
xfs_trans_ail_cursor_first(
struct xfs_ail *ailp,
struct xfs_ail_cursor *cur,
xfs_lsn_t lsn)
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
{
struct xfs_log_item *lip;
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
xfs_trans_ail_cursor_init(ailp, cur);
if (lsn == 0) {
lip = xfs_ail_min(ailp);
goto out;
}
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
list_for_each_entry(lip, &ailp->ail_head, li_ail) {
if (XFS_LSN_CMP(lip->li_lsn, lsn) >= 0)
goto out;
}
return NULL;
out:
if (lip)
cur->item = xfs_ail_next(ailp, lip);
return lip;
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
}
static struct xfs_log_item *
__xfs_trans_ail_cursor_last(
struct xfs_ail *ailp,
xfs_lsn_t lsn)
{
struct xfs_log_item *lip;
list_for_each_entry_reverse(lip, &ailp->ail_head, li_ail) {
if (XFS_LSN_CMP(lip->li_lsn, lsn) <= 0)
return lip;
}
return NULL;
}
/*
* Find the last item in the AIL with the given @lsn by searching in descending
* LSN order and initialise the cursor to point to that item. If there is no
* item with the value of @lsn, then it sets the cursor to the last item with an
* LSN lower than @lsn. Returns NULL if the list is empty.
*/
struct xfs_log_item *
xfs_trans_ail_cursor_last(
struct xfs_ail *ailp,
struct xfs_ail_cursor *cur,
xfs_lsn_t lsn)
{
xfs_trans_ail_cursor_init(ailp, cur);
cur->item = __xfs_trans_ail_cursor_last(ailp, lsn);
return cur->item;
}
/*
* Splice the log item list into the AIL at the given LSN. We splice to the
* tail of the given LSN to maintain insert order for push traversals. The
* cursor is optional, allowing repeated updates to the same LSN to avoid
* repeated traversals. This should not be called with an empty list.
*/
static void
xfs_ail_splice(
struct xfs_ail *ailp,
struct xfs_ail_cursor *cur,
struct list_head *list,
xfs_lsn_t lsn)
{
struct xfs_log_item *lip;
ASSERT(!list_empty(list));
/*
* Use the cursor to determine the insertion point if one is
* provided. If not, or if the one we got is not valid,
* find the place in the AIL where the items belong.
*/
lip = cur ? cur->item : NULL;
if (!lip || (uintptr_t)lip & 1)
lip = __xfs_trans_ail_cursor_last(ailp, lsn);
/*
* If a cursor is provided, we know we're processing the AIL
* in lsn order, and future items to be spliced in will
* follow the last one being inserted now. Update the
* cursor to point to that last item, now while we have a
* reliable pointer to it.
*/
if (cur)
cur->item = list_entry(list->prev, struct xfs_log_item, li_ail);
/*
* Finally perform the splice. Unless the AIL was empty,
* lip points to the item in the AIL _after_ which the new
* items should go. If lip is null the AIL was empty, so
* the new items go at the head of the AIL.
*/
if (lip)
list_splice(list, &lip->li_ail);
else
list_splice(list, &ailp->ail_head);
}
/*
* Delete the given item from the AIL. Return a pointer to the item.
*/
static void
xfs_ail_delete(
struct xfs_ail *ailp,
struct xfs_log_item *lip)
{
xfs_ail_check(ailp, lip);
list_del(&lip->li_ail);
xfs_trans_ail_cursor_clear(ailp, lip);
}
/*
* Requeue a failed buffer for writeback.
*
* We clear the log item failed state here as well, but we have to be careful
* about reference counts because the only active reference counts on the buffer
* may be the failed log items. Hence if we clear the log item failed state
* before queuing the buffer for IO we can release all active references to
* the buffer and free it, leading to use after free problems in
* xfs_buf_delwri_queue. It makes no difference to the buffer or log items which
* order we process them in - the buffer is locked, and we own the buffer list
* so nothing on them is going to change while we are performing this action.
*
* Hence we can safely queue the buffer for IO before we clear the failed log
* item state, therefore always having an active reference to the buffer and
* avoiding the transient zero-reference state that leads to use-after-free.
*/
static inline int
xfsaild_resubmit_item(
struct xfs_log_item *lip,
struct list_head *buffer_list)
{
struct xfs_buf *bp = lip->li_buf;
if (!xfs_buf_trylock(bp))
return XFS_ITEM_LOCKED;
if (!xfs_buf_delwri_queue(bp, buffer_list)) {
xfs_buf_unlock(bp);
return XFS_ITEM_FLUSHING;
}
/* protected by ail_lock */
xfs: pin inode backing buffer to the inode log item When we dirty an inode, we are going to have to write it disk at some point in the near future. This requires the inode cluster backing buffer to be present in memory. Unfortunately, under severe memory pressure we can reclaim the inode backing buffer while the inode is dirty in memory, resulting in stalling the AIL pushing because it has to do a read-modify-write cycle on the cluster buffer. When we have no memory available, the read of the cluster buffer blocks the AIL pushing process, and this causes all sorts of issues for memory reclaim as it requires inode writeback to make forwards progress. Allocating a cluster buffer causes more memory pressure, and results in more cluster buffers to be reclaimed, resulting in more RMW cycles to be done in the AIL context and everything then backs up on AIL progress. Only the synchronous inode cluster writeback in the the inode reclaim code provides some level of forwards progress guarantees that prevent OOM-killer rampages in this situation. Fix this by pinning the inode backing buffer to the inode log item when the inode is first dirtied (i.e. in xfs_trans_log_inode()). This may mean the first modification of an inode that has been held in cache for a long time may block on a cluster buffer read, but we can do that in transaction context and block safely until the buffer has been allocated and read. Once we have the cluster buffer, the inode log item takes a reference to it, pinning it in memory, and attaches it to the log item for future reference. This means we can always grab the cluster buffer from the inode log item when we need it. When the inode is finally cleaned and removed from the AIL, we can drop the reference the inode log item holds on the cluster buffer. Once all inodes on the cluster buffer are clean, the cluster buffer will be unpinned and it will be available for memory reclaim to reclaim again. This avoids the issues with needing to do RMW cycles in the AIL pushing context, and hence allows complete non-blocking inode flushing to be performed by the AIL pushing context. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
2020-06-30 05:49:15 +08:00
list_for_each_entry(lip, &bp->b_li_list, li_bio_list) {
if (bp->b_flags & _XBF_INODES)
clear_bit(XFS_LI_FAILED, &lip->li_flags);
else
xfs_clear_li_failed(lip);
}
xfs_buf_unlock(bp);
return XFS_ITEM_SUCCESS;
}
static inline uint
xfsaild_push_item(
struct xfs_ail *ailp,
struct xfs_log_item *lip)
{
/*
* If log item pinning is enabled, skip the push and track the item as
* pinned. This can help induce head-behind-tail conditions.
*/
if (XFS_TEST_ERROR(false, ailp->ail_log->l_mp, XFS_ERRTAG_LOG_ITEM_PIN))
return XFS_ITEM_PINNED;
/*
* Consider the item pinned if a push callback is not defined so the
* caller will force the log. This should only happen for intent items
* as they are unpinned once the associated done item is committed to
* the on-disk log.
