/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Copyright (C) 2006 Cluster File Systems, Inc.
+ * Copyright (C) 2006-2007 Cluster File Systems, Inc.
* Author: Eric Mei <ericm@clusterfs.com>
*
* This file is part of Lustre, http://www.lustre.org.
#define PTRS_PER_PAGE (CFS_PAGE_SIZE / sizeof(void *))
#define PAGES_PER_POOL (PTRS_PER_PAGE)
+#define IDLE_IDX_MAX (100)
+#define IDLE_IDX_WEIGHT (3)
+
+#define CACHE_QUIESCENCE_PERIOD (20)
+
static struct ptlrpc_enc_page_pool {
/*
* constants
*/
unsigned long epp_max_pages; /* maximum pages can hold, const */
unsigned int epp_max_pools; /* number of pools, const */
- /*
- * users of the pools. the capacity grow as more user added,
- * but doesn't shrink when users gone -- just current policy.
- * during failover there might be user add/remove activities.
- */
- atomic_t epp_users; /* shared by how many users (osc) */
- atomic_t epp_users_gone; /* users removed */
+
/*
* wait queue in case of not enough free pages.
*/
cfs_waitq_t epp_waitq; /* waiting threads */
unsigned int epp_waitqlen; /* wait queue length */
unsigned long epp_pages_short; /* # of pages wanted of in-q users */
- unsigned long epp_adding:1, /* during adding pages */
- epp_full:1; /* pools are all full */
+ unsigned int epp_growing:1; /* during adding pages */
+
+ /*
+ * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
+ * this is counted based on each time when getting pages from
+ * the pools, not based on time. which means in case that system
+ * is idled for a while but the idle_idx might still be low if no
+ * activities happened in the pools.
+ */
+ unsigned long epp_idle_idx;
+
+ /* last shrink time due to mem tight */
+ long epp_last_shrink;
+ long epp_last_access;
+
/*
* in-pool pages bookkeeping
*/
spinlock_t epp_lock; /* protect following fields */
unsigned long epp_total_pages; /* total pages in pools */
unsigned long epp_free_pages; /* current pages available */
+
/*
* statistics
*/
- unsigned int epp_st_adds;
- unsigned int epp_st_failadds; /* # of add pages failures */
- unsigned long epp_st_reqs; /* # of get_pages requests */
- unsigned long epp_st_missings; /* # of cache missing */
- unsigned long epp_st_lowfree; /* lowest free pages ever reached */
- unsigned long epp_st_max_wqlen;/* highest waitqueue length ever */
- cfs_time_t epp_st_max_wait; /* in jeffies */
+ unsigned int epp_st_grows; /* # of grows */
+ unsigned int epp_st_grow_fails; /* # of add pages failures */
+ unsigned int epp_st_shrinks; /* # of shrinks */
+ unsigned long epp_st_access; /* # of access */
+ unsigned long epp_st_missings; /* # of cache missing */
+ unsigned long epp_st_lowfree; /* lowest free pages reached */
+ unsigned long epp_st_max_wqlen; /* highest waitqueue length */
+ cfs_time_t epp_st_max_wait; /* in jeffies */
/*
* pointers to pools
*/
cfs_page_t ***epp_pools;
} page_pools;
+/*
+ * memory shrinker
+ */
+const int pools_shrinker_seeks = DEFAULT_SEEKS;
+static struct shrinker *pools_shrinker = NULL;
+
+
+/*
+ * /proc/fs/lustre/sptlrpc/encrypt_page_pools
+ */
int sptlrpc_proc_read_enc_pool(char *page, char **start, off_t off, int count,
int *eof, void *data)
{
"pages per pool: %lu\n"
"max pages: %lu\n"
"max pools: %u\n"
- "users: %d - %d\n"
- "current waitqueue len: %u\n"
- "current pages in short: %lu\n"
"total pages: %lu\n"
"total free: %lu\n"
- "add page times: %u\n"
- "add page failed times: %u\n"
- "total requests: %lu\n"
+ "idle index: %lu/100\n"
+ "last shrink: %lds\n"
