X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosc%2Fosc_cache.c;h=431df82528d7d5f8d2303a435f16423961fc1ccc;hp=1b35afd9f0a97250cbd9756faaa475d870b783fc;hb=3cce65712d94cffe8f1626545845b95b88aef672;hpb=07ee41661fec73addd3c73bea3d516a99d92d5c8 diff --git a/lustre/osc/osc_cache.c b/lustre/osc/osc_cache.c index 1b35afd..431df82 100644 --- a/lustre/osc/osc_cache.c +++ b/lustre/osc/osc_cache.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2013, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. * */ /* @@ -41,14 +37,15 @@ #define DEBUG_SUBSYSTEM S_OSC -#include "osc_cl_internal.h" +#include + #include "osc_internal.h" static int extent_debug; /* set it to be true for more debug */ static void osc_update_pending(struct osc_object *obj, int cmd, int delta); static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext, - int state); + enum osc_extent_state state); static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli, struct osc_async_page *oap, int sent, int rc); static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap, @@ -58,12 +55,15 @@ static int osc_refresh_count(const struct lu_env *env, static int osc_io_unplug_async(const struct lu_env *env, struct client_obd *cli, struct osc_object *osc); static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages, - unsigned int lost_grant); + unsigned int lost_grant, unsigned int dirty_grant); -static void osc_extent_tree_dump0(int level, struct osc_object *obj, +static void osc_extent_tree_dump0(int mask, struct osc_object *obj, const char *func, int line); -#define osc_extent_tree_dump(lvl, obj) \ - osc_extent_tree_dump0(lvl, obj, __func__, __LINE__) +#define osc_extent_tree_dump(mask, obj) \ + osc_extent_tree_dump0(mask, obj, __func__, __LINE__) + +static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved, + unsigned int unused); /** \addtogroup osc * @{ @@ -76,6 +76,8 @@ static inline char *ext_flags(struct osc_extent *ext, char *flags) *buf++ = ext->oe_rw ? 'r' : 'w'; if (ext->oe_intree) *buf++ = 'i'; + if (ext->oe_sync) + *buf++ = 'S'; if (ext->oe_srvlock) *buf++ = 's'; if (ext->oe_hp) @@ -92,9 +94,9 @@ static inline char *ext_flags(struct osc_extent *ext, char *flags) return flags; } -static inline char list_empty_marker(cfs_list_t *list) +static inline char list_empty_marker(struct list_head *list) { - return cfs_list_empty(list) ? '-' : '+'; + return list_empty(list) ? '-' : '+'; } #define EXTSTR "[%lu -> %lu/%lu]" @@ -102,18 +104,19 @@ static inline char list_empty_marker(cfs_list_t *list) static const char *oes_strings[] = { "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL }; -#define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do { \ +#define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\ + static struct cfs_debug_limit_state cdls; \ struct osc_extent *__ext = (extent); \ char __buf[16]; \ \ - CDEBUG(lvl, \ + __CDEBUG_WITH_LOC(file, func, line, mask, &cdls, \ "extent %p@{" EXTSTR ", " \ "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt, \ /* ----- extent part 0 ----- */ \ __ext, EXTPARA(__ext), \ /* ----- part 1 ----- */ \ - cfs_atomic_read(&__ext->oe_refc), \ - cfs_atomic_read(&__ext->oe_users), \ + atomic_read(&__ext->oe_refc), \ + atomic_read(&__ext->oe_users), \ list_empty_marker(&__ext->oe_link), \ oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \ __ext->oe_obj, \ @@ -121,11 +124,19 @@ static const char *oes_strings[] = { __ext->oe_grants, __ext->oe_nr_pages, \ list_empty_marker(&__ext->oe_pages), \ waitqueue_active(&__ext->oe_waitq) ? '+' : '-', \ - __ext->oe_osclock, __ext->oe_mppr, __ext->oe_owner, \ + __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner, \ /* ----- part 4 ----- */ \ ## __VA_ARGS__); \ + if (mask == D_ERROR && __ext->oe_dlmlock != NULL) \ + LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext); \ + else \ + LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext); \ } while (0) +#define OSC_EXTENT_DUMP(mask, ext, fmt, ...) \ + OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__, \ + mask, ext, fmt, ## __VA_ARGS__) + #undef EASSERTF #define EASSERTF(expr, ext, fmt, args...) do { \ if (!(expr)) { \ @@ -175,7 +186,7 @@ static int osc_extent_sanity_check0(struct osc_extent *ext, { struct osc_object *obj = ext->oe_obj; struct osc_async_page *oap; - int page_count; + size_t page_count; int rc = 0; if (!osc_object_is_locked(obj)) @@ -184,20 +195,20 @@ static int osc_extent_sanity_check0(struct osc_extent *ext, if (ext->oe_state >= OES_STATE_MAX) GOTO(out, rc = 10); - if (cfs_atomic_read(&ext->oe_refc) <= 0) + if (atomic_read(&ext->oe_refc) <= 0) GOTO(out, rc = 20); - if (cfs_atomic_read(&ext->oe_refc) < cfs_atomic_read(&ext->oe_users)) + if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users)) GOTO(out, rc = 30); switch (ext->oe_state) { case OES_INV: - if (ext->oe_nr_pages > 0 || !cfs_list_empty(&ext->oe_pages)) + if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages)) GOTO(out, rc = 35); GOTO(out, rc = 0); break; case OES_ACTIVE: - if (cfs_atomic_read(&ext->oe_users) == 0) + if (atomic_read(&ext->oe_users) == 0) GOTO(out, rc = 40); if (ext->oe_hp) GOTO(out, rc = 50); @@ -210,22 +221,28 @@ static int osc_extent_sanity_check0(struct osc_extent *ext, if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp) GOTO(out, rc = 65); default: - if (cfs_atomic_read(&ext->oe_users) > 0) + if (atomic_read(&ext->oe_users) > 0) GOTO(out, rc = 70); } if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start) GOTO(out, rc = 80); - if (ext->oe_osclock == NULL && ext->oe_grants > 0) + if (ext->oe_sync && ext->oe_grants > 0) GOTO(out, rc = 90); - if (ext->oe_osclock) { - struct cl_lock_descr *descr; - descr = &ext->oe_osclock->cll_descr; - if (!(descr->cld_start <= ext->oe_start && - descr->cld_end >= ext->oe_max_end)) + if (ext->oe_dlmlock != NULL && + ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT && + !ldlm_is_failed(ext->oe_dlmlock)) { + struct ldlm_extent *extent; + + extent = &ext->oe_dlmlock->l_policy_data.l_extent; + if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) && + extent->end >= cl_offset(osc2cl(obj), ext->oe_max_end))) GOTO(out, rc = 100); + + if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))) + GOTO(out, rc = 102); } if (ext->oe_nr_pages > ext->oe_mppr) @@ -240,8 +257,8 @@ static int osc_extent_sanity_check0(struct osc_extent *ext, GOTO(out, rc = 0); page_count = 0; - cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) { - pgoff_t index = oap2cl_page(oap)->cp_index; + list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) { + pgoff_t index = osc_index(oap2osc(oap)); ++page_count; if (index > ext->oe_end || index < ext->oe_start) GOTO(out, rc = 110); @@ -251,9 +268,9 @@ static int osc_extent_sanity_check0(struct osc_extent *ext, out: if (rc != 0) - OSC_EXTENT_DUMP(D_ERROR, ext, - "%s:%d sanity check %p failed with rc = %d\n", - func, line, ext, rc); + OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext, + "sanity check %p failed: rc = %d\n", + ext, rc); return rc; } @@ -309,19 +326,20 @@ static struct osc_extent *osc_extent_alloc(struct osc_object *obj) { struct osc_extent *ext; - OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS); + OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS); if (ext == NULL) return NULL; RB_CLEAR_NODE(&ext->oe_node); ext->oe_obj = obj; - cfs_atomic_set(&ext->oe_refc, 1); - cfs_atomic_set(&ext->oe_users, 0); - CFS_INIT_LIST_HEAD(&ext->oe_link); + cl_object_get(osc2cl(obj)); + atomic_set(&ext->oe_refc, 1); + atomic_set(&ext->oe_users, 0); + INIT_LIST_HEAD(&ext->oe_link); ext->oe_state = OES_INV; - CFS_INIT_LIST_HEAD(&ext->oe_pages); + INIT_LIST_HEAD(&ext->oe_pages); init_waitqueue_head(&ext->oe_waitq); - ext->oe_osclock = NULL; + ext->oe_dlmlock = NULL; return ext; } @@ -333,24 +351,27 @@ static void osc_extent_free(struct osc_extent *ext) static struct osc_extent *osc_extent_get(struct osc_extent *ext) { - LASSERT(cfs_atomic_read(&ext->oe_refc) >= 0); - cfs_atomic_inc(&ext->oe_refc); + LASSERT(atomic_read(&ext->oe_refc) >= 0); + atomic_inc(&ext->oe_refc); return ext; } static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext) { - LASSERT(cfs_atomic_read(&ext->oe_refc) > 0); - if (cfs_atomic_dec_and_test(&ext->oe_refc)) { - LASSERT(cfs_list_empty(&ext->oe_link)); - LASSERT(cfs_atomic_read(&ext->oe_users) == 0); + LASSERT(atomic_read(&ext->oe_refc) > 0); + if (atomic_dec_and_test(&ext->oe_refc)) { + LASSERT(list_empty(&ext->oe_link)); + LASSERT(atomic_read(&ext->oe_users) == 0); LASSERT(ext->oe_state == OES_INV); LASSERT(!ext->oe_intree); - if (ext->oe_osclock) { - cl_lock_put(env, ext->oe_osclock); - ext->oe_osclock = NULL; + if (ext->oe_dlmlock != NULL) { + lu_ref_add(&ext->oe_dlmlock->l_reference, + "osc_extent", ext); + LDLM_LOCK_PUT(ext->oe_dlmlock); + ext->oe_dlmlock = NULL; } + cl_object_put(env, osc2cl(ext->oe_obj)); osc_extent_free(ext); } } @@ -362,9 +383,9 @@ static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext) */ static void osc_extent_put_trust(struct osc_extent *ext) { - LASSERT(cfs_atomic_read(&ext->oe_refc) > 1); + LASSERT(atomic_read(&ext->oe_refc) > 1); LASSERT(osc_object_is_locked(ext->oe_obj)); - cfs_atomic_dec(&ext->oe_refc); + atomic_dec(&ext->oe_refc); } /** @@ -426,7 +447,7 @@ static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext) else if (ext->oe_start > tmp->oe_end) n = &(*n)->rb_right; else - EASSERTF(0, tmp, EXTSTR, EXTPARA(ext)); + EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext)); } rb_link_node(&ext->oe_node, parent, n); rb_insert_color(&ext->oe_node, &obj->oo_root); @@ -457,17 +478,17 @@ static struct osc_extent *osc_extent_hold(struct osc_extent *ext) osc_extent_state_set(ext, OES_ACTIVE); osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages); } - cfs_atomic_inc(&ext->oe_users); - cfs_list_del_init(&ext->oe_link); + atomic_inc(&ext->oe_users); + list_del_init(&ext->oe_link); return osc_extent_get(ext); } static void __osc_extent_remove(struct osc_extent *ext) { LASSERT(osc_object_is_locked(ext->oe_obj)); - LASSERT(cfs_list_empty(&ext->oe_pages)); + LASSERT(list_empty(&ext->oe_pages)); osc_extent_erase(ext); - cfs_list_del_init(&ext->oe_link); + list_del_init(&ext->oe_link); osc_extent_state_set(ext, OES_INV); OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n"); } @@ -483,15 +504,16 @@ static void osc_extent_remove(struct osc_extent *ext) /** * This function is used to merge extents to get better performance. It checks - * if @cur and @victim are contiguous at chunk level. + * if @cur and @victim are contiguous at block level. */ static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur, struct osc_extent *victim) { - struct osc_object *obj = cur->oe_obj; - pgoff_t chunk_start; - pgoff_t chunk_end; - int ppc_bits; + struct osc_object *obj = cur->oe_obj; + struct client_obd *cli = osc_cli(obj); + pgoff_t chunk_start; + pgoff_t chunk_end; + int ppc_bits; LASSERT(cur->oe_state == OES_CACHE); LASSERT(osc_object_is_locked(obj)); @@ -504,25 +526,32 @@ static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur, if (cur->oe_max_end != victim->oe_max_end) return -ERANGE; - LASSERT(cur->oe_osclock == victim->oe_osclock); - ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT; + LASSERT(cur->oe_dlmlock == victim->oe_dlmlock); + ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT; chunk_start = cur->oe_start >> ppc_bits; chunk_end = cur->oe_end >> ppc_bits; if (chunk_start != (victim->oe_end >> ppc_bits) + 1 && chunk_end + 1 != victim->oe_start >> ppc_bits) return -ERANGE; + /* overall extent size should not exceed the max supported limit + * reported by the server */ + if (cur->oe_end - cur->oe_start + 1 + + victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages) + return -ERANGE; + OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur); cur->oe_start = min(cur->oe_start, victim->oe_start); cur->oe_end = max(cur->oe_end, victim->oe_end); - cur->oe_grants += victim->oe_grants; + /* per-extent tax should be accounted only once for the whole extent */ + cur->oe_grants += victim->oe_grants - cli->cl_grant_extent_tax; cur->oe_nr_pages += victim->oe_nr_pages; /* only the following bits are needed to merge */ cur->oe_urgent |= victim->oe_urgent; cur->oe_memalloc |= victim->oe_memalloc; - cfs_list_splice_init(&victim->oe_pages, &cur->oe_pages); - cfs_list_del_init(&victim->oe_link); + list_splice_init(&victim->oe_pages, &cur->oe_pages); + list_del_init(&victim->oe_link); victim->oe_nr_pages = 0; osc_extent_get(victim); @@ -539,14 +568,15 @@ static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur, int osc_extent_release(const struct lu_env *env, struct osc_extent *ext) { struct osc_object *obj = ext->oe_obj; + struct client_obd *cli = osc_cli(obj); int rc = 0; ENTRY; - LASSERT(cfs_atomic_read(&ext->oe_users) > 0); + LASSERT(atomic_read(&ext->oe_users) > 0); LASSERT(sanity_check(ext) == 0); LASSERT(ext->oe_grants > 0); - if (cfs_atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) { + if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) { LASSERT(ext->oe_state == OES_ACTIVE); if (ext->oe_trunc_pending) { /* a truncate process is waiting for this extent. @@ -554,22 +584,33 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext) * osc_cache_truncate_start(). */ osc_extent_state_set(ext, OES_TRUNC); ext->oe_trunc_pending = 0; + osc_object_unlock(obj); } else { + int grant = 0; + osc_extent_state_set(ext, OES_CACHE); osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages); /* try to merge the previous and next extent. */ - osc_extent_merge(env, ext, prev_extent(ext)); - osc_extent_merge(env, ext, next_extent(ext)); + if (osc_extent_merge(env, ext, prev_extent(ext)) == 0) + grant += cli->cl_grant_extent_tax; + if (osc_extent_merge(env, ext, next_extent(ext)) == 0) + grant += cli->cl_grant_extent_tax; if (ext->oe_urgent) - cfs_list_move_tail(&ext->oe_link, - &obj->oo_urgent_exts); + list_move_tail(&ext->oe_link, + &obj->oo_urgent_exts); + else if (ext->oe_nr_pages == ext->oe_mppr) { + list_move_tail(&ext->oe_link, + &obj->oo_full_exts); + } + osc_object_unlock(obj); + if (grant > 0) + osc_unreserve_grant(cli, 0, grant); } - osc_object_unlock(obj); - osc_io_unplug_async(env, osc_cli(obj), obj); + osc_io_unplug_async(env, cli, obj); } osc_extent_put(env, ext); RETURN(rc); @@ -584,23 +625,23 @@ static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2) * Find or create an extent which includes @index, core function to manage * extent tree. */ -struct osc_extent *osc_extent_find(const struct lu_env *env, - struct osc_object *obj, pgoff_t index, - int *grants) - +static struct osc_extent *osc_extent_find(const struct lu_env *env, + struct osc_object *obj, pgoff_t index, + unsigned int *grants) { struct client_obd *cli = osc_cli(obj); - struct cl_lock *lock; + struct osc_lock *olck; + struct cl_lock_descr *descr; struct osc_extent *cur; struct osc_extent *ext; struct osc_extent *conflict = NULL; struct osc_extent *found = NULL; pgoff_t chunk; pgoff_t max_end; - int max_pages; /* max_pages_per_rpc */ - int chunksize; + unsigned int max_pages; /* max_pages_per_rpc */ + unsigned int chunksize; int ppc_bits; /* pages per chunk bits */ - int chunk_mask; + pgoff_t chunk_mask; int rc; ENTRY; @@ -608,38 +649,51 @@ struct osc_extent *osc_extent_find(const struct lu_env *env, if (cur == NULL) RETURN(ERR_PTR(-ENOMEM)); - lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0); - LASSERT(lock != NULL); - LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE); + olck = osc_env_io(env)->oi_write_osclock; + LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index); + LASSERT(olck->ols_state == OLS_GRANTED); + + descr = &olck->ols_cl.cls_lock->cll_descr; + LASSERT(descr->cld_mode >= CLM_WRITE); - LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT); - ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT; + LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT, + "chunkbits: %u\n", cli->cl_chunkbits); + ppc_bits = cli->cl_chunkbits - PAGE_SHIFT; chunk_mask = ~((1 << ppc_bits) - 1); chunksize = 1 << cli->cl_chunkbits; chunk = index >> ppc_bits; - /* align end to rpc edge, rpc size may not be a power 2 integer. */ + /* align end to RPC edge. */ max_pages = cli->cl_max_pages_per_rpc; - LASSERT((max_pages & ~chunk_mask) == 0); + if ((max_pages & ~chunk_mask) != 0) { + CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n", + max_pages, cli->cl_chunkbits, chunk_mask); + RETURN(ERR_PTR(-EINVAL)); + } max_end = index - (index % max_pages) + max_pages - 1; - max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end); + max_end = min_t(pgoff_t, max_end, descr->cld_end); /* initialize new extent by parameters so far */ cur->oe_max_end = max_end; cur->oe_start = index & chunk_mask; cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1; - if (cur->oe_start < lock->cll_descr.cld_start) - cur->oe_start = lock->cll_descr.cld_start; + if (cur->oe_start < descr->cld_start) + cur->oe_start = descr->cld_start; if (cur->oe_end > max_end) cur->oe_end = max_end; - cur->oe_osclock = lock; cur->oe_grants = 0; cur->oe_mppr = max_pages; + if (olck->ols_dlmlock != NULL) { + LASSERT(olck->ols_hold); + cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock); + lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur); + } /* grants has been allocated by caller */ - LASSERTF(*grants >= chunksize + cli->cl_extent_tax, - "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax); - LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur)); + LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax, + "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax); + LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n", + EXTPARA(cur)); restart: osc_object_lock(obj); @@ -647,17 +701,17 @@ restart: if (ext == NULL) ext = first_extent(obj); while (ext != NULL) { - loff_t ext_chk_start = ext->oe_start >> ppc_bits; - loff_t ext_chk_end = ext->oe_end >> ppc_bits; + pgoff_t ext_chk_start = ext->oe_start >> ppc_bits; + pgoff_t ext_chk_end = ext->oe_end >> ppc_bits; LASSERT(sanity_check_nolock(ext) == 0); - if (chunk > ext_chk_end + 1) + if (chunk > ext_chk_end + 1 || chunk < ext_chk_start) break; /* if covering by different locks, no chance to match */ - if (lock != ext->oe_osclock) { + if (olck->ols_dlmlock != ext->oe_dlmlock) { EASSERTF(!overlapped(ext, cur), ext, - EXTSTR, EXTPARA(cur)); + EXTSTR"\n", EXTPARA(cur)); ext = next_extent(ext); continue; @@ -678,7 +732,7 @@ restart: * full contain. */ EASSERTF((ext->oe_start <= cur->oe_start && ext->oe_end >= cur->oe_end), - ext, EXTSTR, EXTPARA(cur)); + ext, EXTSTR"\n", EXTPARA(cur)); if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) { /* for simplicity, we wait for this extent to @@ -710,6 +764,13 @@ restart: continue; } + /* check whether maximum extent size will be hit */ + if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits > + cli->cl_max_extent_pages) { + ext = next_extent(ext); + continue; + } + /* it's required that an extent must be contiguous at chunk * level so that we know the whole extent is covered by grant * (the pages in the extent are NOT required to be contiguous). @@ -724,6 +785,7 @@ restart: /* pull ext's start back to cover cur */ ext->oe_start = cur->oe_start; ext->oe_grants += chunksize; + LASSERT(*grants >= chunksize); *grants -= chunksize; found = osc_extent_hold(ext); @@ -731,13 +793,14 @@ restart: /* rear merge */ ext->oe_end = cur->oe_end; ext->oe_grants += chunksize; + LASSERT(*grants >= chunksize); *grants -= chunksize; /* try to merge with the next one because we just fill * in a gap */ if (osc_extent_merge(env, ext, next_extent(ext)) == 0) /* we can save extent tax from next extent */ - *grants += cli->cl_extent_tax; + *grants += cli->cl_grant_extent_tax; found = osc_extent_hold(ext); } @@ -751,22 +814,22 @@ restart: if (found != NULL) { LASSERT(conflict == NULL); if (!IS_ERR(found)) { - LASSERT(found->oe_osclock == cur->oe_osclock); + LASSERT(found->oe_dlmlock == cur->oe_dlmlock); OSC_EXTENT_DUMP(D_CACHE, found, "found caching ext for %lu.\n", index); } } else if (conflict == NULL) { /* create a new extent */ EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur); - cur->oe_grants = chunksize + cli->cl_extent_tax; + cur->oe_grants = chunksize + cli->cl_grant_extent_tax; + LASSERT(*grants >= cur->oe_grants); *grants -= cur->oe_grants; - LASSERT(*grants >= 0); cur->oe_state = OES_CACHE; found = osc_extent_hold(cur); osc_extent_insert(obj, cur); OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n", - index, lock->cll_descr.cld_end); + index, descr->cld_end); } osc_object_unlock(obj); @@ -787,7 +850,6 @@ restart: out: osc_extent_put(env, cur); - LASSERT(*grants >= 0); return found; } @@ -803,7 +865,7 @@ int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext, int nr_pages = ext->oe_nr_pages; int lost_grant = 0; int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096; - __u64 last_off = 0; + loff_t last_off = 0; int last_count = -1; ENTRY; @@ -811,10 +873,12 @@ int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext, ext->oe_rc = rc ?: ext->oe_nr_pages; EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext); - cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages, + + osc_lru_add_batch(cli, &ext->oe_pages); + list_for_each_entry_safe(oap, tmp, &ext->oe_pages, oap_pending_item) { - cfs_list_del_init(&oap->oap_rpc_item); - cfs_list_del_init(&oap->oap_pending_item); + list_del_init(&oap->oap_rpc_item); + list_del_init(&oap->oap_pending_item); if (last_off <= oap->oap_obj_off) { last_off = oap->oap_obj_off; last_count = oap->oap_count; @@ -827,21 +891,21 @@ int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext, if (!sent) { lost_grant = ext->oe_grants; - } else if (blocksize < PAGE_CACHE_SIZE && - last_count != PAGE_CACHE_SIZE) { + } else if (blocksize < PAGE_SIZE && + last_count != PAGE_SIZE) { /* For short writes we shouldn't count parts of pages that * span a whole chunk on the OST side, or our accounting goes * wrong. Should match the code in filter_grant_check. */ - int offset = oap->oap_page_off & ~CFS_PAGE_MASK; - int count = oap->oap_count + (offset & (blocksize - 1)); - int end = (offset + oap->oap_count) & (blocksize - 1); + int offset = last_off & ~PAGE_MASK; + int count = last_count + (offset & (blocksize - 1)); + int end = (offset + last_count) & (blocksize - 1); if (end) count += blocksize - end; - lost_grant = PAGE_CACHE_SIZE - count; + lost_grant = PAGE_SIZE - count; } if (ext->oe_grants > 0) - osc_free_grant(cli, nr_pages, lost_grant); + osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants); osc_extent_remove(ext); /* put the refcount for RPC */ @@ -849,7 +913,7 @@ int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext, RETURN(0); } -static int extent_wait_cb(struct osc_extent *ext, int state) +static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state) { int ret; @@ -864,7 +928,7 @@ static int extent_wait_cb(struct osc_extent *ext, int state) * Wait for the extent's state to become @state. */ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext, - int state) + enum osc_extent_state state) { struct osc_object *obj = ext->oe_obj; struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL, @@ -893,8 +957,8 @@ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext, rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi); if (rc == -ETIMEDOUT) { OSC_EXTENT_DUMP(D_ERROR, ext, - "%s: wait ext to %d timedout, recovery in progress?\n", - osc_export(obj)->exp_obd->obd_name, state); + "%s: wait ext to %u timedout, recovery in progress?