*/
if (!lip->li_ops->iop_push)
return XFS_ITEM_PINNED;
if (test_bit(XFS_LI_FAILED, &lip->li_flags))
return xfsaild_resubmit_item(lip, &ailp->ail_buf_list);
return lip->li_ops->iop_push(lip, &ailp->ail_buf_list);
}
static long
xfsaild_push(
struct xfs_ail *ailp)
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
{
struct xfs_mount *mp = ailp->ail_log->l_mp;
struct xfs_ail_cursor cur;
struct xfs_log_item *lip;
xfs_lsn_t lsn;
xfs_lsn_t target;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
long tout;
int stuck = 0;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
int flushing = 0;
int count = 0;
/*
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
* If we encountered pinned items or did not finish writing out all
xfs: AIL needs asynchronous CIL forcing The AIL pushing is stalling on log forces when it comes across pinned items. This is happening on removal workloads where the AIL is dominated by stale items that are removed from AIL when the checkpoint that marks the items stale is committed to the journal. This results is relatively few items in the AIL, but those that are are often pinned as directories items are being removed from are still being logged. As a result, many push cycles through the CIL will first issue a blocking log force to unpin the items. This can take some time to complete, with tracing regularly showing push delays of half a second and sometimes up into the range of several seconds. Sequences like this aren't uncommon: .... 399.829437: xfsaild: last lsn 0x11002dd000 count 101 stuck 101 flushing 0 tout 20 <wanted 20ms, got 270ms delay> 400.099622: xfsaild: target 0x11002f3600, prev 0x11002f3600, last lsn 0x0 400.099623: xfsaild: first lsn 0x11002f3600 400.099679: xfsaild: last lsn 0x1100305000 count 16 stuck 11 flushing 0 tout 50 <wanted 50ms, got 500ms delay> 400.589348: xfsaild: target 0x110032e600, prev 0x11002f3600, last lsn 0x0 400.589349: xfsaild: first lsn 0x1100305000 400.589595: xfsaild: last lsn 0x110032e600 count 156 stuck 101 flushing 30 tout 50 <wanted 50ms, got 460ms delay> 400.950341: xfsaild: target 0x1100353000, prev 0x110032e600, last lsn 0x0 400.950343: xfsaild: first lsn 0x1100317c00 400.950436: xfsaild: last lsn 0x110033d200 count 105 stuck 101 flushing 0 tout 20 <wanted 20ms, got 200ms delay> 401.142333: xfsaild: target 0x1100361600, prev 0x1100353000, last lsn 0x0 401.142334: xfsaild: first lsn 0x110032e600 401.142535: xfsaild: last lsn 0x1100353000 count 122 stuck 101 flushing 8 tout 10 <wanted 10ms, got 10ms delay> 401.154323: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x1100353000 401.154328: xfsaild: first lsn 0x1100353000 401.154389: xfsaild: last lsn 0x1100353000 count 101 stuck 101 flushing 0 tout 20 <wanted 20ms, got 300ms delay> 401.451525: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x0 401.451526: xfsaild: first lsn 0x1100353000 401.451804: xfsaild: last lsn 0x1100377200 count 170 stuck 22 flushing 122 tout 50 <wanted 50ms, got 500ms delay> 401.933581: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x0 .... In each of these cases, every AIL pass saw 101 log items stuck on the AIL (pinned) with very few other items being found. Each pass, a log force was issued, and delay between last/first is the sleep time + the sync log force time. Some of these 101 items pinned the tail of the log. The tail of the log does slowly creep forward (first lsn), but the problem is that the log is actually out of reservation space because it's been running so many transactions that stale items that never reach the AIL but consume log space. Hence we have a largely empty AIL, with long term pins on items that pin the tail of the log that don't get pushed frequently enough to keep log space available. The problem is the hundreds of milliseconds that we block in the log force pushing the CIL out to disk. The AIL should not be stalled like this - it needs to run and flush items that are at the tail of the log with minimal latency. What we really need to do is trigger a log flush, but then not wait for it at all - we've already done our waiting for stuff to complete when we backed off prior to the log force being issued. Even if we remove the XFS_LOG_SYNC from the xfs_log_force() call, we still do a blocking flush of the CIL and that is what is causing the issue. Hence we need a new interface for the CIL to trigger an immediate background push of the CIL to get it moving faster but not to wait on that to occur. While the CIL is pushing, the AIL can also be pushing. We already have an internal interface to do this - xlog_cil_push_now() - but we need a wrapper for it to be used externally. xlog_cil_force_seq() can easily be extended to do what we need as it already implements the synchronous CIL push via xlog_cil_push_now(). Add the necessary flags and "push current sequence" semantics to xlog_cil_force_seq() and convert the AIL pushing to use it. One of the complexities here is that the CIL push does not guarantee that the commit record for the CIL checkpoint is written to disk. The current log force ensures this by submitting the current ACTIVE iclog that the commit record was written to. We need the CIL to actually write this commit record to disk for an async push to ensure that the checkpoint actually makes it to disk and unpins the pinned items in the checkpoint on completion. Hence we need to pass down to the CIL push that we are doing an async flush so that it can switch out the commit_iclog if necessary to get written to disk when the commit iclog is finally released. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-by: Allison Henderson <allison.henderson@oracle.com> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
2021-08-11 09:00:44 +08:00
* buffers the last time we ran, force a background CIL push to get the
* items unpinned in the near future. We do not wait on the CIL push as
* that could stall us for seconds if there is enough background IO
* load. Stalling for that long when the tail of the log is pinned and
* needs flushing will hard stop the transaction subsystem when log
* space runs out.
*/
if (ailp->ail_log_flush && ailp->ail_last_pushed_lsn == 0 &&
(!list_empty_careful(&ailp->ail_buf_list) ||
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
xfs_ail_min_lsn(ailp))) {
ailp->ail_log_flush = 0;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
XFS_STATS_INC(mp, xs_push_ail_flush);
xlog_cil_flush(ailp->ail_log);
}
spin_lock(&ailp->ail_lock);
xfs: xfs_ail_push_all_sync() stalls when racing with updates xfs_ail_push_all_sync() has a loop like this: while max_ail_lsn { prepare_to_wait(ail_empty) target = max_ail_lsn wake_up(ail_task); schedule() } Which is designed to sleep until the AIL is emptied. When xfs_ail_update_finish() moves the tail of the log, it does: if (list_empty(&ailp->ail_head)) wake_up_all(&ailp->ail_empty); So it will only wake up the sync push waiter when the AIL goes empty. If, by the time the push waiter has woken, the AIL has more in it, it will reset the target, wake the push task and go back to sleep. The problem here is that if the AIL is having items added to it when xfs_ail_push_all_sync() is called, then they may get inserted into the AIL at a LSN higher than the target LSN. At this point, xfsaild_push() will see that the target is X, the item LSNs are (X+N) and skip over them, hence never pushing the out. The result of this the AIL will not get emptied by the AIL push thread, hence xfs_ail_finish_update() will never see the AIL being empty even if it moves the tail. Hence xfs_ail_push_all_sync() never gets woken and hence cannot update the push target to capture the items beyond the current target on the LSN. This is a TOCTOU type of issue so the way to avoid it is to not use the push target at all for sync pushes. We know that a sync push is being requested by the fact the ail_empty wait queue is active, hence the xfsaild can just set the target to max_ail_lsn on every push that we see the wait queue active. Hence we no longer will leave items on the AIL that are beyond the LSN sampled at the start of a sync push. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Chandan Babu R <chandan.babu@oracle.com> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
2022-03-18 00:09:11 +08:00
/*
* If we have a sync push waiter, we always have to push till the AIL is
* empty. Update the target to point to the end of the AIL so that
* capture updates that occur after the sync push waiter has gone to
* sleep.
*/
if (waitqueue_active(&ailp->ail_empty)) {
lip = xfs_ail_max(ailp);
if (lip)
target = lip->li_lsn;
} else {
/* barrier matches the ail_target update in xfs_ail_push() */
smp_rmb();
target = ailp->ail_target;
ailp->ail_target_prev = target;
}
xfs: drain the buf delwri queue before xfsaild idles xfsaild is racy with respect to transaction abort and shutdown in that the task can idle or exit with an empty AIL but buffers still on the delwri queue. This was partly addressed by cancelling the delwri queue before the task exits to prevent memory leaks, but it's also possible for xfsaild to empty and idle with buffers on the delwri queue. For example, a transaction that pins a buffer that also happens to sit on the AIL delwri queue will explicitly remove the associated log item from the AIL if the transaction aborts. The side effect of this is an unmount hang in xfs_wait_buftarg() as the associated buffers remain held by the delwri queue indefinitely. This is reproduced on repeated runs of generic/531 with an fs format (-mrmapbt=1 -bsize=1k) that happens to also reproduce transaction aborts. Update xfsaild to not idle until both the AIL and associated delwri queue are empty and update the push code to continue delwri queue submission attempts even when the AIL is empty. This allows the AIL to eventually release aborted buffers stranded on the delwri queue when they are unlocked by the associated transaction. This should have no significant effect on normal runtime behavior because the xfsaild currently idles only when the AIL is empty and in practice the AIL is rarely empty with a populated delwri queue. The items must be AIL resident to land in the queue in the first place and generally aren't removed until writeback completes. Note that the pre-existing delwri queue cancel logic in the exit path is retained because task stop is external, could technically come at any point, and xfsaild is still responsible to release its buffer references before it exits. Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
2020-07-16 22:39:29 +08:00
/* we're done if the AIL is empty or our push has reached the end */
lip = xfs_trans_ail_cursor_first(ailp, &cur, ailp->ail_last_pushed_lsn);
xfs: drain the buf delwri queue before xfsaild idles xfsaild is racy with respect to transaction abort and shutdown in that the task can idle or exit with an empty AIL but buffers still on the delwri queue. This was partly addressed by cancelling the delwri queue before the task exits to prevent memory leaks, but it's also possible for xfsaild to empty and idle with buffers on the delwri queue. For example, a transaction that pins a buffer that also happens to sit on the AIL delwri queue will explicitly remove the associated log item from the AIL if the transaction aborts. The side effect of this is an unmount hang in xfs_wait_buftarg() as the associated buffers remain held by the delwri queue indefinitely. This is reproduced on repeated runs of generic/531 with an fs format (-mrmapbt=1 -bsize=1k) that happens to also reproduce transaction aborts. Update xfsaild to not idle until both the AIL and associated delwri queue are empty and update the push code to continue delwri queue submission attempts even when the AIL is empty. This allows the AIL to eventually release aborted buffers stranded on the delwri queue when they are unlocked by the associated transaction. This should have no significant effect on normal runtime behavior because the xfsaild currently idles only when the AIL is empty and in practice the AIL is rarely empty with a populated delwri queue. The items must be AIL resident to land in the queue in the first place and generally aren't removed until writeback completes. Note that the pre-existing delwri queue cancel logic in the exit path is retained because task stop is external, could technically come at any point, and xfsaild is still responsible to release its buffer references before it exits. Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
2020-07-16 22:39:29 +08:00
if (!lip)
goto out_done;
XFS_STATS_INC(mp, xs_push_ail);
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
lsn = lip->li_lsn;
while ((XFS_LSN_CMP(lip->li_lsn, target) <= 0)) {
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
int lock_result;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
/*
* Note that iop_push may unlock and reacquire the AIL lock. We
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
* rely on the AIL cursor implementation to be able to deal with
* the dropped lock.
