This fixes a corner case that is caused by a race of dio write vs dio
read/write.
Here is how the race could happen.
Suppose that no extent map has been loaded into memory yet.
There is a file extent [0, 32K), two jobs are running concurrently
against it, t1 is doing dio write to [8K, 32K) and t2 is doing dio
read from [0, 4K) or [4K, 8K).
t1 goes ahead of t2 and splits em [0, 32K) to em [0K, 8K) and [8K 32K).
------------------------------------------------------
t1 t2
btrfs_get_blocks_direct() btrfs_get_blocks_direct()
-> btrfs_get_extent() -> btrfs_get_extent()
-> lookup_extent_mapping()
-> add_extent_mapping() -> lookup_extent_mapping()
# load [0, 32K)
-> btrfs_new_extent_direct()
-> btrfs_drop_extent_cache()
# split [0, 32K) and
# drop [8K, 32K)
-> add_extent_mapping()
# add [8K, 32K)
-> add_extent_mapping()
# handle -EEXIST when adding
# [0, 32K)
------------------------------------------------------
About how t2(dio read/write) runs into -EEXIST:
a) add_extent_mapping() gets -EEXIST for adding em [0, 32k),
b) search_extent_mapping() then returns [0, 8k) as the existing em,
even though start == existing->start, em is [0, 32k) so that
extent_map_end(em) > extent_map_end(existing), i.e. 32k > 8k,
c) then it goes thru merge_extent_mapping() which tries to add a [8k, 8k)
(with a length 0) and returns -EEXIST as [8k, 32k) is already in tree,
d) so btrfs_get_extent() ends up returning -EEXIST to dio read/write,
which is confusing applications.
Here I conclude all the possible situations,
1) start < existing->start
+-----------+em+-----------+
+--prev---+ | +-------------+ |
| | | | | |
+---------+ + +---+existing++ ++
+
|
+
start
2) start == existing->start
+------------em------------+
| +-------------+ |
| | | |
+ +----existing-+ +
|
|
+
start
3) start > existing->start && start < (existing->start + existing->len)
+------------em------------+
| +-------------+ |
| | | |
+ +----existing-+ +
|
|
+
start
4) start >= (existing->start + existing->len)
+-----------+em+-----------+
| +-------------+ | +--next---+
| | | | | |
+ +---+existing++ + +---------+
+
|
+
start
As we can see, it turns out that if start is within existing em (front
inclusive), then the existing em should be returned as is, otherwise,
we try our best to merge candidate em with sibling ems to form a
larger em (in order to reduce the total number of em).
Reported-by: David Vallender <david.vallender@landmark.co.uk>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
%block_len could be checked on deciding if two em are mergeable.
merge_extent_mapping() has only added the front pad if the front part
of em gets truncated, but it's possible that the end part gets
truncated.
For both compressed extent and inline extent, em->block_len is not
adjusted accordingly, and for regular extent, em->block_len always
equals to em->len, hence this sets em->block_len with em->len.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The reada_lock in struct btrfs_device was only initialised, and not
actually used. That's good because there's another lock also called
reada_lock in the btrfs_fs_info that was quite heavily used. Remove
this one.
Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Before rbio_orig_end_io() goes to free rbio, rbio may get merged with
more bios from other rbios and rbio->bio_list becomes non-empty,
in that case, these newly merged bios don't end properly.
Once unlock_stripe() is done, rbio->bio_list will not be updated any
more and we can call bio_endio() on all queued bios.
It should only happen in error-out cases, the normal path of recover
and full stripe write have already set RBIO_RMW_LOCKED_BIT to disable
merge before doing IO, so rbio_orig_end_io() called by them doesn't
have the above issue.
Reported-by: Jérôme Carretero <cJ-ko@zougloub.eu>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since raid6 recover tries all possible combinations of failed stripes,
- when raid6 rebuild algorithm is used, i.e. raid6_datap_recov() and
raid6_2data_recov(), it may change the in-memory content of failed
stripes, if such a raid bio is cached, a later raid write rmw or recover
can steal @stripe_pages from it instead of reading from disks, such that
it carries the wrong content to do write rmw or recovery and ends up
with corruption or recovery failures.
- when raid5 rebuild algorithm is used, i.e. xor, raid bio can be cached
because the only failed stripe which contains @rbio->bio_pages gets
modified, others remain the same so that their in-memory content is
consistent with their on-disk content.
This adds a check to skip caching rbio if using raid6 recover.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Bio iterated by set_bio_pages_uptodate() is raid56 internal one, so it
will never be a BIO_CLONED bio, and since this is called by end_io
functions, bio->bi_iter.bi_size is zero, we mustn't use
bio_for_each_segment() as that is a no-op if bi_size is zero.
