linux/fs/nfsd/nfscache.c
Trond Myklebust 736c6625de knfsd: Improve lookup performance in the duplicate reply cache using an rbtree
Use an rbtree to ensure the lookup/insert of an entry in a DRC bucket is
O(log(N)).

Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2018-10-29 16:58:04 -04:00

603 lines
15 KiB
C

// SPDX-License-Identifier: GPL-2.0
/*
* Request reply cache. This is currently a global cache, but this may
* change in the future and be a per-client cache.
*
* This code is heavily inspired by the 44BSD implementation, although
* it does things a bit differently.
*
* Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
*/
#include <linux/slab.h>
#include <linux/vmalloc.h>
#include <linux/sunrpc/addr.h>
#include <linux/highmem.h>
#include <linux/log2.h>
#include <linux/hash.h>
#include <net/checksum.h>
#include "nfsd.h"
#include "cache.h"
#define NFSDDBG_FACILITY NFSDDBG_REPCACHE
/*
* We use this value to determine the number of hash buckets from the max
* cache size, the idea being that when the cache is at its maximum number
* of entries, then this should be the average number of entries per bucket.
*/
#define TARGET_BUCKET_SIZE 64
struct nfsd_drc_bucket {
struct rb_root rb_head;
struct list_head lru_head;
spinlock_t cache_lock;
};
static struct nfsd_drc_bucket *drc_hashtbl;
static struct kmem_cache *drc_slab;
/* max number of entries allowed in the cache */
static unsigned int max_drc_entries;
/* number of significant bits in the hash value */
static unsigned int maskbits;
static unsigned int drc_hashsize;
/*
* Stats and other tracking of on the duplicate reply cache. All of these and
* the "rc" fields in nfsdstats are protected by the cache_lock
*/
/* total number of entries */
static atomic_t num_drc_entries;
/* cache misses due only to checksum comparison failures */
static unsigned int payload_misses;
/* amount of memory (in bytes) currently consumed by the DRC */
static unsigned int drc_mem_usage;
/* longest hash chain seen */
static unsigned int longest_chain;
/* size of cache when we saw the longest hash chain */
static unsigned int longest_chain_cachesize;
static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
struct shrink_control *sc);
static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
struct shrink_control *sc);
static struct shrinker nfsd_reply_cache_shrinker = {
.scan_objects = nfsd_reply_cache_scan,
.count_objects = nfsd_reply_cache_count,
.seeks = 1,
};
/*
* Put a cap on the size of the DRC based on the amount of available
* low memory in the machine.
*
* 64MB: 8192
* 128MB: 11585
* 256MB: 16384
* 512MB: 23170
* 1GB: 32768
* 2GB: 46340
* 4GB: 65536
* 8GB: 92681
* 16GB: 131072
*
* ...with a hard cap of 256k entries. In the worst case, each entry will be
* ~1k, so the above numbers should give a rough max of the amount of memory
* used in k.
*/
static unsigned int
nfsd_cache_size_limit(void)
{
unsigned int limit;
unsigned long low_pages = totalram_pages - totalhigh_pages;
limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
return min_t(unsigned int, limit, 256*1024);
}
/*
* Compute the number of hash buckets we need. Divide the max cachesize by
* the "target" max bucket size, and round up to next power of two.
