2019-05-29 01:10:09 +08:00
|
|
|
// SPDX-License-Identifier: GPL-2.0-only
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
/* Copyright (c) 2016 Facebook
|
|
|
|
|
*/
|
|
|
|
|
#include <linux/cpumask.h>
|
|
|
|
|
#include <linux/spinlock.h>
|
|
|
|
|
#include <linux/percpu.h>
|
|
|
|
|
|
|
|
|
|
#include "bpf_lru_list.h"
|
|
|
|
|
|
|
|
|
|
#define LOCAL_FREE_TARGET (128)
|
|
|
|
|
#define LOCAL_NR_SCANS LOCAL_FREE_TARGET
|
|
|
|
|
|
2017-04-15 01:30:29 +08:00
|
|
|
#define PERCPU_FREE_TARGET (4)
|
2016-11-12 02:55:07 +08:00
|
|
|
#define PERCPU_NR_SCANS PERCPU_FREE_TARGET
|
|
|
|
|
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
/* Helpers to get the local list index */
|
|
|
|
|
#define LOCAL_LIST_IDX(t) ((t) - BPF_LOCAL_LIST_T_OFFSET)
|
|
|
|
|
#define LOCAL_FREE_LIST_IDX LOCAL_LIST_IDX(BPF_LRU_LOCAL_LIST_T_FREE)
|
|
|
|
|
#define LOCAL_PENDING_LIST_IDX LOCAL_LIST_IDX(BPF_LRU_LOCAL_LIST_T_PENDING)
|
|
|
|
|
#define IS_LOCAL_LIST_TYPE(t) ((t) >= BPF_LOCAL_LIST_T_OFFSET)
|
|
|
|
|
|
|
|
|
|
static int get_next_cpu(int cpu)
|
|
|
|
|
{
|
|
|
|
|
cpu = cpumask_next(cpu, cpu_possible_mask);
|
|
|
|
|
if (cpu >= nr_cpu_ids)
|
|
|
|
|
cpu = cpumask_first(cpu_possible_mask);
|
|
|
|
|
return cpu;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Local list helpers */
|
|
|
|
|
static struct list_head *local_free_list(struct bpf_lru_locallist *loc_l)
|
|
|
|
|
{
|
|
|
|
|
return &loc_l->lists[LOCAL_FREE_LIST_IDX];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static struct list_head *local_pending_list(struct bpf_lru_locallist *loc_l)
|
|
|
|
|
{
|
|
|
|
|
return &loc_l->lists[LOCAL_PENDING_LIST_IDX];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* bpf_lru_node helpers */
|
|
|
|
|
static bool bpf_lru_node_is_ref(const struct bpf_lru_node *node)
|
|
|
|
|
{
|
2023-05-11 12:37:48 +08:00
|
|
|
return READ_ONCE(node->ref);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bpf_lru_node_clear_ref(struct bpf_lru_node *node)
|
|
|
|
|
{
|
|
|
|
|
WRITE_ONCE(node->ref, 0);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bpf_lru_list_count_inc(struct bpf_lru_list *l,
|
|
|
|
|
enum bpf_lru_list_type type)
|
|
|
|
|
{
|
|
|
|
|
if (type < NR_BPF_LRU_LIST_COUNT)
|
|
|
|
|
l->counts[type]++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bpf_lru_list_count_dec(struct bpf_lru_list *l,
|
|
|
|
|
enum bpf_lru_list_type type)
|
|
|
|
|
{
|
|
|
|
|
if (type < NR_BPF_LRU_LIST_COUNT)
|
|
|
|
|
l->counts[type]--;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void __bpf_lru_node_move_to_free(struct bpf_lru_list *l,
|
|
|
|
|
struct bpf_lru_node *node,
|
|
|
|
|
struct list_head *free_list,
|
|
|
|
|
enum bpf_lru_list_type tgt_free_type)
|
|
|
|
|
{
|
|
|
|
|
if (WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(node->type)))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
/* If the removing node is the next_inactive_rotation candidate,
|
|
|
|
|
* move the next_inactive_rotation pointer also.
|
|
|
|
|
*/
|
|
|
|
|
if (&node->list == l->next_inactive_rotation)
|
|
|
|
|
l->next_inactive_rotation = l->next_inactive_rotation->prev;
|
|
|
|
|
|
|
|
|
|
bpf_lru_list_count_dec(l, node->type);
|
|
|
|
|
|
|
|
|
|
node->type = tgt_free_type;
|
|
|
|
|
list_move(&node->list, free_list);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Move nodes from local list to the LRU list */
|
|
|
|
|
static void __bpf_lru_node_move_in(struct bpf_lru_list *l,
|
|
|
|
|
struct bpf_lru_node *node,
|
|
|
|
|
enum bpf_lru_list_type tgt_type)
|
|
|
|
|
{
|
|
|
|
|
if (WARN_ON_ONCE(!IS_LOCAL_LIST_TYPE(node->type)) ||
|
|
|
|
|
WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(tgt_type)))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
bpf_lru_list_count_inc(l, tgt_type);
|
|
|
|
|
node->type = tgt_type;
|
2023-05-11 12:37:48 +08:00
|
|
|
bpf_lru_node_clear_ref(node);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
list_move(&node->list, &l->lists[tgt_type]);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Move nodes between or within active and inactive list (like
|
|
|
|
|
* active to inactive, inactive to active or tail of active back to
|
|
|
|
|
* the head of active).
