X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Frw.c;h=0cf7ab2e48d0244310532af2366efef3b8773ce0;hp=3b96985f2db5226a88f22b1b74c330f1a993a325;hb=48fcf440fe5f9728e3dcdf411bd7969b01e4449d;hpb=3b6b93c54e1fa99af9d1ef9094de1c68beea8614 diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c index 3b96985..0cf7ab2 100644 --- a/lustre/llite/rw.c +++ b/lustre/llite/rw.c @@ -1,1080 +1,391 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Lustre Lite I/O page cache routines shared by different kernel revs + * 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 (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Intel Corporation. + */ +/* + * 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 #include -#include #include -#include -#include +#include #include #include #include #include -#ifdef HAVE_SEGMENT_H -# include -#endif #include #include -#include +/* current_is_kswapd() */ +#include #define DEBUG_SUBSYSTEM S_LLITE -//#include #include +#include #include "llite_internal.h" #include -#ifndef list_for_each_prev_safe -#define list_for_each_prev_safe(pos, n, head) \ - for (pos = (head)->prev, n = pos->prev; pos != (head); \ - pos = n, n = pos->prev ) -#endif - -cfs_mem_cache_t *ll_async_page_slab = NULL; -size_t ll_async_page_slab_size = 0; - -/* SYNCHRONOUS I/O to object storage for an inode */ -static int ll_brw(int cmd, struct inode *inode, struct obdo *oa, - struct page *page, int flags) +/** + * Finalizes cl-data before exiting typical address_space operation. Dual to + * ll_cl_init(). + */ +static void ll_cl_fini(struct ll_cl_context *lcc) { - struct ll_inode_info *lli = ll_i2info(inode); - struct lov_stripe_md *lsm = lli->lli_smd; - struct obd_info oinfo = { { { 0 } } }; - struct brw_page pg; - int opc, rc; - ENTRY; + struct lu_env *env = lcc->lcc_env; + struct cl_io *io = lcc->lcc_io; + struct cl_page *page = lcc->lcc_page; - pg.pg = page; - pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT; + LASSERT(lcc->lcc_cookie == current); + LASSERT(env != NULL); - if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode))) - pg.count = i_size_read(inode) % CFS_PAGE_SIZE; - else - pg.count = CFS_PAGE_SIZE; - - LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n", - cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, - inode->i_ino, pg.off, pg.off); - if (pg.count == 0) { - CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off " - LPU64"\n", inode->i_ino, inode, i_size_read(inode), - page->mapping->host, i_size_read(page->mapping->host), - page->index, pg.off); + if (page != NULL) { + lu_ref_del(&page->cp_reference, "cl_io", io); + cl_page_put(env, page); } - pg.flag = flags; - - if (cmd & OBD_BRW_WRITE) - ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE, - pg.count); - else - ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ, - pg.count); - oinfo.oi_oa = oa; - oinfo.oi_md = lsm; - /* NB partial write, so we might not have CAPA_OPC_OSS_READ capa */ - opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW; - oinfo.oi_capa = ll_osscapa_get(inode, opc); - rc = obd_brw(cmd, ll_i2dtexp(inode), &oinfo, 1, &pg, NULL); - capa_put(oinfo.oi_capa); - if (rc == 0) - obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS); - else if (rc != -EIO) - CERROR("error from obd_brw: rc = %d\n", rc); - RETURN(rc); + if (io && lcc->lcc_created) { + cl_io_end(env, io); + cl_io_unlock(env, io); + cl_io_iter_fini(env, io); + cl_io_fini(env, io); + } + cl_env_put(env, &lcc->lcc_refcheck); } -/* 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 - * avoid races. - * - * must be called under ->lli_size_sem */ -void ll_truncate(struct inode *inode) +/** + * Initializes common cl-data at the typical address_space operation entry + * point. + */ +static struct ll_cl_context *ll_cl_init(struct file *file, + struct page *vmpage, int create) { - struct ll_inode_info *lli = ll_i2info(inode); - struct obd_info oinfo = { { { 0 } } }; - struct ost_lvb lvb; - struct obdo oa; - int rc; - 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), - i_size_read(inode)); - - ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1); - if (lli->lli_size_sem_owner != current) { - EXIT; - return; - } - - if (!lli->lli_smd) { - CDEBUG(D_INODE, "truncate on inode %lu with no objects\n", - inode->i_ino); - GOTO(out_unlock, 0); - } - - 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)); - 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 - * checksum. */ - struct page *page = find_get_page(inode->i_mapping, - i_size_read(inode) >> - CFS_PAGE_SHIFT); - if (page != NULL) { - struct ll_async_page *llap = llap_cast_private(page); - if (llap != NULL) { - char *kaddr = kmap_atomic(page, KM_USER0); - llap->llap_checksum = - 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); + struct ll_cl_context *lcc; + struct lu_env *env; + struct cl_io *io; + struct cl_object *clob; + struct ccc_io *cio; + + int refcheck; + int result = 0; + + clob = ll_i2info(vmpage->mapping->host)->lli_clob; + LASSERT(clob != NULL); + + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + return ERR_PTR(PTR_ERR(env)); + + lcc = &vvp_env_info(env)->vti_io_ctx; + memset(lcc, 0, sizeof(*lcc)); + lcc->lcc_env = env; + lcc->lcc_refcheck = refcheck; + lcc->lcc_cookie = current; + + cio = ccc_env_io(env); + io = cio->cui_cl.cis_io; + if (io == NULL && create) { + struct inode *inode = vmpage->mapping->host; + loff_t pos; + + if (mutex_trylock(&inode->i_mutex)) { + mutex_unlock(&(inode)->i_mutex); + + /* this is too bad. Someone is trying to write the + * page w/o holding inode mutex. This means we can + * add dirty pages into cache during truncate */ + CERROR("Proc %s is dirting page w/o inode lock, this" + "will break truncate.\n", cfs_current()->comm); + libcfs_debug_dumpstack(NULL); + LBUG(); + return ERR_PTR(-EIO); + } + + /* + * Loop-back driver calls ->prepare_write() and ->sendfile() + * methods directly, bypassing file system ->write() operation, + * so cl_io has to be created here. + */ + io = ccc_env_thread_io(env); + ll_io_init(io, file, 1); + + /* No lock at all for this kind of IO - we can't do it because + * we have held page lock, it would cause deadlock. + * XXX: This causes poor performance to loop device - One page + * per RPC. + * In order to get better performance, users should use + * lloop driver instead. + */ + io->ci_lockreq = CILR_NEVER; + + pos = (vmpage->index << CFS_PAGE_SHIFT); + + /* Create a temp IO to serve write. */ + result = cl_io_rw_init(env, io, CIT_WRITE, pos, CFS_PAGE_SIZE); + if (result == 0) { + cio->cui_fd = LUSTRE_FPRIVATE(file); + cio->cui_iov = NULL; + cio->cui_nrsegs = 0; + result = cl_io_iter_init(env, io); + if (result == 0) { + result = cl_io_lock(env, io); + if (result == 0) + result = cl_io_start(env, io); } - 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); - - 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); - else - obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | - OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME); - EXIT; - return; - - out_unlock: - ll_inode_size_unlock(inode, 0); -} /* ll_truncate */ - -int ll_prepare_write(struct file *file, struct page *page, unsigned from, - unsigned to) -{ - struct inode *inode = page->mapping->host; - struct ll_inode_info *lli = ll_i2info(inode); - struct lov_stripe_md *lsm = lli->lli_smd; - obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT; - struct obd_info oinfo = { { { 0 } } }; - struct brw_page pga; - struct obdo oa; - struct ost_lvb lvb; - int rc = 0; - ENTRY; - - LASSERT(PageLocked(page)); - (void)llap_cast_private(page); /* assertion */ - - /* Check to see if we should return -EIO right away */ - pga.pg = page; - pga.off = offset; - pga.count = CFS_PAGE_SIZE; - pga.flag = 0; - - 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 | - OBD_MD_FLTYPE | OBD_MD_FLGROUP; - obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER); - - oinfo.oi_oa = &oa; - oinfo.oi_md = lsm; - rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL); - if (rc) - RETURN(rc); - - if (PageUptodate(page)) { - LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n"); - RETURN(0); - } - - /* We're completely overwriting an existing page, so _don't_ set it up - * to date until commit_write */ - if (from == 0 && to == CFS_PAGE_SIZE) { - LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n"); - POISON_PAGE(page, 0x11); - RETURN(0); + } else + result = io->ci_result; + lcc->lcc_created = 1; } - /* If are writing to a new page, no need to read old data. The extent - * locking will have updated the KMS, and for our purposes here we can - * treat it like i_size. */ - lov_stripe_lock(lsm); - inode_init_lvb(inode, &lvb); - obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1); - lov_stripe_unlock(lsm); - if (lvb.lvb_size <= offset) { - char *kaddr = kmap_atomic(page, KM_USER0); - LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n", - lvb.lvb_size, offset); - memset(kaddr, 0, CFS_PAGE_SIZE); - kunmap_atomic(kaddr, KM_USER0); - GOTO(prepare_done, rc = 0); + lcc->lcc_io = io; + if (io == NULL) + result = -EIO; + if (result == 0) { + struct cl_page *page; + + LASSERT(io != NULL); + LASSERT(io->ci_state == CIS_IO_GOING); + LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file)); + page = cl_page_find(env, clob, vmpage->index, vmpage, + CPT_CACHEABLE); + if (!IS_ERR(page)) { + lcc->lcc_page = page; + lu_ref_add(&page->cp_reference, "cl_io", io); + result = 0; + } else + result = PTR_ERR(page); } - - /* XXX could be an async ocp read.. read-ahead? */ - rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0); - if (rc == 0) { - /* bug 1598: don't clobber blksize */ - oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ); - obdo_refresh_inode(inode, &oa, oa.o_valid); + if (result) { + ll_cl_fini(lcc); + lcc = ERR_PTR(result); } - EXIT; - prepare_done: - if (rc == 0) - SetPageUptodate(page); - - return rc; + CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n", + vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result, + env, io); + return lcc; } -static int ll_ap_make_ready(void *data, int cmd) +static struct ll_cl_context *ll_cl_get(void) { - struct ll_async_page *llap; - struct page *page; - ENTRY; - - 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)); - - /* if we left PageDirty we might get another writepage call - * in the future. list walkers are bright enough - * to check page dirty so we can leave it on whatever list - * its on. XXX also, we're called with the cli list so if - * we got the page cache list we'd create a lock inversion - * with the removepage path which gets the page lock then the - * 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); - - /* This actually clears the dirty bit in the radix tree.*/ - set_page_writeback(page); - - LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n"); - page_cache_get(page); - - RETURN(0); + struct ll_cl_context *lcc; + struct lu_env *env; + int refcheck; + + env = cl_env_get(&refcheck); + LASSERT(!IS_ERR(env)); + lcc = &vvp_env_info(env)->vti_io_ctx; + LASSERT(env == lcc->lcc_env); + LASSERT(current == lcc->lcc_cookie); + cl_env_put(env, &refcheck); + + /* env has got in ll_cl_init, so it is still usable. */ + return lcc; } -/* We have two reasons for giving llite the opportunity to change the - * write length of a given queued page as it builds the RPC containing - * the page: - * - * 1) Further extending writes may have landed in the page cache - * since a partial write first queued this page requiring us - * to write more from the page cache. (No further races are possible, since - * by the time this is called, the page is locked.) - * 2) We might have raced with truncate and want to avoid performing - * write RPCs that are just going to be thrown away by the - * truncate's punch on the storage targets. - * - * The kms serves these purposes as it is set at both truncate and extending - * writes. +/** + * ->prepare_write() address space operation called by generic_file_write() + * for every page during write. */ -static int ll_ap_refresh_count(void *data, int cmd) -{ - struct ll_inode_info *lli; - struct ll_async_page *llap; - struct lov_stripe_md *lsm; - struct page *page; - struct inode *inode; - struct ost_lvb lvb; - __u64 kms; - ENTRY; - - /* readpage queues with _COUNT_STABLE, shouldn't get here. */ - LASSERT(cmd != OBD_BRW_READ); - - llap = LLAP_FROM_COOKIE(data); - page = llap->llap_page; - inode = page->mapping->host; - lli = ll_i2info(inode); - lsm = lli->lli_smd; - - lov_stripe_lock(lsm); - inode_init_lvb(inode, &lvb); - obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1); - kms = lvb.lvb_size; - lov_stripe_unlock(lsm); - - /* catch race with truncate */ - if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms) - return 0; - - /* catch sub-page write at end of file */ - if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms) - return kms % CFS_PAGE_SIZE; - - return CFS_PAGE_SIZE; -} - -void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa) -{ - struct lov_stripe_md *lsm; - obd_flag valid_flags; - - lsm = ll_i2info(inode)->lli_smd; - - oa->o_id = lsm->lsm_object_id; - oa->o_gr = lsm->lsm_object_gr; - oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP; - valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME; - if (cmd & OBD_BRW_WRITE) { - oa->o_valid |= OBD_MD_FLEPOCH; - oa->o_easize = ll_i2info(inode)->lli_ioepoch; - - valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME | - OBD_MD_FLUID | OBD_MD_FLGID | - OBD_MD_FLFID | OBD_MD_FLGENER; - } - - obdo_from_inode(oa, inode, valid_flags); -} - -static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa) -{ - struct ll_async_page *llap; - ENTRY; - - llap = LLAP_FROM_COOKIE(data); - ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa); - - EXIT; -} - -static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa, - obd_valid