*/
static unsigned int
nfsd_hashsize(unsigned int limit)
{
return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
}
static u32
nfsd_cache_hash(__be32 xid)
{
return hash_32(be32_to_cpu(xid), maskbits);
}
static struct svc_cacherep *
nfsd_reply_cache_alloc(struct svc_rqst *rqstp, __wsum csum)
{
struct svc_cacherep *rp;
rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
if (rp) {
rp->c_state = RC_UNUSED;
rp->c_type = RC_NOCACHE;
RB_CLEAR_NODE(&rp->c_node);
INIT_LIST_HEAD(&rp->c_lru);
memset(&rp->c_key, 0, sizeof(rp->c_key));
rp->c_key.k_xid = rqstp->rq_xid;
rp->c_key.k_proc = rqstp->rq_proc;
rpc_copy_addr((struct sockaddr *)&rp->c_key.k_addr, svc_addr(rqstp));
rpc_set_port((struct sockaddr *)&rp->c_key.k_addr, rpc_get_port(svc_addr(rqstp)));
rp->c_key.k_prot = rqstp->rq_prot;
rp->c_key.k_vers = rqstp->rq_vers;
rp->c_key.k_len = rqstp->rq_arg.len;
rp->c_key.k_csum = csum;
}
return rp;
}
static void
nfsd_reply_cache_free_locked(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
{
if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
drc_mem_usage -= rp->c_replvec.iov_len;
kfree(rp->c_replvec.iov_base);
}
if (rp->c_state != RC_UNUSED) {
rb_erase(&rp->c_node, &b->rb_head);
list_del(&rp->c_lru);
atomic_dec(&num_drc_entries);
drc_mem_usage -= sizeof(*rp);
}
kmem_cache_free(drc_slab, rp);
}
static void
nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
{
spin_lock(&b->cache_lock);
nfsd_reply_cache_free_locked(b, rp);
spin_unlock(&b->cache_lock);
}
int nfsd_reply_cache_init(void)
{
unsigned int hashsize;
unsigned int i;
int status = 0;
max_drc_entries = nfsd_cache_size_limit();
atomic_set(&num_drc_entries, 0);
hashsize = nfsd_hashsize(max_drc_entries);
maskbits = ilog2(hashsize);
status = register_shrinker(&nfsd_reply_cache_shrinker);
if (status)
return status;
drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
0, 0, NULL);
if (!drc_slab)
goto out_nomem;
drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
if (!drc_hashtbl) {
drc_hashtbl = vzalloc(array_size(hashsize,
sizeof(*drc_hashtbl)));
if (!drc_hashtbl)
goto out_nomem;
}
for (i = 0; i < hashsize; i++) {
INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
spin_lock_init(&drc_hashtbl[i].cache_lock);
}
drc_hashsize = hashsize;
return 0;
out_nomem:
printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
nfsd_reply_cache_shutdown();
return -ENOMEM;
}
void nfsd_reply_cache_shutdown(void)
{
struct svc_cacherep *rp;
unsigned int i;
unregister_shrinker(&nfsd_reply_cache_shrinker);
for (i = 0; i < drc_hashsize; i++) {
struct list_head *head = &drc_hashtbl[i].lru_head;
while (!list_empty(head)) {
rp = list_first_entry(head, struct svc_cacherep, c_lru);
nfsd_reply_cache_free_locked(&drc_hashtbl[i], rp);
}
}
kvfree(drc_hashtbl);
drc_hashtbl = NULL;
drc_hashsize = 0;
kmem_cache_destroy(drc_slab);
drc_slab = NULL;
}
/*
* Move cache entry to end of LRU list, and queue the cleaner to run if it's
* not already scheduled.
*/
static void
lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
{
rp->c_timestamp = jiffies;
list_move_tail(&rp->c_lru, &b->lru_head);
}
static long
prune_bucket(struct nfsd_drc_bucket *b)
{
struct svc_cacherep *rp, *tmp;
long freed = 0;
list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
/*
* Don't free entries attached to calls that are still
* in-progress, but do keep scanning the list.
*/
if (rp->c_state == RC_INPROG)
continue;
if (atomic_read(&num_drc_entries) <= max_drc_entries &&
time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
break;
nfsd_reply_cache_free_locked(b, rp);
freed++;
}
return freed;
}
/*
* Walk the LRU list and prune off entries that are older than RC_EXPIRE.
* Also prune the oldest ones when the total exceeds the max number of entries.