*/
lock_result = xfsaild_push_item(ailp, lip);
switch (lock_result) {
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
case XFS_ITEM_SUCCESS:
XFS_STATS_INC(mp, xs_push_ail_success);
trace_xfs_ail_push(lip);
ailp->ail_last_pushed_lsn = lsn;
break;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
case XFS_ITEM_FLUSHING:
/*
* The item or its backing buffer is already being
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
* flushed. The typical reason for that is that an
* inode buffer is locked because we already pushed the
* updates to it as part of inode clustering.
*
* We do not want to stop flushing just because lots
* of items are already being flushed, but we need to
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
* re-try the flushing relatively soon if most of the
* AIL is being flushed.
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
*/
XFS_STATS_INC(mp, xs_push_ail_flushing);
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
trace_xfs_ail_flushing(lip);
flushing++;
ailp->ail_last_pushed_lsn = lsn;
break;
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
case XFS_ITEM_PINNED:
XFS_STATS_INC(mp, xs_push_ail_pinned);
trace_xfs_ail_pinned(lip);
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
stuck++;
ailp->ail_log_flush++;
break;
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
case XFS_ITEM_LOCKED:
XFS_STATS_INC(mp, xs_push_ail_locked);
trace_xfs_ail_locked(lip);
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
stuck++;
break;
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
default:
ASSERT(0);
break;
}
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
count++;
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
/*
* Are there too many items we can't do anything with?
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
*
* If we are skipping too many items because we can't flush
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
* them or they are already being flushed, we back off and
* given them time to complete whatever operation is being
* done. i.e. remove pressure from the AIL while we can't make
* progress so traversals don't slow down further inserts and
* removals to/from the AIL.
*
* The value of 100 is an arbitrary magic number based on
* observation.
*/
if (stuck > 100)
break;
lip = xfs_trans_ail_cursor_next(ailp, &cur);
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
if (lip == NULL)
break;
lsn = lip->li_lsn;
}
xfs: drain the buf delwri queue before xfsaild idles xfsaild is racy with respect to transaction abort and shutdown in that the task can idle or exit with an empty AIL but buffers still on the delwri queue. This was partly addressed by cancelling the delwri queue before the task exits to prevent memory leaks, but it's also possible for xfsaild to empty and idle with buffers on the delwri queue. For example, a transaction that pins a buffer that also happens to sit on the AIL delwri queue will explicitly remove the associated log item from the AIL if the transaction aborts. The side effect of this is an unmount hang in xfs_wait_buftarg() as the associated buffers remain held by the delwri queue indefinitely. This is reproduced on repeated runs of generic/531 with an fs format (-mrmapbt=1 -bsize=1k) that happens to also reproduce transaction aborts. Update xfsaild to not idle until both the AIL and associated delwri queue are empty and update the push code to continue delwri queue submission attempts even when the AIL is empty. This allows the AIL to eventually release aborted buffers stranded on the delwri queue when they are unlocked by the associated transaction. This should have no significant effect on normal runtime behavior because the xfsaild currently idles only when the AIL is empty and in practice the AIL is rarely empty with a populated delwri queue. The items must be AIL resident to land in the queue in the first place and generally aren't removed until writeback completes. Note that the pre-existing delwri queue cancel logic in the exit path is retained because task stop is external, could technically come at any point, and xfsaild is still responsible to release its buffer references before it exits. Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
2020-07-16 22:39:29 +08:00
out_done:
xfs_trans_ail_cursor_done(&cur);
spin_unlock(&ailp->ail_lock);
if (xfs_buf_delwri_submit_nowait(&ailp->ail_buf_list))
ailp->ail_log_flush++;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
if (!count || XFS_LSN_CMP(lsn, target) >= 0) {
/*
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
* We reached the target or the AIL is empty, so wait a bit
* longer for I/O to complete and remove pushed items from the
* AIL before we start the next scan from the start of the AIL.
*/
tout = 50;
ailp->ail_last_pushed_lsn = 0;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
} else if (((stuck + flushing) * 100) / count > 90) {
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
/*
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
* Either there is a lot of contention on the AIL or we are
* stuck due to operations in progress. "Stuck" in this case
* is defined as >90% of the items we tried to push were stuck.
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
*
* Backoff a bit more to allow some I/O to complete before
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
* restarting from the start of the AIL. This prevents us from
* spinning on the same items, and if they are pinned will all
* the restart to issue a log force to unpin the stuck items.
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
*/
tout = 20;
ailp->ail_last_pushed_lsn = 0;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
} else {
/*
* Assume we have more work to do in a short while.