\n", + cli_name(osc_cli(obj)), state); lwi = LWI_INTR(NULL, NULL); rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), @@ -912,20 +976,21 @@ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext, static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index, bool partial) { - struct cl_env_nest nest; struct lu_env *env; struct cl_io *io; struct osc_object *obj = ext->oe_obj; struct client_obd *cli = osc_cli(obj); struct osc_async_page *oap; struct osc_async_page *tmp; + struct pagevec *pvec; int pages_in_chunk = 0; int ppc_bits = cli->cl_chunkbits - - PAGE_CACHE_SHIFT; + PAGE_SHIFT; __u64 trunc_chunk = trunc_index >> ppc_bits; int grants = 0; int nr_pages = 0; int rc = 0; + __u16 refcheck; ENTRY; LASSERT(sanity_check(ext) == 0); @@ -935,39 +1000,44 @@ static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index, /* Request new lu_env. * We can't use that env from osc_cache_truncate_start() because * it's from lov_io_sub and not fully initialized. */ - env = cl_env_nested_get(&nest); - io = &osc_env_info(env)->oti_io; + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + RETURN(PTR_ERR(env)); + + io = osc_env_thread_io(env); io->ci_obj = cl_object_top(osc2cl(obj)); + io->ci_ignore_layout = 1; + pvec = &osc_env_info(env)->oti_pagevec; + ll_pagevec_init(pvec, 0); rc = cl_io_init(env, io, CIT_MISC, io->ci_obj); if (rc < 0) GOTO(out, rc); - /* discard all pages with index greater then trunc_index */ - cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages, + /* discard all pages with index greater than trunc_index */ + list_for_each_entry_safe(oap, tmp, &ext->oe_pages, oap_pending_item) { - struct cl_page *sub = oap2cl_page(oap); - struct cl_page *page = cl_page_top(sub); + pgoff_t index = osc_index(oap2osc(oap)); + struct cl_page *page = oap2cl_page(oap); - LASSERT(cfs_list_empty(&oap->oap_rpc_item)); + LASSERT(list_empty(&oap->oap_rpc_item)); /* only discard the pages with their index greater than * trunc_index, and ... */ - if (sub->cp_index < trunc_index || - (sub->cp_index == trunc_index && partial)) { + if (index < trunc_index || + (index == trunc_index && partial)) { /* accounting how many pages remaining in the chunk * so that we can calculate grants correctly. */ - if (sub->cp_index >> ppc_bits == trunc_chunk) + if (index >> ppc_bits == trunc_chunk) ++pages_in_chunk; continue; } - cfs_list_del_init(&oap->oap_pending_item); + list_del_init(&oap->oap_pending_item); cl_page_get(page); lu_ref_add(&page->cp_reference, "truncate", current); if (cl_page_own(env, io, page) == 0) { - cl_page_unmap(env, io, page); cl_page_discard(env, io, page); cl_page_disown(env, io, page); } else { @@ -976,11 +1046,13 @@ static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index, } lu_ref_del(&page->cp_reference, "truncate", current); - cl_page_put(env, page); + cl_pagevec_put(env, page, pvec); --ext->oe_nr_pages; ++nr_pages; } + pagevec_release(pvec); + EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial, ext->oe_nr_pages == 0), ext, "trunc_index %lu, partial %d\n", trunc_index, partial); @@ -1017,11 +1089,11 @@ static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index, osc_object_unlock(obj); if (grants > 0 || nr_pages > 0) - osc_free_grant(cli, nr_pages, grants); + osc_free_grant(cli, nr_pages, grants, grants); out: cl_io_fini(env, io); - cl_env_nested_put(&nest, env); + cl_env_put(env, &refcheck); RETURN(rc); } @@ -1035,7 +1107,7 @@ static int osc_extent_make_ready(const struct lu_env *env, struct osc_async_page *oap; struct osc_async_page *last = NULL; struct osc_object *obj = ext->oe_obj; - int page_count = 0; + unsigned int page_count = 0; int rc; ENTRY; @@ -1047,7 +1119,7 @@ static int osc_extent_make_ready(const struct lu_env *env, OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n"); - cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) { + list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) { ++page_count; if (last == NULL || last->oap_obj_off < oap->oap_obj_off) last = oap; @@ -1076,9 +1148,10 @@ static int osc_extent_make_ready(const struct lu_env *env, /* the last page is the only one we need to refresh its count by * the size of file. */ if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) { - last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE); - LASSERT(last->oap_count > 0); - LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE); + int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE); + LASSERT(last_oap_count > 0); + LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE); + last->oap_count = last_oap_count; spin_lock(&last->oap_lock); last->oap_async_flags |= ASYNC_COUNT_STABLE; spin_unlock(&last->oap_lock); @@ -1086,9 +1159,9 @@ static int osc_extent_make_ready(const struct lu_env *env, /* for the rest of pages, we don't need to call osf_refresh_count() * because it's known they are not the last page */ - cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) { + list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) { if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) { - oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off; + oap->oap_count = PAGE_SIZE - oap->oap_page_off; spin_lock(&oap->oap_lock); oap->oap_async_flags |= ASYNC_COUNT_STABLE; spin_unlock(&oap->oap_lock); @@ -1109,16 +1182,17 @@ static int osc_extent_make_ready(const struct lu_env *env, * called to expand the extent for the same IO. To expand the extent, the * page index must be in the same or next chunk of ext->oe_end. */ -static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants) +static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, + unsigned int *grants) { struct osc_object *obj = ext->oe_obj; struct client_obd *cli = osc_cli(obj); struct osc_extent *next; - int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT; + int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT; pgoff_t chunk = index >> ppc_bits; pgoff_t end_chunk; pgoff_t end_index; - int chunksize = 1 << cli->cl_chunkbits; + unsigned int chunksize = 1 << cli->cl_chunkbits; int rc = 0; ENTRY; @@ -1133,9 +1207,14 @@ static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants) GOTO(out, rc = 0); LASSERT(end_chunk + 1 == chunk); + /* try to expand this extent to cover @index */ end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1); + /* don't go over the maximum extent size reported by server */ + if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages) + GOTO(out, rc = -ERANGE); + next = next_extent(ext); if (next != NULL && next->oe_start <= end_index) /* complex mode - overlapped with the next extent, @@ -1144,8 +1223,8 @@ static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants) ext->oe_end = end_index; ext->oe_grants += chunksize; + LASSERT(*grants >= chunksize); *grants -= chunksize; - LASSERT(*grants >= 0); EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext, "overlapped after expanding for %lu.\n", index); EXIT; @@ -1155,31 +1234,34 @@ out: RETURN(rc); } -static void osc_extent_tree_dump0(int level, struct osc_object *obj, +static void osc_extent_tree_dump0(int mask, struct osc_object *obj, const char *func, int line) { struct osc_extent *ext; int cnt; - CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n", + if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) + return; + + CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n", obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc); /* osc_object_lock(obj); */ cnt = 1; for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext)) - OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++); + OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++); cnt = 1; - cfs_list_for_each_entry(ext, &obj->oo_hp_exts, oe_link) - OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++); + list_for_each_entry(ext, &obj->oo_hp_exts, oe_link) + OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++); cnt = 1; - cfs_list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link) - OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++); + list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link) + OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++); cnt = 1; - cfs_list_for_each_entry(ext, &obj->oo_reading_exts, oe_link) - OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++); + list_for_each_entry(ext, &obj->oo_reading_exts, oe_link) + OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++); /* osc_object_unlock(obj); */ } @@ -1187,8 +1269,8 @@ static void osc_extent_tree_dump0(int level, struct osc_object *obj, static inline int osc_is_ready(struct osc_object *osc) { - return !cfs_list_empty(&osc->oo_ready_item) || - !cfs_list_empty(&osc->oo_hp_ready_item); + return !list_empty(&osc->oo_ready_item) || + !list_empty(&osc->oo_hp_ready_item); } #define OSC_IO_DEBUG(OSC, STR, args...) \ @@ -1196,10 +1278,10 @@ static inline int osc_is_ready(struct osc_object *osc) (OSC), osc_is_ready(OSC), \ list_empty_marker(&(OSC)->oo_hp_ready_item), \ list_empty_marker(&(OSC)->oo_ready_item), \ - cfs_atomic_read(&(OSC)->oo_nr_writes), \ + atomic_read(&(OSC)->oo_nr_writes), \ list_empty_marker(&(OSC)->oo_hp_exts), \ list_empty_marker(&(OSC)->oo_urgent_exts), \ - cfs_atomic_read(&(OSC)->oo_nr_reads), \ + atomic_read(&(OSC)->oo_nr_reads), \ list_empty_marker(&(OSC)->oo_reading_exts), \ ##args) @@ -1207,7 +1289,7 @@ static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap, int cmd) { struct osc_page *opg = oap2osc_page(oap); - struct cl_page *page = cl_page_top(oap2cl_page(oap)); + struct cl_page *page = oap2cl_page(oap); int result; LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */ @@ -1215,7 +1297,7 @@ static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap, ENTRY; result = cl_page_make_ready(env, page, CRT_WRITE); if (result == 0) - opg->ops_submit_time = cfs_time_current(); + opg->ops_submit_time = ktime_get(); RETURN(result); } @@ -1223,10 +1305,9 @@ static int osc_refresh_count(const struct lu_env *env, struct osc_async_page *oap, int cmd) { struct osc_page *opg = oap2osc_page(oap); - struct cl_page *page = oap2cl_page(oap); + pgoff_t index = osc_index(oap2osc(oap)); struct cl_object *obj; struct cl_attr *attr = &osc_env_info(env)->oti_attr; - int result; loff_t kms; @@ -1241,22 +1322,21 @@ static int osc_refresh_count(const struct lu_env *env, if (result < 0) return result; kms = attr->cat_kms; - if (cl_offset(obj, page->cp_index) >= kms) + if (cl_offset(obj, index) >= kms) /* catch race with truncate */ return 0; - else if (cl_offset(obj, page->cp_index + 1) > kms) + else if (cl_offset(obj, index + 1) > kms) /* catch sub-page write at end of file */ - return kms % PAGE_CACHE_SIZE; + return kms % PAGE_SIZE; else - return PAGE_CACHE_SIZE; + return PAGE_SIZE; } static int osc_completion(const struct lu_env *env, struct osc_async_page *oap, int cmd, int rc) { struct osc_page *opg = oap2osc_page(oap); - struct cl_page *page = cl_page_top(oap2cl_page(oap)); - struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj); + struct cl_page *page = oap2cl_page(oap); enum cl_req_type crt; int srvlock; @@ -1269,33 +1349,18 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap, "cp_state:%u, cmd:%d\n", page->cp_state, cmd); LASSERT(opg->ops_transfer_pinned); - /* - * page->cp_req can be NULL if io submission failed before - * cl_req was allocated. - */ - if (page->cp_req != NULL) - cl_req_page_done(env, page); - LASSERT(page->cp_req == NULL); - crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE; /* Clear opg->ops_transfer_pinned before VM lock is released. */ opg->ops_transfer_pinned = 0; - spin_lock(&obj->oo_seatbelt); - LASSERT(opg->ops_submitter != NULL); - LASSERT(!cfs_list_empty(&opg->ops_inflight)); - cfs_list_del_init(&opg->ops_inflight); - opg->ops_submitter = NULL; - spin_unlock(&obj->oo_seatbelt); - - opg->ops_submit_time = 0; + opg->ops_submit_time = ktime_set(0, 0); srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK; /* statistic */ if (rc == 0 && srvlock) { struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev; struct osc_stats *stats = &lu2osc_dev(ld)->od_stats; - int bytes = oap->oap_count; + size_t bytes = oap->oap_count; if (crt == CRT_READ) stats->os_lockless_reads += bytes; @@ -1311,38 +1376,38 