|
|
|
|
|
*/
|
|
|
|
|
static void __bpf_lru_node_move(struct bpf_lru_list *l,
|
|
|
|
|
struct bpf_lru_node *node,
|
|
|
|
|
enum bpf_lru_list_type tgt_type)
|
|
|
|
|
{
|
|
|
|
|
if (WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(node->type)) ||
|
|
|
|
|
WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(tgt_type)))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (node->type != tgt_type) {
|
|
|
|
|
bpf_lru_list_count_dec(l, node->type);
|
|
|
|
|
bpf_lru_list_count_inc(l, tgt_type);
|
|
|
|
|
node->type = tgt_type;
|
|
|
|
|
}
|
2023-05-11 12:37:48 +08:00
|
|
|
bpf_lru_node_clear_ref(node);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
|
|
|
|
|
/* If the moving node is the next_inactive_rotation candidate,
|
|
|
|
|
* move the next_inactive_rotation pointer also.
|
|
|
|
|
*/
|
|
|
|
|
if (&node->list == l->next_inactive_rotation)
|
|
|
|
|
l->next_inactive_rotation = l->next_inactive_rotation->prev;
|
|
|
|
|
|
|
|
|
|
list_move(&node->list, &l->lists[tgt_type]);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static bool bpf_lru_list_inactive_low(const struct bpf_lru_list *l)
|
|
|
|
|
{
|
|
|
|
|
return l->counts[BPF_LRU_LIST_T_INACTIVE] <
|
|
|
|
|
l->counts[BPF_LRU_LIST_T_ACTIVE];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Rotate the active list:
|
|
|
|
|
* 1. Start from tail
|
|
|
|
|
* 2. If the node has the ref bit set, it will be rotated
|
|
|
|
|
* back to the head of active list with the ref bit cleared.
|
|
|
|
|
* Give this node one more chance to survive in the active list.
|
|
|
|
|
* 3. If the ref bit is not set, move it to the head of the
|
|
|
|
|
* inactive list.
|
|
|
|
|
* 4. It will at most scan nr_scans nodes
|
|
|
|
|
*/
|
|
|
|
|
static void __bpf_lru_list_rotate_active(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_list *l)
|
|
|
|
|
{
|
|
|
|
|
struct list_head *active = &l->lists[BPF_LRU_LIST_T_ACTIVE];
|
|
|
|
|
struct bpf_lru_node *node, *tmp_node, *first_node;
|
|
|
|
|
unsigned int i = 0;
|
|
|
|
|
|
|
|
|
|
first_node = list_first_entry(active, struct bpf_lru_node, list);
|
|
|
|
|
list_for_each_entry_safe_reverse(node, tmp_node, active, list) {
|
|
|
|
|
if (bpf_lru_node_is_ref(node))
|
|
|
|
|
__bpf_lru_node_move(l, node, BPF_LRU_LIST_T_ACTIVE);
|
|
|
|
|
else
|
|
|
|
|
__bpf_lru_node_move(l, node, BPF_LRU_LIST_T_INACTIVE);
|
|
|
|
|
|
|
|
|
|
if (++i == lru->nr_scans || node == first_node)
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Rotate the inactive list. It starts from the next_inactive_rotation
|
|
|
|
|
* 1. If the node has ref bit set, it will be moved to the head
|
|
|
|
|
* of active list with the ref bit cleared.
|
|
|
|
|
* 2. If the node does not have ref bit set, it will leave it
|
|
|
|
|
* at its current location (i.e. do nothing) so that it can
|
|
|
|
|
* be considered during the next inactive_shrink.
|
|
|
|
|
* 3. It will at most scan nr_scans nodes
|
|
|
|
|
*/
|
|
|
|
|
static void __bpf_lru_list_rotate_inactive(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_list *l)
|
|
|
|
|
{
|
|
|
|
|
struct list_head *inactive = &l->lists[BPF_LRU_LIST_T_INACTIVE];
|
2016-11-16 03:00:04 +08:00
|
|
|
struct list_head *cur, *last, *next = inactive;
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
struct bpf_lru_node *node;
|
|
|
|
|
unsigned int i = 0;
|
|
|
|
|
|
|
|
|
|
if (list_empty(inactive))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
last = l->next_inactive_rotation->next;
|
|
|
|
|
if (last == inactive)
|
|
|
|
|
last = last->next;
|
|
|
|
|
|
|
|
|
|
cur = l->next_inactive_rotation;
|
|
|
|
|
while (i < lru->nr_scans) {
|
|
|
|
|
if (cur == inactive) {
|
|
|
|
|
cur = cur->prev;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
node = list_entry(cur, struct bpf_lru_node, list);
|
|
|
|
|
next = cur->prev;
|
|
|
|
|
if (bpf_lru_node_is_ref(node))
|
|
|
|
|
__bpf_lru_node_move(l, node, BPF_LRU_LIST_T_ACTIVE);
|
|
|
|
|
if (cur == last)
|
|
|
|
|
break;
|
|
|
|
|
cur = next;
|
|
|
|
|
i++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
l->next_inactive_rotation = next;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Shrink the inactive list. It starts from the tail of the
|
|
|
|
|
* inactive list and only move the nodes without the ref bit
|
|
|
|
|
* set to the designated free list.