+ "last access: %lds\n"
+ "grows: %u\n"
+ "grows failure: %u\n"
+ "shrinks: %u\n"
+ "cache access: %lu\n"
"cache missing: %lu\n"
- "lowest free pages: %lu\n"
+ "low free mark: %lu\n"
"max waitqueue depth: %lu\n"
- "max wait time: "CFS_TIME_T"\n"
+ "max wait time: "CFS_TIME_T"/%u\n"
,
num_physpages,
PAGES_PER_POOL,
page_pools.epp_max_pages,
page_pools.epp_max_pools,
- atomic_read(&page_pools.epp_users),
- atomic_read(&page_pools.epp_users_gone),
- page_pools.epp_waitqlen,
- page_pools.epp_pages_short,
page_pools.epp_total_pages,
page_pools.epp_free_pages,
- page_pools.epp_st_adds,
- page_pools.epp_st_failadds,
- page_pools.epp_st_reqs,
+ page_pools.epp_idle_idx,
+ cfs_time_current_sec() - page_pools.epp_last_shrink,
+ cfs_time_current_sec() - page_pools.epp_last_access,
+ page_pools.epp_st_grows,
+ page_pools.epp_st_grow_fails,
+ page_pools.epp_st_shrinks,
+ page_pools.epp_st_access,
page_pools.epp_st_missings,
page_pools.epp_st_lowfree,
page_pools.epp_st_max_wqlen,
- page_pools.epp_st_max_wait
+ page_pools.epp_st_max_wait, HZ
);
spin_unlock(&page_pools.epp_lock);
return rc;
}
+static void enc_pools_release_free_pages(long npages)
+{
+ int p_idx, g_idx;
+
+ LASSERT(npages <= page_pools.epp_free_pages);
+
+ p_idx = (page_pools.epp_free_pages - 1) / PAGES_PER_POOL;
+ g_idx = (page_pools.epp_free_pages - 1) % PAGES_PER_POOL;
+ LASSERT(page_pools.epp_pools[p_idx]);
+
+ page_pools.epp_free_pages -= npages;
+ page_pools.epp_total_pages -= npages;
+
+ while (npages-- > 0) {
+ LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
+
+ cfs_free_page(page_pools.epp_pools[p_idx][g_idx]);
+ page_pools.epp_pools[p_idx][g_idx] = NULL;
+
+ if (g_idx-- == 0) {
+ p_idx--;
+ g_idx = PAGES_PER_POOL - 1;
+
+ LASSERT(page_pools.epp_pools[p_idx]);
+ }
+ }
+}
+
+/*
+ * could be called frequently for query (@nr_to_scan == 0)
+ */
+static int enc_pools_shrink(int nr_to_scan, unsigned int gfp_mask)
+{
+ unsigned long ret;
+
+ spin_lock(&page_pools.epp_lock);
+
+ if (nr_to_scan) {
+ if (nr_to_scan > page_pools.epp_free_pages)
+ nr_to_scan = page_pools.epp_free_pages;
+
+ enc_pools_release_free_pages(nr_to_scan);
+ CDEBUG(D_SEC, "released %d pages, %ld left\n",
+ nr_to_scan, page_pools.epp_free_pages);
+
+ page_pools.epp_st_shrinks++;
+ page_pools.epp_last_shrink = cfs_time_current_sec();
+ }
+
+ /*
+ * try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool
+ */
+ if (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES) {
+ ret = 0;
+ goto out_unlock;
+ }
+
+ /*
+ * if no pool access for a long time, we consider it's fully idle
+ */
+ if (cfs_time_current_sec() - page_pools.epp_last_access >
+ CACHE_QUIESCENCE_PERIOD)
+ page_pools.epp_idle_idx = IDLE_IDX_MAX;
+
+ LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
+ ret = (page_pools.epp_free_pages * page_pools.epp_idle_idx /
+ IDLE_IDX_MAX);
+ if (page_pools.epp_free_pages - ret < PTLRPC_MAX_BRW_PAGES)
+ ret = page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES;
+
+out_unlock:
+ spin_unlock(&page_pools.epp_lock);
+ return ret;
+}
+
static inline
int npages_to_npools(unsigned long npages)
{
/*
* return how many pages cleaned up.
*/
-static unsigned long enc_cleanup_pools(cfs_page_t ***pools, int npools)
+static unsigned long enc_pools_cleanup(cfs_page_t ***pools, int npools)
{
unsigned long cleaned = 0;
int i, j;
* we have options to avoid most memory copy with some tricks. but we choose
* the simplest way to avoid complexity. It's not frequently called.
