X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Frw.c;h=3d62c588e1a8dbe0e08bb1c7e16a286e3900e099;hp=8af953fee721e74234dac6d926e3e0192593cb05;hb=c4b74bf9e14a113e531d2d38a294a870ef3e0593;hpb=4a40e3ad61c4fcc72799db0e07dd73f55ab2a1a7 diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c index 8af953f..3d62c58 100644 --- a/lustre/llite/rw.c +++ b/lustre/llite/rw.c @@ -1,28 +1,44 @@ /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- * vim:expandtab:shiftwidth=8:tabstop=8: * - * Lustre Lite I/O page cache routines shared by different kernel revs + * GPL HEADER START + * + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * Copyright (c) 2001-2003 Cluster File Systems, Inc. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * This file is part of Lustre, http://www.lustre.org. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * 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 * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * 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. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * GPL HEADER END */ -#ifndef AUTOCONF_INCLUDED -#include -#endif +/* + * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Use is subject to license terms. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/llite/rw.c + * + * Lustre Lite I/O page cache routines shared by different kernel revs + */ + +#include #include #include #include @@ -37,7 +53,6 @@ #include #include #include -#include #include #include #include @@ -46,6 +61,7 @@ //#include #include +#include #include "llite_internal.h" #include @@ -109,6 +125,47 @@ static int ll_brw(int cmd, struct inode *inode, struct obdo *oa, RETURN(rc); } +int ll_file_punch(struct inode * inode, loff_t new_size, int srvlock) +{ + struct ll_inode_info *lli = ll_i2info(inode); + struct obd_info oinfo = { { { 0 } } }; + struct obdo oa; + int rc; + + ENTRY; + CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n", + lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode)); + + oinfo.oi_md = lli->lli_smd; + oinfo.oi_policy.l_extent.start = new_size; + oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF; + oinfo.oi_oa = &oa; + oa.o_id = lli->lli_smd->lsm_object_id; + oa.o_gr = lli->lli_smd->lsm_object_gr; + oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP; + if (srvlock) { + /* set OBD_MD_FLFLAGS in o_valid, only if we + * set OBD_FL_TRUNCLOCK, otherwise ost_punch + * and filter_setattr get confused, see the comment + * in ost_punch */ + oa.o_flags = OBD_FL_TRUNCLOCK; + oa.o_valid |= OBD_MD_FLFLAGS; + } + obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE | + OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME | + OBD_MD_FLFID | OBD_MD_FLGENER); + + oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC); + rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL); + ll_truncate_free_capa(oinfo.oi_capa); + if (rc) + CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino); + else + obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | + OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME); + RETURN(rc); +} + /* this isn't where truncate starts. roughly: * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to @@ -118,10 +175,8 @@ static int ll_brw(int cmd, struct inode *inode, struct obdo *oa, void ll_truncate(struct inode *inode) { struct ll_inode_info *lli = ll_i2info(inode); - struct obd_info oinfo = { { { 0 } } }; - struct ost_lvb lvb; - struct obdo oa; - int rc; + int srvlock = !!(lli->lli_flags & LLIF_SRVLOCK); + loff_t new_size; ENTRY; CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino, inode->i_generation, inode, i_size_read(inode), @@ -141,22 +196,27 @@ void ll_truncate(struct inode *inode) LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0); - /* XXX I'm pretty sure this is a hack to paper over a more fundamental - * race condition. */ - lov_stripe_lock(lli->lli_smd); - inode_init_lvb(inode, &lvb); - rc = obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0); - if (lvb.lvb_size == i_size_read(inode) && rc == 0) { - CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n", - lli->lli_smd->lsm_object_id, i_size_read(inode), - i_size_read(inode)); + if (!srvlock) { + struct ost_lvb lvb; + int rc; + + /* XXX I'm pretty sure this is a hack to paper + * over a more fundamental race condition. */ + lov_stripe_lock(lli->lli_smd); + inode_init_lvb(inode, &lvb); + rc = obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0); + if (lvb.lvb_size == i_size_read(inode) && rc == 0) { + CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64 + ",%Lu=%#Lx\n", lli->lli_smd->lsm_object_id, + i_size_read(inode), i_size_read(inode)); + lov_stripe_unlock(lli->lli_smd); + GOTO(out_unlock, 0); + } + obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, + i_size_read(inode), 1); lov_stripe_unlock(lli->lli_smd); - GOTO(out_unlock, 0); } - obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, i_size_read(inode), 1); - lov_stripe_unlock(lli->lli_smd); - if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) && (i_size_read(inode) & ~CFS_PAGE_MASK))) { /* If the truncate leaves behind a partial page, update its @@ -169,38 +229,24 @@ void ll_truncate(struct inode *inode) if (llap != NULL) { char *kaddr = kmap_atomic(page, KM_USER0); llap->llap_checksum = - crc32_le(0, kaddr, CFS_PAGE_SIZE); + init_checksum(OSC_DEFAULT_CKSUM); + llap->llap_checksum = + compute_checksum(llap->llap_checksum, + kaddr, CFS_PAGE_SIZE, + OSC_DEFAULT_CKSUM); kunmap_atomic(kaddr, KM_USER0); } page_cache_release(page); } } - CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n", - lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode)); - - oinfo.oi_md = lli->lli_smd; - oinfo.oi_policy.l_extent.start = i_size_read(inode); - oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF; - oinfo.oi_oa = &oa; - oa.o_id = lli->lli_smd->lsm_object_id; - oa.o_gr = lli->lli_smd->lsm_object_gr; - oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP; - - obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE | - OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME | - OBD_MD_FLFID | OBD_MD_FLGENER); - + new_size = i_size_read(inode); ll_inode_size_unlock(inode, 0); - - oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC); - rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL); - ll_truncate_free_capa(oinfo.oi_capa); - if (rc) - CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino); + if (!srvlock) + ll_file_punch(inode, new_size, 0); else - obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | - OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME); + ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LOCKLESS_TRUNC, 1); + EXIT; return; @@ -234,7 +280,7 @@ int ll_prepare_write(struct file *file, struct page *page, unsigned from, oa.o_mode = inode->i_mode; oa.o_id = lsm->lsm_object_id; oa.o_gr = lsm->lsm_object_gr; - oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | + oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE | OBD_MD_FLGROUP; obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER); @@ -289,6 +335,14 @@ int ll_prepare_write(struct file *file, struct page *page, unsigned from, return rc; } +/** + * make page ready for ASYNC write + * \param data - pointer to llap cookie + * \param cmd - is OBD_BRW_* macroses + * + * \retval 0 is page successfully prepared to send + * \retval -EAGAIN is page not need to send + */ static int ll_ap_make_ready(void *data, int cmd) { struct ll_async_page *llap; @@ -298,14 +352,13 @@ static int ll_ap_make_ready(void *data, int cmd) llap = LLAP_FROM_COOKIE(data); page = llap->llap_page; - LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page, - page->mapping->host->i_ino, page->index); - /* we're trying to write, but the page is locked.. come back later */ if (TryLockPage(page)) RETURN(-EAGAIN); - LASSERT(!PageWriteback(page)); + LASSERTF(!