*/
tout = 10;
}
return tout;
}
static int
xfsaild(
void *data)
{
struct xfs_ail *ailp = data;
long tout = 0; /* milliseconds */
xfs: clear PF_MEMALLOC before exiting xfsaild thread Leaving PF_MEMALLOC set when exiting a kthread causes it to remain set during do_exit(). That can confuse things. In particular, if BSD process accounting is enabled, then do_exit() writes data to an accounting file. If that file has FS_SYNC_FL set, then this write occurs synchronously and can misbehave if PF_MEMALLOC is set. For example, if the accounting file is located on an XFS filesystem, then a WARN_ON_ONCE() in iomap_do_writepage() is triggered and the data doesn't get written when it should. Or if the accounting file is located on an ext4 filesystem without a journal, then a WARN_ON_ONCE() in ext4_write_inode() is triggered and the inode doesn't get written. Fix this in xfsaild() by using the helper functions to save and restore PF_MEMALLOC. This can be reproduced as follows in the kvm-xfstests test appliance modified to add the 'acct' Debian package, and with kvm-xfstests's recommended kconfig modified to add CONFIG_BSD_PROCESS_ACCT=y: mkfs.xfs -f /dev/vdb mount /vdb touch /vdb/file chattr +S /vdb/file accton /vdb/file mkfs.xfs -f /dev/vdc mount /vdc umount /vdc It causes: WARNING: CPU: 1 PID: 336 at fs/iomap/buffered-io.c:1534 CPU: 1 PID: 336 Comm: xfsaild/vdc Not tainted 5.6.0-rc5 #3 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20191223_100556-anatol 04/01/2014 RIP: 0010:iomap_do_writepage+0x16b/0x1f0 fs/iomap/buffered-io.c:1534 [...] Call Trace: write_cache_pages+0x189/0x4d0 mm/page-writeback.c:2238 iomap_writepages+0x1c/0x33 fs/iomap/buffered-io.c:1642 xfs_vm_writepages+0x65/0x90 fs/xfs/xfs_aops.c:578 do_writepages+0x41/0xe0 mm/page-writeback.c:2344 __filemap_fdatawrite_range+0xd2/0x120 mm/filemap.c:421 file_write_and_wait_range+0x71/0xc0 mm/filemap.c:760 xfs_file_fsync+0x7a/0x2b0 fs/xfs/xfs_file.c:114 generic_write_sync include/linux/fs.h:2867 [inline] xfs_file_buffered_aio_write+0x379/0x3b0 fs/xfs/xfs_file.c:691 call_write_iter include/linux/fs.h:1901 [inline] new_sync_write+0x130/0x1d0 fs/read_write.c:483 __kernel_write+0x54/0xe0 fs/read_write.c:515 do_acct_process+0x122/0x170 kernel/acct.c:522 slow_acct_process kernel/acct.c:581 [inline] acct_process+0x1d4/0x27c kernel/acct.c:607 do_exit+0x83d/0xbc0 kernel/exit.c:791 kthread+0xf1/0x140 kernel/kthread.c:257 ret_from_fork+0x27/0x50 arch/x86/entry/entry_64.S:352 This bug was originally reported by syzbot at https://lore.kernel.org/r/0000000000000e7156059f751d7b@google.com. Reported-by: syzbot+1f9dc49e8de2582d90c2@syzkaller.appspotmail.com Signed-off-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2020-03-10 23:57:27 +08:00
unsigned int noreclaim_flag;
xfs: clear PF_MEMALLOC before exiting xfsaild thread Leaving PF_MEMALLOC set when exiting a kthread causes it to remain set during do_exit(). That can confuse things. In particular, if BSD process accounting is enabled, then do_exit() writes data to an accounting file. If that file has FS_SYNC_FL set, then this write occurs synchronously and can misbehave if PF_MEMALLOC is set. For example, if the accounting file is located on an XFS filesystem, then a WARN_ON_ONCE() in iomap_do_writepage() is triggered and the data doesn't get written when it should. Or if the accounting file is located on an ext4 filesystem without a journal, then a WARN_ON_ONCE() in ext4_write_inode() is triggered and the inode doesn't get written. Fix this in xfsaild() by using the helper functions to save and restore PF_MEMALLOC. This can be reproduced as follows in the kvm-xfstests test appliance modified to add the 'acct' Debian package, and with kvm-xfstests's recommended kconfig modified to add CONFIG_BSD_PROCESS_ACCT=y: mkfs.xfs -f /dev/vdb mount /vdb touch /vdb/file chattr +S /vdb/file accton /vdb/file mkfs.xfs -f /dev/vdc mount /vdc umount /vdc It causes: WARNING: CPU: 1 PID: 336 at fs/iomap/buffered-io.c:1534 CPU: 1 PID: 336 Comm: xfsaild/vdc Not tainted 5.6.0-rc5 #3 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20191223_100556-anatol 04/01/2014 RIP: 0010:iomap_do_writepage+0x16b/0x1f0 fs/iomap/buffered-io.c:1534 [...] Call Trace: write_cache_pages+0x189/0x4d0 mm/page-writeback.c:2238 iomap_writepages+0x1c/0x33 fs/iomap/buffered-io.c:1642 xfs_vm_writepages+0x65/0x90 fs/xfs/xfs_aops.c:578 do_writepages+0x41/0xe0 mm/page-writeback.c:2344 __filemap_fdatawrite_range+0xd2/0x120 mm/filemap.c:421 file_write_and_wait_range+0x71/0xc0 mm/filemap.c:760 xfs_file_fsync+0x7a/0x2b0 fs/xfs/xfs_file.c:114 generic_write_sync include/linux/fs.h:2867 [inline] xfs_file_buffered_aio_write+0x379/0x3b0 fs/xfs/xfs_file.c:691 call_write_iter include/linux/fs.h:1901 [inline] new_sync_write+0x130/0x1d0 fs/read_write.c:483 __kernel_write+0x54/0xe0 fs/read_write.c:515 do_acct_process+0x122/0x170 kernel/acct.c:522 slow_acct_process kernel/acct.c:581 [inline] acct_process+0x1d4/0x27c kernel/acct.c:607 do_exit+0x83d/0xbc0 kernel/exit.c:791 kthread+0xf1/0x140 kernel/kthread.c:257 ret_from_fork+0x27/0x50 arch/x86/entry/entry_64.S:352 This bug was originally reported by syzbot at https://lore.kernel.org/r/0000000000000e7156059f751d7b@google.com. Reported-by: syzbot+1f9dc49e8de2582d90c2@syzkaller.appspotmail.com Signed-off-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2020-03-10 23:57:27 +08:00
noreclaim_flag = memalloc_noreclaim_save();
xfs: Make xfsaild freezeable again Hendik has reported suspend failures due to xfsaild blocking the freezer to settle down. Jan 17 19:59:56 linux-6380 kernel: PM: Syncing filesystems ... done. Jan 17 19:59:56 linux-6380 kernel: PM: Preparing system for sleep (mem) Jan 17 19:59:56 linux-6380 kernel: Freezing user space processes ... (elapsed 0.001 seconds) done. Jan 17 19:59:56 linux-6380 kernel: Freezing remaining freezable tasks ... Jan 17 19:59:56 linux-6380 kernel: Freezing of tasks failed after 20.002 seconds (1 tasks refusing to freeze, wq_busy=0): Jan 17 19:59:56 linux-6380 kernel: xfsaild/dm-5 S 00000000 0 1293 2 0x00000080 Jan 17 19:59:56 linux-6380 kernel: f0ef5f00 00000046 00000200 00000000 ffff9022 c02d3800 00000000 00000032 Jan 17 19:59:56 linux-6380 kernel: ee0b2400 00000032 f71e0d00 f36fabc0 f0ef2d00 f0ef6000 f0ef2d00 f12f90c0 Jan 17 19:59:56 linux-6380 kernel: f0ef5f0c c0844e44 00000000 f0ef5f6c f811e0be 00000000 00000000 f0ef2d00 Jan 17 19:59:56 linux-6380 kernel: Call Trace: Jan 17 19:59:56 linux-6380 kernel: [<c0844e44>] schedule+0x34/0x90 Jan 17 19:59:56 linux-6380 kernel: [<f811e0be>] xfsaild+0x5de/0x600 [xfs] Jan 17 19:59:56 linux-6380 kernel: [<c0286cbb>] kthread+0x9b/0xb0 Jan 17 19:59:56 linux-6380 kernel: [<c0848a79>] ret_from_kernel_thread+0x21/0x38 The issue has been there for quite some time but it has been made visible by only by 24ba16bb3d49 ("xfs: clear PF_NOFREEZE for xfsaild kthread") because the suspend started seeing xfsaild. The above commit has missed that the !xfs_ail_min branch might call schedule with TASK_INTERRUPTIBLE without calling try_to_freeze so the pm suspend would wake up the kernel thread over and over again without any progress. What we want here is to use freezable_schedule instead to hide the thread from the suspend. While we are here also change schedule_timeout to freezable variant to prevent from spurious wakeups by suspend. [dchinner: re-add set_freezeable call so the freezer will account properly for this kthread. ] Reported-by: Hendrik Woltersdorf <hendrikw@arcor.de> Signed-off-by: Michal Hocko <mhocko@suse.com> Reviewed-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-02-08 11:59:07 +08:00
set_freezable();
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 13:58:39 +08:00
while (1) {
if (tout && tout <= 20)
set_current_state(TASK_KILLABLE);
else
set_current_state(TASK_INTERRUPTIBLE);
/*
xfs: clear ail delwri queued bufs on unmount of shutdown fs In the typical unmount case, the AIL is forced out by the unmount sequence before the xfsaild task is stopped. Since AIL items are removed on writeback completion, this means that the AIL ->ail_buf_list delwri queue has been drained. This is not always true in the shutdown case, however. It's possible for buffers to sit on a delwri queue for a period of time across submission attempts if said items are locked or have been relogged and pinned since first added to the queue. If the attempt to log such an item results in a log I/O error, the error processing can shutdown the fs, remove the item from the AIL, stale the buffer (dropping the LRU reference) and clear its delwri queue state. The latter bit means the buffer will be released from a delwri queue on the next submission attempt, but this might never occur if the filesystem has shutdown and the AIL is empty. This means that such buffers are held indefinitely by the AIL delwri queue across destruction of the AIL. Aside from being a memory leak, these buffers can also hold references to in-core perag structures. The latter problem manifests as a generic/475 failure, reproducing the following asserts at unmount time: XFS: Assertion failed: atomic_read(&pag->pag_ref) == 0, file: fs/xfs/xfs_mount.c, line: 151 XFS: Assertion failed: atomic_read(&pag->pag_ref) == 0, file: fs/xfs/xfs_mount.c, line: 132 To prevent this problem, clear the AIL delwri queue as a final step before xfsaild() exit. The !empty state should never occur in the normal case, so add an assert to catch unexpected problems going forward. [dgc: add comment explaining need for xfs_buf_delwri_cancel() after calling xfs_buf_delwri_submit_nowait().] Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2018-10-18 14:21:49 +08:00
* Check kthread_should_stop() after we set the task state to
* guarantee that we either see the stop bit and exit or the
* task state is reset to runnable such that it's not scheduled
* out indefinitely and detects the stop bit at next iteration.