Fixes: 6592e58c6b ("Btrfs: fix write corruption due to bio cloning on raid5/6")
Cc: <stable@vger.kernel.org> # v4.12-rc6+
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There is no function named btrfs_get_inode_index_count.
Explanation for magic number index_cnt=2 in btrfs_new_inode() is
actually located in btrfs_set_inode_index_count().
So replace 'btrfs_get_inode_index_count' in the comment by
'btrfs_set_inode_index_count'.
Signed-off-by: Su Yue <suy.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.com>
It's not used outside of extent-tree so there is no reason to not be
static.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We call btrfs_free_stale_device() only when we alloc a new struct
btrfs_device (ret=1), so move it closer to where we alloc the new
device. Also drop the comments.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
I've noticed that the updated item checker stack consumption increased
dramatically in 542f5385e20cf97447 ("btrfs: tree-checker: Add checker
for dir item")
tree-checker.c:check_leaf +552 (176 -> 728)
The array is 255 bytes long, dynamic allocation would slow down the
sanity checks so it's more reasonable to keep it on-stack. Moving the
variable to the scope of use reduces the stack usage again
tree-checker.c:check_leaf -264 (728 -> 464)
Reviewed-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
gcc-8 reports:
fs/btrfs/ioctl.c: In function 'btrfs_ioctl':
./include/linux/string.h:245:9: warning: '__builtin_strncpy' specified
bound 1024 equals destination size [-Wstringop-truncation]
We need one less byte or call strlcpy() to make it a nul-terminated
string. This is done on the next line anyway, but we want to avoid the
warning.
Signed-off-by: Xiongfeng Wang <xiongfeng.wang@linaro.org>
Reviewed-by: David Sterba <dsterba@suse.com>
[ update changelog ]
Signed-off-by: David Sterba <dsterba@suse.com>
It appears from the original commit [1] that there isn't any design
specific reason not to fail the mount instead of just warning. This
patch will change it to fail.
[1]
commit 319e4d0661
btrfs: Enhance super validation check
Fixes: 319e4d0661 ("btrfs: Enhance super validation check")
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The UUID change by btrfstune sets SUPER_FLAG_CHANGING_FSID and resets it
only when changing fsid is complete. Its not a good idea to mount the
device anything in between, reading metadata blocks would fail with UUID
mismatch.
This patch doesn't add SUPER_FLAG_CHANGING_FSID into
BTRFS_SUPER_FLAG_SUPP list, so mount will fail (along with the fix in
the next patch) when SUPER_FLAG_CHANGING_FSID is set.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ update changelog ]
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs-progs uses super flag bit BTRFS_SUPER_FLAG_METADUMP_V2 (1ULL << 34).
So just define that in kernel so that we know its been used.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We've avoided data losing raid profile when doing balance, but it
turns out that deleting a device could also result in the same
problem.
Say we have 3 disks, and they're created with '-d raid1' profile.
- We have chunk P (the only data chunk on the empty btrfs).
- Suppose that chunk P's two raid1 copies reside in disk A and disk B.
- Now, 'btrfs device remove disk B'
btrfs_rm_device()
-> btrfs_shrink_device()
-> btrfs_relocate_chunk() #relocate any chunk on disk B
to other places.
- Chunk P will be removed and a new chunk will be created to hold
those data, but as chunk P is the only one holding raid1 profile,
after it goes away, the new chunk will be created as single profile
which is our default profile.
This fixes the problem by creating an empty data chunk before
relocating the data chunk.
Metadata/System chunk are supposed to have non-zero bytes all the time
so their raid profile is preserved.
Reported-by: James Alandt <James.Alandt@wdc.com>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For a fallocate's zero range operation that targets a range with an end
that is not aligned to the sector size, we can end up not updating the
inode's i_size. This happens when the last page of the range maps to an
unwritten (prealloc) extent and before that last page we have either a
hole or a written extent. This is because in this scenario we relied
on a call to btrfs_prealloc_file_range() to update the inode's i_size,
however it can only update the i_size to the "down aligned" end of the
range.
Example:
$ mkfs.btrfs -f /dev/sdc
$ mount /dev/sdc /mnt
$ xfs_io -f -c "pwrite -S 0xff 0 428K" /mnt/foobar
$ xfs_io -c "falloc -k 428K 4K" /mnt/foobar
$ xfs_io -c "fzero 0 430K" /mnt/foobar
$ du --bytes /mnt/foobar
438272 /mnt/foobar
The inode's i_size was left as 428Kb (438272 bytes) when it should have
been updated to 430Kb (440320 bytes).
Fix this by always updating the inode's i_size explicitly after zeroing
the range.
