X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosc%2Fosc_page.c;h=eb60998012bc0c4c5b47b57e5d6236d3ee7ac909;hp=9dbad530e33d04709f0b41377cfce5fd5f328084;hb=3c6a1e94c652685fac7c69bf530e05d27b7f477c;hpb=8151daa2c8d97917b37b9271c5d6936a75b48328 diff --git a/lustre/osc/osc_page.c b/lustre/osc/osc_page.c index 9dbad53..eb60998 100644 --- a/lustre/osc/osc_page.c +++ b/lustre/osc/osc_page.c @@ -27,7 +27,6 @@ */ /* * This file is part of Lustre, http://www.lustre.org/ - * Lustre is a trademark of Sun Microsystems, Inc. * * Implementation of cl_page for OSC layer. * @@ -150,7 +149,7 @@ static int osc_page_print(const struct lu_env *env, cli->cl_r_in_flight, cli->cl_w_in_flight, cli->cl_max_rpcs_in_flight, cli->cl_avail_grant, - list_empty_marker(&cli->cl_cache_waiters), + waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-', list_empty_marker(&cli->cl_loi_ready_list), list_empty_marker(&cli->cl_loi_hp_ready_list), list_empty_marker(&cli->cl_loi_write_list), @@ -214,7 +213,11 @@ static void osc_page_clip(const struct lu_env *env, struct osc_async_page *oap = &opg->ops_oap; opg->ops_from = from; - opg->ops_to = to; + /* argument @to is exclusive, but @ops_to is inclusive */ + opg->ops_to = to - 1; + /* This isn't really necessary for transient pages, but we also don't + * call clip on transient pages often, so it's OK. + */ spin_lock(&oap->oap_lock); oap->oap_async_flags |= ASYNC_COUNT_STABLE; spin_unlock(&oap->oap_lock); @@ -249,30 +252,28 @@ static const struct cl_page_operations osc_page_ops = { }; int osc_page_init(const struct lu_env *env, struct cl_object *obj, - struct cl_page *page, pgoff_t index) + struct cl_page *cl_page, pgoff_t index) { struct osc_object *osc = cl2osc(obj); - struct osc_page *opg = cl_object_page_slice(obj, page); + struct osc_page *opg = cl_object_page_slice(obj, cl_page); struct osc_io *oio = osc_env_io(env); int result; opg->ops_from = 0; - opg->ops_to = PAGE_SIZE; + opg->ops_to = PAGE_SIZE - 1; INIT_LIST_HEAD(&opg->ops_lru); - result = osc_prep_async_page(osc, opg, page->cp_vmpage, - cl_offset(obj, index)); + result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index)); if (result != 0) return result; opg->ops_srvlock = osc_io_srvlock(oio); - cl_page_slice_add(page, &opg->ops_cl, obj, index, - &osc_page_ops); - + cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops); + cl_page->cp_osc_index = index; /* reserve an LRU space for this page */ - if (page->cp_type == CPT_CACHEABLE) { + if (cl_page->cp_type == CPT_CACHEABLE) { result = osc_lru_alloc(env, osc_cli(osc), opg); if (result == 0) { result = radix_tree_preload(GFP_NOFS); @@ -300,7 +301,7 @@ EXPORT_SYMBOL(osc_page_init); * transfer (i.e., transferred synchronously). */ void osc_page_submit(const struct lu_env *env, struct osc_page *opg, - enum cl_req_type crt, int brw_flags) + enum cl_req_type crt, int brw_flags, ktime_t submit_time) { struct osc_io *oio = osc_env_io(env); struct osc_async_page *oap = &opg->ops_oap; @@ -310,17 +311,17 @@ void osc_page_submit(const struct lu_env *env, struct osc_page *opg, LASSERT(oap->oap_async_flags & ASYNC_READY); LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE); - oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ; - oap->oap_page_off = opg->ops_from; - oap->oap_count = opg->ops_to - opg->ops_from; + oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ; + oap->oap_page_off = opg->ops_from; + oap->oap_count = opg->ops_to - opg->ops_from + 1; oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags; if (oio->oi_cap_sys_resource) { - oap->oap_brw_flags |= OBD_BRW_NOQUOTA; - oap->oap_cmd |= OBD_BRW_NOQUOTA; + oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE; + oap->oap_cmd |= OBD_BRW_SYS_RESOURCE; } - opg->ops_submit_time = ktime_get(); + opg->ops_submit_time = submit_time; osc_page_transfer_get(opg, "transfer\0imm"); osc_page_transfer_add(env, opg, crt); } @@ -597,8 +598,8 @@ long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli, if (--maxscan < 0) break; - opg = list_entry(cli->cl_lru_list.next, struct osc_page, - ops_lru); + opg = list_first_entry(&cli->cl_lru_list, struct osc_page, + ops_lru); page = opg->ops_cl.cpl_page; if (lru_page_busy(cli, page)) { list_move_tail(&opg->ops_lru, &cli->cl_lru_list); @@ -679,7 +680,7 @@ long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli, atomic_dec(&cli->cl_lru_shrinkers); if (count > 0) { atomic_long_add(count, cli->cl_lru_left); - wake_up_all(&osc_lru_waitq); + wake_up(&osc_lru_waitq); } RETURN(count > 0 ? count : rc); } @@ -695,6 +696,7 @@ static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages) { struct lu_env *env; struct cl_client_cache *cache = cli->cl_cache; + struct client_obd *scan; int max_scans; __u16 refcheck; long rc = 0; @@ -733,20 +735,20 @@ static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages) list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru); max_scans = atomic_read(&cache->ccc_users) - 2; - while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) { - cli = list_entry(cache->ccc_lru.next, struct