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap, lu_ref_del(&page->cp_reference, "transfer", page); cl_page_completion(env, page, crt, rc); + cl_page_put(env, page); RETURN(0); } -#define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do { \ - struct client_obd *__tmp = (cli); \ - CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %d/%d " \ - "unstable_pages: %d/%d dropped: %ld avail: %ld, " \ - "reserved: %ld, flight: %d } lru {in list: %d, " \ - "left: %d, waiters: %d }" fmt, \ - __tmp->cl_import->imp_obd->obd_name, \ - __tmp->cl_dirty, __tmp->cl_dirty_max, \ - cfs_atomic_read(&obd_dirty_pages), obd_max_dirty_pages, \ - cfs_atomic_read(&obd_unstable_pages), obd_max_dirty_pages, \ - __tmp->cl_lost_grant, __tmp->cl_avail_grant, \ - __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \ - cfs_atomic_read(&__tmp->cl_lru_in_list), \ - cfs_atomic_read(&__tmp->cl_lru_busy), \ - cfs_atomic_read(&__tmp->cl_lru_shrinkers), ##args); \ +#define OSC_DUMP_GRANT(mask, cli, fmt, args...) do { \ + struct client_obd *__tmp = (cli); \ + CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \ + "dropped: %ld avail: %ld, dirty_grant: %ld, " \ + "reserved: %ld, flight: %d } lru {in list: %ld, " \ + "left: %ld, waiters: %d }" fmt "\n", \ + cli_name(__tmp), \ + __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \ + atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \ + __tmp->cl_lost_grant, __tmp->cl_avail_grant, \ + __tmp->cl_dirty_grant, \ + __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \ + atomic_long_read(&__tmp->cl_lru_in_list), \ + atomic_long_read(&__tmp->cl_lru_busy), \ + atomic_read(&__tmp->cl_lru_shrinkers), ##args); \ } while (0) /* caller must hold loi_list_lock */ static void osc_consume_write_grant(struct client_obd *cli, struct brw_page *pga) { - LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock)); + assert_spin_locked(&cli->cl_loi_list_lock); LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT)); - cfs_atomic_inc(&obd_dirty_pages); - cli->cl_dirty += PAGE_CACHE_SIZE; + cli->cl_dirty_pages++; pga->flag |= OBD_BRW_FROM_GRANT; CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n", - PAGE_CACHE_SIZE, pga, pga->pg); + PAGE_SIZE, pga, pga->pg); osc_update_next_shrink(cli); } @@ -1353,19 +1418,19 @@ static void osc_release_write_grant(struct client_obd *cli, { ENTRY; - LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock)); + assert_spin_locked(&cli->cl_loi_list_lock); if (!(pga->flag & OBD_BRW_FROM_GRANT)) { EXIT; return; } pga->flag &= ~OBD_BRW_FROM_GRANT; - cfs_atomic_dec(&obd_dirty_pages); - cli->cl_dirty -= PAGE_CACHE_SIZE; + atomic_long_dec(&obd_dirty_pages); + cli->cl_dirty_pages--; if (pga->flag & OBD_BRW_NOCACHE) { pga->flag &= ~OBD_BRW_NOCACHE; - cfs_atomic_dec(&obd_dirty_transit_pages); - cli->cl_dirty_transit -= PAGE_CACHE_SIZE; + atomic_long_dec(&obd_dirty_transit_pages); + cli->cl_dirty_transit--; } EXIT; } @@ -1399,19 +1464,28 @@ static void __osc_unreserve_grant(struct client_obd *cli, if (unused > reserved) { cli->cl_avail_grant += reserved; cli->cl_lost_grant += unused - reserved; + cli->cl_dirty_grant -= unused - reserved; } else { cli->cl_avail_grant += unused; + cli->cl_dirty_grant += reserved - unused; } } -void osc_unreserve_grant(struct client_obd *cli, - unsigned int reserved, unsigned int unused) +static void osc_unreserve_grant_nolock(struct client_obd *cli, + unsigned int reserved, + unsigned int unused) { - client_obd_list_lock(&cli->cl_loi_list_lock); __osc_unreserve_grant(cli, reserved, unused); if (unused > 0) osc_wake_cache_waiters(cli); - client_obd_list_unlock(&cli->cl_loi_list_lock); +} + +static void osc_unreserve_grant(struct client_obd *cli, + unsigned int reserved, unsigned int unused) +{ + spin_lock(&cli->cl_loi_list_lock); + osc_unreserve_grant_nolock(cli, reserved, unused); + spin_unlock(&cli->cl_loi_list_lock); } /** @@ -1421,21 +1495,24 @@ void osc_unreserve_grant(struct client_obd *cli, * used, we should return these grants to OST. There're two cases where grants * can be lost: * 1. truncate; - * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was + * 2. blocksize at OST is less than PAGE_SIZE and a partial page was * written. In this case OST may use less chunks to serve this partial * write. OSTs don't actually know the page size on the client side. so * clients have to calculate lost grant by the blocksize on the OST. * See filter_grant_check() for details. */ static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages, - unsigned int lost_grant) + unsigned int lost_grant, unsigned int dirty_grant) { - int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax; + unsigned long grant; - client_obd_list_lock(&cli->cl_loi_list_lock); - cfs_atomic_sub(nr_pages, &obd_dirty_pages); - cli->cl_dirty -= nr_pages << PAGE_CACHE_SHIFT; + grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax; + + spin_lock(&cli->cl_loi_list_lock); + atomic_long_sub(nr_pages, &obd_dirty_pages); + cli->cl_dirty_pages -= nr_pages; cli->cl_lost_grant += lost_grant; + cli->cl_dirty_grant -= dirty_grant; if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) { /* borrow some grant from truncate to avoid the case that * truncate uses up all avail grant */ @@ -1443,10 +1520,11 @@ static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages, cli->cl_avail_grant += grant; } osc_wake_cache_waiters(cli); - client_obd_list_unlock(&cli->cl_loi_list_lock); - CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n", + spin_unlock(&cli->cl_loi_list_lock); + CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n", lost_grant, cli->cl_lost_grant, - cli->cl_avail_grant, cli->cl_dirty); + cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT, + cli->cl_dirty_grant); } /** @@ -1455,9 +1533,9 @@ static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages, */ static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap) { - client_obd_list_lock(&cli->cl_loi_list_lock); + spin_lock(&cli->cl_loi_list_lock); osc_release_write_grant(cli, &oap->oap_brw_page); - client_obd_list_unlock(&cli->cl_loi_list_lock); + spin_unlock(&cli->cl_loi_list_lock); } /** @@ -1470,35 +1548,38 @@ static int osc_enter_cache_try(struct client_obd *cli, { int rc; - OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes); + OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes); rc = osc_reserve_grant(cli, bytes); if (rc < 0) return 0; - if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max && - cfs_atomic_read(&obd_unstable_pages) + 1 + - cfs_atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) { - osc_consume_write_grant(cli, &oap->oap_brw_page); - if (transient) { - cli->cl_dirty_transit += PAGE_CACHE_SIZE; - cfs_atomic_inc(&obd_dirty_transit_pages); - oap->oap_brw_flags |= OBD_BRW_NOCACHE; - } - rc = 1; - } else { - __osc_unreserve_grant(cli, bytes, bytes); - rc = 0; + if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) { + if (atomic_long_add_return(1, &obd_dirty_pages) <= + obd_max_dirty_pages) { + osc_consume_write_grant(cli, &oap->oap_brw_page); + if (transient) { + cli->cl_dirty_transit++; + atomic_long_inc(&obd_dirty_transit_pages); + oap->oap_brw_flags |= OBD_BRW_NOCACHE; + } + rc = 1; + goto out; + } else + atomic_long_dec(&obd_dirty_pages); } + __osc_unreserve_grant(cli, bytes, bytes); + +out: return rc; } static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw) { int rc; - client_obd_list_lock(&cli->cl_loi_list_lock); - rc = cfs_list_empty(&ocw->ocw_entry); - client_obd_list_unlock(&cli->cl_loi_list_lock); + spin_lock(&cli->cl_loi_list_lock); + rc = list_empty(&ocw->ocw_entry); + spin_unlock(&cli->cl_loi_list_lock); return rc; } @@ -1512,28 +1593,35 @@ static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw) static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli, struct osc_async_page *oap, int bytes) { - struct osc_object *osc = oap->oap_obj; - struct lov_oinfo *loi = osc->oo_oinfo; - struct osc_cache_waiter ocw; - struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL, - LWI_ON_SIGNAL_NOOP, NULL); - int rc = -EDQUOT; + struct osc_object *osc = oap->oap_obj; + struct lov_oinfo *loi = osc->oo_oinfo; + struct osc_cache_waiter ocw; + struct l_wait_info lwi; + int rc = -EDQUOT; ENTRY; - OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes); + lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max), + NULL, LWI_ON_SIGNAL_NOOP, NULL); - client_obd_list_lock(&cli->cl_loi_list_lock); + OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes); + + spin_lock(&cli->cl_loi_list_lock); /* force the caller to try sync io. this can jump the list * of queued writes and create a discontiguous rpc stream */ if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) || - cli->cl_dirty_max < PAGE_CACHE_SIZE || - cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) + cli->cl_dirty_max_pages == 0 || + cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) { + OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n"); GOTO(out, rc = -EDQUOT); + } /* Hopefully normal case - cache space and write credits available */ - if (osc_enter_cache_try(cli, oap, bytes, 0)) + if (list_empty(&cli->cl_cache_waiters) && + osc_enter_cache_try(cli, oap, bytes, 0)) { + OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n"); GOTO(out, rc = 0); + } /* We can get here for two reasons: too many dirty pages in cache, or * run out of grants. In both cases we should write dirty pages out. @@ -1544,98 +1632,101 @@ static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli, init_waitqueue_head(&ocw.ocw_waitq); ocw.ocw_oap = oap; ocw.ocw_grant = bytes; - while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) { - cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters); + while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) { + list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters); ocw.ocw_rc = 0; - client_obd_list_unlock(&cli->cl_loi_list_lock); + spin_unlock(&cli->cl_loi_list_lock); osc_io_unplug_async(env, cli, NULL); CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n", - cli->cl_import->imp_obd->obd_name, &ocw, oap); + cli_name(cli), &ocw, oap); rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi); - client_obd_list_lock(&cli->cl_loi_list_lock); + spin_lock(&cli->cl_loi_list_lock); - /* l_wait_event is interrupted by signal, or timed out */ if (rc < 0) { - switch (rc) { - case -ETIMEDOUT: - OSC_DUMP_GRANT(D_ERROR, cli, - "try to reserve %d.\n", bytes); - osc_extent_tree_dump(D_ERROR, osc); - rc = -EDQUOT; - break; - case -EINTR: - /* Ensures restartability - LU-3581 */ - rc = -ERESTARTSYS; - break; - default: - CDEBUG(D_CACHE, "%s: event for cache space @" - " %p never arrived due to %d\n", - cli->cl_import->imp_obd->obd_name, - &ocw, rc); - break; - } - cfs_list_del_init(&ocw.ocw_entry); - GOTO(out, rc); + /* l_wait_event is interrupted by signal or timed out */ + list_del_init(&ocw.ocw_entry); + break; } - - LASSERT(cfs_list_empty(&ocw.ocw_entry)); + LASSERT(list_empty(&ocw.ocw_entry)); rc = ocw.ocw_rc; if (rc != -EDQUOT) - GOTO(out, rc); - if (osc_enter_cache_try(cli, oap, bytes, 0)) - GOTO(out, rc = 0); + break; + if (osc_enter_cache_try(cli, oap, bytes, 0)) { + rc = 0; + break; + } + } + + switch (rc) { + case 0: + OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n"); + break; + case -ETIMEDOUT: + OSC_DUMP_GRANT(D_CACHE, cli, + "timeout, fall back to sync i/o\n"); + osc_extent_tree_dump(D_CACHE, osc); + /* fall back to synchronous I/O */ + rc = -EDQUOT; + break; + case -EINTR: + /* Ensures restartability - LU-3581 */ + OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n"); + rc = -ERESTARTSYS; + break; + case -EDQUOT: + OSC_DUMP_GRANT(D_CACHE, cli, + "no grant space, fall back to sync i/o\n"); + break; + default: + CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived " + "due to %d, fall back to sync i/o\n", + cli_name(cli), &ocw, rc); + break; } EXIT; out: - client_obd_list_unlock(&cli->cl_loi_list_lock); - OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc); + spin_unlock(&cli->cl_loi_list_lock); RETURN(rc); } /* caller must hold loi_list_lock */ void osc_wake_cache_waiters(struct client_obd *cli) { - cfs_list_t *l, *tmp; + struct list_head *l, *tmp; struct osc_cache_waiter *ocw; ENTRY; - cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) { - ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry); - cfs_list_del_init(&ocw->ocw_entry); + list_for_each_safe(l, tmp, &cli->cl_cache_waiters) { + ocw = list_entry(l, struct osc_cache_waiter, ocw_entry); ocw->ocw_rc = -EDQUOT; - /* we can't dirty more */ - if ((cli->cl_dirty + PAGE_CACHE_SIZE > cli->cl_dirty_max) || - (cfs_atomic_read(&obd_unstable_pages) + 1 + - cfs_atomic_read(&obd_dirty_pages) > obd_max_dirty_pages)) { - CDEBUG(D_CACHE, "no dirty room: dirty: %ld " - "osc max %ld, sys max %d\n", cli->cl_dirty, - cli->cl_dirty_max, obd_max_dirty_pages); - goto wakeup; - } - ocw->ocw_rc = 0; - if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0)) - ocw->ocw_rc = -EDQUOT; + if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0)) + ocw->ocw_rc = 0; -wakeup: - CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n", - ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc); + if (ocw->ocw_rc == 0 || + !