mirror of
https://github.com/torvalds/linux.git
synced 2024-11-28 07:01:32 +00:00
3a08c2fd76
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>
81 lines
2.1 KiB
C
81 lines
2.1 KiB
C
/* Copyright (c) 2016 Facebook
|
|
*
|
|
* This program is free software; you can redistribute it and/or
|
|
* modify it under the terms of version 2 of the GNU General Public
|
|
* License as published by the Free Software Foundation.
|
|
*/
|
|
#ifndef __BPF_LRU_LIST_H_
|
|
#define __BPF_LRU_LIST_H_
|
|
|
|
#include <linux/list.h>
|
|
#include <linux/spinlock_types.h>
|
|
|
|
#define NR_BPF_LRU_LIST_T (3)
|
|
#define NR_BPF_LRU_LIST_COUNT (2)
|
|
#define NR_BPF_LRU_LOCAL_LIST_T (2)
|
|
#define BPF_LOCAL_LIST_T_OFFSET NR_BPF_LRU_LIST_T
|
|
|
|
enum bpf_lru_list_type {
|
|
BPF_LRU_LIST_T_ACTIVE,
|
|
BPF_LRU_LIST_T_INACTIVE,
|
|
BPF_LRU_LIST_T_FREE,
|
|
BPF_LRU_LOCAL_LIST_T_FREE,
|
|
BPF_LRU_LOCAL_LIST_T_PENDING,
|
|
};
|
|
|
|
struct bpf_lru_node {
|
|
struct list_head list;
|
|
u16 cpu;
|
|
u8 type;
|
|
u8 ref;
|
|
};
|
|
|
|
struct bpf_lru_list {
|
|
struct list_head lists[NR_BPF_LRU_LIST_T];
|
|
unsigned int counts[NR_BPF_LRU_LIST_COUNT];
|
|
/* The next inacitve list rotation starts from here */
|
|
struct list_head *next_inactive_rotation;
|
|
|
|
raw_spinlock_t lock ____cacheline_aligned_in_smp;
|
|
};
|
|
|
|
struct bpf_lru_locallist {
|
|
struct list_head lists[NR_BPF_LRU_LOCAL_LIST_T];
|
|
u16 next_steal;
|
|
raw_spinlock_t lock;
|
|
};
|
|
|
|
struct bpf_common_lru {
|
|
struct bpf_lru_list lru_list;
|
|
struct bpf_lru_locallist __percpu *local_list;
|
|
};
|
|
|
|
typedef bool (*del_from_htab_func)(void *arg, struct bpf_lru_node *node);
|
|
|
|
struct bpf_lru {
|
|
struct bpf_common_lru common_lru;
|
|
del_from_htab_func del_from_htab;
|
|
void *del_arg;
|
|
unsigned int hash_offset;
|
|
unsigned int nr_scans;
|
|
};
|
|
|
|
static inline void bpf_lru_node_set_ref(struct bpf_lru_node *node)
|
|
{
|
|
/* ref is an approximation on access frequency. It does not
|
|
* have to be very accurate. Hence, no protection is used.
|
|
*/
|
|
node->ref = 1;
|
|
}
|
|
|
|
int bpf_lru_init(struct bpf_lru *lru, u32 hash_offset,
|
|
del_from_htab_func del_from_htab, void *delete_arg);
|
|
void bpf_lru_populate(struct bpf_lru *lru, void *buf, u32 node_offset,
|
|
u32 elem_size, u32 nr_elems);
|
|
void bpf_lru_destroy(struct bpf_lru *lru);
|
|
struct bpf_lru_node *bpf_lru_pop_free(struct bpf_lru *lru, u32 hash);
|
|
void bpf_lru_push_free(struct bpf_lru *lru, struct bpf_lru_node *node);
|
|
void bpf_lru_promote(struct bpf_lru *lru, struct bpf_lru_node *node);
|
|
|
|
#endif
|