* A memory barrier is included in above task state set to
* serialize again kthread_stop().
*/
if (kthread_should_stop()) {
__set_current_state(TASK_RUNNING);
xfs: clear ail delwri queued bufs on unmount of shutdown fs In the typical unmount case, the AIL is forced out by the unmount sequence before the xfsaild task is stopped. Since AIL items are removed on writeback completion, this means that the AIL ->ail_buf_list delwri queue has been drained. This is not always true in the shutdown case, however. It's possible for buffers to sit on a delwri queue for a period of time across submission attempts if said items are locked or have been relogged and pinned since first added to the queue. If the attempt to log such an item results in a log I/O error, the error processing can shutdown the fs, remove the item from the AIL, stale the buffer (dropping the LRU reference) and clear its delwri queue state. The latter bit means the buffer will be released from a delwri queue on the next submission attempt, but this might never occur if the filesystem has shutdown and the AIL is empty. This means that such buffers are held indefinitely by the AIL delwri queue across destruction of the AIL. Aside from being a memory leak, these buffers can also hold references to in-core perag structures. The latter problem manifests as a generic/475 failure, reproducing the following asserts at unmount time: XFS: Assertion failed: atomic_read(&pag->pag_ref) == 0, file: fs/xfs/xfs_mount.c, line: 151 XFS: Assertion failed: atomic_read(&pag->pag_ref) == 0, file: fs/xfs/xfs_mount.c, line: 132 To prevent this problem, clear the AIL delwri queue as a final step before xfsaild() exit. The !empty state should never occur in the normal case, so add an assert to catch unexpected problems going forward. [dgc: add comment explaining need for xfs_buf_delwri_cancel() after calling xfs_buf_delwri_submit_nowait().] Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2018-10-18 14:21:49 +08:00
/*
* The caller forces out the AIL before stopping the
* thread in the common case, which means the delwri
* queue is drained. In the shutdown case, the queue may
* still hold relogged buffers that haven't been
* submitted because they were pinned since added to the
* queue.
*
* Log I/O error processing stales the underlying buffer
* and clears the delwri state, expecting the buf to be
* removed on the next submission attempt. That won't
* happen if we're shutting down, so this is the last
* opportunity to release such buffers from the queue.
*/
ASSERT(list_empty(&ailp->ail_buf_list) ||
xlog_is_shutdown(ailp->ail_log));
xfs: clear ail delwri queued bufs on unmount of shutdown fs In the typical unmount case, the AIL is forced out by the unmount sequence before the xfsaild task is stopped. Since AIL items are removed on writeback completion, this means that the AIL ->ail_buf_list delwri queue has been drained. This is not always true in the shutdown case, however. It's possible for buffers to sit on a delwri queue for a period of time across submission attempts if said items are locked or have been relogged and pinned since first added to the queue. If the attempt to log such an item results in a log I/O error, the error processing can shutdown the fs, remove the item from the AIL, stale the buffer (dropping the LRU reference) and clear its delwri queue state. The latter bit means the buffer will be released from a delwri queue on the next submission attempt, but this might never occur if the filesystem has shutdown and the AIL is empty. This means that such buffers are held indefinitely by the AIL delwri queue across destruction of the AIL. Aside from being a memory leak, these buffers can also hold references to in-core perag structures. The latter problem manifests as a generic/475 failure, reproducing the following asserts at unmount time: XFS: Assertion failed: atomic_read(&pag->pag_ref) == 0, file: fs/xfs/xfs_mount.c, line: 151 XFS: Assertion failed: atomic_read(&pag->pag_ref) == 0, file: fs/xfs/xfs_mount.c, line: 132 To prevent this problem, clear the AIL delwri queue as a final step before xfsaild() exit. The !empty state should never occur in the normal case, so add an assert to catch unexpected problems going forward. [dgc: add comment explaining need for xfs_buf_delwri_cancel() after calling xfs_buf_delwri_submit_nowait().] Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2018-10-18 14:21:49 +08:00
xfs_buf_delwri_cancel(&ailp->ail_buf_list);
break;
}
spin_lock(&ailp->ail_lock);
/*
* Idle if the AIL is empty and we are not racing with a target
* update. We check the AIL after we set the task to a sleep
* state to guarantee that we either catch an ail_target update
* or that a wake_up resets the state to TASK_RUNNING.
* Otherwise, we run the risk of sleeping indefinitely.
*
* The barrier matches the ail_target update in xfs_ail_push().
*/
smp_rmb();
if (!xfs_ail_min(ailp) &&
xfs: drain the buf delwri queue before xfsaild idles xfsaild is racy with respect to transaction abort and shutdown in that the task can idle or exit with an empty AIL but buffers still on the delwri queue. This was partly addressed by cancelling the delwri queue before the task exits to prevent memory leaks, but it's also possible for xfsaild to empty and idle with buffers on the delwri queue. For example, a transaction that pins a buffer that also happens to sit on the AIL delwri queue will explicitly remove the associated log item from the AIL if the transaction aborts. The side effect of this is an unmount hang in xfs_wait_buftarg() as the associated buffers remain held by the delwri queue indefinitely. This is reproduced on repeated runs of generic/531 with an fs format (-mrmapbt=1 -bsize=1k) that happens to also reproduce transaction aborts. Update xfsaild to not idle until both the AIL and associated delwri queue are empty and update the push code to continue delwri queue submission attempts even when the AIL is empty. This allows the AIL to eventually release aborted buffers stranded on the delwri queue when they are unlocked by the associated transaction. This should have no significant effect on normal runtime behavior because the xfsaild currently idles only when the AIL is empty and in practice the AIL is rarely empty with a populated delwri queue. The items must be AIL resident to land in the queue in the first place and generally aren't removed until writeback completes. Note that the pre-existing delwri queue cancel logic in the exit path is retained because task stop is external, could technically come at any point, and xfsaild is still responsible to release its buffer references before it exits. Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
2020-07-16 22:39:29 +08:00
ailp->ail_target == ailp->ail_target_prev &&
list_empty(&ailp->ail_buf_list)) {
spin_unlock(&ailp->ail_lock);