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) { + list_del_init(&ocw->ocw_entry); + CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant " + "%ld, %d\n", ocw, ocw->ocw_oap, + cli->cl_avail_grant, ocw->ocw_rc); - wake_up(&ocw->ocw_waitq); + wake_up(&ocw->ocw_waitq); + } } EXIT; } +EXPORT_SYMBOL(osc_wake_cache_waiters); static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc) { - int hprpc = !!cfs_list_empty(&osc->oo_hp_exts); + int hprpc = !!list_empty(&osc->oo_hp_exts); return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc; } @@ -1656,39 +1747,40 @@ static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc, invalid_import = 1; if (cmd & OBD_BRW_WRITE) { - if (cfs_atomic_read(&osc->oo_nr_writes) == 0) + if (atomic_read(&osc->oo_nr_writes) == 0) RETURN(0); if (invalid_import) { CDEBUG(D_CACHE, "invalid import forcing RPC\n"); RETURN(1); } - if (!cfs_list_empty(&osc->oo_hp_exts)) { + if (!list_empty(&osc->oo_hp_exts)) { CDEBUG(D_CACHE, "high prio request forcing RPC\n"); RETURN(1); } - if (!cfs_list_empty(&osc->oo_urgent_exts)) { + if (!list_empty(&osc->oo_urgent_exts)) { CDEBUG(D_CACHE, "urgent request forcing RPC\n"); RETURN(1); } /* trigger a write rpc stream as long as there are dirtiers * waiting for space. as they're waiting, they're not going to * create more pages to coalesce with what's waiting.. */ - if (!cfs_list_empty(&cli->cl_cache_waiters)) { + if (!list_empty(&cli->cl_cache_waiters)) { CDEBUG(D_CACHE, "cache waiters forcing RPC\n"); RETURN(1); } - if (cfs_atomic_read(&osc->oo_nr_writes) >= - cli->cl_max_pages_per_rpc) + if (!list_empty(&osc->oo_full_exts)) { + CDEBUG(D_CACHE, "full extent ready, make an RPC\n"); RETURN(1); + } } else { - if (cfs_atomic_read(&osc->oo_nr_reads) == 0) + if (atomic_read(&osc->oo_nr_reads) == 0) RETURN(0); if (invalid_import) { CDEBUG(D_CACHE, "invalid import forcing RPC\n"); RETURN(1); } /* all read are urgent. */ - if (!cfs_list_empty(&osc->oo_reading_exts)) + if (!list_empty(&osc->oo_reading_exts)) RETURN(1); } @@ -1699,28 +1791,29 @@ static void osc_update_pending(struct osc_object *obj, int cmd, int delta) { struct client_obd *cli = osc_cli(obj); if (cmd & OBD_BRW_WRITE) { - cfs_atomic_add(delta, &obj->oo_nr_writes); - cfs_atomic_add(delta, &cli->cl_pending_w_pages); - LASSERT(cfs_atomic_read(&obj->oo_nr_writes) >= 0); + atomic_add(delta, &obj->oo_nr_writes); + atomic_add(delta, &cli->cl_pending_w_pages); + LASSERT(atomic_read(&obj->oo_nr_writes) >= 0); } else { - cfs_atomic_add(delta, &obj->oo_nr_reads); - cfs_atomic_add(delta, &cli->cl_pending_r_pages); - LASSERT(cfs_atomic_read(&obj->oo_nr_reads) >= 0); + atomic_add(delta, &obj->oo_nr_reads); + atomic_add(delta, &cli->cl_pending_r_pages); + LASSERT(atomic_read(&obj->oo_nr_reads) >= 0); } OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta); } static int osc_makes_hprpc(struct osc_object *obj) { - return !cfs_list_empty(&obj->oo_hp_exts); + return !list_empty(&obj->oo_hp_exts); } -static void on_list(cfs_list_t *item, cfs_list_t *list, int should_be_on) +static void on_list(struct list_head *item, struct list_head *list, + int should_be_on) { - if (cfs_list_empty(item) && should_be_on) - cfs_list_add_tail(item, list); - else if (!cfs_list_empty(item) && !should_be_on) - cfs_list_del_init(item); + if (list_empty(item) && should_be_on) + list_add_tail(item, list); + else if (!list_empty(item) && !should_be_on) + list_del_init(item); } /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc @@ -1739,10 +1832,10 @@ static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc) } on_list(&osc->oo_write_item, &cli->cl_loi_write_list, - cfs_atomic_read(&osc->oo_nr_writes) > 0); + atomic_read(&osc->oo_nr_writes) > 0); on_list(&osc->oo_read_item, &cli->cl_loi_read_list, - cfs_atomic_read(&osc->oo_nr_reads) > 0); + atomic_read(&osc->oo_nr_reads) > 0); return osc_is_ready(osc); } @@ -1751,9 +1844,9 @@ static int osc_list_maint(struct client_obd *cli, struct osc_object *osc) { int is_ready; - client_obd_list_lock(&cli->cl_loi_list_lock); + spin_lock(&cli->cl_loi_list_lock); is_ready = __osc_list_maint(cli, osc); - client_obd_list_unlock(&cli->cl_loi_list_lock); + spin_unlock(&cli->cl_loi_list_lock); return is_ready; } @@ -1780,91 +1873,6 @@ static void osc_process_ar(struct osc_async_rc *ar, __u64 xid, ar->ar_force_sync = 0; } -/* Performs "unstable" page accounting. This function balances the - * increment operations performed in osc_inc_unstable_pages. It is - * registered as the RPC request callback, and is executed when the - * bulk RPC is committed on the server. Thus at this point, the pages - * involved in the bulk transfer are no longer considered unstable. */ -void osc_dec_unstable_pages(struct ptlrpc_request *req) -{ - struct ptlrpc_bulk_desc *desc = req->rq_bulk; - struct client_obd *cli = &req->rq_import->imp_obd->u.cli; - obd_count page_count = desc->bd_iov_count; - int i; - - /* No unstable page tracking */ - if (cli->cl_cache == NULL) - return; - - LASSERT(page_count >= 0); - - for (i = 0; i < page_count; i++) - dec_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS); - - cfs_atomic_sub(page_count, &cli->cl_cache->ccc_unstable_nr); - LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0); - - cfs_atomic_sub(page_count, &cli->cl_unstable_count); - LASSERT(cfs_atomic_read(&cli->cl_unstable_count) >= 0); - - cfs_atomic_sub(page_count, &obd_unstable_pages); - LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0); - - spin_lock(&req->rq_lock); - req->rq_committed = 1; - req->rq_unstable = 0; - spin_unlock(&req->rq_lock); - - wake_up_all(&cli->cl_cache->ccc_unstable_waitq); -} - -/* "unstable" page accounting. See: osc_dec_unstable_pages. */ -void osc_inc_unstable_pages(struct ptlrpc_request *req) -{ - struct ptlrpc_bulk_desc *desc = req->rq_bulk; - struct client_obd *cli = &req->rq_import->imp_obd->u.cli; - obd_count page_count = desc->bd_iov_count; - int i; - - /* No unstable page tracking */ - if (cli->cl_cache == NULL) - return; - - LASSERT(page_count >= 0); - - for (i = 0; i < page_count; i++) - inc_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS); - - LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0); - cfs_atomic_add(page_count, &cli->cl_cache->ccc_unstable_nr); - - LASSERT(cfs_atomic_read(&cli->cl_unstable_count) >= 0); - cfs_atomic_add(page_count, &cli->cl_unstable_count); - - LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0); - cfs_atomic_add(page_count, &obd_unstable_pages); - - spin_lock(&req->rq_lock); - - /* If the request has already been committed (i.e. brw_commit - * called via rq_commit_cb), we need to undo the unstable page - * increments we just performed because rq_commit_cb wont be - * called again. Otherwise, just set the commit callback so the - * unstable page accounting is properly updated when the request - * is committed */ - if (req->rq_committed) { - /* Drop lock before calling osc_dec_unstable_pages */ - spin_unlock(&req->rq_lock); - osc_dec_unstable_pages(req); - spin_lock(&req->rq_lock); - } else { - req->rq_unstable = 1; - req->rq_commit_cb = osc_dec_unstable_pages; - } - - spin_unlock(&req->rq_lock); -} - /* this must be called holding the loi list lock to give coverage to exit_cache, * async_flag maintenance, and oap_request */ static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli, @@ -1876,9 +1884,6 @@ static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli, ENTRY; if (oap->oap_request != NULL) { - if (rc == 0) - osc_inc_unstable_pages(oap->oap_request); - xid = ptlrpc_req_xid(oap->oap_request); ptlrpc_req_finished(oap->oap_request); oap->oap_request = NULL; @@ -1891,10 +1896,10 @@ static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli, oap->oap_interrupted = 0; if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) { - client_obd_list_lock(&cli->cl_loi_list_lock); + spin_lock(&cli->cl_loi_list_lock); osc_process_ar(&cli->cl_ar, xid, rc); osc_process_ar(&loi->loi_ar, xid, rc); - client_obd_list_unlock(&cli->cl_loi_list_lock); + spin_unlock(&cli->cl_loi_list_lock); } rc = osc_completion(env, oap, oap->oap_cmd, rc); @@ -1905,26 +1910,67 @@ static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli, EXIT; } +struct extent_rpc_data { + struct list_head *erd_rpc_list; + unsigned int erd_page_count; + unsigned int erd_max_pages; + unsigned int erd_max_chunks; + unsigned int erd_max_extents; +}; + +static inline unsigned osc_extent_chunks(const struct osc_extent *ext) +{ + struct client_obd *cli = osc_cli(ext->oe_obj); + unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT; + + return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1; +} + /** * Try to add extent to one RPC. We need to think about the following things: * - # of pages must not be over max_pages_per_rpc * - extent must be compatible with previous ones */ static int try_to_add_extent_for_io(struct client_obd *cli, - struct osc_extent *ext, cfs_list_t *rpclist, - int *pc, unsigned int *max_pages) + struct osc_extent *ext, + struct extent_rpc_data *data) { struct osc_extent *tmp; + unsigned int chunk_count; + struct osc_async_page *oap = list_first_entry(&ext->oe_pages, + struct osc_async_page, + oap_pending_item); ENTRY; EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE), ext); + OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n"); - *max_pages = max(ext->oe_mppr, *max_pages); - if (*pc + ext->oe_nr_pages > *max_pages) + if (data->erd_max_extents == 0) RETURN(0); - cfs_list_for_each_entry(tmp, rpclist, oe_link) { + chunk_count = osc_extent_chunks(ext); + EASSERTF(data->erd_page_count != 0 || + chunk_count <= data->erd_max_chunks, ext, + "The first extent to be fit in a RPC contains %u chunks, " + "which is over the limit %u.\n", chunk_count, + data->erd_max_chunks); + if (chunk_count > data->erd_max_chunks) + RETURN(0); + + data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages); + EASSERTF(data->erd_page_count != 0 || + ext->oe_nr_pages <= data->erd_max_pages, ext, + "The first extent to be fit in a RPC contains %u pages, " + "which is over the limit %u.