Fixes: ba6d5887946ff86d93dc ("Btrfs: add support for fallocate's zero range operation")
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
During a buffered IO write, we can have an extent state that we got when
we locked the range (if the range starts at an offset lower than eof), so
always pass it to btrfs_dirty_pages() so that setting the delalloc bit
in the range does not need to do a full search in the inode's io tree,
saving time and reducing the amount of time we hold the io tree's lock.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
This implements support the zero range operation of fallocate. For now
at least it's as simple as possible while reusing most of the existing
fallocate and hole punching infrastructure.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since fail stripe index in rbio would be used to decide which
algorithm reconstruction would be run, we cannot merge rbios if
their's fail striped indexes are different, otherwise, one of the two
reconstructions would fail.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Given the above
'
if (last->operation != cur->operation)
return 0;
',
it's guaranteed that two operations are same.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Assign ret = -EINVAL where it is actually required.
Remove { } around single line if else code.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_device::scrub_device is not a device which is being scrubbed,
but it holds the scrub context, so rename to reflect the same. No
functional changes here.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There is no other consumer for btrfs_handle_error() other than
__btrfs_handle_fs_error(), further this function quite small.
Merge it into its parent.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
[ reformat comment ]
Signed-off-by: David Sterba <dsterba@suse.com>
__btrfs_handle_fs_error() sets BTRFS_FS_STATE_ERROR, and calls
btrfs_handle_error() so no need to check if the BTRFS_FS_STATE_ERROR
is set in btrfs_handle_error(). And there is no other user of
btrfs_handle_error() as well.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There is a scenario that can end up with rebuild process failing to
return good content, i.e.
suppose that all disks can be read without problems and if the content
that was read out doesn't match its checksum, currently for raid6
btrfs at most retries twice,
- the 1st retry is to rebuild with all other stripes, it'll eventually
be a raid5 xor rebuild,
- if the 1st fails, the 2nd retry will deliberately fail parity p so
that it will do raid6 style rebuild,
however, the chances are that another non-parity stripe content also
has something corrupted, so that the above retries are not able to
return correct content, and users will think of this as data loss.
More seriouly, if the loss happens on some important internal btree
roots, it could refuse to mount.
This extends btrfs to do more retries and each retry fails only one
stripe. Since raid6 can tolerate 2 disk failures, if there is one
more failure besides the failure on which we're recovering, this can
always work.
The worst case is to retry as many times as the number of raid6 disks,
but given the fact that such a scenario is really rare in practice,
it's still acceptable.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The raid6 corruption is that,
suppose that all disks can be read without problems and if the content
that was read out doesn't match its checksum, currently for raid6
btrfs at most retries twice,
- the 1st retry is to rebuild with all other stripes, it'll eventually
be a raid5 xor rebuild,
- if the 1st fails, the 2nd retry will deliberately fail parity p so
that it will do raid6 style rebuild,
however, the chances are that another non-parity stripe content also
has something corrupted, so that the above retries are not able to
return correct content.
We've fixed normal reads to rebuild raid6 correctly with more retries
in Patch "Btrfs: make raid6 rebuild retry more"[1], this is to fix
scrub to do the exactly same rebuild process.
[1]: https://patchwork.kernel.org/patch/10091755/
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Update btrfs_check_rw_degradable() to check against the given device if
its lost.
We can use this function to know if the volume is going to be in
degraded mode OR failed state, when the given device fails. Which is
needed when we are handling the device failed state.
A preparatory patch does not affect the flow as such.
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
[ enhance comment ]
Signed-off-by: David Sterba <dsterba@suse.com>
All callers pass either GFP_NOFS or GFP_KERNEL now, so we can sink the
parameter to the function, though we lose some of the slightly better
semantics of GFP_KERNEL in some places, it's worth cleaning up the
callchains.
Signed-off-by: David Sterba <dsterba@suse.com>
There's only one instance where we pass different gfp mask to
unlock_extent_cached. Add a separate helper for that and then we can
drop the gfp parameter from unlock_extent_cached.
Signed-off-by: David Sterba <dsterba@suse.com>
Recent patches reworking the mount path left some unused parameters. We
pass a vfsmount to mount_subvol, the flags and data (ie. mount options)
have been already applied and we will not need them.
Signed-off-by: David Sterba <dsterba@suse.com>
Long ago, commit edf24abe51 ("btrfs: sanity mount option parsing and
early mount code") split the btrfs_parse_options() into two parts
(btrfs_parse_early_options() and btrfs_parse_options()). As a result,
btrfs_parse_optins no longer gets called twice and is the last one to
parse mount option string. Therefore there is no need to dup it.