xfs: Make xfsaild freezeable again Hendik has reported suspend failures due to xfsaild blocking the freezer to settle down. Jan 17 19:59:56 linux-6380 kernel: PM: Syncing filesystems ... done. Jan 17 19:59:56 linux-6380 kernel: PM: Preparing system for sleep (mem) Jan 17 19:59:56 linux-6380 kernel: Freezing user space processes ... (elapsed 0.001 seconds) done. Jan 17 19:59:56 linux-6380 kernel: Freezing remaining freezable tasks ... Jan 17 19:59:56 linux-6380 kernel: Freezing of tasks failed after 20.002 seconds (1 tasks refusing to freeze, wq_busy=0): Jan 17 19:59:56 linux-6380 kernel: xfsaild/dm-5 S 00000000 0 1293 2 0x00000080 Jan 17 19:59:56 linux-6380 kernel: f0ef5f00 00000046 00000200 00000000 ffff9022 c02d3800 00000000 00000032 Jan 17 19:59:56 linux-6380 kernel: ee0b2400 00000032 f71e0d00 f36fabc0 f0ef2d00 f0ef6000 f0ef2d00 f12f90c0 Jan 17 19:59:56 linux-6380 kernel: f0ef5f0c c0844e44 00000000 f0ef5f6c f811e0be 00000000 00000000 f0ef2d00 Jan 17 19:59:56 linux-6380 kernel: Call Trace: Jan 17 19:59:56 linux-6380 kernel: [<c0844e44>] schedule+0x34/0x90 Jan 17 19:59:56 linux-6380 kernel: [<f811e0be>] xfsaild+0x5de/0x600 [xfs] Jan 17 19:59:56 linux-6380 kernel: [<c0286cbb>] kthread+0x9b/0xb0 Jan 17 19:59:56 linux-6380 kernel: [<c0848a79>] ret_from_kernel_thread+0x21/0x38 The issue has been there for quite some time but it has been made visible by only by 24ba16bb3d49 ("xfs: clear PF_NOFREEZE for xfsaild kthread") because the suspend started seeing xfsaild. The above commit has missed that the !xfs_ail_min branch might call schedule with TASK_INTERRUPTIBLE without calling try_to_freeze so the pm suspend would wake up the kernel thread over and over again without any progress. What we want here is to use freezable_schedule instead to hide the thread from the suspend. While we are here also change schedule_timeout to freezable variant to prevent from spurious wakeups by suspend. [dchinner: re-add set_freezeable call so the freezer will account properly for this kthread. ] Reported-by: Hendrik Woltersdorf <hendrikw@arcor.de> Signed-off-by: Michal Hocko <mhocko@suse.com> Reviewed-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-02-08 11:59:07 +08:00
freezable_schedule();
tout = 0;
continue;
}
spin_unlock(&ailp->ail_lock);
if (tout)
xfs: Make xfsaild freezeable again Hendik has reported suspend failures due to xfsaild blocking the freezer to settle down. Jan 17 19:59:56 linux-6380 kernel: PM: Syncing filesystems ... done. Jan 17 19:59:56 linux-6380 kernel: PM: Preparing system for sleep (mem) Jan 17 19:59:56 linux-6380 kernel: Freezing user space processes ... (elapsed 0.001 seconds) done. Jan 17 19:59:56 linux-6380 kernel: Freezing remaining freezable tasks ... Jan 17 19:59:56 linux-6380 kernel: Freezing of tasks failed after 20.002 seconds (1 tasks refusing to freeze, wq_busy=0): Jan 17 19:59:56 linux-6380 kernel: xfsaild/dm-5 S 00000000 0 1293 2 0x00000080 Jan 17 19:59:56 linux-6380 kernel: f0ef5f00 00000046 00000200 00000000 ffff9022 c02d3800 00000000 00000032 Jan 17 19:59:56 linux-6380 kernel: ee0b2400 00000032 f71e0d00 f36fabc0 f0ef2d00 f0ef6000 f0ef2d00 f12f90c0 Jan 17 19:59:56 linux-6380 kernel: f0ef5f0c c0844e44 00000000 f0ef5f6c f811e0be 00000000 00000000 f0ef2d00 Jan 17 19:59:56 linux-6380 kernel: Call Trace: Jan 17 19:59:56 linux-6380 kernel: [<c0844e44>] schedule+0x34/0x90 Jan 17 19:59:56 linux-6380 kernel: [<f811e0be>] xfsaild+0x5de/0x600 [xfs] Jan 17 19:59:56 linux-6380 kernel: [<c0286cbb>] kthread+0x9b/0xb0 Jan 17 19:59:56 linux-6380 kernel: [<c0848a79>] ret_from_kernel_thread+0x21/0x38 The issue has been there for quite some time but it has been made visible by only by 24ba16bb3d49 ("xfs: clear PF_NOFREEZE for xfsaild kthread") because the suspend started seeing xfsaild. The above commit has missed that the !xfs_ail_min branch might call schedule with TASK_INTERRUPTIBLE without calling try_to_freeze so the pm suspend would wake up the kernel thread over and over again without any progress. What we want here is to use freezable_schedule instead to hide the thread from the suspend. While we are here also change schedule_timeout to freezable variant to prevent from spurious wakeups by suspend. [dchinner: re-add set_freezeable call so the freezer will account properly for this kthread. ] Reported-by: Hendrik Woltersdorf <hendrikw@arcor.de> Signed-off-by: Michal Hocko <mhocko@suse.com> Reviewed-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-02-08 11:59:07 +08:00
freezable_schedule_timeout(msecs_to_jiffies(tout));
__set_current_state(TASK_RUNNING);
try_to_freeze();
tout = xfsaild_push(ailp);
}
xfs: clear PF_MEMALLOC before exiting xfsaild thread Leaving PF_MEMALLOC set when exiting a kthread causes it to remain set during do_exit(). That can confuse things. In particular, if BSD process accounting is enabled, then do_exit() writes data to an accounting file. If that file has FS_SYNC_FL set, then this write occurs synchronously and can misbehave if PF_MEMALLOC is set. For example, if the accounting file is located on an XFS filesystem, then a WARN_ON_ONCE() in iomap_do_writepage() is triggered and the data doesn't get written when it should. Or if the accounting file is located on an ext4 filesystem without a journal, then a WARN_ON_ONCE() in ext4_write_inode() is triggered and the inode doesn't get written. Fix this in xfsaild() by using the helper functions to save and restore PF_MEMALLOC. This can be reproduced as follows in the kvm-xfstests test appliance modified to add the 'acct' Debian package, and with kvm-xfstests's recommended kconfig modified to add CONFIG_BSD_PROCESS_ACCT=y: mkfs.xfs -f /dev/vdb mount /vdb touch /vdb/file chattr +S /vdb/file accton /vdb/file mkfs.xfs -f /dev/vdc mount /vdc umount /vdc It causes: WARNING: CPU: 1 PID: 336 at fs/iomap/buffered-io.c:1534 CPU: 1 PID: 336 Comm: xfsaild/vdc Not tainted 5.6.0-rc5 #3 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20191223_100556-anatol 04/01/2014 RIP: 0010:iomap_do_writepage+0x16b/0x1f0 fs/iomap/buffered-io.c:1534 [...] Call Trace: write_cache_pages+0x189/0x4d0 mm/page-writeback.c:2238 iomap_writepages+0x1c/0x33 fs/iomap/buffered-io.c:1642 xfs_vm_writepages+0x65/0x90 fs/xfs/xfs_aops.c:578 do_writepages+0x41/0xe0 mm/page-writeback.c:2344 __filemap_fdatawrite_range+0xd2/0x120 mm/filemap.c:421 file_write_and_wait_range+0x71/0xc0 mm/filemap.c:760 xfs_file_fsync+0x7a/0x2b0 fs/xfs/xfs_file.c:114 generic_write_sync include/linux/fs.h:2867 [inline] xfs_file_buffered_aio_write+0x379/0x3b0 fs/xfs/xfs_file.c:691 call_write_iter include/linux/fs.h:1901 [inline] new_sync_write+0x130/0x1d0 fs/read_write.c:483 __kernel_write+0x54/0xe0 fs/read_write.c:515 do_acct_process+0x122/0x170 kernel/acct.c:522 slow_acct_process kernel/acct.c:581 [inline] acct_process+0x1d4/0x27c kernel/acct.c:607 do_exit+0x83d/0xbc0 kernel/exit.c:791 kthread+0xf1/0x140 kernel/kthread.c:257 ret_from_fork+0x27/0x50 arch/x86/entry/entry_64.S:352 This bug was originally reported by syzbot at https://lore.kernel.org/r/0000000000000e7156059f751d7b@google.com. Reported-by: syzbot+1f9dc49e8de2582d90c2@syzkaller.appspotmail.com Signed-off-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2020-03-10 23:57:27 +08:00
memalloc_noreclaim_restore(noreclaim_flag);
return 0;
}
/*
* This routine is called to move the tail of the AIL forward. It does this by
* trying to flush items in the AIL whose lsns are below the given
* threshold_lsn.
*
* The push is run asynchronously in a workqueue, which means the caller needs
* to handle waiting on the async flush for space to become available.
* We don't want to interrupt any push that is in progress, hence we only queue
* work if we set the pushing bit appropriately.
*
* We do this unlocked - we only need to know whether there is anything in the
* AIL at the time we are called. We don't need to access the contents of
* any of the objects, so the lock is not needed.
*/
void
xfs_ail_push(
struct xfs_ail *ailp,
xfs_lsn_t threshold_lsn)
{
struct xfs_log_item *lip;
lip = xfs_ail_min(ailp);
if (!lip || xlog_is_shutdown(ailp->ail_log) ||
XFS_LSN_CMP(threshold_lsn, ailp->ail_target) <= 0)
return;
/*
* Ensure that the new target is noticed in push code before it clears
* the XFS_AIL_PUSHING_BIT.
*/
smp_wmb();
xfs_trans_ail_copy_lsn(ailp, &ailp->ail_target, &threshold_lsn);
smp_wmb();
wake_up_process(ailp->ail_task);
}
/*
* Push out all items in the AIL immediately
*/
void
xfs_ail_push_all(
struct xfs_ail *ailp)
{
xfs_lsn_t threshold_lsn = xfs_ail_max_lsn(ailp);
if (threshold_lsn)
xfs_ail_push(ailp, threshold_lsn);
}
/*
* Push out all items in the AIL immediately and wait until the AIL is empty.