(cmd & OBD_BRW_READ) || !PageWriteback(page), + "cmd %x page %p ino %lu index %lu fl %lx\n", cmd, page, + page->mapping->host->i_ino, page->index, page->flags); /* if we left PageDirty we might get another writepage call * in the future. list walkers are bright enough @@ -316,7 +369,10 @@ static int ll_ap_make_ready(void *data, int cmd) * cli lock */ LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page, page->mapping->host->i_ino, page->index); - clear_page_dirty_for_io(page); + if(!clear_page_dirty_for_io(page)) { + unlock_page(page); + RETURN(-EAGAIN); + } /* This actually clears the dirty bit in the radix tree.*/ set_page_writeback(page); @@ -453,50 +509,36 @@ struct ll_async_page *llap_cast_private(struct page *page) return llap; } -/* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction. +/* Try to reap @target pages in the specific @cpu's async page list. * * There is an llap attached onto every page in lustre, linked off @sbi. * We add an llap to the list so we don't lose our place during list walking. * If llaps in the list are being moved they will only move to the end * of the LRU, and we aren't terribly interested in those pages here (we - * start at the beginning of the list where the least-used llaps are. - */ -int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction) + * start at the beginning of the list where the least-used llaps are. */ +static inline int llap_shrink_cache_internal(struct ll_sb_info *sbi, + int cpu, int target) { struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a }; - unsigned long total, want, count = 0; - - total = sbi->ll_async_page_count; - - /* There can be a large number of llaps (600k or more in a large - * memory machine) so the VM 1/6 shrink ratio is likely too much. - * Since we are freeing pages also, we don't necessarily want to - * shrink so much. Limit to 40MB of pages + llaps per call. */ - if (shrink_fraction == 0) - want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32; - else - want = (total + shrink_fraction - 1) / shrink_fraction; - - if (want > 40 << (20 - CFS_PAGE_SHIFT)) - want = 40 << (20 - CFS_PAGE_SHIFT); - - CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n", - want, total, shrink_fraction); - - spin_lock(&sbi->ll_lock); - list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist); - - while (--total >= 0 && count < want) { + struct ll_pglist_data *pd; + struct list_head *head; + int count = 0; + + pd = ll_pglist_cpu_lock(sbi, cpu); + head = &pd->llpd_list; + list_add(&dummy_llap.llap_pglist_item, head); + while (count < target) { struct page *page; int keep; if (unlikely(need_resched())) { - spin_unlock(&sbi->ll_lock); + ll_pglist_cpu_unlock(sbi, cpu); cond_resched(); - spin_lock(&sbi->ll_lock); + ll_pglist_cpu_lock(sbi, cpu); } - llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item); + llap = llite_pglist_next_llap(head, + &dummy_llap.llap_pglist_item); list_del_init(&dummy_llap.llap_pglist_item); if (llap == NULL) break; @@ -532,7 +574,7 @@ int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction) } page_cache_get(page); - spin_unlock(&sbi->ll_lock); + ll_pglist_cpu_unlock(sbi, cpu); if (page->mapping != NULL) { ll_teardown_mmaps(page->mapping, @@ -554,24 +596,158 @@ int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction) unlock_page(page); page_cache_release(page); - spin_lock(&sbi->ll_lock); + ll_pglist_cpu_lock(sbi, cpu); } list_del(&dummy_llap.llap_pglist_item); - spin_unlock(&sbi->ll_lock); + ll_pglist_cpu_unlock(sbi, cpu); + + CDEBUG(D_CACHE, "shrank %d, expected %d however. \n", count, target); + return count; +} + + +/* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction. + * + * At first, this code calculates total pages wanted by @shrink_fraction, then + * it deduces how many pages should be reaped from each cpu in proportion as + * their own # of page count(llpd_count). + */ +int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction) +{ + unsigned long total, want, percpu_want, count = 0; + int cpu, nr_cpus; + + total = lcounter_read(&sbi->ll_async_page_count); + if (total == 0) + return 0; + +#ifdef HAVE_SHRINKER_CACHE + want = shrink_fraction; + if (want == 0) + return total; +#else + /* There can be a large number of llaps (600k or more in a large + * memory machine) so the VM 1/6 shrink ratio is likely too much. + * Since we are freeing pages also, we don't necessarily want to + * shrink so much. Limit to 40MB of pages + llaps per call. */ + if (shrink_fraction <= 0) + want = total - sbi->ll_async_page_max + 32*num_online_cpus(); + else + want = (total + shrink_fraction - 1) / shrink_fraction; +#endif + + if (want > 40 << (20 - CFS_PAGE_SHIFT)) + want = 40 << (20 - CFS_PAGE_SHIFT); + + CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n", + want, total, shrink_fraction); + + nr_cpus = num_possible_cpus(); + cpu = sbi->ll_async_page_clock_hand; + /* we at most do one round */ + do { + int c; + + cpu = (cpu + 1) % nr_cpus; + c = LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_count; + if (!cpu_online(cpu)) + percpu_want = c; + else + percpu_want = want / ((total / (c + 1)) + 1); + if (percpu_want == 0) + continue; + + count += llap_shrink_cache_internal(sbi, cpu, percpu_want); + if (count >= want) + sbi->ll_async_page_clock_hand = cpu; + } while (cpu != sbi->ll_async_page_clock_hand); CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n", count, want, total); +#ifdef HAVE_SHRINKER_CACHE + return lcounter_read(&sbi->ll_async_page_count); +#else return count; +#endif +} + +/* Rebalance the async page queue len for each cpu. We hope that the cpu + * which do much IO job has a relative longer queue len. + * This function should be called with preempt disabled. + */ +static inline int llap_async_cache_rebalance(struct ll_sb_info *sbi) +{ + unsigned long sample = 0, *cpu_sample, bias, slice; + struct ll_pglist_data *pd; + cpumask_t mask; + int cpu, surplus; + int w1 = 7, w2 = 3, base = (w1 + w2); /* weight value */ + atomic_t *pcnt; + + if (!spin_trylock(&sbi->ll_async_page_reblnc_lock)) { + /* someone else is doing the job */ + return 1; + } + + pcnt = &LL_PGLIST_DATA(sbi)->llpd_sample_count; + if (!atomic_read(pcnt)) { + /* rare case, somebody else has gotten this job done */ + spin_unlock(&sbi->ll_async_page_reblnc_lock); + return 1; + } + + sbi->ll_async_page_reblnc_count++; + cpu_sample = sbi->ll_async_page_sample; + memset(cpu_sample, 0, num_possible_cpus() * sizeof(unsigned long)); + for_each_online_cpu(cpu) { + pcnt = &LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_sample_count; + cpu_sample[cpu] = atomic_read(pcnt); + atomic_set(pcnt, 0); + sample += cpu_sample[cpu]; + } + + cpus_clear(mask); + surplus = sbi->ll_async_page_max; + slice = surplus / sample + 1; + sample /= num_online_cpus(); + bias = sample >> 4; + for_each_online_cpu(cpu) { + pd = LL_PGLIST_DATA_CPU(sbi, cpu); + if (labs((long int)sample - cpu_sample[cpu]) > bias) { + unsigned long budget = pd->llpd_budget; + /* weighted original queue length and expected queue + * length to avoid thrashing. */ + pd->llpd_budget = (budget * w1) / base + + (slice * cpu_sample[cpu]) * w2 / base; + cpu_set(cpu, mask); + } + surplus -= pd->llpd_budget; + } + surplus /= cpus_weight(mask) ?: 1; + for_each_cpu_mask(cpu, mask) + LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_budget += surplus; + spin_unlock(&sbi->ll_async_page_reblnc_lock); + + /* TODO: do we really need to call llap_shrink_cache_internal + * for every cpus with its page_count greater than budget? + * for_each_cpu_mask(cpu, mask) + * ll_shrink_cache_internal(...) + */ + + return 0; } -struct ll_async_page *llap_from_page(struct page *page, unsigned origin) +static struct ll_async_page *llap_from_page_with_lockh(struct page *page, + unsigned origin, + struct lustre_handle *lockh) { struct ll_async_page *llap; struct obd_export *exp; struct inode *inode = page->mapping->host; struct ll_sb_info *sbi; - int rc; + struct ll_pglist_data *pd; + int rc, cpu, target; ENTRY; if (!inode) { @@ -594,11 +770,30 @@ struct ll_async_page *llap_from_page(struct page *page, unsigned origin) /* move to end of LRU list, except when page is just about to * die */ if (origin != LLAP_ORIGIN_REMOVEPAGE) { - spin_lock(&sbi->ll_lock); - sbi->ll_pglist_gen++; - list_del_init(&llap->llap_pglist_item); - list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist); - spin_unlock(&sbi->ll_lock); + int old_cpu = llap->llap_pglist_cpu; + struct ll_pglist_data *old_pd; + + pd = ll_pglist_double_lock(sbi, old_cpu, &old_pd); + pd->llpd_hit++; + while (old_cpu != llap->llap_pglist_cpu) { + /* rarely case, someone else is touching this + * page too. */ + ll_pglist_double_unlock(sbi, old_cpu); + old_cpu = llap->llap_pglist_cpu; + pd=ll_pglist_double_lock(sbi, old_cpu, &old_pd); + } + + list_move(&llap->llap_pglist_item, + &pd->llpd_list); + old_pd->llpd_gen++; + if (pd->llpd_cpu != old_cpu) { + pd->llpd_count++; + old_pd->llpd_count--; + old_pd->llpd_gen++; + llap->llap_pglist_cpu = pd->llpd_cpu; + pd->llpd_cross++; + } + ll_pglist_double_unlock(sbi, old_cpu); } GOTO(out, llap); } @@ -608,8 +803,28 @@ struct ll_async_page *llap_from_page(struct page *page, unsigned origin) RETURN(ERR_PTR(-EINVAL)); /* limit the number of lustre-cached pages */ - if (sbi->ll_async_page_count >= sbi->ll_async_page_max) - llap_shrink_cache(sbi, 0); + cpu = get_cpu(); + pd = LL_PGLIST_DATA(sbi); + target = pd->llpd_count - pd->llpd_budget; + if (target > 0) { + rc = 0; + atomic_inc(&pd->llpd_sample_count); + if (atomic_read(&pd->llpd_sample_count) > + sbi->ll_async_page_sample_max) { + pd->llpd_reblnc_count++; + rc = llap_async_cache_rebalance(sbi); + if (rc == 0) + target = pd->llpd_count - pd->llpd_budget; + } + /* if rc equals 1, it means other cpu is doing the rebalance + * job, and our budget # would be modified when we read it. + * Furthermore, it is much likely being increased because + * we have already reached the rebalance threshold. In this + * case, we skip to shrink cache here. */ + if ((rc == 0) && target > 0) + llap_shrink_cache_internal(sbi, cpu, target + 32); + } + put_cpu(); OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD, ll_async_page_slab_size); @@ -618,9 +833,14 @@ struct ll_async_page *llap_from_page(struct page *page, unsigned origin) llap->llap_magic = LLAP_MAGIC; llap->llap_cookie = (void *)llap + size_round(sizeof(*llap)); + /* XXX: for bug 11270 - check for lockless origin here! */ + if (origin == LLAP_ORIGIN_LOCKLESS_IO) + llap->llap_nocache = 1; + rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page, (obd_off)page->index << CFS_PAGE_SHIFT, - &ll_async_page_ops, llap, &llap->llap_cookie); + &ll_async_page_ops, llap, &llap->llap_cookie, + llap->llap_nocache, lockh); if (rc) { OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size); @@ -632,21 +852,28 @@ struct ll_async_page *llap_from_page(struct page *page, unsigned origin) /* also zeroing the PRIVBITS low order bitflags */ __set_page_ll_data(page, llap); llap->llap_page = page; - spin_lock(&sbi->ll_lock); - sbi->ll_pglist_gen++; - sbi->ll_async_page_count++; - list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist); + + lcounter_inc(&sbi->ll_async_page_count); + pd = ll_pglist_lock(sbi); + list_add_tail(&llap->llap_pglist_item, &pd->llpd_list); INIT_LIST_HEAD(&llap->llap_pending_write); - spin_unlock(&sbi->ll_lock); + pd->llpd_count++; + pd->llpd_gen++; + pd->llpd_miss++; + llap->llap_pglist_cpu = pd->llpd_cpu; + ll_pglist_unlock(sbi); out: if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) { - __u32 csum = 0; + __u32 csum; char *kaddr = kmap_atomic(page, KM_USER0); - csum = crc32_le(csum, kaddr, CFS_PAGE_SIZE); + csum = init_checksum(OSC_DEFAULT_CKSUM); + csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE, + OSC_DEFAULT_CKSUM); kunmap_atomic(kaddr, KM_USER0); if (origin == LLAP_ORIGIN_READAHEAD || - origin == LLAP_ORIGIN_READPAGE) { + origin == LLAP_ORIGIN_READPAGE || + origin == LLAP_ORIGIN_LOCKLESS_IO) { llap->llap_checksum = 0; } else if (origin == LLAP_ORIGIN_COMMIT_WRITE || llap->llap_checksum == 0) { @@ -667,6 +894,12 @@ struct ll_async_page *llap_from_page(struct page *page, unsigned origin) RETURN(llap); } +struct ll_async_page *llap_from_page(struct page *page, + unsigned origin) +{ + return llap_from_page_with_lockh(page, origin, NULL); +} + static int queue_or_sync_write(struct obd_export *exp, struct inode *inode, struct ll_async_page *llap, unsigned to, obd_flag async_flags) @@ -690,7 +923,7 @@ static int queue_or_sync_write(struct obd_export *exp, struct inode *inode, llap->llap_write_queued = 0; /* Do not pass llap here as it is sync write. */ llap_write_pending(inode, NULL); - + rc = oig_init(&oig); if (rc) GOTO(out, rc); @@ -713,10 +946,12 @@ static int queue_or_sync_write(struct obd_export *exp, struct inode *inode, /* compare the checksum once before the page leaves llite */ if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) && llap->llap_checksum != 0)) { - __u32 csum = 0; + __u32 csum; struct page *page = llap->llap_page; char *kaddr = kmap_atomic(page, KM_USER0); - csum = crc32_le(csum, kaddr, CFS_PAGE_SIZE); + csum = init_checksum(OSC_DEFAULT_CKSUM); + csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE, + OSC_DEFAULT_CKSUM); kunmap_atomic(kaddr, KM_USER0); if (llap->llap_checksum == csum) { CDEBUG(D_PAGE, "page %p cksum %x confirmed\n", @@ -742,9 +977,8 @@ static int queue_or_sync_write(struct obd_export *exp, struct inode *inode, if (!rc && async_flags & ASYNC_READY) { unlock_page(llap->llap_page); - if (PageWriteback(llap->llap_page)) { + if (PageWriteback(llap->llap_page)) end_page_writeback(llap->llap_page); - } } if (rc == 0 && llap_write_complete(inode, llap)) @@ -766,12 +1000,14 @@ out: int ll_commit_write(struct file *file, struct page *page, unsigned from, unsigned to) { + struct ll_file_data *fd = LUSTRE_FPRIVATE(file); struct inode *inode = page->mapping->host; struct ll_inode_info *lli = ll_i2info(inode); struct lov_stripe_md *lsm = lli->lli_smd; struct obd_export *exp; struct ll_async_page *llap; loff_t size; + struct lustre_handle *lockh = NULL; int rc = 0; ENTRY; @@ -782,7 +1018,10 @@ int ll_commit_write(struct file *file, struct page *page, unsigned from, CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n", inode, page, from, to, page->index); - llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE); + if (fd->fd_flags & LL_FILE_GROUP_LOCKED) + lockh = &fd->fd_cwlockh; + + llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_COMMIT_WRITE, lockh); if (IS_ERR(llap)) RETURN(PTR_ERR(llap)); @@ -848,28 +1087,40 @@ out: RETURN(rc); } +static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which); + +/* WARNING: This algorithm is used to reduce the contention on + * sbi->ll_lock. It should work well if the ra_max_pages is much + * greater than the single file's read-ahead window. + * + * TODO: There may exist a `global sync problem' in this implementation. + * Considering the global ra window is 100M, and each file's ra window is 10M, + * there are over 10 files trying to get its ra budget and reach + * ll_ra_count_get at the exactly same time. All of them will get a zero ra + * window, although the global window is 100M. -jay + */ static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len) { struct ll_ra_info *ra = &sbi->ll_ra_info; unsigned long ret; ENTRY; - spin_lock(&sbi->ll_lock); - ret = min(ra->ra_max_pages - ra->ra_cur_pages, len); - ra->ra_cur_pages += ret; - spin_unlock(&sbi->ll_lock); + ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), len); + if ((int)ret < 0) + GOTO(out, ret = 0); + if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) { + atomic_sub(ret, &ra->ra_cur_pages); + ret = 0; + } +out: RETURN(ret); } static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len) { struct ll_ra_info *ra = &sbi->ll_ra_info; - spin_lock(&sbi->ll_lock); - LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n", - ra->ra_cur_pages, len); - ra->ra_cur_pages -= len; - spin_unlock(&sbi->ll_lock); + atomic_sub(len, &ra->ra_cur_pages); } /* called for each page in a completed rpc.*/ @@ -884,7 +1135,7 @@ int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc) page = llap->llap_page; LASSERT(PageLocked(page)); LASSERT(CheckWriteback(page,cmd)); - + LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc); if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate) @@ -909,48 +1160,39 @@ int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc) set_bit(AS_EIO, &page->mapping->flags); } + /* be carefull about clear WB. + * if WB will cleared after page lock is released - paralel IO can be + * started before ap_make_ready is finished - so we will be have page + * with PG_Writeback set from ->writepage() and completed READ which + * clear this flag */ + if ((cmd & OBD_BRW_WRITE) && PageWriteback(page)) + end_page_writeback(page); + unlock_page(page); if (cmd & OBD_BRW_WRITE) { /* Only rc == 0, write succeed, then this page could be deleted - * from the pending_writing list + * from the pending_writing list */ if (rc == 0 && llap_write_complete(page->mapping->host, llap)) ll_queue_done_writing(page->mapping->host, 0); } - if (PageWriteback(page)) { - end_page_writeback(page); - } page_cache_release(page); RETURN(ret); } -/* the kernel calls us here when a page is unhashed from the page cache. - * the page will be locked and the kernel is holding a spinlock, so - * we need to be careful. we're just tearing down our book-keeping - * here. */ -void ll_removepage(struct page *page) +static void __ll_put_llap(struct page *page) { struct inode *inode = page->mapping->host; struct obd_export *exp; struct ll_async_page *llap; struct ll_sb_info *sbi = ll_i2sbi(inode); - int rc; + struct ll_pglist_data *pd; + int rc, cpu; ENTRY; - LASSERT(!in_interrupt()); - - /* sync pages or failed read pages can leave pages in the page - * cache that don't have our data associated with them anymore */ - if (page_private(page) == 0) { - EXIT; - return; - } - - LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n"); - exp = ll_i2dtexp(inode); if (exp == NULL) { CERROR("page %p ind %lu gave null export\n", page, page->index); @@ -978,38 +1220,41 @@ void ll_removepage(struct page *page) * is providing exclusivity to memory pressure/truncate/writeback..*/ __clear_page_ll_data(page); - spin_lock(&sbi->ll_lock); + lcounter_dec(&sbi->ll_async_page_count); + cpu = llap->llap_pglist_cpu; + pd = ll_pglist_cpu_lock(sbi, cpu); + pd->llpd_gen++; + pd->llpd_count--; if (!list_empty(&llap->llap_pglist_item)) list_del_init(&llap->llap_pglist_item); - sbi->ll_pglist_gen++; - sbi->ll_async_page_count--; - spin_unlock(&sbi->ll_lock); + ll_pglist_cpu_unlock(sbi, cpu); OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size); EXIT; } -static int ll_page_matches(struct page *page, int fd_flags) +/* the kernel calls us here when a page is unhashed from the page cache. + * the page will be locked and the kernel is holding a spinlock, so + * we need to be careful. we're just tearing down our book-keeping + * here. */ +void ll_removepage(struct page *page) { - struct lustre_handle match_lockh = {0}; - struct inode *inode = page->mapping->host; - ldlm_policy_data_t page_extent; - int flags, matches; + struct ll_async_page *llap = llap_cast_private(page); ENTRY; - if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED)) - RETURN(1); - - page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT; - page_extent.l_extent.end = - page_extent.l_extent.start + CFS_PAGE_SIZE - 1; - flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED; - if (!(fd_flags & LL_FILE_READAHEAD)) - flags |= LDLM_FL_CBPENDING; - matches = obd_match(ll_i2sbi(inode)->ll_dt_exp, - ll_i2info(inode)->lli_smd, LDLM_EXTENT, - &page_extent, LCK_PR | LCK_PW, &flags, inode, - &match_lockh); - RETURN(matches); + LASSERT(!in_interrupt()); + + /* sync pages or failed read pages can leave pages in the page + * cache that don't have our data associated with them anymore */ + if (page_private(page) == 0) { + EXIT; + return; + } + + LASSERT(!llap->llap_lockless_io_page); + LASSERT(!llap->llap_nocache); + LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n"); + __ll_put_llap(page); + EXIT; } static int ll_issue_page_read(struct obd_export *exp, @@ -1033,20 +1278,16 @@ static int ll_issue_page_read(struct obd_export *exp, RETURN(rc); } -static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which) +static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which) { LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which); - ra->ra_stats[which]++; + lprocfs_counter_incr(sbi->ll_ra_stats, which); } static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which) { struct ll_sb_info *sbi = ll_i2sbi(mapping->host); - struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info; - - spin_lock(&sbi->ll_lock); - ll_ra_stats_inc_unlocked(ra, which); - spin_unlock(&sbi->ll_lock); + ll_ra_stats_inc_sbi(sbi, which); } void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping) @@ -1059,11 +1300,14 @@ void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping) #define RAS_CDEBUG(ras) \ CDEBUG(D_READA, \ - "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu\n", \ + "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \ + "csr %lu sf %lu sp %lu sl %lu \n", \ ras->ras_last_readpage, ras->ras_consecutive_requests, \ ras->ras_consecutive_pages, ras->ras_window_start, \ ras->ras_window_len, ras->ras_next_readahead, \ - ras->ras_requests, ras->ras_request_index); + ras->ras_requests, ras->ras_request_index, \ + ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \ + ras->ras_stride_pages, ras->ras_stride_length) static int index_in_window(unsigned long index, unsigned long point, unsigned long before, unsigned long after) @@ -1137,20 +1381,211 @@ struct ll_ra_read *ll_ra_read_get(struct file *f) return bead; } +static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig, + int index, struct address_space *mapping) +{ + struct ll_async_page *llap; + struct page *page; + unsigned int gfp_mask = 0; + int rc = 0; + + gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT; +#ifdef __GFP_NOWARN + gfp_mask |= __GFP_NOWARN; +#endif + page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask); + if (page == NULL) { + ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE); + CDEBUG(D_READA, "g_c_p_n failed\n"); + return 0; + } + + /* Check if page was truncated or reclaimed */ + if (page->mapping != mapping) { + ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE); + CDEBUG(D_READA, "g_c_p_n returned invalid page\n"); + GOTO(unlock_page, rc = 0); + } + + /* we do this first so that we can see the page in the /proc + * accounting */ + llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD); + if (IS_ERR(llap) || llap->llap_defer_uptodate) { + if (PTR_ERR(llap) == -ENOLCK) { + ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH); + CDEBUG(D_READA | D_PAGE, + "Adding page to cache failed index " + "%d\n", index); + CDEBUG(D_READA, "nolock page\n"); + GOTO(unlock_page, rc = -ENOLCK); + } + CDEBUG(D_READA, "read-ahead page\n"); + GOTO(unlock_page, rc = 0); + } + + /* skip completed pages */ + if (Page_Uptodate(page)) + GOTO(unlock_page, rc = 0); + + /* bail out when we hit the end of the lock. */ + rc = ll_issue_page_read(exp, llap, oig, 1); + if (rc == 0) { + LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n"); + rc = 1; + } else { +unlock_page: + unlock_page(page); + LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n"); + } + page_cache_release(page); + return rc; +} + +/* ra_io_arg will be filled in the beginning of ll_readahead with + * ras_lock, then the following ll_read_ahead_pages will read RA + * pages according to this arg, all the items in this structure are + * counted by page index. + */ +struct ra_io_arg { + unsigned long ria_start; /* start offset of read-ahead*/ + unsigned long ria_end; /* end offset of read-ahead*/ + /* If stride read pattern is detected, ria_stoff means where + * stride read is started. Note: for normal read-ahead, the + * value here is meaningless, and also it will not be accessed*/ + pgoff_t ria_stoff; + /* ria_length and ria_pages are the length and pages length in the + * stride I/O mode. And they will also be used to check whether + * it is stride I/O read-ahead in the read-ahead pages*/ + unsigned long ria_length; + unsigned long ria_pages; +}; + +#define RIA_DEBUG(ria) \ + CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \ + ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\ + ria->ria_pages) + +#define RAS_INCREASE_STEP (1024 * 1024 >> CFS_PAGE_SHIFT) + +static inline int stride_io_mode(struct ll_readahead_state *ras) +{ + return ras->ras_consecutive_stride_requests > 1; +} + +/* The function calculates how much pages will be read in + * [off, off + length], which will be read by stride I/O mode, + * stride_offset = st_off, stride_lengh = st_len, + * stride_pages = st_pgs + */ +static unsigned long +stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs, + unsigned long off, unsigned length) +{ + unsigned long cont_len = st_off > off ? st_off - off : 0; + __u64 stride_len = length + off > st_off ? + length + off + 1 - st_off : 0; + unsigned long left, pg_count; + + if (st_len == 0 || length == 0) + return length; + + left = do_div(stride_len, st_len); + left = min(left, st_pgs); + + pg_count = left + stride_len * st_pgs + cont_len; + + LASSERT(pg_count >= left); + + CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u" + "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count); + + return pg_count; +} + +static int ria_page_count(struct ra_io_arg *ria) +{ + __u64 length = ria->ria_end >= ria->ria_start ? + ria->ria_end - ria->ria_start + 1 : 0; + + return stride_pg_count(ria->ria_stoff, ria->ria_length, + ria->ria_pages, ria->ria_start, + length); +} + +/*Check whether the index is in the defined ra-window */ +static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria) +{ + /* If ria_length == ria_pages, it means non-stride I/O mode, + * idx should always inside read-ahead window in this case + * For stride I/O mode, just check whether the idx is inside + * the ria_pages. */ + return ria->ria_length == 0 || ria->ria_length == ria->ria_pages || + (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages; +} + +static int ll_read_ahead_pages(struct obd_export *exp, + struct obd_io_group *oig, + struct ra_io_arg *ria, + unsigned long *reserved_pages, + struct address_space *mapping, + unsigned long *ra_end) +{ + int rc, count = 0, stride_ria; + unsigned long page_idx; + + LASSERT(ria != NULL); + RIA_DEBUG(ria); + + stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0; + for (page_idx = ria->ria_start; page_idx <= ria->ria_end && + *reserved_pages > 0; page_idx++) { + if (ras_inside_ra_window(page_idx, ria)) { + /* If the page is inside the read-ahead window*/ + rc = ll_read_ahead_page(exp, oig, page_idx, mapping); + if (rc == 1) { + (*reserved_pages)--; + count ++; + } else if (rc == -ENOLCK) + break; + } else if (stride_ria) { + /* If it is not in the read-ahead window, and it is + * read-ahead mode, then check whether it should skip + * the stride gap */ + pgoff_t offset; + /* FIXME: This assertion only is valid when it is for + * forward read-ahead, it will be fixed when backward + * read-ahead is implemented */ + LASSERTF(page_idx > ria->ria_stoff, "since %lu in the" + " gap of ra window,it should bigger than stride" + " offset %lu \n", page_idx, ria->ria_stoff); + + offset = page_idx - ria->ria_stoff; + offset = offset % (ria->ria_length); + if (offset > ria->ria_pages) { + page_idx += ria->ria_length - offset; + CDEBUG(D_READA, "i %lu skip %lu \n", page_idx, + ria->ria_length - offset); + continue; + } + } + } + *ra_end = page_idx; + return count; +} + static int ll_readahead(struct ll_readahead_state *ras, struct obd_export *exp, struct address_space *mapping, struct obd_io_group *oig, int flags) { - unsigned long i, start = 0, end = 0, reserved; - struct ll_async_page *llap; - struct page *page; - int