Signed-off-by: Tomohiro Misono <misono.tomohiro@jp.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In fact nobody is waiting on @wait's waitqueue, it can be safely
removed.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The bio is not referenced after it has been submitted and the endio is
going to consume the sole reference on successful submission. On error,
the callers of __btrfs_submit_dio_bio do invoke bio_put so we don't
leak it either.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The bio is never referenced after it has been submitted so there is no
point in getting an extra reference.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The bio that is passsed is the newly created repair bio which already
has a reference count of 1, which is going to be consumed by the
endio routine on successful submission. On error the handler also
calls bio_put.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
bio_get/set is necessary only if the bio is going to be referenced
following submissions. In the code paths where such calls are made
we don't really need them since the bio is referenced only if
btrfs_map_bio returns an error. And this function can return an error
prior to submission only. So referencing the bio is safe. Furthermore
we do call bio_endio which will consume the last reference. So let's
remove the redundant calls.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
As it's a single instance and local to the file, we don't need to pass
it as an argument.
Reviewed-by: Timofey Titovets <nefelim4ag@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The callback is trivial and we don't need the abstraction for our
purposes. Let's open code it.
Reviewed-by: Timofey Titovets <nefelim4ag@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The callback is trivial and we don't need the abstraction for our
purposes. Let's open code it and also make the array types explicit.
Reviewed-by: Timofey Titovets <nefelim4ag@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Remove unused arg 'holder' from parse_subvol_options(), which has been
forgotten to be cleaned in the commit b99beb110e2d ("btrfs: split
parse_early_options() in two").
Signed-off-by: Tomohiro Misono <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since setup_root_args() is not used anymore, just remove it.
Signed-off-by: Tomohiro Misono <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now parse_early_options() is used by both btrfs_mount() and
btrfs_mount_root(). However, the former only needs subvol related part
and the latter needs the others.
Therefore extract the subvol related parts from parse_early_options() and
move it to new parse function (parse_subvol_options()).
Signed-off-by: Tomohiro Misono <misono.tomohiro@jp.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Cleanup btrfs_mount() by using btrfs_mount_root(). This avoids getting
btrfs_mount() called twice in mount path.
Old btrfs_mount() will do:
0. VFS layer calls vfs_kern_mount() with registered file_system_type
(for btrfs, btrfs_fs_type). btrfs_mount() is called on the way.
1. btrfs_parse_early_options() parses "subvolid=" mount option and set the
value to subvol_objectid. Otherwise, subvol_objectid has the initial
value of 0
2. check subvol_objectid is 5 or not. Assume this time id is not 5, then
btrfs_mount() returns by calling mount_subvol()
3. In mount_subvol(), original mount options are modified to contain
"subvolid=0" in setup_root_args(). Then, vfs_kern_mount() is called with
btrfs_fs_type and new options
4. btrfs_mount() is called again
5. btrfs_parse_early_options() parses "subvolid=0" and set 5 (instead of 0)
to subvol_objectid
6. check subvol_objectid is 5 or not. This time id is 5 and mount_subvol()
is not called. btrfs_mount() finishes mounting a root
7. (in mount_subvol()) with using a return vale of vfs_kern_mount(), it
calls mount_subtree()
8. return subvolume's dentry
Reusing the same file_system_type (and btrfs_mount()) for vfs_kern_mount()
is the cause of complication.
Instead, new btrfs_mount() will do:
1. parse subvol id related options for later use in mount_subvol()
2. mount device's root by calling vfs_kern_mount() with
btrfs_root_fs_type, which is not registered to VFS by
register_filesystem(). As a result, btrfs_mount_root() is called
3. return by calling mount_subvol()
The code of 2. is moved from the first part of mount_subvol().
The semantics of device holder changes from btrfs_fs_type to
btrfs_root_fs_type and has to be used in all contexts. Otherwise we'd
get wrong results when mount and dev scan would not check the same
thing. (this has been found indendently and the fix is folded into this
patch)
Signed-off-by: Tomohiro Misono <misono.tomohiro@jp.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ fold the btrfs_control_ioctl fixup, extend the comment ]
Signed-off-by: David Sterba <dsterba@suse.com>
Add btrfs_mount_root() and new file_system_type for preparation of cleanup
of btrfs_mount(). Code path is not changed yet.
btrfs_mount_root() is almost the same as current btrfs_mount(), but doesn't
have subvolume related part.
Signed-off-by: Tomohiro Misono <misono.tomohiro@jp.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Functions called from extent_write_cache_pages used void* as generic
callback data, but all of them convert it to extent_page_data, or use it
directly.
Signed-off-by: David Sterba <dsterba@suse.com>
The function extent_write_cache_pages is modelled after
write_cache_pages which is a generic interface and the writepage
parameter makes sense there. In btrfs we know exactly which callback
we're going to use, so we can pass it directly.
Signed-off-by: David Sterba <dsterba@suse.com>