valid) -{ - struct ll_async_page *llap; - ENTRY; - - llap = LLAP_FROM_COOKIE(data); - obdo_from_inode(oa, llap->llap_page->mapping->host, valid); - - EXIT; -} - -static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd) -{ - struct ll_async_page *llap = LLAP_FROM_COOKIE(data); - int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW; - - return ll_osscapa_get(llap->llap_page->mapping->host, opc); -} - -static struct obd_async_page_ops ll_async_page_ops = { - .ap_make_ready = ll_ap_make_ready, - .ap_refresh_count = ll_ap_refresh_count, - .ap_fill_obdo = ll_ap_fill_obdo, - .ap_update_obdo = ll_ap_update_obdo, - .ap_completion = ll_ap_completion, - .ap_lookup_capa = ll_ap_lookup_capa, -}; - -struct ll_async_page *llap_cast_private(struct page *page) -{ - struct ll_async_page *llap = (struct ll_async_page *)page_private(page); - - LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC, - "page %p private %lu gave magic %d which != %d\n", - page, page_private(page), llap->llap_magic, LLAP_MAGIC); - - return llap; -} - -/* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction. - * - * 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) -{ - 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 page *page; - int keep; - - if (unlikely(need_resched())) { - spin_unlock(&sbi->ll_lock); - cond_resched(); - spin_lock(&sbi->ll_lock); - } - - llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item); - list_del_init(&dummy_llap.llap_pglist_item); - if (llap == NULL) - break; - - page = llap->llap_page; - LASSERT(page != NULL); - - list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item); - - /* Page needs/undergoing IO */ - if (TryLockPage(page)) { - LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n"); - continue; - } - - keep = (llap->llap_write_queued || PageDirty(page) || - PageWriteback(page) || (!PageUptodate(page) && - llap->llap_origin != LLAP_ORIGIN_READAHEAD)); - - LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n", - keep ? "keep" : "drop", - llap->llap_write_queued ? "wq " : "", - PageDirty(page) ? "pd " : "", - PageUptodate(page) ? "" : "!pu ", - PageWriteback(page) ? "wb" : "", - llap->llap_defer_uptodate ? "" : "!du", - llap_origins[llap->llap_origin]); - - /* If page is dirty or undergoing IO don't discard it */ - if (keep) { - unlock_page(page); - continue; - } - - page_cache_get(page); - spin_unlock(&sbi->ll_lock); - - if (page->mapping != NULL) { - ll_teardown_mmaps(page->mapping, - (__u64)page->index << CFS_PAGE_SHIFT, - ((__u64)page->index << CFS_PAGE_SHIFT)| - ~CFS_PAGE_MASK); - if (!PageDirty(page) && !page_mapped(page)) { - ll_ra_accounting(llap, page->mapping); - ll_truncate_complete_page(page); - ++count; - } else { - LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page" - " because it is " - "%s\n", - PageDirty(page)? - "dirty":"mapped"); - } - } - unlock_page(page); - page_cache_release(page); - - spin_lock(&sbi->ll_lock); - } - list_del(&dummy_llap.llap_pglist_item); - spin_unlock(&sbi->ll_lock); - - CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n", - count, want, total); - - return count; -} - -struct ll_async_page *llap_from_page(struct page *page, unsigned origin) -{ - struct ll_async_page *llap; - struct obd_export *exp; - struct inode *inode = page->mapping->host; - struct ll_sb_info *sbi; - int rc; - ENTRY; - - if (!inode) { - static int triggered; - - if (!triggered) { - LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon " - "page received\n"); - libcfs_debug_dumpstack(NULL); - triggered = 1; - } - RETURN(ERR_PTR(-EINVAL)); - } - sbi = ll_i2sbi(inode); - LASSERT(ll_async_page_slab); - LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin); - - llap = llap_cast_private(page); - if (llap != NULL) { - /* 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); - } - GOTO(out, llap); - } - - exp = ll_i2dtexp(page->mapping->host); - if (exp == NULL) - 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); - - OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD, - ll_async_page_slab_size); - if (llap == NULL) - RETURN(ERR_PTR(-ENOMEM)); - llap->llap_magic = LLAP_MAGIC; - llap->llap_cookie = (void *)llap + size_round(sizeof(*llap)); - - 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); - if (rc) { - OBD_SLAB_FREE(llap, ll_async_page_slab, - ll_async_page_slab_size); - RETURN(ERR_PTR(rc)); - } - - CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap, - page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT); - /* 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); - INIT_LIST_HEAD(&llap->llap_pending_write); - spin_unlock(&sbi->ll_lock); - - out: - if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) { - __u32 csum; - char *kaddr = kmap_atomic(page, KM_USER0); - 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) { - llap->llap_checksum = 0; - } else if (origin == LLAP_ORIGIN_COMMIT_WRITE || - llap->llap_checksum == 0) { - llap->llap_checksum = csum; - CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum); - } else if (llap->llap_checksum == csum) { - /* origin == LLAP_ORIGIN_WRITEPAGE */ - CDEBUG(D_PAGE, "page %p cksum %x confirmed\n", - page, csum); - } else { - /* origin == LLAP_ORIGIN_WRITEPAGE */ - LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new " - "%x!\n", llap->llap_checksum, csum); - } - } - - llap->llap_origin = origin; - RETURN(llap); -} - -static int queue_or_sync_write(struct obd_export *exp, struct inode *inode, - struct ll_async_page *llap, - unsigned to, obd_flag async_flags) +int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from, + unsigned to) { - unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT; - struct obd_io_group *oig; - struct ll_sb_info *sbi = ll_i2sbi(inode); - int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0; + struct ll_cl_context *lcc; + int result; ENTRY; - /* _make_ready only sees llap once we've unlocked the page */ - llap->llap_write_queued = 1; - rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL, - llap->llap_cookie, OBD_BRW_WRITE | noquot, - 0, 0, 0, async_flags); - if (rc == 0) { - LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n"); - GOTO(out, 0); - } - - 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); - - /* make full-page requests if we are not at EOF (bug 4410) */ - if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) { - LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, - "sync write before EOF: size_index %lu, to %d\n", - size_index, to); - to = CFS_PAGE_SIZE; - } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) { - int size_to = i_size_read(inode) & ~CFS_PAGE_MASK; - LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, - "sync write at EOF: size_index %lu, to %d/%d\n", - size_index, to, size_to); - if (to < size_to) - to = size_to; - } - - /* compare the checksum once before the page leaves llite */ - if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) && - llap->llap_checksum != 0)) { - __u32 csum; - struct page *page = llap->llap_page; - char *kaddr = kmap_atomic(page, KM_USER0); - 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", - page, csum); + lcc = ll_cl_init(file, vmpage, 1); + if (!IS_ERR(lcc)) { + struct lu_env *env = lcc->lcc_env; + struct cl_io *io = lcc->lcc_io; + struct cl_page *page = lcc->lcc_page; + + cl_page_assume(env, io, page); + + result = cl_io_prepare_write(env, io, page, from, to); + if (result == 0) { + /* + * Add a reference, so that page is not evicted from + * the cache until ->commit_write() is called. + */ + cl_page_get(page); + lu_ref_add(&page->cp_reference, "prepare_write", + cfs_current()); } else { - CERROR("page %p old cksum %x != new cksum %x!