rc, ret = 0, match_failed = 0; - __u64 kms; - unsigned int gfp_mask; + unsigned long start = 0, end = 0, reserved; + unsigned long ra_end, len; struct inode *inode; struct lov_stripe_md *lsm; struct ll_ra_read *bead; struct ost_lvb lvb; + struct ra_io_arg ria = { 0 }; + int ret = 0; + __u64 kms; ENTRY; inode = mapping->host; @@ -1174,7 +1609,7 @@ static int ll_readahead(struct ll_readahead_state *ras, ras->ras_window_len = bead->lrr_start + bead->lrr_count - ras->ras_window_start; } - /* Reserve a part of the read-ahead window that we'll be issuing */ + /* Reserve a part of the read-ahead window that we'll be issuing */ if (ras->ras_window_len) { start = ras->ras_next_readahead; end = ras->ras_window_start + ras->ras_window_len - 1; @@ -1185,77 +1620,38 @@ static int ll_readahead(struct ll_readahead_state *ras, ras->ras_next_readahead = max(end, end + 1); RAS_CDEBUG(ras); } + ria.ria_start = start; + ria.ria_end = end; + /* If stride I/O mode is detected, get stride window*/ + if (stride_io_mode(ras)) { + ria.ria_stoff = ras->ras_stride_offset; + ria.ria_length = ras->ras_stride_length; + ria.ria_pages = ras->ras_stride_pages; + } spin_unlock(&ras->ras_lock); if (end == 0) { ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW); RETURN(0); } + len = ria_page_count(&ria); + if (len == 0) + RETURN(0); - reserved = ll_ra_count_get(ll_i2sbi(inode), end - start + 1); - if (reserved < end - start + 1) - ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT); - - gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT; -#ifdef __GFP_NOWARN - gfp_mask |= __GFP_NOWARN; -#endif - - for (i = start; reserved > 0 && !match_failed && i <= end; i++) { - /* skip locked pages from previous readpage calls */ - page = grab_cache_page_nowait_gfp(mapping, i, gfp_mask); - if (page == NULL) { - ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE); - CDEBUG(D_READA, "g_c_p_n failed\n"); - continue; - } - - /* Check if page was truncated or reclaimed */ - if (page->mapping != mapping) { - ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE); - CDEBUG(D_READA, "g_c_p_n returned invalid page\n"); - goto next_page; - } - - /* we do this first so that we can see the page in the /proc - * accounting */ - llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD); - if (IS_ERR(llap) || llap->llap_defer_uptodate) - goto next_page; - - /* skip completed pages */ - if (Page_Uptodate(page)) - goto next_page; + reserved = ll_ra_count_get(ll_i2sbi(inode), len); - /* bail when we hit the end of the lock. */ - if ((rc = ll_page_matches(page, flags|LL_FILE_READAHEAD)) <= 0){ - LL_CDEBUG_PAGE(D_READA | D_PAGE, page, - "lock match failed: rc %d\n", rc); - ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH); - match_failed = 1; - goto next_page; - } - - rc = ll_issue_page_read(exp, llap, oig, 1); - if (rc == 0) { - reserved--; - ret++; - LL_CDEBUG_PAGE(D_READA| D_PAGE, page, - "started read-ahead\n"); - } else { - next_page: - LL_CDEBUG_PAGE(D_READA | D_PAGE, page, - "skipping read-ahead\n"); + if (reserved < len) + ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT); - unlock_page(page); - } - page_cache_release(page); - } + CDEBUG(D_READA, "reserved page %lu \n", reserved); + + ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end); LASSERTF(reserved >= 0, "reserved %lu\n", reserved); if (reserved != 0) ll_ra_count_put(ll_i2sbi(inode), reserved); - if (i == end + 1 && end == (kms >> CFS_PAGE_SHIFT)) + + if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT)) ll_ra_stats_inc(mapping, RA_STAT_EOF); /* if we didn't get to the end of the region we reserved from @@ -1263,13 +1659,16 @@ static int ll_readahead(struct ll_readahead_state *ras, * next read-ahead tries from where we left off. we only do so * if the region we failed to issue read-ahead on is still ahead * of the app and behind the next index to start read-ahead from */ - if (i != end + 1) { + CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n", + ra_end, end, ria.ria_end); + + if (ra_end != (end + 1)) { spin_lock(&ras->ras_lock); - if (i < ras->ras_next_readahead && - index_in_window(i, ras->ras_window_start, 0, + if (ra_end < ras->ras_next_readahead && + index_in_window(ra_end, ras->ras_window_start, 0, ras->ras_window_len)) { - ras->ras_next_readahead = i; - RAS_CDEBUG(ras); + ras->ras_next_readahead = ra_end; + RAS_CDEBUG(ras); } spin_unlock(&ras->ras_lock); } @@ -1279,7 +1678,7 @@ static int ll_readahead(struct ll_readahead_state *ras, static void ras_set_start(struct ll_readahead_state *ras, unsigned long index) { - ras->ras_window_start = index & (~((1024 * 1024 >> CFS_PAGE_SHIFT) - 1)); + ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1)); } /* called with the ras_lock held or from places where it doesn't matter */ @@ -1295,6 +1694,13 @@ static void ras_reset(struct ll_readahead_state *ras, unsigned long index) RAS_CDEBUG(ras); } +/* called with the ras_lock held or from places where it doesn't matter */ +static void ras_stride_reset(struct ll_readahead_state *ras) +{ + ras->ras_consecutive_stride_requests = 0; + RAS_CDEBUG(ras); +} + void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) { spin_lock_init(&ras->ras_lock); @@ -1303,18 +1709,123 @@ void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) INIT_LIST_HEAD(&ras->ras_read_beads); } +/* Check whether the read request is in the stride window. + * If it is in the stride window, return 1, otherwise return 0. + * and also update stride_gap and stride_pages. + */ +static int index_in_stride_window(unsigned long index, + struct ll_readahead_state *ras, + struct inode *inode) +{ + int stride_gap = index - ras->ras_last_readpage - 1; + + LASSERT(stride_gap != 0); + + if (ras->ras_consecutive_pages == 0) + return 0; + + /*Otherwise check the stride by itself */ + if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap && + ras->ras_consecutive_pages == ras->ras_stride_pages) + return 1; + + if (stride_gap >= 0) { + /* + * only set stride_pages, stride_length if + * it is forward reading ( stride_gap > 0) + */ + ras->ras_stride_pages = ras->ras_consecutive_pages; + ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages; + } else { + /* + * If stride_gap < 0,(back_forward reading), + * reset the stride_pages/length. + * FIXME:back_ward stride I/O read. + * + */ + ras->ras_stride_pages = 0; + ras->ras_stride_length = 0; + } + RAS_CDEBUG(ras); + + return 0; +} + +static unsigned long +stride_page_count(struct ll_readahead_state *ras, unsigned long len) +{ + return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length, + ras->ras_stride_pages, ras->ras_stride_offset, + len); +} + +/* Stride Read-ahead window will be increased inc_len according to + * stride I/O pattern */ +static void ras_stride_increase_window(struct ll_readahead_state *ras, + struct ll_ra_info *ra, + unsigned long inc_len) +{ + unsigned long left, step, window_len; + unsigned long stride_len; + + LASSERT(ras->ras_stride_length > 0); + + stride_len = ras->ras_window_start + ras->ras_window_len - + ras->ras_stride_offset; + + LASSERTF(stride_len >= 0, "window_start %lu, window_len %lu" + " stride_offset %lu\n", ras->ras_window_start, + ras->ras_window_len, ras->ras_stride_offset); + + left = stride_len % ras->ras_stride_length; + + window_len = ras->ras_window_len - left; + + if (left < ras->ras_stride_pages) + left += inc_len; + else + left = ras->ras_stride_pages + inc_len; + + LASSERT(ras->ras_stride_pages != 0); + + step = left / ras->ras_stride_pages; + left %= ras->ras_stride_pages; + + window_len += step * ras->ras_stride_length + left; + + if (stride_page_count(ras, window_len) <= ra->ra_max_pages) + ras->ras_window_len = window_len; + + RAS_CDEBUG(ras); +} + +/* Set stride I/O read-ahead window start offset */ +static void ras_set_stride_offset(struct ll_readahead_state *ras) +{ + unsigned long window_len = ras->ras_next_readahead - + ras->ras_window_start; + unsigned long left; + + LASSERT(ras->ras_stride_length != 0); + + left = window_len % ras->ras_stride_length; + + ras->ras_stride_offset = ras->ras_next_readahead - left; + + RAS_CDEBUG(ras); +} + static void ras_update(struct ll_sb_info *sbi, struct inode *inode, struct ll_readahead_state *ras, unsigned long index, unsigned hit) { struct ll_ra_info *ra = &sbi->ll_ra_info; - int zero = 0; + int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0; ENTRY; - spin_lock(&sbi->ll_lock); spin_lock(&ras->ras_lock); - ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS); + ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS); /* reset the read-ahead window in two cases. First when the app seeks * or reads to some other part of the file. Secondly if we get a @@ -1323,13 +1834,22 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode, * reclaiming it before we get to it. */ if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) { zero = 1; - ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE); + ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE); + /* check whether it is in stride I/O mode*/ + if (!index_in_stride_window(index, ras, inode)) + stride_zero = 1; } else if (!hit && ras->ras_window_len && index < ras->ras_next_readahead && index_in_window(index, ras->ras_window_start, 0, ras->ras_window_len)) { zero = 1; - ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW); + ra_miss = 1; + /* If it hits read-ahead miss and the stride I/O is still + * not detected, reset stride stuff to re-detect the whole + * stride I/O mode to avoid complication */ + if (!stride_io_mode(ras)) + stride_zero = 1; + ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW); } /* On the second access to a file smaller than the tunable @@ -1359,8 +1879,36 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode, } if (zero) { - ras_reset(ras, index); - GOTO(out_unlock, 0); + /* If it is discontinuous read, check + * whether it is stride I/O mode*/ + if (stride_zero) { + ras_reset(ras, index); + ras->ras_consecutive_pages++; + ras_stride_reset(ras); + RAS_CDEBUG(ras); + GOTO(out_unlock, 0); + } else { + /* The read is still in stride window or + * it hits read-ahead miss */ + + /* If ra-window miss is hitted, which probably means VM + * pressure, and some read-ahead pages were reclaimed.So + * the length of ra-window will not increased, but also + * not reset to avoid redetecting the stride I/O mode.*/ + ras->ras_consecutive_requests = 0; + if (!ra_miss) { + ras->ras_consecutive_pages = 0; + if (++ras->ras_consecutive_stride_requests > 1) + stride_detect = 1; + } + RAS_CDEBUG(ras); + } + } else if (ras->ras_consecutive_stride_requests > 1) { + /* If this is contiguous read but in stride I/O mode + * currently, check whether stride step still is valid, + * if invalid, it will reset the stride ra window*/ + if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages) + ras_stride_reset(ras); } ras->ras_last_readpage = index; @@ -1368,29 +1916,36 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode, ras_set_start(ras, index); ras->ras_next_readahead = max(ras->ras_window_start, ras->ras_next_readahead); + RAS_CDEBUG(ras); /* Trigger RA in the mmap case where ras_consecutive_requests * is not incremented and thus can't be used to trigger RA */ - if (!ras->ras_window_len && ras->ras_consecutive_pages == 3) { - ras->ras_window_len = 1024 * 1024 >> CFS_PAGE_SHIFT; + if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) { + ras->ras_window_len = RAS_INCREASE_STEP; GOTO(out_unlock, 0); } + /* Initially reset the stride window offset to next_readahead*/ + if (ras->ras_consecutive_stride_requests == 2 && stride_detect) + ras_set_stride_offset(ras); + /* The initial ras_window_len is set to the request size. To avoid * uselessly reading and discarding pages for random IO the window is * only increased once per consecutive request received. */ - if (ras->ras_consecutive_requests > 1 && !ras->ras_request_index) { - ras->ras_window_len = min(ras->ras_window_len + - (1024 * 1024 >> CFS_PAGE_SHIFT), - ra->ra_max_pages); + if ((ras->ras_consecutive_requests > 1 && + !ras->ras_request_index) || stride_detect) { + if (stride_io_mode(ras)) + ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP); + else + ras->ras_window_len = min(ras->ras_window_len + + RAS_INCREASE_STEP, + ra->ra_max_pages); } - EXIT; out_unlock: RAS_CDEBUG(ras); ras->ras_request_index++; spin_unlock(&ras->ras_lock); - spin_unlock(&sbi->ll_lock); return; } @@ -1403,7 +1958,6 @@ int ll_writepage(struct page *page) int rc = 0; ENTRY; - LASSERT(!PageDirty(page)); LASSERT(PageLocked(page)); exp = ll_i2dtexp(inode); @@ -1414,6 +1968,7 @@ int ll_writepage(struct page *page) if (IS_ERR(llap)) GOTO(out, rc = PTR_ERR(llap)); + LASSERT(!llap->llap_nocache); LASSERT(!PageWriteback(page)); set_page_writeback(page); @@ -1427,19 +1982,22 @@ int ll_writepage(struct page *page) rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE, ASYNC_READY | ASYNC_URGENT); } - if (rc) - page_cache_release(page); -out: if (rc) { - if (!lli->lli_async_rc) - lli->lli_async_rc = rc; /* re-dirty page on error so it retries write */ - if (PageWriteback(page)) { + if (PageWriteback(page)) end_page_writeback(page); - } + /* resend page only for not started IO*/ if (!PageError(page)) ll_redirty_page(page); + + page_cache_release(page); + } +out: + if (rc) { + if (!lli->lli_async_rc) + lli->lli_async_rc = rc; + /* resend page only for not started IO*/ unlock_page(page); } RETURN(rc); @@ -1461,6 +2019,7 @@ int ll_readpage(struct file *filp, struct page *page) struct obd_export *exp; struct ll_async_page *llap; struct obd_io_group *oig = NULL; + struct lustre_handle *lockh = NULL; int rc; ENTRY; @@ -1492,9 +2051,19 @@ int ll_readpage(struct file *filp, struct page *page) if (exp == NULL) GOTO(out, rc = -EINVAL); - llap = llap_from_page(page, LLAP_ORIGIN_READPAGE); - if (IS_ERR(llap)) + if (fd->fd_flags & LL_FILE_GROUP_LOCKED) + lockh = &fd->fd_cwlockh; + + llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_READPAGE, lockh); + if (IS_ERR(llap)) { + if (PTR_ERR(llap) == -ENOLCK) { + CWARN("ino %lu page %lu (%llu) not covered by " + "a lock (mmap?). check debug logs.