\n", ext->oe_nr_pages, + data->erd_max_pages); + if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages) + RETURN(0); + + list_for_each_entry(tmp, data->erd_rpc_list, oe_link) { + struct osc_async_page *oap2; + oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page, + oap_pending_item); EASSERT(tmp->oe_owner == current, tmp); #if 0 if (overlapped(tmp, ext)) { @@ -1932,17 +1978,26 @@ static int try_to_add_extent_for_io(struct client_obd *cli, EASSERT(0, ext); } #endif + if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) { + CDEBUG(D_CACHE, "Do not permit different types of IO " + "in one RPC\n"); + RETURN(0); + } if (tmp->oe_srvlock != ext->oe_srvlock || - !tmp->oe_grants != !ext->oe_grants) + !tmp->oe_grants != !ext->oe_grants || + tmp->oe_ndelay != ext->oe_ndelay || + tmp->oe_no_merge || ext->oe_no_merge) RETURN(0); /* remove break for strict check */ break; } - *pc += ext->oe_nr_pages; - cfs_list_move_tail(&ext->oe_link, rpclist); + data->erd_max_extents--; + data->erd_max_chunks -= chunk_count; + data->erd_page_count += ext->oe_nr_pages; + list_move_tail(&ext->oe_link, data->erd_rpc_list); ext->oe_owner = current; RETURN(1); } @@ -1960,77 +2015,81 @@ static int try_to_add_extent_for_io(struct client_obd *cli, * 5. Traverse the extent tree from the 1st extent; * 6. Above steps exit if there is no space in this RPC. */ -static int get_write_extents(struct osc_object *obj, cfs_list_t *rpclist) +static unsigned int get_write_extents(struct osc_object *obj, + struct list_head *rpclist) { struct client_obd *cli = osc_cli(obj); struct osc_extent *ext; - int page_count = 0; - unsigned int max_pages = cli->cl_max_pages_per_rpc; + struct extent_rpc_data data = { + .erd_rpc_list = rpclist, + .erd_page_count = 0, + .erd_max_pages = cli->cl_max_pages_per_rpc, + .erd_max_chunks = osc_max_write_chunks(cli), + .erd_max_extents = 256, + }; LASSERT(osc_object_is_locked(obj)); - while (!cfs_list_empty(&obj->oo_hp_exts)) { - ext = cfs_list_entry(obj->oo_hp_exts.next, struct osc_extent, - oe_link); + while (!list_empty(&obj->oo_hp_exts)) { + ext = list_entry(obj->oo_hp_exts.next, struct osc_extent, + oe_link); LASSERT(ext->oe_state == OES_CACHE); - if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count, - &max_pages)) - return page_count; - EASSERT(ext->oe_nr_pages <= max_pages, ext); - } - if (page_count == max_pages) - return page_count; - - while (!cfs_list_empty(&obj->oo_urgent_exts)) { - ext = cfs_list_entry(obj->oo_urgent_exts.next, - struct osc_extent, oe_link); - if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count, - &max_pages)) - return page_count; - - if (!ext->oe_intree) - continue; - - while ((ext = next_extent(ext)) != NULL) { - if ((ext->oe_state != OES_CACHE) || - (!cfs_list_empty(&ext->oe_link) && - ext->oe_owner != NULL)) - continue; - - if (!try_to_add_extent_for_io(cli, ext, rpclist, - &page_count, &max_pages)) - return page_count; - } + if (!try_to_add_extent_for_io(cli, ext, &data)) + return data.erd_page_count; + EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext); + } + if (data.erd_page_count == data.erd_max_pages) + return data.erd_page_count; + + while (!list_empty(&obj->oo_urgent_exts)) { + ext = list_entry(obj->oo_urgent_exts.next, + struct osc_extent, oe_link); + if (!try_to_add_extent_for_io(cli, ext, &data)) + return data.erd_page_count; + } + if (data.erd_page_count == data.erd_max_pages) + return data.erd_page_count; + + /* One key difference between full extents and other extents: full + * extents can usually only be added if the rpclist was empty, so if we + * can't add one, we continue on to trying to add normal extents. This + * is so we don't miss adding extra extents to an RPC containing high + * priority or urgent extents. */ + while (!list_empty(&obj->oo_full_exts)) { + ext = list_entry(obj->oo_full_exts.next, + struct osc_extent, oe_link); + if (!try_to_add_extent_for_io(cli, ext, &data)) + break; } - if (page_count == max_pages) - return page_count; + if (data.erd_page_count == data.erd_max_pages) + return data.erd_page_count; ext = first_extent(obj); while (ext != NULL) { if ((ext->oe_state != OES_CACHE) || /* this extent may be already in current rpclist */ - (!cfs_list_empty(&ext->oe_link) && ext->oe_owner != NULL)) { + (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) { ext = next_extent(ext); continue; } - if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count, - &max_pages)) - return page_count; + if (!try_to_add_extent_for_io(cli, ext, &data)) + return data.erd_page_count; ext = next_extent(ext); } - return page_count; + return data.erd_page_count; } static int osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli, - struct osc_object *osc, pdl_policy_t pol) + struct osc_object *osc) +__must_hold(osc) { - CFS_LIST_HEAD(rpclist); + struct list_head rpclist = LIST_HEAD_INIT(rpclist); struct osc_extent *ext; struct osc_extent *tmp; struct osc_extent *first = NULL; - obd_count page_count = 0; + unsigned int page_count = 0; int srvlock = 0; int rc = 0; ENTRY; @@ -2038,14 +2097,14 @@ osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli, LASSERT(osc_object_is_locked(osc)); page_count = get_write_extents(osc, &rpclist); - LASSERT(equi(page_count == 0, cfs_list_empty(&rpclist))); + LASSERT(equi(page_count == 0, list_empty(&rpclist))); - if (cfs_list_empty(&rpclist)) + if (list_empty(&rpclist)) RETURN(0); osc_update_pending(osc, OBD_BRW_WRITE, -page_count); - cfs_list_for_each_entry(ext, &rpclist, oe_link) { + list_for_each_entry(ext, &rpclist, oe_link) { LASSERT(ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE); if (ext->oe_state == OES_CACHE) @@ -2058,11 +2117,11 @@ osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli, * lock order is page lock -> object lock. */ osc_object_unlock(osc); - cfs_list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) { + list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) { if (ext->oe_state == OES_LOCKING) { rc = osc_extent_make_ready(env, ext); if (unlikely(rc < 0)) { - cfs_list_del_init(&ext->oe_link); + list_del_init(&ext->oe_link); osc_extent_finish(env, ext, 0, rc); continue; } @@ -2075,10 +2134,10 @@ osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli, } } - if (!cfs_list_empty(&rpclist)) { + if (!list_empty(&rpclist)) { LASSERT(page_count > 0); - rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol); - LASSERT(cfs_list_empty(&rpclist)); + rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE); + LASSERT(list_empty(&rpclist)); } osc_object_lock(osc); @@ -2097,36 +2156,39 @@ osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli, */ static int osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli, - struct osc_object *osc, pdl_policy_t pol) + struct osc_object *osc) +__must_hold(osc) { struct osc_extent *ext; struct osc_extent *next; - CFS_LIST_HEAD(rpclist); - int page_count = 0; - unsigned int max_pages = cli->cl_max_pages_per_rpc; + struct list_head rpclist = LIST_HEAD_INIT(rpclist); + struct extent_rpc_data data = { + .erd_rpc_list = &rpclist, + .erd_page_count = 0, + .erd_max_pages = cli->cl_max_pages_per_rpc, + .erd_max_chunks = UINT_MAX, + .erd_max_extents = UINT_MAX, + }; int rc = 0; ENTRY; LASSERT(osc_object_is_locked(osc)); - cfs_list_for_each_entry_safe(ext, next, - &osc->oo_reading_exts, oe_link) { + list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) { EASSERT(ext->oe_state == OES_LOCK_DONE, ext); - if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count, - &max_pages)) + if (!try_to_add_extent_for_io(cli, ext, &data)) break; osc_extent_state_set(ext, OES_RPC); - EASSERT(ext->oe_nr_pages <= max_pages, ext); + EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext); } - LASSERT(page_count <= max_pages); + LASSERT(data.erd_page_count <= data.erd_max_pages); - osc_update_pending(osc, OBD_BRW_READ, -page_count); + osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count); - if (!cfs_list_empty(&rpclist)) { + if (!list_empty(&rpclist)) { osc_object_unlock(osc); - LASSERT(page_count > 0); - rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol); - LASSERT(cfs_list_empty(&rpclist)); + rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ); + LASSERT(list_empty(&rpclist)); osc_object_lock(osc); } @@ -2134,9 +2196,9 @@ osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli, } #define list_to_obj(list, item) ({ \ - cfs_list_t *__tmp = (list)->next; \ - cfs_list_del_init(__tmp); \ - cfs_list_entry(__tmp, struct osc_object, oo_##item); \ + struct list_head *__tmp = (list)->next; \ + list_del_init(__tmp); \ + list_entry(__tmp, struct osc_object, oo_##item); \ }) /* This is called by osc_check_rpcs() to find which objects have pages that @@ -2148,26 +2210,26 @@ static struct osc_object *osc_next_obj(struct client_obd *cli) /* First return objects that have blocked locks so that they * will be flushed quickly and other clients can get the lock, * then objects which have pages ready to be stuffed into RPCs */ - if (!cfs_list_empty(&cli->cl_loi_hp_ready_list)) + if (!list_empty(&cli->cl_loi_hp_ready_list)) RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item)); - if (!cfs_list_empty(&cli->cl_loi_ready_list)) + if (!list_empty(&cli->cl_loi_ready_list)) RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item)); /* then if we have cache waiters, return all objects with queued * writes. This is especially important when many small files * have filled up the cache and not been fired into rpcs because * they don't pass the nr_pending/object threshhold */ - if (!cfs_list_empty(&cli->cl_cache_waiters) && - !cfs_list_empty(&cli->cl_loi_write_list)) + if (!list_empty(&cli->cl_cache_waiters) && + !list_empty(&cli->cl_loi_write_list)) RETURN(list_to_obj(&cli->cl_loi_write_list, write_item)); /* then return all queued objects when we have an invalid import * so that they get flushed */ if (cli->cl_import == NULL || cli->cl_import->imp_invalid) { - if (!cfs_list_empty(&cli->cl_loi_write_list)) + if (!list_empty(&cli->cl_loi_write_list)) RETURN(list_to_obj(&cli->cl_loi_write_list, write_item)); - if (!cfs_list_empty(&cli->cl_loi_read_list)) + if (!list_empty(&cli->cl_loi_read_list)) RETURN(list_to_obj(&cli->cl_loi_read_list, read_item)); } @@ -2175,8 +2237,8 @@ static struct osc_object *osc_next_obj(struct client_obd *cli) } /* called with the loi list lock held */ -static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli, - pdl_policy_t pol) +static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli) +__must_hold(&cli->cl_loi_list_lock) { struct osc_object *osc; int rc = 0; @@ -2194,9 +2256,8 @@ static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli, } cl_object_get(obj); - client_obd_list_unlock(&cli->cl_loi_list_lock); - lu_object_ref_add_at(&obj->co_lu, &link, "check", - current); + spin_unlock(&cli->cl_loi_list_lock); + lu_object_ref_add_at(&obj->co_lu, &link, "check", current); /* attempt some read/write balancing by alternating between * reads and writes in an object. The makes_rpc checks here @@ -2206,7 +2267,7 @@ static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli, * do io on writes while there are cache waiters */ osc_object_lock(osc); if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) { - rc = osc_send_write_rpc(env, cli, osc, pol); + rc = osc_send_write_rpc(env, cli, osc); if (rc < 0) { CERROR("Write request failed with %d\n", rc); @@ -2230,23 +2291,22 @@ static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli, } } if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) { - rc = osc_send_read_rpc(env, cli, osc, pol); + rc = osc_send_read_rpc(env, cli, osc); if (rc < 0) CERROR("Read request failed with %d\n", rc); } osc_object_unlock(osc); osc_list_maint(cli, osc); - lu_object_ref_del_at(&obj->co_lu, &link, "check", - current); + lu_object_ref_del_at(&obj->co_lu, &link, "check", current); cl_object_put(env, obj); - client_obd_list_lock(&cli->cl_loi_list_lock); + spin_lock(&cli->cl_loi_list_lock); } } -static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli, - struct osc_object *osc, pdl_policy_t pol, int async) +int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli, + struct osc_object *osc, int async) { int rc = 0; @@ -2254,13 +2314,9 @@ static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli, return 0; if (!async) { - /* disable osc_lru_shrink() temporarily to avoid - * potential stack overrun problem. LU-2859 */ - cfs_atomic_inc(&cli->cl_lru_shrinkers); - client_obd_list_lock(&cli->cl_loi_list_lock); - osc_check_rpcs(env, cli, pol); - client_obd_list_unlock(&cli->cl_loi_list_lock); - cfs_atomic_dec(&cli->cl_lru_shrinkers); + spin_lock(&cli->cl_loi_list_lock); + osc_check_rpcs(env, cli); + spin_unlock(&cli->cl_loi_list_lock); } else { CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli); LASSERT(cli->cl_writeback_work != NULL); @@ -2268,19 +2324,7 @@ static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli, } return rc; } - -static int osc_io_unplug_async(const struct lu_env *env, - struct client_obd *cli, struct osc_object *osc) -{ - /* XXX: policy is no use actually. */ - return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1); -} - -void osc_io_unplug(const struct lu_env *env, struct client_obd *cli, - struct osc_object *osc, pdl_policy_t pol) -{ - (void)osc_io_unplug0(env, cli, osc, pol, 0); -} +EXPORT_SYMBOL(osc_io_unplug0); int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops, struct page *page, loff_t offset) @@ -2298,19 +2342,17 @@ int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops, oap->oap_page = page; oap->oap_obj_off = offset; - LASSERT(!