|
|
|
|
|
*/
|
|
|
|
|
static unsigned int
|
|
|
|
|
__bpf_lru_list_shrink_inactive(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_list *l,
|
|
|
|
|
unsigned int tgt_nshrink,
|
|
|
|
|
struct list_head *free_list,
|
|
|
|
|
enum bpf_lru_list_type tgt_free_type)
|
|
|
|
|
{
|
|
|
|
|
struct list_head *inactive = &l->lists[BPF_LRU_LIST_T_INACTIVE];
|
2017-01-10 22:02:07 +08:00
|
|
|
struct bpf_lru_node *node, *tmp_node;
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
unsigned int nshrinked = 0;
|
|
|
|
|
unsigned int i = 0;
|
|
|
|
|
|
|
|
|
|
list_for_each_entry_safe_reverse(node, tmp_node, inactive, list) {
|
|
|
|
|
if (bpf_lru_node_is_ref(node)) {
|
|
|
|
|
__bpf_lru_node_move(l, node, BPF_LRU_LIST_T_ACTIVE);
|
|
|
|
|
} else if (lru->del_from_htab(lru->del_arg, node)) {
|
|
|
|
|
__bpf_lru_node_move_to_free(l, node, free_list,
|
|
|
|
|
tgt_free_type);
|
|
|
|
|
if (++nshrinked == tgt_nshrink)
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (++i == lru->nr_scans)
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return nshrinked;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* 1. Rotate the active list (if needed)
|
|
|
|
|
* 2. Always rotate the inactive list
|
|
|
|
|
*/
|
|
|
|
|
static void __bpf_lru_list_rotate(struct bpf_lru *lru, struct bpf_lru_list *l)
|
|
|
|
|
{
|
|
|
|
|
if (bpf_lru_list_inactive_low(l))
|
|
|
|
|
__bpf_lru_list_rotate_active(lru, l);
|
|
|
|
|
|
|
|
|
|
__bpf_lru_list_rotate_inactive(lru, l);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Calls __bpf_lru_list_shrink_inactive() to shrink some
|
|
|
|
|
* ref-bit-cleared nodes and move them to the designated
|
|
|
|
|
* free list.
|
|
|
|
|
*
|
|
|
|
|
* If it cannot get a free node after calling
|
|
|
|
|
* __bpf_lru_list_shrink_inactive(). It will just remove
|
|
|
|
|
* one node from either inactive or active list without
|
|
|
|
|
* honoring the ref-bit. It prefers inactive list to active
|
|
|
|
|
* list in this situation.
|
|
|
|
|
*/
|
|
|
|
|
static unsigned int __bpf_lru_list_shrink(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_list *l,
|
|
|
|
|
unsigned int tgt_nshrink,
|
|
|
|
|
struct list_head *free_list,
|
|
|
|
|
enum bpf_lru_list_type tgt_free_type)
|
|
|
|
|
|
|
|
|
|
{
|
|
|
|
|
struct bpf_lru_node *node, *tmp_node;
|
|
|
|
|
struct list_head *force_shrink_list;
|
|
|
|
|
unsigned int nshrinked;
|
|
|
|
|
|
|
|
|
|
nshrinked = __bpf_lru_list_shrink_inactive(lru, l, tgt_nshrink,
|
|
|
|
|
free_list, tgt_free_type);
|
|
|
|
|
if (nshrinked)
|
|
|
|
|
return nshrinked;
|
|
|
|
|
|
|
|
|
|
/* Do a force shrink by ignoring the reference bit */
|
|
|
|
|
if (!list_empty(&l->lists[BPF_LRU_LIST_T_INACTIVE]))
|
|
|
|
|
force_shrink_list = &l->lists[BPF_LRU_LIST_T_INACTIVE];
|
|
|
|
|
else
|
|
|
|
|
force_shrink_list = &l->lists[BPF_LRU_LIST_T_ACTIVE];
|
|
|
|
|
|
|
|
|
|
list_for_each_entry_safe_reverse(node, tmp_node, force_shrink_list,
|
|
|
|
|
list) {
|
|
|
|
|
if (lru->del_from_htab(lru->del_arg, node)) {
|
|
|
|
|
__bpf_lru_node_move_to_free(l, node, free_list,
|
|
|
|
|
tgt_free_type);