*/
-static void enc_insert_pool(cfs_page_t ***pools, int npools, int npages)
+static void enc_pools_insert(cfs_page_t ***pools, int npools, int npages)
{
int freeslot;
int op_idx, np_idx, og_idx, ng_idx;
page_pools.epp_free_pages += npages;
page_pools.epp_st_lowfree = page_pools.epp_free_pages;
- if (page_pools.epp_total_pages == page_pools.epp_max_pages)
- page_pools.epp_full = 1;
-
CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
page_pools.epp_total_pages);
int npools, alloced = 0;
int i, j, rc = -ENOMEM;
+ if (npages < PTLRPC_MAX_BRW_PAGES)
+ npages = PTLRPC_MAX_BRW_PAGES;
+
down(&sem_add_pages);
- if (npages > page_pools.epp_max_pages - page_pools.epp_total_pages)
+ if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
- if (npages == 0) {
- rc = 0;
- goto out;
- }
+ LASSERT(npages > 0);
- page_pools.epp_st_adds++;
+ page_pools.epp_st_grows++;
npools = npages_to_npools(npages);
OBD_ALLOC(pools, npools * sizeof(*pools));
}
}
- enc_insert_pool(pools, npools, npages);
- CDEBUG(D_SEC, "add %d pages into enc page pools\n", npages);
+ enc_pools_insert(pools, npools, npages);
+ CDEBUG(D_SEC, "added %d pages into pools\n", npages);
rc = 0;
out_pools:
- enc_cleanup_pools(pools, npools);
+ enc_pools_cleanup(pools, npools);
OBD_FREE(pools, npools * sizeof(*pools));
out:
if (rc) {
- page_pools.epp_st_failadds++;
- CERROR("Failed to pre-allocate %d enc pages\n", npages);
+ page_pools.epp_st_grow_fails++;
+ CERROR("Failed to allocate %d enc pages\n", npages);
}
up(&sem_add_pages);
return rc;
}
-/*
- * both "max bulk rpcs inflight" and "lnet MTU" are tunable, we use the
- * default fixed value initially.
- */
-int sptlrpc_enc_pool_add_user(void)
+static inline void enc_pools_wakeup(void)
{
- int page_plus = PTLRPC_MAX_BRW_PAGES * OSC_MAX_RIF_DEFAULT;
- int users, users_gone, shift, rc;
+ if (unlikely(page_pools.epp_waitqlen)) {
+ LASSERT(page_pools.epp_waitqlen > 0);
+ LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
+ cfs_waitq_broadcast(&page_pools.epp_waitq);
+ }
+}
- LASSERT(!in_interrupt());
- LASSERT(atomic_read(&page_pools.epp_users) >= 0);
+static int enc_pools_should_grow(int page_needed, long now)
+{
+ /* don't grow if someone else is growing the pools right now,
+ * or the pools has reached its full capacity
+ */
+ if (page_pools.epp_growing ||
+ page_pools.epp_total_pages == page_pools.epp_max_pages)
+ return 0;
- users_gone = atomic_dec_return(&page_pools.epp_users_gone);
- if (users_gone >= 0) {
- CWARN("%d users gone, skip\n", users_gone + 1);
+ /* if total pages is not enough, we need to grow */
+ if (page_pools.epp_total_pages < page_needed)
+ return 1;
+
+ /* if we just did a shrink due to memory tight, we'd better
+ * wait a while to grow again.
+ */
+ if (now - page_pools.epp_last_shrink < 2)
return 0;
- }
- atomic_inc(&page_pools.epp_users_gone);
/*
- * prepare full pages for first 2 users; 1/2 for next 2 users;
- * 1/4 for next 4 users; 1/8 for next 8 users; 1/16 for next 16 users;
- * ...
+ * here we perhaps need consider other factors like wait queue
+ * length, idle index, etc. ?