(offset & ~CFS_PAGE_MASK)); + LASSERT(!(offset & ~PAGE_MASK)); - if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE)) - oap->oap_brw_flags = OBD_BRW_NOQUOTA; - - CFS_INIT_LIST_HEAD(&oap->oap_pending_item); - CFS_INIT_LIST_HEAD(&oap->oap_rpc_item); + INIT_LIST_HEAD(&oap->oap_pending_item); + INIT_LIST_HEAD(&oap->oap_rpc_item); spin_lock_init(&oap->oap_lock); - CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n", + CDEBUG(D_INFO, "oap %p page %p obj off %llu\n", oap, page, oap->oap_obj_off); RETURN(0); } +EXPORT_SYMBOL(osc_prep_async_page); int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, struct osc_page *ops) @@ -2321,8 +2363,9 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, struct client_obd *cli = oap->oap_cli; struct osc_object *osc = oap->oap_obj; pgoff_t index; - int grants = 0; - int brw_flags = OBD_BRW_ASYNC; + unsigned int tmp; + unsigned int grants = 0; + u32 brw_flags = OBD_BRW_ASYNC; int cmd = OBD_BRW_WRITE; int need_release = 0; int rc = 0; @@ -2334,14 +2377,13 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, if (cli->cl_import == NULL || cli->cl_import->imp_invalid) RETURN(-EIO); - if (!cfs_list_empty(&oap->oap_pending_item) || - !cfs_list_empty(&oap->oap_rpc_item)) + if (!list_empty(&oap->oap_pending_item) || + !list_empty(&oap->oap_rpc_item)) RETURN(-EBUSY); /* Set the OBD_BRW_SRVLOCK before the page is queued. */ brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0; - if (!client_is_remote(osc_export(osc)) && - cfs_capable(CFS_CAP_SYS_RESOURCE)) { + if (oio->oi_cap_sys_resource) { brw_flags |= OBD_BRW_NOQUOTA; cmd |= OBD_BRW_NOQUOTA; } @@ -2350,7 +2392,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, if (!(cmd & OBD_BRW_NOQUOTA)) { struct cl_object *obj; struct cl_attr *attr; - unsigned int qid[MAXQUOTAS]; + unsigned int qid[LL_MAXQUOTAS]; obj = cl_object_top(&osc->oo_cl); attr = &osc_env_info(env)->oti_attr; @@ -2361,6 +2403,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, qid[USRQUOTA] = attr->cat_uid; qid[GRPQUOTA] = attr->cat_gid; + qid[PRJQUOTA] = attr->cat_projid; if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA) rc = -EDQUOT; if (rc) @@ -2378,7 +2421,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n", oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK); - index = oap2cl_page(oap)->cp_index; + index = osc_index(oap2osc(oap)); /* Add this page into extent by the following steps: * 1. if there exists an active extent for this IO, mostly this page @@ -2390,19 +2433,18 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) { /* one chunk plus extent overhead must be enough to write this * page */ - grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax; + grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax; if (ext->oe_end >= index) grants = 0; /* it doesn't need any grant to dirty this page */ - client_obd_list_lock(&cli->cl_loi_list_lock); + spin_lock(&cli->cl_loi_list_lock); rc = osc_enter_cache_try(cli, oap, grants, 0); - client_obd_list_unlock(&cli->cl_loi_list_lock); if (rc == 0) { /* try failed */ grants = 0; need_release = 1; } else if (ext->oe_end < index) { - int tmp = grants; + tmp = grants; /* try to expand this extent */ rc = osc_extent_expand(ext, index, &tmp); if (rc < 0) { @@ -2411,10 +2453,11 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, } else { OSC_EXTENT_DUMP(D_CACHE, ext, "expanded for %lu.\n", index); - osc_unreserve_grant(cli, grants, tmp); + osc_unreserve_grant_nolock(cli, grants, tmp); grants = 0; } } + spin_unlock(&cli->cl_loi_list_lock); rc = 0; } else if (ext != NULL) { /* index is located outside of active extent */ @@ -2427,7 +2470,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, } if (ext == NULL) { - int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax; + tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax; /* try to find new extent to cover this page */ LASSERT(oio->oi_active == NULL); @@ -2471,9 +2514,13 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io, else LASSERT(ext->oe_srvlock == ops->ops_srvlock); ++ext->oe_nr_pages; - cfs_list_add_tail(&oap->oap_pending_item, &ext->oe_pages); + list_add_tail(&oap->oap_pending_item, &ext->oe_pages); osc_object_unlock(osc); + + if (!ext->oe_layout_version) + ext->oe_layout_version = io->ci_layout_version; } + RETURN(rc); } @@ -2481,33 +2528,34 @@ int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj, struct osc_page *ops) { struct osc_async_page *oap = &ops->ops_oap; - struct osc_extent *ext = NULL; int rc = 0; ENTRY; LASSERT(oap->oap_magic == OAP_MAGIC); CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n", - oap, ops, oap2cl_page(oap)->cp_index); + oap, ops, osc_index(oap2osc(oap))); - osc_object_lock(obj); - if (!cfs_list_empty(&oap->oap_rpc_item)) { + if (!list_empty(&oap->oap_rpc_item)) { CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap); rc = -EBUSY; - } else if (!cfs_list_empty(&oap->oap_pending_item)) { - ext = osc_extent_lookup(obj, oap2cl_page(oap)->cp_index); + } else if (!list_empty(&oap->oap_pending_item)) { + struct osc_extent *ext = NULL; + + osc_object_lock(obj); + ext = osc_extent_lookup(obj, osc_index(oap2osc(oap))); + osc_object_unlock(obj); /* only truncated pages are allowed to be taken out. * See osc_extent_truncate() and osc_cache_truncate_start() * for details. */ if (ext != NULL && ext->oe_state != OES_TRUNC) { OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n", - oap2cl_page(oap)->cp_index); + osc_index(oap2osc(oap))); rc = -EBUSY; } + if (ext != NULL) + osc_extent_put(env, ext); } - osc_object_unlock(obj); - if (ext != NULL) - osc_extent_put(env, ext); RETURN(rc); } @@ -2524,7 +2572,7 @@ int osc_flush_async_page(const struct lu_env *env, struct cl_io *io, struct osc_extent *ext = NULL; struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj); struct cl_page *cp = ops->ops_cl.cpl_page; - pgoff_t index = cp->cp_index; + pgoff_t index = osc_index(ops); struct osc_async_page *oap = &ops->ops_oap; bool unplug = false; int rc = 0; @@ -2540,8 +2588,7 @@ int osc_flush_async_page(const struct lu_env *env, struct cl_io *io, switch (ext->oe_state) { case OES_RPC: case OES_LOCK_DONE: - CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(cp), - "flush an in-rpc page?\n"); + CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n"); LASSERT(0); break; case OES_LOCKING: @@ -2553,12 +2600,18 @@ int osc_flush_async_page(const struct lu_env *env, struct cl_io *io, * really sending the RPC. */ case OES_TRUNC: /* race with truncate, page will be redirtied */ + case OES_ACTIVE: + /* The extent is active so we need to abort and let the caller + * re-dirty the page. If we continued on here, and we were the + * one making the extent active, we could deadlock waiting for + * the page writeback to clear but it won't because the extent + * is active and won't be written out. */ GOTO(out, rc = -EAGAIN); default: break; } - rc = cl_page_prep(env, io, cl_page_top(cp), CRT_WRITE); + rc = cl_page_prep(env, io, cp, CRT_WRITE); if (rc) GOTO(out, rc); @@ -2573,8 +2626,8 @@ int osc_flush_async_page(const struct lu_env *env, struct cl_io *io, if (ext->oe_state == OES_CACHE) { OSC_EXTENT_DUMP(D_CACHE, ext, "flush page %p make it urgent.\n", oap); - if (cfs_list_empty(&ext->oe_link)) - cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts); + if (list_empty(&ext->oe_link)) + list_add_tail(&ext->oe_link, &obj->oo_urgent_exts); unplug = true; } rc = 0; @@ -2602,8 +2655,8 @@ int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops) struct client_obd *cli = osc_cli(obj); struct osc_extent *ext; struct osc_extent *found = NULL; - cfs_list_t *plist; - pgoff_t index = oap2cl_page(oap)->cp_index; + struct list_head *plist; + pgoff_t index = osc_index(ops); int rc = -EBUSY; int cmd; ENTRY; @@ -2620,7 +2673,7 @@ int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops) plist = &obj->oo_reading_exts; cmd = OBD_BRW_READ; } - cfs_list_for_each_entry(ext, plist, oe_link) { + list_for_each_entry(ext, plist, oe_link) { if (ext->oe_start <= index && ext->oe_end >= index) { LASSERT(ext->oe_state == OES_LOCK_DONE); /* For OES_LOCK_DONE state extent, it has already held @@ -2630,7 +2683,7 @@ int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops) } } if (found != NULL) { - cfs_list_del_init(&found->oe_link); + list_del_init(&found->oe_link); osc_update_pending(obj, cmd, -found->oe_nr_pages); osc_object_unlock(obj); @@ -2653,74 +2706,87 @@ int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops) RETURN(rc); } -int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj, - cfs_list_t *list, int cmd, int brw_flags) +int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io, + struct osc_object *obj, struct list_head *list, + int brw_flags) { struct client_obd *cli = osc_cli(obj); struct osc_extent *ext; struct osc_async_page *oap; int page_count = 0; int mppr = cli->cl_max_pages_per_rpc; + bool can_merge = true; pgoff_t start = CL_PAGE_EOF; pgoff_t end = 0; ENTRY; - cfs_list_for_each_entry(oap, list, oap_pending_item) { - struct cl_page *cp = oap2cl_page(oap); - if (cp->cp_index > end) - end = cp->cp_index; - if (cp->cp_index < start) - start = cp->cp_index; + list_for_each_entry(oap, list, oap_pending_item) { + struct osc_page *opg = oap2osc_page(oap); + pgoff_t index = osc_index(opg); + + if (index > end) + end = index; + if (index < start) + start = index; ++page_count; mppr <<= (page_count > mppr); + + if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE)) + can_merge = false; } ext = osc_extent_alloc(obj); if (ext == NULL) { - cfs_list_for_each_entry(oap, list, oap_pending_item) { - cfs_list_del_init(&oap->oap_pending_item); + struct osc_async_page *tmp; + + list_for_each_entry_safe(oap, tmp, list, oap_pending_item) { + list_del_init(&oap->oap_pending_item); osc_ap_completion(env, cli, oap, 0, -ENOMEM); } RETURN(-ENOMEM); } - ext->oe_rw = !!(cmd & OBD_BRW_READ); + ext->oe_rw = !!