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Flush the nodes from the local pending list to the LRU list */
|
|
|
|
|
static void __local_list_flush(struct bpf_lru_list *l,
|
|
|
|
|
struct bpf_lru_locallist *loc_l)
|
|
|
|
|
{
|
|
|
|
|
struct bpf_lru_node *node, *tmp_node;
|
|
|
|
|
|
|
|
|
|
list_for_each_entry_safe_reverse(node, tmp_node,
|
|
|
|
|
local_pending_list(loc_l), list) {
|
|
|
|
|
if (bpf_lru_node_is_ref(node))
|
|
|
|
|
__bpf_lru_node_move_in(l, node, BPF_LRU_LIST_T_ACTIVE);
|
|
|
|
|
else
|
|
|
|
|
__bpf_lru_node_move_in(l, node,
|
|
|
|
|
BPF_LRU_LIST_T_INACTIVE);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bpf_lru_list_push_free(struct bpf_lru_list *l,
|
|
|
|
|
struct bpf_lru_node *node)
|
|
|
|
|
{
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
|
|
if (WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(node->type)))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_irqsave(&l->lock, flags);
|
|
|
|
|
__bpf_lru_node_move(l, node, BPF_LRU_LIST_T_FREE);
|
|
|
|
|
raw_spin_unlock_irqrestore(&l->lock, flags);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bpf_lru_list_pop_free_to_local(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_locallist *loc_l)
|
|
|
|
|
{
|
|
|
|
|
struct bpf_lru_list *l = &lru->common_lru.lru_list;
|
|
|
|
|
struct bpf_lru_node *node, *tmp_node;
|
|
|
|
|
unsigned int nfree = 0;
|
|
|
|
|
|
|
|
|
|
raw_spin_lock(&l->lock);
|
|
|
|
|
|
|
|
|
|
__local_list_flush(l, loc_l);
|
|
|
|
|
|
|
|
|
|
__bpf_lru_list_rotate(lru, l);
|
|
|
|
|
|
|
|
|
|
list_for_each_entry_safe(node, tmp_node, &l->lists[BPF_LRU_LIST_T_FREE],
|
|
|
|
|
list) {
|
|
|
|
|
__bpf_lru_node_move_to_free(l, node, local_free_list(loc_l),
|
|
|
|
|
BPF_LRU_LOCAL_LIST_T_FREE);
|
|
|
|
|
if (++nfree == LOCAL_FREE_TARGET)
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (nfree < LOCAL_FREE_TARGET)
|
|
|
|
|
__bpf_lru_list_shrink(lru, l, LOCAL_FREE_TARGET - nfree,
|
|
|
|
|
local_free_list(loc_l),
|
|
|
|
|
BPF_LRU_LOCAL_LIST_T_FREE);
|
|
|
|
|
|
|
|
|
|
raw_spin_unlock(&l->lock);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void __local_list_add_pending(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_locallist *loc_l,
|
|
|
|
|
int cpu,
|
|
|
|
|
struct bpf_lru_node *node,
|
|
|
|
|
u32 hash)
|
|
|
|
|
{
|
|
|
|
|
*(u32 *)((void *)node + lru->hash_offset) = hash;
|
|
|
|
|
node->cpu = cpu;
|
|
|
|
|
node->type = BPF_LRU_LOCAL_LIST_T_PENDING;
|
2023-05-11 12:37:48 +08:00
|
|
|
bpf_lru_node_clear_ref(node);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
list_add(&node->list, local_pending_list(loc_l));
|
|
|
|
|
}
|
|
|
|
|
|
2017-01-10 22:02:16 +08:00
|
|
|
static struct bpf_lru_node *
|
|
|
|
|
__local_list_pop_free(struct bpf_lru_locallist *loc_l)
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
{
|
|
|
|
|
struct bpf_lru_node *node;
|
|
|
|
|
|
|
|
|
|
node = list_first_entry_or_null(local_free_list(loc_l),
|
|
|
|
|
struct bpf_lru_node,
|
|
|
|
|
list);
|
|
|
|
|
if (node)
|
|
|
|
|
list_del(&node->list);
|
|
|
|
|
|
|
|
|
|
return node;
|
|
|
|
|
}
|
|
|
|
|
|
2017-01-10 22:02:16 +08:00
|