*/
- users = atomic_add_return(1, &page_pools.epp_users);
- shift = fls(users - 1);
- shift = shift > 1 ? shift - 1 : 0;
- page_plus = page_plus >> shift;
- page_plus = page_plus > 2 ? page_plus : 2;
- rc = enc_pools_add_pages(page_plus);
- return 0;
-}
-EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
-
-int sptlrpc_enc_pool_del_user(void)
-{
- atomic_inc(&page_pools.epp_users_gone);
- return 0;
+ /* grow the pools in any other cases */
+ return 1;
}
-EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
/*
* we allocate the requested pages atomically.
int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
{
cfs_waitlink_t waitlink;
- cfs_time_t tick1 = 0, tick2;
+ unsigned long this_idle = -1;
+ cfs_time_t tick = 0;
+ long now;
int p_idx, g_idx;
int i;
LASSERT(desc->bd_max_iov > 0);
- LASSERT(desc->bd_max_iov <= page_pools.epp_total_pages);
+ LASSERT(desc->bd_max_iov <= page_pools.epp_max_pages);
/* resent bulk, enc pages might have been allocated previously */
if (desc->bd_enc_pages != NULL)
return -ENOMEM;
spin_lock(&page_pools.epp_lock);
-again:
- page_pools.epp_st_reqs++;
+ page_pools.epp_st_access++;
+again:
if (unlikely(page_pools.epp_free_pages < desc->bd_max_iov)) {
- if (tick1 == 0)
- tick1 = cfs_time_current();
+ if (tick == 0)
+ tick = cfs_time_current();
+
+ now = cfs_time_current_sec();
page_pools.epp_st_missings++;
page_pools.epp_pages_short += desc->bd_max_iov;
- if (++page_pools.epp_waitqlen > page_pools.epp_st_max_wqlen)
- page_pools.epp_st_max_wqlen = page_pools.epp_waitqlen;
-
- /* we just wait if someone else is adding more pages, or
- * wait queue length is not deep enough. otherwise try to
- * add more pages in the pools.
- *
- * FIXME the policy of detecting resource tight & growing pool
- * need to be reconsidered. */
- if (page_pools.epp_adding || page_pools.epp_waitqlen < 2 ||
- page_pools.epp_full) {
- set_current_state(TASK_UNINTERRUPTIBLE);
- cfs_waitlink_init(&waitlink);
- cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
+ if (enc_pools_should_grow(desc->bd_max_iov, now)) {
+ page_pools.epp_growing = 1;
spin_unlock(&page_pools.epp_lock);
- cfs_schedule();
+ enc_pools_add_pages(page_pools.epp_pages_short / 2);
spin_lock(&page_pools.epp_lock);
+
+ page_pools.epp_growing = 0;
} else {
- page_pools.epp_adding = 1;
+ if (++page_pools.epp_waitqlen >
+ page_pools.epp_st_max_wqlen)
+ page_pools.epp_st_max_wqlen =
+ page_pools.epp_waitqlen;
+
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ cfs_waitlink_init(&waitlink);
+ cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
spin_unlock(&page_pools.epp_lock);
- enc_pools_add_pages(page_pools.epp_pages_short / 2);
+ cfs_schedule();
spin_lock(&page_pools.epp_lock);
- page_pools.epp_adding = 0;
+ LASSERT(page_pools.epp_waitqlen > 0);
+ page_pools.epp_waitqlen--;
}
LASSERT(page_pools.epp_pages_short >= desc->bd_max_iov);
- LASSERT(page_pools.epp_waitqlen > 0);
page_pools.epp_pages_short -= desc->bd_max_iov;
- page_pools.epp_waitqlen--;
+ this_idle = 0;
goto again;
}
/* record max wait time */
- if (unlikely(tick1 != 0)) {
- tick2 = cfs_time_current();
- if (tick2 - tick1 > page_pools.epp_st_max_wait)
- page_pools.epp_st_max_wait = tick2 - tick1;
+ if (unlikely(tick != 0)) {
+ tick = cfs_time_current() - tick;
+ if (tick > page_pools.epp_st_max_wait)
+ page_pools.epp_st_max_wait = tick;
}
/* proceed with rest of allocation */
if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
page_pools.epp_st_lowfree = page_pools.epp_free_pages;
+ /*
+ * new idle index = (old * weight + new) / (weight + 1)
+ */
+ if (this_idle == -1) {
+ this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
+ page_pools.epp_total_pages;
+ }
+ page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
+ this_idle) /
+ (IDLE_IDX_WEIGHT + 1);
+
+ page_pools.epp_last_access = cfs_time_current_sec();
+
spin_unlock(&page_pools.epp_lock);
return 0;
}
page_pools.epp_free_pages += desc->bd_max_iov;
- if (unlikely(page_pools.epp_waitqlen)) {
- LASSERT(page_pools.epp_waitqlen > 0);
- LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
- cfs_waitq_broadcast(&page_pools.epp_waitq);
- }
+ enc_pools_wakeup();
spin_unlock(&page_pools.epp_lock);
}
EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
+/*
+ * we don't do much stuff for add_user/del_user anymore, except adding some
+ * initial pages in add_user() if current pools are empty, rest would be
+ * handled by the pools's self-adaption.
