X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosc%2Fosc_page.c;h=cc8df4312e305d5575fcdecdd4d4ea7e9840ffcd;hb=6d20df8d6f9cbb6e1fe5debad62a84d130369894;hp=e72e6d2d9460f1e34fda873cb64c3c9a7a47c279;hpb=84495601f9401b0288c8a8acab656a94a64408c2;p=fs%2Flustre-release.git diff --git a/lustre/osc/osc_page.c b/lustre/osc/osc_page.c index e72e6d2..cc8df43 100644 --- a/lustre/osc/osc_page.c +++ b/lustre/osc/osc_page.c @@ -38,12 +38,22 @@ * Author: Nikita Danilov */ -/** \addtogroup osc osc @{ */ - #define DEBUG_SUBSYSTEM S_OSC #include "osc_cl_internal.h" +/** \addtogroup osc + * @{ + */ + +/* + * Comment out osc_page_protected because it may sleep inside the + * the client_obd_list_lock. + * client_obd_list_lock -> osc_ap_completion -> osc_completion -> + * -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base + * -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep. + */ +#if 0 static int osc_page_is_dlocked(const struct lu_env *env, const struct osc_page *opg, enum cl_lock_mode mode, int pending, int unref) @@ -57,6 +67,8 @@ static int osc_page_is_dlocked(const struct lu_env *env, ldlm_mode_t dlmmode; int flags; + cfs_might_sleep(); + info = osc_env_info(env); resname = &info->oti_resname; policy = &info->oti_policy; @@ -75,6 +87,10 @@ static int osc_page_is_dlocked(const struct lu_env *env, dlmmode, &flags, NULL, lockh, unref); } +/** + * Checks an invariant that a page in the cache is covered by a lock, as + * needed. + */ static int osc_page_protected(const struct lu_env *env, const struct osc_page *opg, enum cl_lock_mode mode, int unref) @@ -87,17 +103,26 @@ static int osc_page_protected(const struct lu_env *env, LINVRNT(!opg->ops_temp); + page = opg->ops_cl.cpl_page; + if (page->cp_owner != NULL && + cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER) + /* + * If IO is done without locks (liblustre, or lloop), lock is + * not required. + */ + result = 1; + else + /* otherwise check for a DLM lock */ result = osc_page_is_dlocked(env, opg, mode, 1, unref); if (result == 0) { /* maybe this page is a part of a lockless io? */ hdr = cl_object_header(opg->ops_cl.cpl_obj); - page = opg->ops_cl.cpl_page; descr = &osc_env_info(env)->oti_descr; descr->cld_mode = mode; descr->cld_start = page->cp_index; descr->cld_end = page->cp_index; - spin_lock(&hdr->coh_lock_guard); - list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) { + cfs_spin_lock(&hdr->coh_lock_guard); + cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) { /* * Lock-less sub-lock has to be either in HELD state * (when io is actively going on), or in CACHED state, @@ -114,10 +139,18 @@ static int osc_page_protected(const struct lu_env *env, break; } } - spin_unlock(&hdr->coh_lock_guard); + cfs_spin_unlock(&hdr->coh_lock_guard); } return result; } +#else +static int osc_page_protected(const struct lu_env *env, + const struct osc_page *opg, + enum cl_lock_mode mode, int unref) +{ + return 1; +} +#endif /***************************************************************************** * @@ -167,19 +200,18 @@ static void osc_page_transfer_add(const struct lu_env *env, LINVRNT(cl_page_is_vmlocked(env, opg->ops_cl.cpl_page)); obj = cl2osc(opg->ops_cl.cpl_obj); - spin_lock(&obj->oo_seatbelt); - list_add(&opg->ops_inflight, &obj->oo_inflight[crt]); + cfs_spin_lock(&obj->oo_seatbelt); + cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]); opg->ops_submitter = cfs_current(); - spin_unlock(&obj->oo_seatbelt); + cfs_spin_unlock(&obj->oo_seatbelt); } static int osc_page_cache_add(const struct lu_env *env, const struct cl_page_slice *slice, - struct cl_io *_) + struct cl_io *unused) { struct osc_page *opg = cl2osc_page(slice); struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj); - struct osc_io *oio = osc_env_io(env); int result; int brw_flags; int noquota = 0; @@ -188,7 +220,7 @@ static int osc_page_cache_add(const struct lu_env *env, ENTRY; /* Set the OBD_BRW_SRVLOCK before the page is queued. */ - brw_flags = oio->oi_lockless ? OBD_BRW_SRVLOCK : 0; + brw_flags = opg->ops_srvlock ? OBD_BRW_SRVLOCK : 0; if (!client_is_remote(osc_export(obj)) && cfs_capable(CFS_CAP_SYS_RESOURCE)) { brw_flags |= OBD_BRW_NOQUOTA; @@ -216,7 +248,7 @@ void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj, static int osc_page_is_under_lock(const struct lu_env *env, const struct cl_page_slice *slice, - struct cl_io *_) + struct cl_io *unused) { struct cl_lock *lock; int result; @@ -233,7 +265,8 @@ static int osc_page_is_under_lock(const struct lu_env *env, } static int osc_page_fail(const struct lu_env *env, - const struct cl_page_slice *slice, struct cl_io *_) + const struct cl_page_slice *slice, + struct cl_io *unused) { /* * Cached read? @@ -243,9 +276,17 @@ static int osc_page_fail(const struct lu_env *env, } -static const char *osc_list(struct list_head *head) +static const char *osc_list(cfs_list_t *head) { - return list_empty(head) ? "-" : "+"; + return cfs_list_empty(head) ? "-" : "+"; +} + +static inline cfs_time_t osc_submit_duration(struct osc_page *opg) +{ + if (opg->ops_submit_time == 0) + return 0; + + return (cfs_time_current() - opg->ops_submit_time); } static int osc_page_print(const struct lu_env *env, @@ -254,18 +295,53 @@ static int osc_page_print(const struct lu_env *env, { struct osc_page *opg = cl2osc_page(slice); struct osc_async_page *oap = &opg->ops_oap; + struct osc_object *obj = cl2osc(slice->cpl_obj); + struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli; + struct lov_oinfo *loi = obj->oo_oinfo; return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: " - "%#x %d %u %s %s %s %llu %u %#x %p %p %p %p %p\n", - opg, oap->oap_magic, oap->oap_cmd, + "1< %#x %d %u %s %s %s > " + "2< "LPU64" %u %#x %#x | %p %p %p %p %p > " + "3< %s %p %d %lu %d > " + "4< %d %d %d %lu %s | %s %s %s %s > " + "5< %s %s %s %s | %d %s %s | %d %s %s>\n", + opg, + /* 1 */ + oap->oap_magic, oap->oap_cmd, oap->oap_interrupted, osc_list(&oap->oap_pending_item), osc_list(&oap->oap_urgent_item), osc_list(&oap->oap_rpc_item), + /* 2 */ oap->oap_obj_off, oap->oap_page_off, - oap->oap_async_flags, oap->oap_request, + oap->oap_async_flags, oap->oap_brw_flags, + oap->oap_request, oap->oap_cli, oap->oap_loi, oap->oap_caller_ops, - oap->oap_caller_data); + oap->oap_caller_data, + /* 3 */ + osc_list(&opg->ops_inflight), + opg->ops_submitter, opg->ops_transfer_pinned, + osc_submit_duration(opg), opg->ops_srvlock, + /* 4 */ + cli->cl_r_in_flight, cli->cl_w_in_flight, + cli->cl_max_rpcs_in_flight, + cli->cl_avail_grant, + osc_list(&cli->cl_cache_waiters), + osc_list(&cli->cl_loi_ready_list), + osc_list(&cli->cl_loi_hp_ready_list), + osc_list(&cli->cl_loi_write_list), + osc_list(&cli->cl_loi_read_list), + /* 5 */ + osc_list(&loi->loi_ready_item), + osc_list(&loi->loi_hp_ready_item), + osc_list(&loi->loi_write_item), + osc_list(&loi->loi_read_item), + loi->loi_read_lop.lop_num_pending, + osc_list(&loi->loi_read_lop.lop_pending), + osc_list(&loi->loi_read_lop.lop_urgent), + loi->loi_write_lop.lop_num_pending, + osc_list(&loi->loi_write_lop.lop_pending), + osc_list(&loi->loi_write_lop.lop_urgent)); } static void osc_page_delete(const struct lu_env *env, @@ -282,10 +358,14 @@ static void osc_page_delete(const struct lu_env *env, CDEBUG(D_TRACE, "%p\n", opg); osc_page_transfer_put(env, opg); rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap); - LASSERTF(rc == 0, "%i\n", rc); - spin_lock(&obj->oo_seatbelt); - list_del_init(&opg->ops_inflight); - spin_unlock(&obj->oo_seatbelt); + if (rc) { + CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page), + "Trying to teardown failed: %d\n", rc); + LASSERT(0); + } + cfs_spin_lock(&obj->oo_seatbelt); + cfs_list_del_init(&opg->ops_inflight); + cfs_spin_unlock(&obj->oo_seatbelt); EXIT; } @@ -299,7 +379,9 @@ void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice, opg->ops_from = from; opg->ops_to = to; + cfs_spin_lock(&oap->oap_lock); oap->oap_async_flags |= ASYNC_COUNT_STABLE; + cfs_spin_unlock(&oap->oap_lock); } static int osc_page_cancel(const struct lu_env *env, @@ -350,6 +432,8 @@ static int osc_make_ready(const struct lu_env *env, void *data, int cmd) ENTRY; result = cl_page_make_ready(env, page, CRT_WRITE); + if (result == 0) + opg->ops_submit_time = cfs_time_current(); RETURN(result); } @@ -415,17 +499,21 @@ static int osc_completion(const struct lu_env *env, LASSERT(page->cp_req == NULL); /* As the transfer for this page is being done, clear the flags */ + cfs_spin_lock(&oap->oap_lock); oap->oap_async_flags = 0; + cfs_spin_unlock(&oap->oap_lock); 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); + cfs_spin_lock(&obj->oo_seatbelt); LASSERT(opg->ops_submitter != NULL); - LASSERT(!list_empty(&opg->ops_inflight)); - list_del_init(&opg->ops_inflight); - spin_unlock(&obj->oo_seatbelt); + LASSERT(!cfs_list_empty(&opg->ops_inflight)); + cfs_list_del_init(&opg->ops_inflight); + cfs_spin_unlock(&obj->oo_seatbelt); + + opg->ops_submit_time = 0; cl_page_completion(env, page, crt, rc); @@ -470,7 +558,7 @@ struct cl_page *osc_page_init(const struct lu_env *env, struct osc_page *opg; int result; - OBD_SLAB_ALLOC_PTR(opg, osc_page_kmem); + OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO); if (opg != NULL) { void *oap = &opg->ops_oap; @@ -482,9 +570,12 @@ struct cl_page *osc_page_init(const struct lu_env *env, cl_offset(obj, page->cp_index), &osc_async_page_ops, opg, (void **)&oap, 1, NULL); - if (result == 0) + if (result == 0) { + struct osc_io *oio = osc_env_io(env); + opg->ops_srvlock = osc_io_srvlock(oio); cl_page_slice_add(page, &opg->ops_cl, obj, &osc_page_ops); + } /* * Cannot assert osc_page_protected() here as read-ahead * creates temporary pages outside of a lock. @@ -508,14 +599,18 @@ void osc_io_submit_page(const struct lu_env *env, { struct osc_async_page *oap = &opg->ops_oap; struct client_obd *cli = oap->oap_cli; + int flags = 0; LINVRNT(osc_page_protected(env, opg, crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1)); oap->oap_page_off = opg->ops_from; oap->oap_count = opg->ops_to - opg->ops_from; + /* Give a hint to OST that requests are coming from kswapd - bug19529 */ + if (libcfs_memory_pressure_get()) + oap->oap_brw_flags |= OBD_BRW_MEMALLOC; oap->oap_brw_flags |= OBD_BRW_SYNC; - if (oio->oi_lockless) + if (osc_io_srvlock(oio)) oap->oap_brw_flags |= OBD_BRW_SRVLOCK; oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ; @@ -525,12 +620,15 @@ void osc_io_submit_page(const struct lu_env *env, oap->oap_cmd |= OBD_BRW_NOQUOTA; } - oap->oap_async_flags |= OSC_FLAGS; if (oap->oap_cmd & OBD_BRW_READ) - oap->oap_async_flags |= ASYNC_COUNT_STABLE; + flags = ASYNC_COUNT_STABLE; else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT)) osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1); + cfs_spin_lock(&oap->oap_lock); + oap->oap_async_flags |= OSC_FLAGS | flags; + cfs_spin_unlock(&oap->oap_lock); + osc_oap_to_pending(oap); osc_page_transfer_get(opg, "transfer\0imm"); osc_page_transfer_add(env, opg, crt);