(brw_flags & OBD_BRW_READ); + ext->oe_sync = 1; + ext->oe_no_merge = !can_merge; ext->oe_urgent = 1; ext->oe_start = start; ext->oe_end = ext->oe_max_end = end; ext->oe_obj = obj; ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK); + ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY); ext->oe_nr_pages = page_count; ext->oe_mppr = mppr; - cfs_list_splice_init(list, &ext->oe_pages); + list_splice_init(list, &ext->oe_pages); + ext->oe_layout_version = io->ci_layout_version; osc_object_lock(obj); /* Reuse the initial refcount for RPC, don't drop it */ osc_extent_state_set(ext, OES_LOCK_DONE); - if (cmd & OBD_BRW_WRITE) { - cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts); + if (!ext->oe_rw) { /* write */ + list_add_tail(&ext->oe_link, &obj->oo_urgent_exts); osc_update_pending(obj, OBD_BRW_WRITE, page_count); } else { - cfs_list_add_tail(&ext->oe_link, &obj->oo_reading_exts); + list_add_tail(&ext->oe_link, &obj->oo_reading_exts); osc_update_pending(obj, OBD_BRW_READ, page_count); } osc_object_unlock(obj); - osc_io_unplug(env, cli, obj, PDL_POLICY_ROUND); + osc_io_unplug_async(env, cli, obj); RETURN(0); } /** * Called by osc_io_setattr_start() to freeze and destroy covering extents. */ -int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio, - struct osc_object *obj, __u64 size) +int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj, + __u64 size, struct osc_extent **extp) { struct client_obd *cli = osc_cli(obj); struct osc_extent *ext; struct osc_extent *waiting = NULL; pgoff_t index; - CFS_LIST_HEAD(list); + struct list_head list = LIST_HEAD_INIT(list); int result = 0; bool partial; ENTRY; @@ -2744,14 +2810,13 @@ again: * a page already having been flushed by write_page(). * We have to wait for this extent because we can't * truncate that page. */ - LASSERT(!ext->oe_hp); OSC_EXTENT_DUMP(D_CACHE, ext, "waiting for busy extent\n"); waiting = osc_extent_get(ext); break; } - OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size); + OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size); osc_extent_get(ext); if (ext->oe_state == OES_ACTIVE) { @@ -2767,8 +2832,12 @@ again: osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages); } - EASSERT(cfs_list_empty(&ext->oe_link), ext); - cfs_list_add_tail(&ext->oe_link, &list); + /* This extent could be on the full extents list, that's OK */ + EASSERT(!ext->oe_hp && !ext->oe_urgent, ext); + if (!list_empty(&ext->oe_link)) + list_move_tail(&ext->oe_link, &list); + else + list_add_tail(&ext->oe_link, &list); ext = next_extent(ext); } @@ -2776,11 +2845,11 @@ again: osc_list_maint(cli, obj); - while (!cfs_list_empty(&list)) { + while (!list_empty(&list)) { int rc; - ext = cfs_list_entry(list.next, struct osc_extent, oe_link); - cfs_list_del_init(&ext->oe_link); + ext = list_entry(list.next, struct osc_extent, oe_link); + list_del_init(&ext->oe_link); /* extent may be in OES_ACTIVE state because inode mutex * is released before osc_io_end() in file write case */ @@ -2808,11 +2877,13 @@ again: /* we need to hold this extent in OES_TRUNC state so * that no writeback will happen. This is to avoid - * BUG 17397. */ - LASSERT(oio->oi_trunc == NULL); - oio->oi_trunc = osc_extent_get(ext); + * BUG 17397. + * Only partial truncate can reach here, if @size is + * not zero, the caller should provide a valid @extp. */ + LASSERT(*extp == NULL); + *extp = osc_extent_get(ext); OSC_EXTENT_DUMP(D_CACHE, ext, - "trunc at "LPU64"\n", size); + "trunc at %llu\n", size); } osc_extent_put(env, ext); } @@ -2831,17 +2902,15 @@ again: } RETURN(result); } +EXPORT_SYMBOL(osc_cache_truncate_start); /** * Called after osc_io_setattr_end to add oio->oi_trunc back to cache. */ -void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio, - struct osc_object *obj) +void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext) { - struct osc_extent *ext = oio->oi_trunc; - - oio->oi_trunc = NULL; if (ext != NULL) { + struct osc_object *obj = ext->oe_obj; bool unplug = false; EASSERT(ext->oe_nr_pages > 0, ext); @@ -2853,7 +2922,7 @@ void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio, osc_extent_state_set(ext, OES_CACHE); if (ext->oe_fsync_wait && !ext->oe_urgent) { ext->oe_urgent = 1; - cfs_list_move_tail(&ext->oe_link, &obj->oo_urgent_exts); + list_move_tail(&ext->oe_link, &obj->oo_urgent_exts); unplug = true; } osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages); @@ -2920,6 +2989,7 @@ again: OSC_IO_DEBUG(obj, "sync file range.\n"); RETURN(result); } +EXPORT_SYMBOL(osc_cache_wait_range); /** * Called to write out a range of osc object. @@ -2934,7 +3004,7 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj, pgoff_t start, pgoff_t end, int hp, int discard) { struct osc_extent *ext; - CFS_LIST_HEAD(discard_list); + struct list_head discard_list = LIST_HEAD_INIT(discard_list); bool unplug = false; int result = 0; ENTRY; @@ -2954,7 +3024,7 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj, case OES_CACHE: result += ext->oe_nr_pages; if (!discard) { - cfs_list_t *list = NULL; + struct list_head *list = NULL; if (hp) { EASSERT(!ext->oe_hp, ext); ext->oe_hp = 1; @@ -2964,16 +3034,31 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj, list = &obj->oo_urgent_exts; } if (list != NULL) - cfs_list_move_tail(&ext->oe_link, list); + list_move_tail(&ext->oe_link, list); unplug = true; } else { + struct client_obd *cli = osc_cli(obj); + int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT; + pgoff_t align_by = (1 << pcc_bits); + pgoff_t a_start = round_down(start, align_by); + pgoff_t a_end = round_up(end, align_by); + + /* overflow case */ + if (end && !a_end) + a_end = CL_PAGE_EOF; /* the only discarder is lock cancelling, so - * [start, end] must contain this extent */ - EASSERT(ext->oe_start >= start && - ext->oe_max_end <= end, ext); + * [start, end], aligned by chunk size, must + * contain this extent */ + LASSERTF(ext->oe_start >= a_start && + ext->oe_end <= a_end, + "ext [%lu, %lu] reg [%lu, %lu] " + "orig [%lu %lu] align %lu bits " + "%d\n", ext->oe_start, ext->oe_end, + a_start, a_end, start, end, + align_by, pcc_bits); osc_extent_state_set(ext, OES_LOCKING); ext->oe_owner = current; - cfs_list_move_tail(&ext->oe_link, + list_move_tail(&ext->oe_link, &discard_list); osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages); @@ -2998,14 +3083,14 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj, } osc_object_unlock(obj); - LASSERT(ergo(!discard, cfs_list_empty(&discard_list))); - if (!cfs_list_empty(&discard_list)) { + LASSERT(ergo(!discard, list_empty(&discard_list))); + if (!list_empty(&discard_list)) { struct osc_extent *tmp; int rc; osc_list_maint(osc_cli(obj), obj); - cfs_list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) { - cfs_list_del_init(&ext->oe_link); + list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) { + list_del_init(&ext->oe_link); EASSERT(ext->oe_state == OES_LOCKING, ext); /* Discard caching pages. We don't actually write this @@ -3024,7 +3109,7 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj, } if (unplug) - osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND); + osc_io_unplug(env, osc_cli(obj), obj); if (hp || discard) { int rc; @@ -3036,5 +3121,212 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj, OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result); RETURN(result); } +EXPORT_SYMBOL(osc_cache_writeback_range); + +/** + * Returns a list of pages by a given [start, end] of \a obj. + * + * \param resched If not NULL, then we give up before hogging CPU for too + * long and set *resched = 1, in that case caller should implement a retry + * logic. + * + * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely + * crucial in the face of [offset, EOF] locks. + * + * Return at least one page in @queue unless there is no covered page. + */ +int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io, + struct osc_object *osc, pgoff_t start, pgoff_t end, + osc_page_gang_cbt cb, void *cbdata) +{ + struct osc_page *ops; + struct pagevec *pagevec; + void **pvec; + pgoff_t idx; + unsigned int nr; + unsigned int i; + unsigned int j; + int res = CLP_GANG_OKAY; + bool tree_lock = true; + ENTRY; + + idx = start; + pvec = osc_env_info(env)->oti_pvec; + pagevec = &osc_env_info(env)->oti_pagevec; + ll_pagevec_init(pagevec, 0); + spin_lock(&osc->oo_tree_lock); + while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec, + idx, OTI_PVEC_SIZE)) > 0) { + struct cl_page *page; + bool end_of_region = false; + + for (i = 0, j = 0; i < nr; ++i) { + ops = pvec[i]; + pvec[i] = NULL; + + idx = osc_index(ops); + if (idx > end) { + end_of_region = true; + break; + } + + page = ops->ops_cl.cpl_page; + LASSERT(page->cp_type == CPT_CACHEABLE); + if (page->cp_state == CPS_FREEING) + continue; + + cl_page_get(page); + lu_ref_add_atomic(&page->cp_reference, + "gang_lookup", current); + pvec[j++] = ops; + } + ++idx; + + /* + * Here a delicate locking dance is performed. Current thread + * holds a reference to a page, but has to own it before it + * can be placed into queue. Owning implies waiting, so + * radix-tree lock is to be released. After a wait one has to + * check that pages weren't truncated (cl_page_own() returns + * error in the latter case). + */ + spin_unlock(&osc->oo_tree_lock); + tree_lock = false; + + for (i = 0; i < j; ++i) { + ops = pvec[i]; + if (res == CLP_GANG_OKAY) + res = (*cb)(env, io, ops, cbdata); + + page = ops->ops_cl.cpl_page; + lu_ref_del(&page->cp_reference, "gang_lookup", current); + cl_pagevec_put(env, page, pagevec); + } + pagevec_release(pagevec); + + if (nr < OTI_PVEC_SIZE || end_of_region) + break; + + if (res == CLP_GANG_OKAY && need_resched()) + res = CLP_GANG_RESCHED; + if (res != CLP_GANG_OKAY) + break; + + spin_lock(&osc->oo_tree_lock); + tree_lock = true; + } + if (tree_lock) + spin_unlock(&osc->oo_tree_lock); + RETURN(res); +} +EXPORT_SYMBOL(osc_page_gang_lookup); + +/** + * Check if page @page is covered by an extra lock or discard it. + */ +static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io, + struct osc_page *ops, void *cbdata) +{ + struct osc_thread_info *info = osc_env_info(env); + struct osc_object *osc = cbdata; + pgoff_t index; + + index = osc_index(ops); + if (index >= info->oti_fn_index) { + struct ldlm_lock *tmp; + struct cl_page *page = ops->ops_cl.cpl_page; + + /* refresh non-overlapped index */ + tmp = osc_dlmlock_at_pgoff(env, osc, index, + OSC_DAP_FL_TEST_LOCK); + if (tmp != NULL) { + __u64 end = tmp->l_policy_data.l_extent.end; + /* Cache the first-non-overlapped index so as to skip + * all pages within [index, oti_fn_index). This is safe + * because if tmp lock is canceled, it will discard + * these pages. */ + info->oti_fn_index = cl_index(osc2cl(osc), end + 1); + if (end == OBD_OBJECT_EOF) + info->oti_fn_index = CL_PAGE_EOF; + LDLM_LOCK_PUT(tmp); + } else if (cl_page_own(env, io, page) == 0) { + /* discard the page */ + cl_page_discard(env, io, page); + cl_page_disown(env, io, page); + } else { + LASSERT(page->cp_state == CPS_FREEING); + } + } + + info->oti_next_index = index + 1; + return CLP_GANG_OKAY; +} + +int osc_discard_cb(const struct lu_env *env, struct cl_io *io, + struct osc_page *ops, void *cbdata) +{ + struct osc_thread_info *info = osc_env_info(env); + struct cl_page *page = ops->ops_cl.cpl_page; + + /* page is top page. */ + info->oti_next_index = osc_index(ops) + 1; + if (cl_page_own(env, io, page) == 0) { + if (!ergo(page->cp_type == CPT_CACHEABLE, + !PageDirty(cl_page_vmpage(page)))) + CL_PAGE_DEBUG(D_ERROR, env, page, + "discard dirty page?\n"); + + /* discard the page */ + cl_page_discard(env, io, page); + cl_page_disown(env, io, page); + } else { + LASSERT(page->cp_state == CPS_FREEING); + } + + return CLP_GANG_OKAY; +} +EXPORT_SYMBOL(osc_discard_cb); + +/** + * Discard pages protected by the given lock. This function traverses radix + * tree to find all covering pages and discard them. If a page is being covered + * by other locks, it should remain in cache. + * + * If error happens on any step, the process continues anyway (the reasoning + * behind this being that lock cancellation cannot be delayed indefinitely). + */ +int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc, + pgoff_t start, pgoff_t end, bool discard) +{ + struct osc_thread_info *info = osc_env_info(env); + struct cl_io *io = osc_env_thread_io(env); + osc_page_gang_cbt cb; + int res; + int result; + + ENTRY; + + io->ci_obj = cl_object_top(osc2cl(osc)); + io->ci_ignore_layout = 1; + result = cl_io_init(env, io, CIT_MISC, io->ci_obj); + if (result != 0) + GOTO(out, result); + + cb = discard ? osc_discard_cb : check_and_discard_cb; + info->oti_fn_index = info->oti_next_index = start; + do { + res = osc_page_gang_lookup(env, io, osc, + info->oti_next_index, end, cb, osc); + if (info->oti_next_index > end) + break; + + if (res == CLP_GANG_RESCHED) + cond_resched(); + } while (res != CLP_GANG_OKAY); +out: + cl_io_fini(env, io); + RETURN(result); +} + /** @} osc */