*/
void
xfs_ail_push_all_sync(
struct xfs_ail *ailp)
{
struct xfs_log_item *lip;
DEFINE_WAIT(wait);
spin_lock(&ailp->ail_lock);
while ((lip = xfs_ail_max(ailp)) != NULL) {
prepare_to_wait(&ailp->ail_empty, &wait, TASK_UNINTERRUPTIBLE);
wake_up_process(ailp->ail_task);
spin_unlock(&ailp->ail_lock);
schedule();
spin_lock(&ailp->ail_lock);
}
spin_unlock(&ailp->ail_lock);
finish_wait(&ailp->ail_empty, &wait);
}
void
xfs_ail_update_finish(
struct xfs_ail *ailp,
xfs_lsn_t old_lsn) __releases(ailp->ail_lock)
{
struct xlog *log = ailp->ail_log;
/* if the tail lsn hasn't changed, don't do updates or wakeups. */
if (!old_lsn || old_lsn == __xfs_ail_min_lsn(ailp)) {
spin_unlock(&ailp->ail_lock);
return;
}
if (!xlog_is_shutdown(log))
xlog_assign_tail_lsn_locked(log->l_mp);
if (list_empty(&ailp->ail_head))
wake_up_all(&ailp->ail_empty);
spin_unlock(&ailp->ail_lock);
xfs_log_space_wake(log->l_mp);
}
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
/*
* xfs_trans_ail_update - bulk AIL insertion operation.
*
* @xfs_trans_ail_update takes an array of log items that all need to be
* positioned at the same LSN in the AIL. If an item is not in the AIL, it will
* be added. Otherwise, it will be repositioned by removing it and re-adding
* it to the AIL. If we move the first item in the AIL, update the log tail to
* match the new minimum LSN in the AIL.
*
* This function takes the AIL lock once to execute the update operations on
* all the items in the array, and as such should not be called with the AIL
* lock held. As a result, once we have the AIL lock, we need to check each log
* item LSN to confirm it needs to be moved forward in the AIL.
*
* To optimise the insert operation, we delete all the items from the AIL in
* the first pass, moving them into a temporary list, then splice the temporary
* list into the correct position in the AIL. This avoids needing to do an
* insert operation on every item.
*
* This function must be called with the AIL lock held. The lock is dropped
* before returning.
*/
void
xfs_trans_ail_update_bulk(
struct xfs_ail *ailp,
struct xfs_ail_cursor *cur,
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
struct xfs_log_item **log_items,
int nr_items,
xfs_lsn_t lsn) __releases(ailp->ail_lock)
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
{
struct xfs_log_item *mlip;
xfs_lsn_t tail_lsn = 0;
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
int i;
LIST_HEAD(tmp);
ASSERT(nr_items > 0); /* Not required, but true. */
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
mlip = xfs_ail_min(ailp);
for (i = 0; i < nr_items; i++) {
struct xfs_log_item *lip = log_items[i];
if (test_and_set_bit(XFS_LI_IN_AIL, &lip->li_flags)) {
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
/* check if we really need to move the item */
if (XFS_LSN_CMP(lsn, lip->li_lsn) <= 0)
continue;
trace_xfs_ail_move(lip, lip->li_lsn, lsn);
if (mlip == lip && !tail_lsn)
tail_lsn = lip->li_lsn;
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
xfs_ail_delete(ailp, lip);
} else {
trace_xfs_ail_insert(lip, 0, lsn);
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
}
lip->li_lsn = lsn;
list_add(&lip->li_ail, &tmp);
}
if (!list_empty(&tmp))
xfs_ail_splice(ailp, cur, &tmp, lsn);
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
xfs_ail_update_finish(ailp, tail_lsn);
}
/* Insert a log item into the AIL. */
void
xfs_trans_ail_insert(
struct xfs_ail *ailp,
struct xfs_log_item *lip,
xfs_lsn_t lsn)
{
spin_lock(&ailp->ail_lock);
xfs_trans_ail_update_bulk(ailp, NULL, &lip, 1, lsn);
xfs: bulk AIL insertion during transaction commit When inserting items into the AIL from the transaction committed callbacks, we take the AIL lock for every single item that is to be inserted. For a CIL checkpoint commit, this can be tens of thousands of individual inserts, yet almost all of the items will be inserted at the same point in the AIL because they have the same index. To reduce the overhead and contention on the AIL lock for such operations, introduce a "bulk insert" operation which allows a list of log items with the same LSN to be inserted in a single operation via a list splice. To do this, we need to pre-sort the log items being committed into a temporary list for insertion. The complexity is that not every log item will end up with the same LSN, and not every item is actually inserted into the AIL. Items that don't match the commit LSN will be inserted and unpinned as per the current one-at-a-time method (relatively rare), while items that are not to be inserted will be unpinned and freed immediately. Items that are to be inserted at the given commit lsn are placed in a temporary array and inserted into the AIL in bulk each time the array fills up. As a result of this, we trade off AIL hold time for a significant reduction in traffic. lock_stat output shows that the worst case hold time is unchanged, but contention from AIL inserts drops by an order of magnitude and the number of lock traversal decreases significantly. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2010-12-20 09:02:19 +08:00
}
/*
* Delete one log item from the AIL.
*
* If this item was at the tail of the AIL, return the LSN of the log item so
* that we can use it to check if the LSN of the tail of the log has moved
* when finishing up the AIL delete process in xfs_ail_update_finish().
*/
xfs_lsn_t
xfs_ail_delete_one(
struct xfs_ail *ailp,
struct xfs_log_item *lip)
{
struct xfs_log_item *mlip = xfs_ail_min(ailp);
xfs_lsn_t lsn = lip->li_lsn;
trace_xfs_ail_delete(lip, mlip->li_lsn, lip->li_lsn);
xfs_ail_delete(ailp, lip);
clear_bit(XFS_LI_IN_AIL, &lip->li_flags);
lip->li_lsn = 0;
if (mlip == lip)
return lsn;
return 0;
}
void
xfs_trans_ail_delete(
struct xfs_log_item *lip,
int shutdown_type)
{
struct xfs_ail *ailp = lip->li_ailp;
xfs: log shutdown triggers should only shut down the log We've got a mess on our hands. 1. xfs_trans_commit() cannot cancel transactions because the mount is shut down - that causes dirty, aborted, unlogged log items to sit unpinned in memory and potentially get written to disk before the log is shut down. Hence xfs_trans_commit() can only abort transactions when xlog_is_shutdown() is true. 2. xfs_force_shutdown() is used in places to cause the current modification to be aborted via xfs_trans_commit() because it may be impractical or impossible to cancel the transaction directly, and hence xfs_trans_commit() must cancel transactions when xfs_is_shutdown() is true in this situation. But we can't do that because of #1. 3. Log IO errors cause log shutdowns by calling xfs_force_shutdown() to shut down the mount and then the log from log IO completion. 4. xfs_force_shutdown() can result in a log force being issued, which has to wait for log IO completion before it will mark the log as shut down. If #3 races with some other shutdown trigger that runs a log force, we rely on xfs_force_shutdown() silently ignoring #3 and avoiding shutting down the log until the failed log force completes. 5. To ensure #2 always works, we have to ensure that xfs_force_shutdown() does not return until the the log is shut down. But in the case of #4, this will result in a deadlock because the log Io completion will block waiting for a log force to complete which is blocked waiting for log IO to complete.... So the very first thing we have to do here to untangle this mess is dissociate log shutdown triggers from mount shutdowns. We already have xlog_forced_shutdown, which will atomically transistion to the log a shutdown state. Due to internal asserts it cannot be called multiple times, but was done simply because the only place that could call it was xfs_do_force_shutdown() (i.e. the mount shutdown!) and that could only call it once and once only. So the first thing we do is remove the asserts. We then convert all the internal log shutdown triggers to call xlog_force_shutdown() directly instead of xfs_force_shutdown(). This allows the log shutdown triggers to shut down the log without needing to care about mount based shutdown constraints. This means we shut down the log independently of the mount and the mount may not notice this until it's next attempt to read or modify metadata. At that point (e.g. xfs_trans_commit()) it will see that the log is shutdown, error out and shutdown the mount. To ensure that all the unmount behaviours and asserts track correctly as a result of a log shutdown, propagate the shutdown up to the mount if it is not already set. This keeps the mount and log state in sync, and saves a huge amount of hassle where code fails because of a log shutdown but only checks for mount shutdowns and hence ends up doing the wrong thing. Cleaning up that mess is an exercise for another day. This enables us to address the other problems noted above in followup patches. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
2022-03-30 09:22:01 +08:00
struct xlog *log = ailp->ail_log;
xfs_lsn_t tail_lsn;
spin_lock(&ailp->ail_lock);
if (!test_bit(XFS_LI_IN_AIL, &lip->li_flags)) {
spin_unlock(&ailp->ail_lock);