client_obd, - cl_lru_osc); - + while (--max_scans > 0 && + (scan = list_first_entry_or_null(&cache->ccc_lru, + struct client_obd, + cl_lru_osc)) != NULL) { CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n", - cli_name(cli), cli, - atomic_long_read(&cli->cl_lru_in_list), - atomic_long_read(&cli->cl_lru_busy)); + cli_name(scan), scan, + atomic_long_read(&scan->cl_lru_in_list), + atomic_long_read(&scan->cl_lru_busy)); - list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru); - if (osc_cache_too_much(cli) > 0) { + list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru); + if (osc_cache_too_much(scan) > 0) { spin_unlock(&cache->ccc_lru_lock); - rc = osc_lru_shrink(env, cli, npages, true); + rc = osc_lru_shrink(env, scan, npages, true); spin_lock(&cache->ccc_lru_lock); if (rc >= npages) break; @@ -759,7 +761,7 @@ static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages) out: cl_env_put(env, &refcheck); CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n", - cli_name(cli), cli, rc); + cli_name(cli), cli, rc); return rc; } @@ -794,6 +796,13 @@ static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli, break; if (rc > 0) continue; + /* IO issued by readahead, don't try hard */ + if (oio->oi_is_readahead) { + if (atomic_long_read(cli->cl_lru_left) > 0) + continue; + rc = -EBUSY; + break; + } cond_resched(); rc = l_wait_event_abortable( @@ -826,16 +835,23 @@ unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages) unsigned long reserved = 0; unsigned long max_pages; unsigned long c; + int rc; - /* reserve a full RPC window at most to avoid that a thread accidentally - * consumes too many LRU slots */ - max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight; - if (npages > max_pages) - npages = max_pages; - +again: c = atomic_long_read(cli->cl_lru_left); if (c < npages && osc_lru_reclaim(cli, npages) > 0) c = atomic_long_read(cli->cl_lru_left); + + if (c < npages) { + /* + * Trigger writeback in the hope some LRU slot could + * be freed. + */ + rc = ptlrpcd_queue_work(cli->cl_writeback_work); + if (rc) + return 0; + } + while (c >= npages) { if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) { reserved = npages; @@ -843,6 +859,16 @@ unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages) } c = atomic_long_read(cli->cl_lru_left); } + + if (reserved != npages) { + cond_resched(); + rc = l_wait_event_abortable( + osc_lru_waitq, + atomic_long_read(cli->cl_lru_left) > 0); + goto again; + } + + max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight; if (atomic_long_read(cli->cl_lru_left) < max_pages) { /* If there aren't enough pages in the per-OSC LRU then * wake up the LRU thread to try and clear out space, so @@ -866,7 +892,7 @@ unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages) void osc_lru_unreserve(struct client_obd *cli, unsigned long npages) { atomic_long_add(npages, cli->cl_lru_left); - wake_up_all(&osc_lru_waitq); + wake_up(&osc_lru_waitq); } /** @@ -875,6 +901,13 @@ void osc_lru_unreserve(struct client_obd *cli, unsigned long npages) * In practice this can work pretty good because the pages in the same RPC * are likely from the same page zone. */ +#ifdef HAVE_NR_UNSTABLE_NFS +/* Old kernels use a separate counter for unstable pages, + * newer kernels treat them like any other writeback. + */ +#define NR_WRITEBACK NR_UNSTABLE_NFS +#endif + static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc, struct osc_brw_async_args *aa, int factor) @@ -893,7 +926,7 @@ static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc, for (i = 0; i < page_count; i++) { void *pz; if (desc) - pz = page_zone(BD_GET_KIOV(desc, i).kiov_page); + pz = page_zone(desc->bd_vec[i].bv_page); else pz = page_zone(aa->aa_ppga[i]->pg); @@ -903,7 +936,7 @@ static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc, } if (count > 0) { - mod_zone_page_state(zone, NR_UNSTABLE_NFS, + mod_zone_page_state(zone, NR_WRITEBACK, factor * count); count = 0; } @@ -911,7 +944,7 @@ static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc, ++count; } if (count > 0) - mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count); + mod_zone_page_state(zone, NR_WRITEBACK, factor * count); } static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc, @@ -962,7 +995,7 @@ void osc_dec_unstable_pages(struct ptlrpc_request *req) &cli->cl_cache->ccc_unstable_nr); LASSERT(unstable_count >= 0); if (unstable_count == 0) - wake_up_all(&cli->cl_cache->ccc_unstable_waitq); + wake_up(&cli->cl_cache->ccc_unstable_waitq); if (waitqueue_active(&osc_lru_waitq)) (void)ptlrpcd_queue_work(cli->cl_lru_work); @@ -1091,10 +1124,9 @@ unsigned long osc_cache_shrink_scan(struct shrinker *sk, return SHRINK_STOP; spin_lock(&osc_shrink_lock); - while (!list_empty(&osc_shrink_list)) { - cli = list_entry(osc_shrink_list.next, struct client_obd, - cl_shrink_list); - + while ((cli = list_first_entry_or_null(&osc_shrink_list, + struct client_obd, + cl_shrink_list)) != NULL) { if (stop_anchor == NULL) stop_anchor = cli; else if (cli == stop_anchor)