|
|
static struct bpf_lru_node *
|
|
|
|
|
__local_list_pop_pending(struct bpf_lru *lru, struct bpf_lru_locallist *loc_l)
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
{
|
|
|
|
|
struct bpf_lru_node *node;
|
|
|
|
|
bool force = false;
|
|
|
|
|
|
|
|
|
|
ignore_ref:
|
|
|
|
|
/* Get from the tail (i.e. older element) of the pending list. */
|
|
|
|
|
list_for_each_entry_reverse(node, local_pending_list(loc_l),
|
|
|
|
|
list) {
|
|
|
|
|
if ((!bpf_lru_node_is_ref(node) || force) &&
|
|
|
|
|
lru->del_from_htab(lru->del_arg, node)) {
|
|
|
|
|
list_del(&node->list);
|
|
|
|
|
return node;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!force) {
|
|
|
|
|
force = true;
|
|
|
|
|
goto ignore_ref;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
static struct bpf_lru_node *bpf_percpu_lru_pop_free(struct bpf_lru *lru,
|
|
|
|
|
u32 hash)
|
|
|
|
|
{
|
|
|
|
|
struct list_head *free_list;
|
|
|
|
|
struct bpf_lru_node *node = NULL;
|
|
|
|
|
struct bpf_lru_list *l;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
int cpu = raw_smp_processor_id();
|
|
|
|
|
|
|
|
|
|
l = per_cpu_ptr(lru->percpu_lru, cpu);
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_irqsave(&l->lock, flags);
|
|
|
|
|
|
|
|
|
|
__bpf_lru_list_rotate(lru, l);
|
|
|
|
|
|
|
|
|
|
free_list = &l->lists[BPF_LRU_LIST_T_FREE];
|
|
|
|
|
if (list_empty(free_list))
|
|
|
|
|
__bpf_lru_list_shrink(lru, l, PERCPU_FREE_TARGET, free_list,
|
|
|
|
|
BPF_LRU_LIST_T_FREE);
|
|
|
|
|
|
|
|
|
|
if (!list_empty(free_list)) {
|
|
|
|
|
node = list_first_entry(free_list, struct bpf_lru_node, list);
|
|
|
|
|
*(u32 *)((void *)node + lru->hash_offset) = hash;
|
2023-05-11 12:37:48 +08:00
|
|
|
bpf_lru_node_clear_ref(node);
|
2016-11-12 02:55:07 +08:00
|
|
|
__bpf_lru_node_move(l, node, BPF_LRU_LIST_T_INACTIVE);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
raw_spin_unlock_irqrestore(&l->lock, flags);
|
|
|
|
|
|
|
|
|
|
return node;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static struct bpf_lru_node *bpf_common_lru_pop_free(struct bpf_lru *lru,
|
|
|
|
|
u32 hash)
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
{
|
|
|
|
|
struct bpf_lru_locallist *loc_l, *steal_loc_l;
|
|
|
|
|
struct bpf_common_lru *clru = &lru->common_lru;
|
|
|
|
|
struct bpf_lru_node *node;
|
|
|
|
|
int steal, first_steal;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
int cpu = raw_smp_processor_id();
|
|
|
|
|
|
|
|
|
|
loc_l = per_cpu_ptr(clru->local_list, cpu);
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_irqsave(&loc_l->lock, flags);
|
|
|
|
|
|
|
|
|
|
node = __local_list_pop_free(loc_l);
|
|
|
|
|
if (!node) {
|
|
|
|
|
bpf_lru_list_pop_free_to_local(lru, loc_l);
|
|
|
|
|
node = __local_list_pop_free(loc_l);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (node)
|
|
|
|
|
__local_list_add_pending(lru, loc_l, cpu, node, hash);
|
|
|
|
|
|
|
|
|
|
raw_spin_unlock_irqrestore(&loc_l->lock, flags);
|
|
|
|
|
|
|
|
|
|
if (node)
|
|
|
|
|
return node;
|
|
|
|
|
|
|
|
|
|
/* No free nodes found from the local free list and
|
|
|
|
|
* the global LRU list.