xfs: log shutdown triggers should only shut down the log We've got a mess on our hands. 1. xfs_trans_commit() cannot cancel transactions because the mount is shut down - that causes dirty, aborted, unlogged log items to sit unpinned in memory and potentially get written to disk before the log is shut down. Hence xfs_trans_commit() can only abort transactions when xlog_is_shutdown() is true. 2. xfs_force_shutdown() is used in places to cause the current modification to be aborted via xfs_trans_commit() because it may be impractical or impossible to cancel the transaction directly, and hence xfs_trans_commit() must cancel transactions when xfs_is_shutdown() is true in this situation. But we can't do that because of #1. 3. Log IO errors cause log shutdowns by calling xfs_force_shutdown() to shut down the mount and then the log from log IO completion. 4. xfs_force_shutdown() can result in a log force being issued, which has to wait for log IO completion before it will mark the log as shut down. If #3 races with some other shutdown trigger that runs a log force, we rely on xfs_force_shutdown() silently ignoring #3 and avoiding shutting down the log until the failed log force completes. 5. To ensure #2 always works, we have to ensure that xfs_force_shutdown() does not return until the the log is shut down. But in the case of #4, this will result in a deadlock because the log Io completion will block waiting for a log force to complete which is blocked waiting for log IO to complete.... So the very first thing we have to do here to untangle this mess is dissociate log shutdown triggers from mount shutdowns. We already have xlog_forced_shutdown, which will atomically transistion to the log a shutdown state. Due to internal asserts it cannot be called multiple times, but was done simply because the only place that could call it was xfs_do_force_shutdown() (i.e. the mount shutdown!) and that could only call it once and once only. So the first thing we do is remove the asserts. We then convert all the internal log shutdown triggers to call xlog_force_shutdown() directly instead of xfs_force_shutdown(). This allows the log shutdown triggers to shut down the log without needing to care about mount based shutdown constraints. This means we shut down the log independently of the mount and the mount may not notice this until it's next attempt to read or modify metadata. At that point (e.g. xfs_trans_commit()) it will see that the log is shutdown, error out and shutdown the mount. To ensure that all the unmount behaviours and asserts track correctly as a result of a log shutdown, propagate the shutdown up to the mount if it is not already set. This keeps the mount and log state in sync, and saves a huge amount of hassle where code fails because of a log shutdown but only checks for mount shutdowns and hence ends up doing the wrong thing. Cleaning up that mess is an exercise for another day. This enables us to address the other problems noted above in followup patches. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
2022-03-30 09:22:01 +08:00
if (shutdown_type && !xlog_is_shutdown(log)) {
xfs_alert_tag(log->l_mp, XFS_PTAG_AILDELETE,
"%s: attempting to delete a log item that is not in the AIL",
__func__);
xfs: log shutdown triggers should only shut down the log We've got a mess on our hands. 1. xfs_trans_commit() cannot cancel transactions because the mount is shut down - that causes dirty, aborted, unlogged log items to sit unpinned in memory and potentially get written to disk before the log is shut down. Hence xfs_trans_commit() can only abort transactions when xlog_is_shutdown() is true. 2. xfs_force_shutdown() is used in places to cause the current modification to be aborted via xfs_trans_commit() because it may be impractical or impossible to cancel the transaction directly, and hence xfs_trans_commit() must cancel transactions when xfs_is_shutdown() is true in this situation. But we can't do that because of #1. 3. Log IO errors cause log shutdowns by calling xfs_force_shutdown() to shut down the mount and then the log from log IO completion. 4. xfs_force_shutdown() can result in a log force being issued, which has to wait for log IO completion before it will mark the log as shut down. If #3 races with some other shutdown trigger that runs a log force, we rely on xfs_force_shutdown() silently ignoring #3 and avoiding shutting down the log until the failed log force completes. 5. To ensure #2 always works, we have to ensure that xfs_force_shutdown() does not return until the the log is shut down. But in the case of #4, this will result in a deadlock because the log Io completion will block waiting for a log force to complete which is blocked waiting for log IO to complete.... So the very first thing we have to do here to untangle this mess is dissociate log shutdown triggers from mount shutdowns. We already have xlog_forced_shutdown, which will atomically transistion to the log a shutdown state. Due to internal asserts it cannot be called multiple times, but was done simply because the only place that could call it was xfs_do_force_shutdown() (i.e. the mount shutdown!) and that could only call it once and once only. So the first thing we do is remove the asserts. We then convert all the internal log shutdown triggers to call xlog_force_shutdown() directly instead of xfs_force_shutdown(). This allows the log shutdown triggers to shut down the log without needing to care about mount based shutdown constraints. This means we shut down the log independently of the mount and the mount may not notice this until it's next attempt to read or modify metadata. At that point (e.g. xfs_trans_commit()) it will see that the log is shutdown, error out and shutdown the mount. To ensure that all the unmount behaviours and asserts track correctly as a result of a log shutdown, propagate the shutdown up to the mount if it is not already set. This keeps the mount and log state in sync, and saves a huge amount of hassle where code fails because of a log shutdown but only checks for mount shutdowns and hence ends up doing the wrong thing. Cleaning up that mess is an exercise for another day. This enables us to address the other problems noted above in followup patches. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
2022-03-30 09:22:01 +08:00
xlog_force_shutdown(log, shutdown_type);
}
return;
}
/* xfs_ail_update_finish() drops the AIL lock */
xfs_clear_li_failed(lip);
tail_lsn = xfs_ail_delete_one(ailp, lip);
xfs_ail_update_finish(ailp, tail_lsn);
}
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
int
xfs_trans_ail_init(
xfs_mount_t *mp)
{
struct xfs_ail *ailp;
ailp = kmem_zalloc(sizeof(struct xfs_ail), KM_MAYFAIL);
if (!ailp)
return -ENOMEM;
ailp->ail_log = mp->m_log;
INIT_LIST_HEAD(&ailp->ail_head);
INIT_LIST_HEAD(&ailp->ail_cursors);
spin_lock_init(&ailp->ail_lock);
INIT_LIST_HEAD(&ailp->ail_buf_list);
init_waitqueue_head(&ailp->ail_empty);
ailp->ail_task = kthread_run(xfsaild, ailp, "xfsaild/%s",
mp->m_super->s_id);
if (IS_ERR(ailp->ail_task))
goto out_free_ailp;
mp->m_ail = ailp;
return 0;
out_free_ailp:
kmem_free(ailp);
return -ENOMEM;
[XFS] Move AIL pushing into it's own thread When many hundreds to thousands of threads all try to do simultaneous transactions and the log is in a tail-pushing situation (i.e. full), we can get multiple threads walking the AIL list and contending on the AIL lock. The AIL push is, in effect, a simple I/O dispatch algorithm complicated by the ordering constraints placed on it by the transaction subsystem. It really does not need multiple threads to push on it - even when only a single CPU is pushing the AIL, it can push the I/O out far faster that pretty much any disk subsystem can handle. So, to avoid contention problems stemming from multiple list walkers, move the list walk off into another thread and simply provide a "target" to push to. When a thread requires a push, it sets the target and wakes the push thread, then goes to sleep waiting for the required amount of space to become available in the log. This mechanism should also be a lot fairer under heavy load as the waiters will queue in arrival order, rather than queuing in "who completed a push first" order. Also, by moving the pushing to a separate thread we can do more effectively overload detection and prevention as we can keep context from loop iteration to loop iteration. That is, we can push only part of the list each loop and not have to loop back to the start of the list every time we run. This should also help by reducing the number of items we try to lock and/or push items that we cannot move. Note that this patch is not intended to solve the inefficiencies in the AIL structure and the associated issues with extremely large list contents. That needs to be addresses separately; parallel access would cause problems to any new structure as well, so I'm only aiming to isolate the structure from unbounded parallelism here. SGI-PV: 972759 SGI-Modid: xfs-linux-melb:xfs-kern:30371a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
2008-02-05 09:13:32 +08:00
}
void
xfs_trans_ail_destroy(
xfs_mount_t *mp)
{
struct xfs_ail *ailp = mp->m_ail;
kthread_stop(ailp->ail_task);
kmem_free(ailp);
}