*/
static long
prune_cache_entries(void)
{
unsigned int i;
long freed = 0;
for (i = 0; i < drc_hashsize; i++) {
struct nfsd_drc_bucket *b = &drc_hashtbl[i];
if (list_empty(&b->lru_head))
continue;
spin_lock(&b->cache_lock);
freed += prune_bucket(b);
spin_unlock(&b->cache_lock);
}
return freed;
}
static unsigned long
nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
{
return atomic_read(&num_drc_entries);
}
static unsigned long
nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
{
return prune_cache_entries();
}
/*
* Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
*/
static __wsum
nfsd_cache_csum(struct svc_rqst *rqstp)
{
int idx;
unsigned int base;
__wsum csum;
struct xdr_buf *buf = &rqstp->rq_arg;
const unsigned char *p = buf->head[0].iov_base;
size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
RC_CSUMLEN);
size_t len = min(buf->head[0].iov_len, csum_len);
/* rq_arg.head first */
csum = csum_partial(p, len, 0);
csum_len -= len;
/* Continue into page array */
idx = buf->page_base / PAGE_SIZE;
base = buf->page_base & ~PAGE_MASK;
while (csum_len) {
p = page_address(buf->pages[idx]) + base;
len = min_t(size_t, PAGE_SIZE - base, csum_len);
csum = csum_partial(p, len, csum);
csum_len -= len;
base = 0;
++idx;
}
return csum;
}
static int
nfsd_cache_key_cmp(const struct svc_cacherep *key, const struct svc_cacherep *rp)
{
if (key->c_key.k_xid == rp->c_key.k_xid &&
key->c_key.k_csum != rp->c_key.k_csum)
++payload_misses;
return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
}
/*
* Search the request hash for an entry that matches the given rqstp.
* Must be called with cache_lock held. Returns the found entry or
* inserts an empty key on failure.
*/
static struct svc_cacherep *
nfsd_cache_insert(struct nfsd_drc_bucket *b, struct svc_cacherep *key)
{
struct svc_cacherep *rp, *ret = key;
struct rb_node **p = &b->rb_head.rb_node,
*parent = NULL;
unsigned int entries = 0;
int cmp;
while (*p != NULL) {
++entries;
parent = *p;
rp = rb_entry(parent, struct svc_cacherep, c_node);
cmp = nfsd_cache_key_cmp(key, rp);
if (cmp < 0)
p = &parent->rb_left;
else if (cmp > 0)
p = &parent->rb_right;
else {
ret = rp;
goto out;
}
}
rb_link_node(&key->c_node, parent, p);
rb_insert_color(&key->c_node, &b->rb_head);
out:
/* tally hash chain length stats */
if (entries > longest_chain) {
longest_chain = entries;
longest_chain_cachesize = atomic_read(&num_drc_entries);
} else if (entries == longest_chain) {
/* prefer to keep the smallest cachesize possible here */
longest_chain_cachesize = min_t(unsigned int,
longest_chain_cachesize,
atomic_read(&num_drc_entries));
}
lru_put_end(b, ret);
return ret;
}
/*
* Try to find an entry matching the current call in the cache. When none
* is found, we try to grab the oldest expired entry off the LRU list. If
* a suitable one isn't there, then drop the cache_lock and allocate a
* new one, then search again in case one got inserted while this thread
* didn't hold the lock.
*/
int
nfsd_cache_lookup(struct svc_rqst *rqstp)
{
struct svc_cacherep *rp, *found;
__be32 xid = rqstp->rq_xid;
__wsum csum;
u32 hash = nfsd_cache_hash(xid);
struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
int type = rqstp->rq_cachetype;
int rtn = RC_DOIT;
rqstp->rq_cacherep = NULL;
if (type == RC_NOCACHE) {
nfsdstats.rcnocache++;
return rtn;
}
csum = nfsd_cache_csum(rqstp);
/*
* Since the common case is a cache miss followed by an insert,
* preallocate an entry.