|
|
|
|
|
*
|
|
|
|
|
* Steal from the local free/pending list of the
|
|
|
|
|
* current CPU and remote CPU in RR. It starts
|
|
|
|
|
* with the loc_l->next_steal CPU.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
first_steal = loc_l->next_steal;
|
|
|
|
|
steal = first_steal;
|
|
|
|
|
do {
|
|
|
|
|
steal_loc_l = per_cpu_ptr(clru->local_list, steal);
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_irqsave(&steal_loc_l->lock, flags);
|
|
|
|
|
|
|
|
|
|
node = __local_list_pop_free(steal_loc_l);
|
|
|
|
|
if (!node)
|
|
|
|
|
node = __local_list_pop_pending(lru, steal_loc_l);
|
|
|
|
|
|
|
|
|
|
raw_spin_unlock_irqrestore(&steal_loc_l->lock, flags);
|
|
|
|
|
|
|
|
|
|
steal = get_next_cpu(steal);
|
|
|
|
|
} while (!node && steal != first_steal);
|
|
|
|
|
|
|
|
|
|
loc_l->next_steal = steal;
|
|
|
|
|
|
|
|
|
|
if (node) {
|
|
|
|
|
raw_spin_lock_irqsave(&loc_l->lock, flags);
|
|
|
|
|
__local_list_add_pending(lru, loc_l, cpu, node, hash);
|
|
|
|
|
raw_spin_unlock_irqrestore(&loc_l->lock, flags);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return node;
|
|
|
|
|
}
|
|
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
struct bpf_lru_node *bpf_lru_pop_free(struct bpf_lru *lru, u32 hash)
|
|
|
|
|
{
|
|
|
|
|
if (lru->percpu)
|
|
|
|
|
return bpf_percpu_lru_pop_free(lru, hash);
|
|
|
|
|
else
|
|
|
|
|
return bpf_common_lru_pop_free(lru, hash);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bpf_common_lru_push_free(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_node *node)
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
{
|
2021-02-09 19:27:01 +08:00
|
|
|
u8 node_type = READ_ONCE(node->type);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
unsigned long flags;
|
|
|
|
|
|
2021-02-09 19:27:01 +08:00
|
|
|
if (WARN_ON_ONCE(node_type == BPF_LRU_LIST_T_FREE) ||
|
|
|
|
|
WARN_ON_ONCE(node_type == BPF_LRU_LOCAL_LIST_T_FREE))
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
return;
|
|
|
|
|
|
2021-02-09 19:27:01 +08:00
|
|
|
if (node_type == BPF_LRU_LOCAL_LIST_T_PENDING) {
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
struct bpf_lru_locallist *loc_l;
|
|
|
|
|
|
|
|
|
|
loc_l = per_cpu_ptr(lru->common_lru.local_list, node->cpu);
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_irqsave(&loc_l->lock, flags);
|
|
|
|
|
|
|
|
|
|
if (unlikely(node->type != BPF_LRU_LOCAL_LIST_T_PENDING)) {
|
|
|
|
|
raw_spin_unlock_irqrestore(&loc_l->lock, flags);
|
|
|
|
|
goto check_lru_list;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
node->type = BPF_LRU_LOCAL_LIST_T_FREE;
|
2023-05-11 12:37:48 +08:00
|
|
|
bpf_lru_node_clear_ref(node);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
list_move(&node->list, local_free_list(loc_l));
|
|
|
|
|
|
|
|
|
|
raw_spin_unlock_irqrestore(&loc_l->lock, flags);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
check_lru_list:
|
|
|
|
|
bpf_lru_list_push_free(&lru->common_lru.lru_list, node);
|
|
|
|
|
}
|
|
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
static void bpf_percpu_lru_push_free(struct bpf_lru *lru,
|
|
|
|
|
struct bpf_lru_node *node)
|
|
|
|
|
{
|
|
|
|
|
struct bpf_lru_list *l;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
|
|
l = per_cpu_ptr(lru->percpu_lru, node->cpu);
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_irqsave(&l->lock, flags);
|
|
|
|
|
|
|
|
|
|
__bpf_lru_node_move(l, node, BPF_LRU_LIST_T_FREE);
|
|
|
|
|
|
|
|
|
|
raw_spin_unlock_irqrestore(&l->lock, flags);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void bpf_lru_push_free(struct bpf_lru *lru, struct bpf_lru_node *node)
|
|
|
|
|
{
|
|
|
|
|
if (lru->percpu)
|
|
|
|
|
bpf_percpu_lru_push_free(lru, node);
|
|
|
|
|
else
|
|
|
|
|
bpf_common_lru_push_free(lru, node);
|
|
|
|
|
}
|
|
|
|
|
|
2017-01-10 22:02:16 +08:00
|
|
|
static void bpf_common_lru_populate(struct bpf_lru *lru, void *buf,
|
|
|
|
|
u32 node_offset, u32 elem_size,
|
|
|
|
|
u32 nr_elems)
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
{
|
|
|
|
|
struct bpf_lru_list *l = &lru->common_lru.lru_list;
|
|
|
|
|
u32 i;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < nr_elems; i++) {
|
|
|
|
|
struct bpf_lru_node *node;
|
|
|
|
|
|
|
|
|
|
node = (struct bpf_lru_node *)(buf + node_offset);
|
|
|
|
|
node->type = BPF_LRU_LIST_T_FREE;
|
2023-05-11 12:37:48 +08:00
|
|
|
bpf_lru_node_clear_ref(node);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