+ */
+int sptlrpc_enc_pool_add_user(void)
+{
+ int need_grow = 0;
+
+ spin_lock(&page_pools.epp_lock);
+ if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
+ page_pools.epp_growing = 1;
+ need_grow = 1;
+ }
+ spin_unlock(&page_pools.epp_lock);
+
+ if (need_grow) {
+ enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES);
+
+ spin_lock(&page_pools.epp_lock);
+ page_pools.epp_growing = 0;
+ enc_pools_wakeup();
+ spin_unlock(&page_pools.epp_lock);
+ }
+ return 0;
+}
+EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
+
+int sptlrpc_enc_pool_del_user(void)
+{
+ return 0;
+}
+EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
+
+static inline void enc_pools_alloc(void)
+{
+ LASSERT(page_pools.epp_max_pools);
+ /*
+ * on system with huge memory but small page size, this might lead to
+ * high-order allocation. but it's not common, and we suppose memory
+ * be not too much fragmented at module loading time.
+ */
+ OBD_ALLOC(page_pools.epp_pools,
+ page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
+}
+
+static inline void enc_pools_free(void)
+{
+ LASSERT(page_pools.epp_max_pools);
+ LASSERT(page_pools.epp_pools);
+
+ OBD_FREE(page_pools.epp_pools,
+ page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
+}
+
int sptlrpc_enc_pool_init(void)
{
- /* constants */
- page_pools.epp_max_pages = num_physpages / 4;
+ /*
+ * maximum capacity is 1/8 of total physical memory.
+ * is the 1/8 a good number?
+ */
+ page_pools.epp_max_pages = num_physpages / 8;
page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
- atomic_set(&page_pools.epp_users, 0);
- atomic_set(&page_pools.epp_users_gone, 0);
-
cfs_waitq_init(&page_pools.epp_waitq);
page_pools.epp_waitqlen = 0;
page_pools.epp_pages_short = 0;
- page_pools.epp_adding = 0;
- page_pools.epp_full = 0;
+ page_pools.epp_growing = 0;
+
+ page_pools.epp_idle_idx = 0;
+ page_pools.epp_last_shrink = cfs_time_current_sec();
+ page_pools.epp_last_access = cfs_time_current_sec();
spin_lock_init(&page_pools.epp_lock);
page_pools.epp_total_pages = 0;
page_pools.epp_free_pages = 0;
- page_pools.epp_st_adds = 0;
- page_pools.epp_st_failadds = 0;
- page_pools.epp_st_reqs = 0;
+ page_pools.epp_st_grows = 0;
+ page_pools.epp_st_grow_fails = 0;
+ page_pools.epp_st_shrinks = 0;
+ page_pools.epp_st_access = 0;
page_pools.epp_st_missings = 0;
page_pools.epp_st_lowfree = 0;
page_pools.epp_st_max_wqlen = 0;
page_pools.epp_st_max_wait = 0;
- OBD_ALLOC(page_pools.epp_pools,
- page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
+ enc_pools_alloc();
if (page_pools.epp_pools == NULL)
return -ENOMEM;
+ pools_shrinker = set_shrinker(pools_shrinker_seeks, enc_pools_shrink);
+ if (pools_shrinker == NULL) {
+ enc_pools_free();
+ return -ENOMEM;
+ }
+
return 0;
}
{
unsigned long cleaned, npools;
+ LASSERT(pools_shrinker);
LASSERT(page_pools.epp_pools);
LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
+ remove_shrinker(pools_shrinker);
+
npools = npages_to_npools(page_pools.epp_total_pages);
- cleaned = enc_cleanup_pools(page_pools.epp_pools, npools);
+ cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
LASSERT(cleaned == page_pools.epp_total_pages);
- OBD_FREE(page_pools.epp_pools,
- page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