*/
rp = nfsd_reply_cache_alloc(rqstp, csum);
if (!rp) {
dprintk("nfsd: unable to allocate DRC entry!\n");
return rtn;
}
spin_lock(&b->cache_lock);
found = nfsd_cache_insert(b, rp);
if (found != rp) {
nfsd_reply_cache_free_locked(NULL, rp);
rp = found;
goto found_entry;
}
nfsdstats.rcmisses++;
rqstp->rq_cacherep = rp;
rp->c_state = RC_INPROG;
atomic_inc(&num_drc_entries);
drc_mem_usage += sizeof(*rp);
/* go ahead and prune the cache */
prune_bucket(b);
out:
spin_unlock(&b->cache_lock);
return rtn;
found_entry:
/* We found a matching entry which is either in progress or done. */
nfsdstats.rchits++;
rtn = RC_DROPIT;
/* Request being processed */
if (rp->c_state == RC_INPROG)
goto out;
/* From the hall of fame of impractical attacks:
* Is this a user who tries to snoop on the cache? */
rtn = RC_DOIT;
if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
goto out;
/* Compose RPC reply header */
switch (rp->c_type) {
case RC_NOCACHE:
break;
case RC_REPLSTAT:
svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
rtn = RC_REPLY;
break;
case RC_REPLBUFF:
if (!nfsd_cache_append(rqstp, &rp->c_replvec))
goto out; /* should not happen */
rtn = RC_REPLY;
break;
default:
printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
nfsd_reply_cache_free_locked(b, rp);
}
goto out;
}
/*
* Update a cache entry. This is called from nfsd_dispatch when
* the procedure has been executed and the complete reply is in
* rqstp->rq_res.
*
* We're copying around data here rather than swapping buffers because
* the toplevel loop requires max-sized buffers, which would be a waste
* of memory for a cache with a max reply size of 100 bytes (diropokres).
*
* If we should start to use different types of cache entries tailored
* specifically for attrstat and fh's, we may save even more space.
*
* Also note that a cachetype of RC_NOCACHE can legally be passed when
* nfsd failed to encode a reply that otherwise would have been cached.
* In this case, nfsd_cache_update is called with statp == NULL.
*/
void
nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
{
struct svc_cacherep *rp = rqstp->rq_cacherep;
struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
u32 hash;
struct nfsd_drc_bucket *b;
int len;
size_t bufsize = 0;
if (!rp)
return;
hash = nfsd_cache_hash(rp->c_key.k_xid);
b = &drc_hashtbl[hash];
len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
len >>= 2;
/* Don't cache excessive amounts of data and XDR failures */
if (!statp || len > (256 >> 2)) {
nfsd_reply_cache_free(b, rp);
return;
}
switch (cachetype) {
case RC_REPLSTAT:
if (len != 1)
printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
rp->c_replstat = *statp;
break;
case RC_REPLBUFF:
cachv = &rp->c_replvec;
bufsize = len << 2;
cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
if (!cachv->iov_base) {
nfsd_reply_cache_free(b, rp);
return;
}
cachv->iov_len = bufsize;
memcpy(cachv->iov_base, statp, bufsize);
break;
case RC_NOCACHE:
nfsd_reply_cache_free(b, rp);
return;
}
spin_lock(&b->cache_lock);
drc_mem_usage += bufsize;
lru_put_end(b, rp);
rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
rp->c_type = cachetype;
rp->c_state = RC_DONE;
spin_unlock(&b->cache_lock);
return;
}
/*
* Copy cached reply to current reply buffer. Should always fit.
* FIXME as reply is in a page, we should just attach the page, and
* keep a refcount....
*/
static int
nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
{
struct kvec *vec = &rqstp->rq_res.head[0];
if (vec->iov_len + data->iov_len > PAGE_SIZE) {
printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
data->iov_len);
return 0;
}
memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
vec->iov_len += data->iov_len;
return 1;
}
/*
* Note that fields may be added, removed or reordered in the future. Programs
* scraping this file for info should test the labels to ensure they're
* getting the correct field.
*/
static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
{
seq_printf(m, "max entries: %u\n", max_drc_entries);
seq_printf(m, "num entries: %u\n",
atomic_read(&num_drc_entries));
seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
seq_printf(m, "mem usage: %u\n", drc_mem_usage);
seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
seq_printf(m, "payload misses: %u\n", payload_misses);
seq_printf(m, "longest chain len: %u\n", longest_chain);
seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
return 0;
}
int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
{
return single_open(file, nfsd_reply_cache_stats_show, NULL);
}