list_add(&node->list, &l->lists[BPF_LRU_LIST_T_FREE]);
|
|
|
|
|
buf += elem_size;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2017-01-10 22:02:16 +08:00
|
|
|
static void bpf_percpu_lru_populate(struct bpf_lru *lru, void *buf,
|
|
|
|
|
u32 node_offset, u32 elem_size,
|
|
|
|
|
u32 nr_elems)
|
2016-11-12 02:55:07 +08:00
|
|
|
{
|
|
|
|
|
u32 i, pcpu_entries;
|
|
|
|
|
int cpu;
|
|
|
|
|
struct bpf_lru_list *l;
|
|
|
|
|
|
|
|
|
|
pcpu_entries = nr_elems / num_possible_cpus();
|
|
|
|
|
|
|
|
|
|
i = 0;
|
|
|
|
|
|
|
|
|
|
for_each_possible_cpu(cpu) {
|
|
|
|
|
struct bpf_lru_node *node;
|
|
|
|
|
|
|
|
|
|
l = per_cpu_ptr(lru->percpu_lru, cpu);
|
|
|
|
|
again:
|
|
|
|
|
node = (struct bpf_lru_node *)(buf + node_offset);
|
|
|
|
|
node->cpu = cpu;
|
|
|
|
|
node->type = BPF_LRU_LIST_T_FREE;
|
2023-05-11 12:37:48 +08:00
|
|
|
bpf_lru_node_clear_ref(node);
|
2016-11-12 02:55:07 +08:00
|
|
|
list_add(&node->list, &l->lists[BPF_LRU_LIST_T_FREE]);
|
|
|
|
|
i++;
|
|
|
|
|
buf += elem_size;
|
|
|
|
|
if (i == nr_elems)
|
|
|
|
|
break;
|
|
|
|
|
if (i % pcpu_entries)
|
|
|
|
|
goto again;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void bpf_lru_populate(struct bpf_lru *lru, void *buf, u32 node_offset,
|
|
|
|
|
u32 elem_size, u32 nr_elems)
|
|
|
|
|
{
|
|
|
|
|
if (lru->percpu)
|
|
|
|
|
bpf_percpu_lru_populate(lru, buf, node_offset, elem_size,
|
|
|
|
|
nr_elems);
|
|
|
|
|
else
|
|
|
|
|
bpf_common_lru_populate(lru, buf, node_offset, elem_size,
|
|
|
|
|
nr_elems);
|
|
|
|
|
}
|
|
|
|
|
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
static void bpf_lru_locallist_init(struct bpf_lru_locallist *loc_l, int cpu)
|
|
|
|
|
{
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < NR_BPF_LRU_LOCAL_LIST_T; i++)
|
|
|
|
|
INIT_LIST_HEAD(&loc_l->lists[i]);
|
|
|
|
|
|
|
|
|
|
loc_l->next_steal = cpu;
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_init(&loc_l->lock);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bpf_lru_list_init(struct bpf_lru_list *l)
|
|
|
|
|
{
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < NR_BPF_LRU_LIST_T; i++)
|
|
|
|
|
INIT_LIST_HEAD(&l->lists[i]);
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < NR_BPF_LRU_LIST_COUNT; i++)
|
|
|
|
|
l->counts[i] = 0;
|
|
|
|
|
|
|
|
|
|
l->next_inactive_rotation = &l->lists[BPF_LRU_LIST_T_INACTIVE];
|
|
|
|
|
|
|
|
|
|
raw_spin_lock_init(&l->lock);
|
|
|
|
|
}
|
|
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
int bpf_lru_init(struct bpf_lru *lru, bool percpu, u32 hash_offset,
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
del_from_htab_func del_from_htab, void *del_arg)
|
|
|
|
|
{
|
|
|
|
|
int cpu;
|
|
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
if (percpu) {
|
|
|
|
|
lru->percpu_lru = alloc_percpu(struct bpf_lru_list);
|
|
|
|
|
if (!lru->percpu_lru)
|
|
|
|
|
return -ENOMEM;
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
for_each_possible_cpu(cpu) {
|
|
|
|
|
struct bpf_lru_list *l;
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
l = per_cpu_ptr(lru->percpu_lru, cpu);
|
|
|
|
|
bpf_lru_list_init(l);
|
|
|
|
|
}
|
|
|
|
|
lru->nr_scans = PERCPU_NR_SCANS;
|
|
|
|
|
} else {
|
|
|
|
|
struct bpf_common_lru *clru = &lru->common_lru;
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
clru->local_list = alloc_percpu(struct bpf_lru_locallist);
|
|
|
|
|
if (!clru->local_list)
|
|
|
|
|
return -ENOMEM;
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
|
2016-11-12 02:55:07 +08:00
|
|
|
for_each_possible_cpu(cpu) {
|
|
|
|
|
struct bpf_lru_locallist *loc_l;
|
|
|
|
|
|
|
|
|
|
loc_l = per_cpu_ptr(clru->local_list, cpu);
|
|
|
|
|
bpf_lru_locallist_init(loc_l, cpu);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
bpf_lru_list_init(&clru->lru_list);
|
|
|
|
|
lru->nr_scans = LOCAL_NR_SCANS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
lru->percpu = percpu;
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
lru->del_from_htab = del_from_htab;
|
|
|
|
|
lru->del_arg = del_arg;
|
|
|
|
|
lru->hash_offset = hash_offset;
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void bpf_lru_destroy(struct bpf_lru *lru)
|
|
|
|
|
{
|
2016-11-12 02:55:07 +08:00
|
|
|
if (lru->percpu)
|
|
|
|
|
free_percpu(lru->percpu_lru);
|
|
|
|
|
else
|
|
|
|
|
free_percpu(lru->common_lru.local_list);
|
bpf: LRU List
Introduce bpf_lru_list which will provide LRU capability to
the bpf_htab in the later patch.