+ enc_pools_free();
}
#else /* !__KERNEL__ */
[BULK_CSUM_ALG_SHA512] = { "sha512", 64 },
};
-const char * sptlrpc_bulk_csum_alg2name(__u32 csum_alg)
+const char * sptlrpc_bulk_csum_alg2name(__u8 csum_alg)
{
if (csum_alg < BULK_CSUM_ALG_MAX)
return csum_types[csum_alg].name;
- return "unknown_cksum";
+ return "unknown";
}
EXPORT_SYMBOL(sptlrpc_bulk_csum_alg2name);
-int bulk_sec_desc_size(__u32 csum_alg, int request, int read)
+int bulk_sec_desc_size(__u8 csum_alg, int request, int read)
{
int size = sizeof(struct ptlrpc_bulk_sec_desc);
if (lustre_msg_swabbed(msg)) {
__swab32s(&bsd->bsd_version);
- __swab32s(&bsd->bsd_pad);
- __swab32s(&bsd->bsd_csum_alg);
- __swab32s(&bsd->bsd_priv_alg);
+ __swab16s(&bsd->bsd_pad);
}
if (bsd->bsd_version != 0) {
LASSERT(rsize >= sizeof(*bsdr));
LASSERT(alg < BULK_CSUM_ALG_MAX);
- if (read)
+ if (read) {
bsdr->bsd_csum_alg = alg;
- else {
+ } else {
rc = generate_bulk_csum(desc, alg, bsdr, rsize);
- if (rc) {
- CERROR("client bulk write: failed to perform "
+ if (rc)
+ CERROR("bulk write: client failed to compute "
"checksum: %d\n", rc);
- }
+
+ /* For sending we only compute the wrong checksum instead
+ * of corrupting the data so it is still correct on a redo */
+ if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND) &&
+ bsdr->bsd_csum_alg != BULK_CSUM_ALG_NULL)
+ bsdr->bsd_csum[0] ^= 0x1;
}
return rc;
if (!memcmp(cli, srv, csum_size)) {
/* checksum confirmed */
CDEBUG(D_SEC, "bulk write checksum (%s) confirmed\n",
- csum_types[bsdr->bsd_csum_alg].name);
+ csum_types[bsdr->bsd_csum_alg].name);
return 0;
}
}
EXPORT_SYMBOL(bulk_csum_cli_reply);
+#ifdef __KERNEL__
+static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
+{
+ char *ptr;
+ unsigned int off, i;
+
+ for (i = 0; i < desc->bd_iov_count; i++) {
+ if (desc->bd_iov[i].kiov_len == 0)
+ continue;
+
+ ptr = cfs_kmap(desc->bd_iov[i].kiov_page);
+ off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
+ ptr[off] ^= 0x1;
+ cfs_kunmap(desc->bd_iov[i].kiov_page);
+ return;
+ }
+}
+#else
+static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
+{
+}
+#endif /* __KERNEL__ */
+
int bulk_csum_svc(struct ptlrpc_bulk_desc *desc, int read,
struct ptlrpc_bulk_sec_desc *bsdv, int vsize,
struct ptlrpc_bulk_sec_desc *bsdr, int rsize)
CERROR("bulk read: server failed to generate %s "
"checksum: %d\n",
csum_types[bsdv->bsd_csum_alg].name, rc);
- } else
+
+ /* corrupt the data after we compute the checksum, to
+ * simulate an OST->client data error */
+ if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
+ corrupt_bulk_data(desc);
+ } else {
rc = verify_bulk_csum(desc, 0, bsdv, vsize, bsdr, rsize);
+ }
return rc;
}
[BULK_PRIV_ALG_ARC4] = { "arc4", 0 },
};
-const char * sptlrpc_bulk_priv_alg2name(__u32 priv_alg)
+const char * sptlrpc_bulk_priv_alg2name(__u8 priv_alg)
{
if (priv_alg < BULK_PRIV_ALG_MAX)
return priv_types[priv_alg].name;
- return "unknown_priv";
+ return "unknown";
}
EXPORT_SYMBOL(sptlrpc_bulk_priv_alg2name);
-__u32 sptlrpc_bulk_priv_alg2flags(__u32 priv_alg)
+__u32 sptlrpc_bulk_priv_alg2flags(__u8 priv_alg)
{
if (priv_alg < BULK_PRIV_ALG_MAX)
return priv_types[priv_alg].flags;
git://git.whamcloud.com / fs / lustre-release.git / blobdiff