\n", + inode->i_ino, page->index, + (long long)page->index << PAGE_CACHE_SHIFT); + } GOTO(out, rc = PTR_ERR(llap)); + } if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages) ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index, @@ -1515,22 +2084,6 @@ int ll_readpage(struct file *filp, struct page *page) GOTO(out_oig, rc = 0); } - if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) { - rc = ll_page_matches(page, fd->fd_flags); - if (rc < 0) { - LL_CDEBUG_PAGE(D_ERROR, page, - "lock match failed: rc %d\n", rc); - GOTO(out, rc); - } - - if (rc == 0) { - CWARN("ino %lu page %lu (%llu) not covered by " - "a lock (mmap?). check debug logs.\n", - inode->i_ino, page->index, - (long long)page->index << CFS_PAGE_SHIFT); - } - } - rc = ll_issue_page_read(exp, llap, oig, 0); if (rc) GOTO(out, rc); @@ -1552,3 +2105,274 @@ out_oig: oig_release(oig); RETURN(rc); } + +static void ll_file_put_pages(struct page **pages, int numpages) +{ + int i; + struct page **pp; + ENTRY; + + for (i = 0, pp = pages; i < numpages; i++, pp++) { + if (*pp) { + LL_CDEBUG_PAGE(D_PAGE, (*pp), "free\n"); + __ll_put_llap(*pp); + if (page_private(*pp)) + CERROR("the llap wasn't freed\n"); + (*pp)->mapping = NULL; + if (page_count(*pp) != 1) + CERROR("page %p, flags %#lx, count %i, private %p\n", + (*pp), (unsigned long)(*pp)->flags, page_count(*pp), + (void*)page_private(*pp)); + __free_pages(*pp, 0); + } + } + OBD_FREE(pages, numpages * sizeof(struct page*)); + EXIT; +} + +static struct page **ll_file_prepare_pages(int numpages, struct inode *inode, + unsigned long first) +{ + struct page **pages; + int i; + int rc = 0; + ENTRY; + + OBD_ALLOC(pages, sizeof(struct page *) * numpages); + if (pages == NULL) + RETURN(ERR_PTR(-ENOMEM)); + for (i = 0; i < numpages; i++) { + struct page *page; + struct ll_async_page *llap; + + page = alloc_pages(GFP_HIGHUSER, 0); + if (page == NULL) + GOTO(err, rc = -ENOMEM); + pages[i] = page; + /* llap_from_page needs page index and mapping to be set */ + page->index = first++; + page->mapping = inode->i_mapping; + llap = llap_from_page(page, LLAP_ORIGIN_LOCKLESS_IO); + if (IS_ERR(llap)) + GOTO(err, rc = PTR_ERR(llap)); + llap->llap_lockless_io_page = 1; + } + RETURN(pages); +err: + ll_file_put_pages(pages, numpages); + RETURN(ERR_PTR(rc)); + } + +static ssize_t ll_file_copy_pages(struct page **pages, int numpages, + char *buf, loff_t pos, size_t count, int rw) +{ + ssize_t amount = 0; + int i; + int updatechecksum = ll_i2sbi(pages[0]->mapping->host)->ll_flags & + LL_SBI_CHECKSUM; + ENTRY; + + for (i = 0; i < numpages; i++) { + unsigned offset, bytes, left; + char *vaddr; + + vaddr = kmap(pages[i]); + offset = pos & (CFS_PAGE_SIZE - 1); + bytes = min_t(unsigned, CFS_PAGE_SIZE - offset, count); + LL_CDEBUG_PAGE(D_PAGE, pages[i], "op = %s, addr = %p, " + "buf = %p, bytes = %u\n", + (rw == WRITE) ? "CFU" : "CTU", + vaddr + offset, buf, bytes); + if (rw == WRITE) { + left = copy_from_user(vaddr + offset, buf, bytes); + if (updatechecksum) { + struct ll_async_page *llap; + + llap = llap_cast_private(pages[i]); + llap->llap_checksum = crc32_le(0, vaddr, + CFS_PAGE_SIZE); + } + } else { + left = copy_to_user(buf, vaddr + offset, bytes); + } + kunmap(pages[i]); + amount += bytes; + if (left) { + amount -= left; + break; + } + buf += bytes; + count -= bytes; + pos += bytes; + } + if (amount == 0) + RETURN(-EFAULT); + RETURN(amount); +} + +static int ll_file_oig_pages(struct inode * inode, struct page **pages, + int numpages, loff_t pos, size_t count, int rw) +{ + struct obd_io_group *oig; + struct ll_inode_info *lli = ll_i2info(inode); + struct obd_export *exp; + loff_t org_pos = pos; + obd_flag brw_flags; + int rc; + int i; + ENTRY; + + exp = ll_i2dtexp(inode); + if (exp == NULL) + RETURN(-EINVAL); + rc = oig_init(&oig); + if (rc) + RETURN(rc); + brw_flags = OBD_BRW_SRVLOCK; + if (capable(CAP_SYS_RESOURCE)) + brw_flags |= OBD_BRW_NOQUOTA; + + for (i = 0; i < numpages; i++) { + struct ll_async_page *llap; + unsigned from, bytes; + + from = pos & (CFS_PAGE_SIZE - 1); + bytes = min_t(unsigned, CFS_PAGE_SIZE - from, + count - pos + org_pos); + llap = llap_cast_private(pages[i]); + LASSERT(llap); + + lock_page(pages[i]); + + LL_CDEBUG_PAGE(D_PAGE, pages[i], "offset "LPU64"," + " from %u, bytes = %u\n", + (__u64)pos, from, bytes); + LASSERTF(pos >> CFS_PAGE_SHIFT == pages[i]->index, + "wrong page index %lu (%lu)\n", + pages[i]->index, + (unsigned long)(pos >> CFS_PAGE_SHIFT)); + rc = obd_queue_group_io(exp, lli->lli_smd, NULL, oig, + llap->llap_cookie, + (rw == WRITE) ? + OBD_BRW_WRITE:OBD_BRW_READ, + from, bytes, brw_flags, + ASYNC_READY | ASYNC_URGENT | + ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC); + if (rc) { + i++; + GOTO(out, rc); + } + pos += bytes; + } + rc = obd_trigger_group_io(exp, lli->lli_smd, NULL, oig); + if (rc) + GOTO(out, rc); + rc = oig_wait(oig); +out: + while(--i >= 0) + unlock_page(pages[i]); + oig_release(oig); + RETURN(rc); +} + +ssize_t ll_file_lockless_io(struct file *file, char *buf, size_t count, + loff_t *ppos, int rw) +{ + loff_t pos; + struct inode *inode = file->f_dentry->d_inode; + ssize_t rc = 0; + int max_pages; + size_t amount = 0; + unsigned long first, last; + ENTRY; + + if (rw == READ) { + loff_t isize; + + ll_inode_size_lock(inode, 0); + isize = i_size_read(inode); + ll_inode_size_unlock(inode, 0); + if (*ppos >= isize) + GOTO(out, rc = 0); + if (*ppos + count >= isize) + count -= *ppos + count - isize; + if (count == 0) + GOTO(out, rc); + } else { + rc = generic_write_checks(file, ppos, &count, 0); + if (rc) + GOTO(out, rc); + rc = ll_remove_suid(file->f_dentry, file->f_vfsmnt); + if (rc) + GOTO(out, rc); + } + pos = *ppos; + first = pos >> CFS_PAGE_SHIFT; + last = (pos + count - 1) >> CFS_PAGE_SHIFT; + max_pages = PTLRPC_MAX_BRW_PAGES * + ll_i2info(inode)->lli_smd->lsm_stripe_count; + CDEBUG(D_INFO, "%u, stripe_count = %u\n", + PTLRPC_MAX_BRW_PAGES /* max_pages_per_rpc */, + ll_i2info(inode)->lli_smd->lsm_stripe_count); + + while (first <= last && rc >= 0) { + int pages_for_io; + struct page **pages; + size_t bytes = count - amount; + + pages_for_io = min_t(int, last - first + 1, max_pages); + pages = ll_file_prepare_pages(pages_for_io, inode, first); + if (IS_ERR(pages)) { + rc = PTR_ERR(pages); + break; + } + if (rw == WRITE) { + rc = ll_file_copy_pages(pages, pages_for_io, buf, + pos + amount, bytes, rw); + if (rc < 0) + GOTO(put_pages, rc); + bytes = rc; + } + rc = ll_file_oig_pages(inode, pages, pages_for_io, + pos + amount, bytes, rw); + if (rc) + GOTO(put_pages, rc); + if (rw == READ) { + rc = ll_file_copy_pages(pages, pages_for_io, buf, + pos + amount, bytes, rw); + if (rc < 0) + GOTO(put_pages, rc); + bytes = rc; + } + amount += bytes; + buf += bytes; +put_pages: + ll_file_put_pages(pages, pages_for_io); + first += pages_for_io; + /* a short read/write check */ + if (pos + amount < ((loff_t)first << CFS_PAGE_SHIFT)) + break; + } + /* NOTE: don't update i_size and KMS in absence of LDLM locks even + * write makes the file large */ + file_accessed(file); + if (rw == READ && amount < count && rc == 0) { + unsigned long not_cleared; + + not_cleared = clear_user(buf, count - amount); + amount = count - not_cleared; + if (not_cleared) + rc = -EFAULT; + } + if (amount > 0) { + lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, + (rw == WRITE) ? + LPROC_LL_LOCKLESS_WRITE : + LPROC_LL_LOCKLESS_READ, + (long)amount); + *ppos += amount; + RETURN(amount); + } +out: + RETURN(rc); +}