* General Thoughts:
1. Target use case. Read is more often than update.
(i.e. bpf_lookup_elem() is more often than bpf_update_elem()).
If bpf_prog does a bpf_lookup_elem() first and then an in-place
update, it still counts as a read operation to the LRU list concern.
2. It may be useful to think of it as a LRU cache
3. Optimize the read case
3.1 No lock in read case
3.2 The LRU maintenance is only done during bpf_update_elem()
4. If there is a percpu LRU list, it will lose the system-wise LRU
property. A completely isolated percpu LRU list has the best
performance but the memory utilization is not ideal considering
the work load may be imbalance.
5. Hence, this patch starts the LRU implementation with a global LRU
list with batched operations before accessing the global LRU list.
As a LRU cache, #read >> #update/#insert operations, it will work well.
6. There is a local list (for each cpu) which is named
'struct bpf_lru_locallist'. This local list is not used to sort
the LRU property. Instead, the local list is to batch enough
operations before acquiring the lock of the global LRU list. More
details on this later.
7. In the later patch, it allows a percpu LRU list by specifying a
map-attribute for scalability reason and for use cases that need to
prepare for the worst (and pathological) case like DoS attack.
The percpu LRU list is completely isolated from each other and the
LRU nodes (including free nodes) cannot be moved across the list. The
following description is for the global LRU list but mostly applicable
to the percpu LRU list also.
* Global LRU List:
1. It has three sub-lists: active-list, inactive-list and free-list.
2. The two list idea, active and inactive, is borrowed from the
page cache.
3. All nodes are pre-allocated and all sit at the free-list (of the
global LRU list) at the beginning. The pre-allocation reasoning
is similar to the existing BPF_MAP_TYPE_HASH. However,
opting-out prealloc (BPF_F_NO_PREALLOC) is not supported in
the LRU map.
* Active/Inactive List (of the global LRU list):
1. The active list, as its name says it, maintains the active set of
the nodes. We can think of it as the working set or more frequently
accessed nodes. The access frequency is approximated by a ref-bit.
The ref-bit is set during the bpf_lookup_elem().
2. The inactive list, as its name also says it, maintains a less
active set of nodes. They are the candidates to be removed
from the bpf_htab when we are running out of free nodes.
3. The ordering of these two lists is acting as a rough clock.
The tail of the inactive list is the older nodes and
should be released first if the bpf_htab needs free element.
* Rotating the Active/Inactive List (of the global LRU list):
1. It is the basic operation to maintain the LRU property of
the global list.
2. The active list is only rotated when the inactive list is running
low. This idea is similar to the current page cache.
Inactive running low is currently defined as
"# of inactive < # of active".
3. The active list rotation always starts from the tail. It moves
node without ref-bit set to the head of the inactive list.
It moves node with ref-bit set back to the head of the active
list and then clears its ref-bit.
4. The inactive rotation is pretty simply.
It walks the inactive list and moves the nodes back to the head of
active list if its ref-bit is set. The ref-bit is cleared after moving
to the active list.
If the node does not have ref-bit set, it just leave it as it is
because it is already in the inactive list.
* Shrinking the Inactive List (of the global LRU list):
1. Shrinking is the operation to get free nodes when the bpf_htab is
full.
2. It usually only shrinks the inactive list to get free nodes.
3. During shrinking, it will walk the inactive list from the tail,
delete the nodes without ref-bit set from bpf_htab.
4. If no free node found after step (3), it will forcefully get
one node from the tail of inactive or active list. Forcefully is
in the sense that it ignores the ref-bit.
* Local List:
1. Each CPU has a 'struct bpf_lru_locallist'. The purpose is to
batch enough operations before acquiring the lock of the
global LRU.
2. A local list has two sub-lists, free-list and pending-list.
3. During bpf_update_elem(), it will try to get from the free-list
of (the current CPU local list).
4. If the local free-list is empty, it will acquire from the
global LRU list. The global LRU list can either satisfy it
by its global free-list or by shrinking the global inactive
list. Since we have acquired the global LRU list lock,
it will try to get at most LOCAL_FREE_TARGET elements
to the local free list.
5. When a new element is added to the bpf_htab, it will
first sit at the pending-list (of the local list) first.
The pending-list will be flushed to the global LRU list
when it needs to acquire free nodes from the global list
next time.
* Lock Consideration:
The LRU list has a lock (lru_lock). Each bucket of htab has a
lock (buck_lock). If both locks need to be acquired together,
the lock order is always lru_lock -> buck_lock and this only
happens in the bpf_lru_list.c logic.
In hashtab.c, both locks are not acquired together (i.e. one
lock is always released first before acquiring another lock).
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-12 02:55:06 +08:00
|
|
|
}
|