\n", - page, llap->llap_checksum, csum); - } - } - - rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig, - llap->llap_cookie, OBD_BRW_WRITE | noquot, - 0, to, 0, ASYNC_READY | ASYNC_URGENT | - ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC); - if (rc) - GOTO(free_oig, rc); - - rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig); - if (rc) - GOTO(free_oig, rc); - - rc = oig_wait(oig); - - if (!rc && async_flags & ASYNC_READY) { - unlock_page(llap->llap_page); - if (PageWriteback(llap->llap_page)) { - end_page_writeback(llap->llap_page); + cl_page_unassume(env, io, page); + ll_cl_fini(lcc); } + /* returning 0 in prepare assumes commit must be called + * afterwards */ + } else { + result = PTR_ERR(lcc); } - - if (rc == 0 && llap_write_complete(inode, llap)) - ll_queue_done_writing(inode, 0); - - LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc); - -free_oig: - oig_release(oig); -out: - RETURN(rc); + RETURN(result); } -/* update our write count to account for i_size increases that may have - * happened since we've queued the page for io. */ - -/* be careful not to return success without setting the page Uptodate or - * the next pass through prepare_write will read in stale data from disk. */ -int ll_commit_write(struct file *file, struct page *page, unsigned from, +int ll_commit_write(struct file *file, struct page *vmpage, unsigned from, unsigned to) { - 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; - int rc = 0; + struct ll_cl_context *lcc; + struct lu_env *env; + struct cl_io *io; + struct cl_page *page; + int result = 0; ENTRY; - SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */ - LASSERT(inode == file->f_dentry->d_inode); - LASSERT(PageLocked(page)); + lcc = ll_cl_get(); + env = lcc->lcc_env; + page = lcc->lcc_page; + io = lcc->lcc_io; - 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 (IS_ERR(llap)) - RETURN(PTR_ERR(llap)); - - exp = ll_i2dtexp(inode); - if (exp == NULL) - RETURN(-EINVAL); - - llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE); + LASSERT(cl_page_is_owned(page, io)); + LASSERT(from <= to); + if (from != to) /* handle short write case. */ + result = cl_io_commit_write(env, io, page, from, to); + if (cl_page_is_owned(page, io)) + cl_page_unassume(env, io, page); /* - * queue a write for some time in the future the first time we - * dirty the page. - * - * This is different from what other file systems do: they usually - * just mark page (and some of its buffers) dirty and rely on - * balance_dirty_pages() to start a write-back. Lustre wants write-back - * to be started earlier for the following reasons: - * - * (1) with a large number of clients we need to limit the amount - * of cached data on the clients a lot; - * - * (2) large compute jobs generally want compute-only then io-only - * and the IO should complete as quickly as possible; - * - * (3) IO is batched up to the RPC size and is async until the - * client max cache is hit - * (/proc/fs/lustre/osc/OSC.../max_dirty_mb) - * + * Release reference acquired by ll_prepare_write(). */ - if (!PageDirty(page)) { - ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1); - - rc = queue_or_sync_write(exp, inode, llap, to, 0); - if (rc) - GOTO(out, rc); - } else { - ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1); - } - - /* put the page in the page cache, from now on ll_removepage is - * responsible for cleaning up the llap. - * only set page dirty when it's queued to be write out */ - if (llap->llap_write_queued) - set_page_dirty(page); - -out: - size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to; - ll_inode_size_lock(inode, 0); - if (rc == 0) { - lov_stripe_lock(lsm); - obd_adjust_kms(exp, lsm, size, 0); - lov_stripe_unlock(lsm); - if (size > i_size_read(inode)) - i_size_write(inode, size); - SetPageUptodate(page); - } else if (size > i_size_read(inode)) { - /* this page beyond the pales of i_size, so it can't be - * truncated in ll_p_r_e during lock revoking. we must - * teardown our book-keeping here. */ - ll_removepage(page); - } - ll_inode_size_unlock(inode, 0); - RETURN(rc); + lu_ref_del(&page->cp_reference, "prepare_write", cfs_current()); + cl_page_put(env, page); + ll_cl_fini(lcc); + RETURN(result); } -static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len) +struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt) { - 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); + __u64 opc; - RETURN(ret); + opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW; + return ll_osscapa_get(inode, opc); } -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); -} +static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which); -/* called for each page in a completed rpc.*/ -int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc) +/** + * Get readahead pages from the filesystem readahead pool of the client for a + * thread. + * + * /param sbi superblock for filesystem readahead state ll_ra_info + * /param ria per-thread readahead state + * /param pages number of pages requested for readahead for the thread. + * + * WARNING: This algorithm is used to reduce 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, and not too many threads contending for + * these readahead pages. + * + * TODO: There may be a 'global sync problem' if many threads are trying + * to get an ra budget that is larger than the remaining readahead pages + * and reach here at exactly the same time. They will compute /a ret to + * consume the remaining pages, but will fail at atomic_add_return() and + * get a zero ra window, although there is still ra space remaining. - Jay */ + +static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, + struct ra_io_arg *ria, + unsigned long pages) { - struct ll_async_page *llap; - struct page *page; - int ret = 0; + struct ll_ra_info *ra = &sbi->ll_ra_info; + long ret; ENTRY; - llap = LLAP_FROM_COOKIE(data); - 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) - ll_ra_count_put(ll_i2sbi(page->mapping->host), 1); - - if (rc == 0) { - if (cmd & OBD_BRW_READ) { - if (!llap->llap_defer_uptodate) - SetPageUptodate(page); - } else { - llap->llap_write_queued = 0; - } - ClearPageError(page); - } else { - if (cmd & OBD_BRW_READ) { - llap->llap_defer_uptodate = 0; - } - SetPageError(page); - if (rc == -ENOSPC) - set_bit(AS_ENOSPC, &page->mapping->flags); - else - set_bit(AS_EIO, &page->mapping->flags); - } - - unlock_page(page); - - if (cmd & OBD_BRW_WRITE) { - /* Only rc == 0, write succeed, then this page could be deleted - * from the pending_writing list - */ - if (rc == 0 && llap_write_complete(page->mapping->host, llap)) - ll_queue_done_writing(page->mapping->host, 0); + /* If read-ahead pages left are less than 1M, do not do read-ahead, + * otherwise it will form small read RPC(< 1M), which hurt server + * performance a lot. */ + ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), pages); + if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages)) + GOTO(out, ret = 0); + + /* If the non-strided (ria_pages == 0) readahead window + * (ria_start + ret) has grown across an RPC boundary, then trim + * readahead size by the amount beyond the RPC so it ends on an + * RPC boundary. If the readahead window is already ending on + * an RPC boundary (beyond_rpc == 0), or smaller than a full + * RPC (beyond_rpc < ret) the readahead size is unchanged. + * The (beyond_rpc != 0) check is skipped since the conditional + * branch is more expensive than subtracting zero from the result. + * + * Strided read is left unaligned to avoid small fragments beyond + * the RPC boundary from needing an extra read RPC. */ + if (ria->ria_pages == 0) { + long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES; + if (/* beyond_rpc != 0 && */ beyond_rpc < ret) + ret -= beyond_rpc; } - if (PageWriteback(page)) { - end_page_writeback(page); + if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) { + cfs_atomic_sub(ret, &ra->ra_cur_pages); + ret = 0; } - page_cache_release(page); +out: 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) -{ - 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; - 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); - EXIT; - return; - } - - llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE); - if (IS_ERR(llap)) { - CERROR("page %p ind %lu couldn't find llap: %ld\n", page, - page->index, PTR_ERR(llap)); - EXIT; - return; - } - - if (llap_write_complete(inode, llap)) - ll_queue_done_writing(inode, 0); - - rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL, - llap->llap_cookie); - if (rc != 0) - CERROR("page %p ind %lu failed: %d\n", page, page->index, rc); - - /* this unconditional free is only safe because the page lock - * is providing exclusivity to memory pressure/truncate/writeback..*/ - __clear_page_ll_data(page); - - spin_lock(&sbi->ll_lock); - 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); - 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) -{ - struct lustre_handle match_lockh = {0}; - struct inode *inode = page->mapping->host; - ldlm_policy_data_t page_extent; - int flags, matches; - 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); -} - -static int ll_issue_page_read(struct obd_export *exp, - struct ll_async_page *llap, - struct obd_io_group *oig, int defer) +void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len) { - struct page *page = llap->llap_page; - int rc; - - page_cache_get(page); - llap->llap_defer_uptodate = defer; - llap->llap_ra_used = 0; - rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd, - NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0, - CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE | - ASYNC_READY | ASYNC_URGENT); - if (rc) { - LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc); - page_cache_release(page); - } - RETURN(rc); + struct ll_ra_info *ra = &sbi->ll_ra_info; + cfs_atomic_sub(len, &ra->ra_cur_pages); } -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) +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); -} - -void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping) -{ - if (!llap->llap_defer_uptodate || llap->llap_ra_used) - return; - - ll_ra_stats_inc(mapping, RA_STAT_DISCARDED); + ll_ra_stats_inc_sbi(sbi, which); } #define RAS_CDEBUG(ras) \ CDEBUG(D_READA, \ "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", \ + "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) @@ -1101,36 +412,36 @@ static struct ll_readahead_state *ll_ras_get(struct file *f) void ll_ra_read_in(struct file *f, struct ll_ra_read *rar) { - struct ll_readahead_state *ras; + struct ll_readahead_state *ras; - ras = ll_ras_get(f); + ras = ll_ras_get(f); - spin_lock(&ras->ras_lock); - ras->ras_requests++; - ras->ras_request_index = 0; - ras->ras_consecutive_requests++; - rar->lrr_reader = current; + spin_lock(&ras->ras_lock); + ras->ras_requests++; + ras->ras_request_index = 0; + ras->ras_consecutive_requests++; + rar->lrr_reader = current; - list_add(&rar->lrr_linkage, &ras->ras_read_beads); - spin_unlock(&ras->ras_lock); + cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads); + spin_unlock(&ras->ras_lock); } void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar) { - struct ll_readahead_state *ras; + struct ll_readahead_state *ras; - ras = ll_ras_get(f); + ras = ll_ras_get(f); - spin_lock(&ras->ras_lock); - list_del_init(&rar->lrr_linkage); - spin_unlock(&ras->ras_lock); + spin_lock(&ras->ras_lock); + cfs_list_del_init(&rar->lrr_linkage); + spin_unlock(&ras->ras_lock); } static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras) { struct ll_ra_read *scan; - list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) { + cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) { if (scan->lrr_reader == current) return scan; } @@ -1139,144 +450,195 @@ static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras) struct ll_ra_read *ll_ra_read_get(struct file *f) { - struct ll_readahead_state *ras; - struct ll_ra_read *bead; + struct ll_readahead_state *ras; + struct ll_ra_read *bead; + + ras = ll_ras_get(f); + + spin_lock(&ras->ras_lock); + bead = ll_ra_read_get_locked(ras); + spin_unlock(&ras->ras_lock); + return bead; +} + +static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io, + struct cl_page_list *queue, struct cl_page *page, + struct page *vmpage) +{ + struct ccc_page *cp; + int rc; - ras = ll_ras_get(f); + ENTRY; - spin_lock(&ras->ras_lock); - bead = ll_ra_read_get_locked(ras); - spin_unlock(&ras->ras_lock); - return bead; + rc = 0; + cl_page_assume(env, io, page); + lu_ref_add(&page->cp_reference, "ra", cfs_current()); + cp = cl2ccc_page(cl_page_at(page, &vvp_device_type)); + if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) { + rc = cl_page_is_under_lock(env, io, page); + if (rc == -EBUSY) { + cp->cpg_defer_uptodate = 1; + cp->cpg_ra_used = 0; + cl_page_list_add(queue, page); + rc = 1; + } else { + cl_page_delete(env, page); + rc = -ENOLCK; + } + } else { + /* skip completed pages */ + cl_page_unassume(env, io, page); + } + lu_ref_del(&page->cp_reference, "ra", cfs_current()); + cl_page_put(env, page); + RETURN(rc); } -static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig, - int index, struct address_space *mapping) +/** + * Initiates read-ahead of a page with given index. + * + * \retval +ve: page was added to \a queue. + * + * \retval -ENOLCK: there is no extent lock for this part of a file, stop + * read-ahead. + * + * \retval -ve, 0: page wasn't added to \a queue for other reason. + */ +static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io, + struct cl_page_list *queue, + pgoff_t index, struct address_space *mapping) { - struct ll_async_page *llap; - struct page *page; - unsigned int gfp_mask = 0; - int rc = 0; - + struct page *vmpage; + struct cl_object *clob = ll_i2info(mapping->host)->lli_clob; + struct cl_page *page; + enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */ + unsigned int gfp_mask; + int rc = 0; + const char *msg = NULL; + + ENTRY; + 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); + vmpage = grab_cache_page_nowait(mapping, index); + if (vmpage != NULL) { + /* Check if vmpage was truncated or reclaimed */ + if (vmpage->mapping == mapping) { + page = cl_page_find(env, clob, vmpage->index, + vmpage, CPT_CACHEABLE); + if (!IS_ERR(page)) { + rc = cl_read_ahead_page(env, io, queue, + page, vmpage); + if (rc == -ENOLCK) { + which = RA_STAT_FAILED_MATCH; + msg = "lock match failed"; + } + } else { + which = RA_STAT_FAILED_GRAB_PAGE; + msg = "cl_page_find failed"; + } + } else { + which = RA_STAT_WRONG_GRAB_PAGE; + msg = "g_c_p_n returned invalid page"; } - 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; + if (rc != 1) + unlock_page(vmpage); + page_cache_release(vmpage); } else { -unlock_page: - unlock_page(page); - LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n"); + which = RA_STAT_FAILED_GRAB_PAGE; + msg = "g_c_p_n failed"; + } + if (msg != NULL) { + ll_ra_stats_inc(mapping, which); + CDEBUG(D_READA, "%s\n", msg); } - page_cache_release(page); - return rc; + 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) \ +#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) +/* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't + * know what the actual RPC size is. If this needs to change, it makes more + * sense to tune the i_blkbits value for the file based on the OSTs it is + * striped over, rather than having a constant value for all files here. */ + +/* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - CFS_PAGE_SHIFT)). + * Temprarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled + * by default, this should be adjusted corresponding with max_read_ahead_mb + * and max_read_ahead_per_file_mb otherwise the readahead budget can be used + * up quickly which will affect read performance siginificantly. See LU-2816 */ +#define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> CFS_PAGE_SHIFT) static inline int stride_io_mode(struct ll_readahead_state *ras) { - return ras->ras_consecutive_stride_requests > 1; + 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, +/* The function calculates how much pages will be read in + * [off, off + length], in such stride IO area, + * stride_offset = st_off, stride_lengh = st_len, * stride_pages = st_pgs - */ + * + * |------------------|*****|------------------|*****|------------|*****|.... + * st_off + * |--- st_pgs ---| + * |----- st_len -----| + * + * How many pages it should read in such pattern + * |-------------------------------------------------------------| + * off + * |<------ length ------->| + * + * = |<----->| + |-------------------------------------| + |---| + * start_left st_pgs * i end_left + */ static unsigned long -stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs, - unsigned long off, unsigned length) +stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs, + unsigned long off, unsigned long length) { - unsigned long cont_len = st_off > off ? st_off - off : 0; - unsigned long stride_len = length + off > st_off ? - length + off + 1 - st_off : 0; - unsigned long left, pg_count; + __u64 start = off > st_off ? off - st_off : 0; + __u64 end = off + length > st_off ? off + length - st_off : 0; + unsigned long start_left = 0; + unsigned long end_left = 0; + unsigned long pg_count; - if (st_len == 0 || length == 0) + if (st_len == 0 || length == 0 || end == 0) return length; - left = do_div(stride_len, st_len); - left = min(left, st_pgs); + start_left = do_div(start, st_len); + if (start_left < st_pgs) + start_left = st_pgs - start_left; + else + start_left = 0; + + end_left = do_div(end, st_len); + if (end_left > st_pgs) + end_left = st_pgs; - pg_count = left + stride_len * st_pgs + cont_len; + CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n", + start, end, start_left, end_left); - LASSERT(pg_count >= left); + if (start == end) + pg_count = end_left - (st_pgs - start_left); + else + pg_count = start_left + st_pgs * (end - start - 1) + end_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); + CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu" + "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 ? + __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, + return stride_pg_count(ria->ria_stoff, ria->ria_length, ria->ria_pages, ria->ria_start, length); } @@ -1285,16 +647,17 @@ static int ria_page_count(struct ra_io_arg *ria) 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 + * 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; + return ria->ria_length == 0 || ria->ria_length == ria->ria_pages || + (idx >= ria->ria_stoff && (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, +static int ll_read_ahead_pages(const struct lu_env *env, + struct cl_io *io, struct cl_page_list *queue, + struct ra_io_arg *ria, unsigned long *reserved_pages, struct address_space *mapping, unsigned long *ra_end) @@ -1304,35 +667,36 @@ static int ll_read_ahead_pages(struct obd_export *exp, 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; + rc = ll_read_ahead_page(env, io, queue, + 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 + /* 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 + * 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); - + LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu" + "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx, + ria->ria_start, ria->ria_end, ria->ria_stoff, + ria->ria_length, ria->ria_pages); 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, + 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; } @@ -1342,78 +706,113 @@ static int ll_read_ahead_pages(struct obd_export *exp, 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) +int ll_readahead(const struct lu_env *env, struct cl_io *io, + struct ll_readahead_state *ras, struct address_space *mapping, + struct cl_page_list *queue, int flags) { + struct vvp_io *vio = vvp_env_io(env); + struct vvp_thread_info *vti = vvp_env_info(env); + struct cl_attr *attr = ccc_env_thread_attr(env); unsigned long start = 0, end = 0, reserved; - unsigned long ra_end, len; + 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; + struct ra_io_arg *ria = &vti->vti_ria; + struct ll_inode_info *lli; + struct cl_object *clob; + int ret = 0; __u64 kms; ENTRY; inode = mapping->host; - lsm = ll_i2info(inode)->lli_smd; + lli = ll_i2info(inode); + clob = lli->lli_clob; + + memset(ria, 0, sizeof *ria); - lov_stripe_lock(lsm); - inode_init_lvb(inode, &lvb); - obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1); - kms = lvb.lvb_size; - lov_stripe_unlock(lsm); + cl_object_attr_lock(clob); + ret = cl_object_attr_get(env, clob, attr); + cl_object_attr_unlock(clob); + + if (ret != 0) + RETURN(ret); + kms = attr->cat_kms; if (kms == 0) { ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN); RETURN(0); } - spin_lock(&ras->ras_lock); - bead = ll_ra_read_get_locked(ras); + spin_lock(&ras->ras_lock); + if (vio->cui_ra_window_set) + bead = &vio->cui_bead; + else + bead = NULL; + /* Enlarge the RA window to encompass the full read */ if (bead != NULL && ras->ras_window_start + ras->ras_window_len < bead->lrr_start + bead->lrr_count) { 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; } if (end != 0) { + unsigned long rpc_boundary; + /* + * Align RA window to an optimal boundary. + * + * XXX This would be better to align to cl_max_pages_per_rpc + * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may + * be aligned to the RAID stripe size in the future and that + * is more important than the RPC size. + */ + /* Note: we only trim the RPC, instead of extending the RPC + * to the boundary, so to avoid reading too much pages during + * random reading. */ + rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1))); + if (rpc_boundary > 0) + rpc_boundary--; + + if (rpc_boundary > start) + end = rpc_boundary; + /* Truncate RA window to end of file */ end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT)); + 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_length = ras->ras_stride_length; - ria.ria_pages = ras->ras_stride_pages; - } - spin_unlock(&ras->ras_lock); + 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); + len = ria_page_count(ria); if (len == 0) RETURN(0); - - reserved = ll_ra_count_get(ll_i2sbi(inode), len); - if (reserved < end - start + 1) + reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len); + if (reserved < len) ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT); - CDEBUG(D_READA, "reserved page %lu \n", reserved); - - ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end); + CDEBUG(D_READA, "reserved page %lu ra_cur %d ra_max %lu\n", reserved, + cfs_atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages), + ll_i2sbi(inode)->ll_ra_info.ra_max_pages); + + ret = ll_read_ahead_pages(env, io, queue, + ria, &reserved, mapping, &ra_end); LASSERTF(reserved >= 0, "reserved %lu\n", reserved); if (reserved != 0) @@ -1428,95 +827,107 @@ static int ll_readahead(struct ll_readahead_state *ras, * 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 */ 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 (ra_end < ras->ras_next_readahead && - index_in_window(ra_end, ras->ras_window_start, 0, - ras->ras_window_len)) { - ras->ras_next_readahead = ra_end; - RAS_CDEBUG(ras); - } - spin_unlock(&ras->ras_lock); - } + ra_end, end, ria->ria_end); + + if (ra_end != end + 1) { + spin_lock(&ras->ras_lock); + 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 = ra_end; + RAS_CDEBUG(ras); + } + spin_unlock(&ras->ras_lock); + } - RETURN(ret); + RETURN(ret); } -static void ras_set_start(struct ll_readahead_state *ras, unsigned long index) +static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras, + unsigned long index) { - ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1)); + ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1)); } /* called with the ras_lock held or from places where it doesn't matter */ -static void ras_reset(struct ll_readahead_state *ras, unsigned long index) +static void ras_reset(struct inode *inode, struct ll_readahead_state *ras, + unsigned long index) { - ras->ras_last_readpage = index; - ras->ras_consecutive_requests = 0; - ras->ras_consecutive_pages = 0; - ras->ras_window_len = 0; - ras_set_start(ras, index); - ras->ras_next_readahead = max(ras->ras_window_start, index); - - RAS_CDEBUG(ras); + ras->ras_last_readpage = index; + ras->ras_consecutive_requests = 0; + ras->ras_consecutive_pages = 0; + ras->ras_window_len = 0; + ras_set_start(inode, ras, index); + ras->ras_next_readahead = max(ras->ras_window_start, 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->ras_stride_length = 0; + ras->ras_stride_pages = 0; RAS_CDEBUG(ras); } void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) { - spin_lock_init(&ras->ras_lock); - ras_reset(ras, 0); - ras->ras_requests = 0; - INIT_LIST_HEAD(&ras->ras_read_beads); + spin_lock_init(&ras->ras_lock); + ras_reset(inode, ras, 0); + ras->ras_requests = 0; + CFS_INIT_LIST_HEAD(&ras->ras_read_beads); } -/* Check whether the read request is in the stride window. +/* + * 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) +static int index_in_stride_window(struct ll_readahead_state *ras, + unsigned long index) { - 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) - */ + unsigned long stride_gap; + + if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 || + ras->ras_stride_pages == ras->ras_stride_length) + return 0; + + stride_gap = index - ras->ras_last_readpage - 1; + + /* If it is contiguous read */ + if (stride_gap == 0) + return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages; + + /* Otherwise check the stride by itself */ + return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap && + ras->ras_consecutive_pages == ras->ras_stride_pages; +} + +static void ras_update_stride_detector(struct ll_readahead_state *ras, + unsigned long index) +{ + unsigned long stride_gap = index - ras->ras_last_readpage - 1; + + if (!stride_io_mode(ras) && (stride_gap != 0 || + ras->ras_consecutive_stride_requests == 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->ras_stride_length = stride_gap +ras->ras_consecutive_pages; } - RAS_CDEBUG(ras); + LASSERT(ras->ras_request_index == 0); + LASSERT(ras->ras_consecutive_stride_requests == 0); + + if (index <= ras->ras_last_readpage) { + /*Reset stride window for forward read*/ + ras_stride_reset(ras); + return; + } + + ras->ras_stride_pages = ras->ras_consecutive_pages; + ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages; - return 0; + RAS_CDEBUG(ras); + return; } static unsigned long @@ -1529,7 +940,7 @@ stride_page_count(struct ll_readahead_state *ras, unsigned long 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, +static void ras_stride_increase_window(struct ll_readahead_state *ras, struct ll_ra_info *ra, unsigned long inc_len) { @@ -1537,22 +948,21 @@ static void ras_stride_increase_window(struct ll_readahead_state *ras, unsigned long stride_len; LASSERT(ras->ras_stride_length > 0); + LASSERTF(ras->ras_window_start + ras->ras_window_len + >= ras->ras_stride_offset, "window_start %lu, window_len %lu" + " stride_offset %lu\n", ras->ras_window_start, + ras->ras_window_len, ras->ras_stride_offset); 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; + left = ras->ras_stride_pages + inc_len; LASSERT(ras->ras_stride_pages != 0); @@ -1561,40 +971,39 @@ static void ras_stride_increase_window(struct ll_readahead_state *ras, window_len += step * ras->ras_stride_length + left; - if (stride_page_count(ras, window_len) <= ra->ra_max_pages) + if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file) 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) +static void ras_increase_window(struct inode *inode, + struct ll_readahead_state *ras, + struct ll_ra_info *ra) { - 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); + /* The stretch of ra-window should be aligned with max rpc_size + * but current clio architecture does not support retrieve such + * information from lower layer. FIXME later + */ + if (stride_io_mode(ras)) + ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode)); + else + ras->ras_window_len = min(ras->ras_window_len + + RAS_INCREASE_STEP(inode), + ra->ra_max_pages_per_file); } -static void ras_update(struct ll_sb_info *sbi, struct inode *inode, - struct ll_readahead_state *ras, unsigned long index, - unsigned hit) +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, stride_zero = 0, stride_detect = 0, ra_miss = 0; - ENTRY; + struct ll_ra_info *ra = &sbi->ll_ra_info; + int zero = 0, stride_detect = 0, ra_miss = 0; + ENTRY; - spin_lock(&sbi->ll_lock); - spin_lock(&ras->ras_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 @@ -1603,28 +1012,19 @@ 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); - /* check whether it is in stride I/O mode*/ - if (!index_in_stride_window(index, ras, inode)) - stride_zero = 1; + ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE); } 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; - 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_unlocked(ra, RA_STAT_MISS_IN_WINDOW); + ra_miss = 1; + ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW); } /* On the second access to a file smaller than the tunable * ra_max_read_ahead_whole_pages trigger RA on all pages in the - * file up to ra_max_pages. This is simply a best effort and - * only occurs once per open file. Normal RA behavior is reverted + * file up to ra_max_pages_per_file. This is simply a best effort + * and only occurs once per open file. Normal RA behavior is reverted * to for subsequent IO. The mmap case does not increment * ras_requests and thus can never trigger this behavior. */ if (ras->ras_requests == 2 && !ras->ras_request_index) { @@ -1634,245 +1034,275 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode, CFS_PAGE_SHIFT; CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages, - ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages); + ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file); if (kms_pages && kms_pages <= ra->ra_max_read_ahead_whole_pages) { ras->ras_window_start = 0; ras->ras_last_readpage = 0; ras->ras_next_readahead = 0; - ras->ras_window_len = min(ra->ra_max_pages, + ras->ras_window_len = min(ra->ra_max_pages_per_file, ra->ra_max_read_ahead_whole_pages); GOTO(out_unlock, 0); } } + if (zero) { + /* check whether it is in stride I/O mode*/ + if (!index_in_stride_window(ras, index)) { + if (ras->ras_consecutive_stride_requests == 0 && + ras->ras_request_index == 0) { + ras_update_stride_detector(ras, index); + ras->ras_consecutive_stride_requests++; + } else { + ras_stride_reset(ras); + } + ras_reset(inode, ras, index); + ras->ras_consecutive_pages++; + GOTO(out_unlock, 0); + } else { + ras->ras_consecutive_pages = 0; + ras->ras_consecutive_requests = 0; + if (++ras->ras_consecutive_stride_requests > 1) + stride_detect = 1; + RAS_CDEBUG(ras); + } + } else { + if (ra_miss) { + if (index_in_stride_window(ras, index) && + stride_io_mode(ras)) { + /*If stride-RA hit cache miss, the stride dector + *will not be reset to avoid the overhead of + *redetecting read-ahead mode */ + if (index != ras->ras_last_readpage + 1) + ras->ras_consecutive_pages = 0; + ras_reset(inode, ras, index); + RAS_CDEBUG(ras); + } else { + /* Reset both stride window and normal RA + * window */ + ras_reset(inode, ras, index); + ras->ras_consecutive_pages++; + ras_stride_reset(ras); + GOTO(out_unlock, 0); + } + } else if (stride_io_mode(ras)) { + /* 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 (!index_in_stride_window(ras, index)) { + /* Shrink stride read-ahead window to be zero */ + ras_stride_reset(ras); + ras->ras_window_len = 0; + ras->ras_next_readahead = index; + } + } + } + ras->ras_consecutive_pages++; + ras->ras_last_readpage = index; + ras_set_start(inode, ras, index); + + if (stride_io_mode(ras)) + /* Since stride readahead is sentivite to the offset + * of read-ahead, so we use original offset here, + * instead of ras_window_start, which is RPC aligned */ + ras->ras_next_readahead = max(index, ras->ras_next_readahead); + else + 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 == 4) { + ras->ras_window_len = RAS_INCREASE_STEP(inode); + GOTO(out_unlock, 0); + } - if (zero) { - /* 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; - ras->ras_consecutive_pages++; - 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 == 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) { + /** + * Once stride IO mode is detected, next_readahead should be + * reset to make sure next_readahead > stride offset + */ + ras->ras_next_readahead = max(index, ras->ras_next_readahead); + ras->ras_stride_offset = index; + ras->ras_window_len = RAS_INCREASE_STEP(inode); + } - /* 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) || 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; + /* 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 || stride_detect) && + !ras->ras_request_index) + ras_increase_window(inode, ras, ra); + EXIT; out_unlock: - RAS_CDEBUG(ras); - ras->ras_request_index++; - spin_unlock(&ras->ras_lock); - spin_unlock(&sbi->ll_lock); - return; + RAS_CDEBUG(ras); + ras->ras_request_index++; + spin_unlock(&ras->ras_lock); + return; } -int ll_writepage(struct page *page) +int ll_writepage(struct page *vmpage, struct writeback_control *wbc) { - struct inode *inode = page->mapping->host; - struct ll_inode_info *lli = ll_i2info(inode); - struct obd_export *exp; - struct ll_async_page *llap; - int rc = 0; + struct inode *inode = vmpage->mapping->host; + struct ll_inode_info *lli = ll_i2info(inode); + struct lu_env *env; + struct cl_io *io; + struct cl_page *page; + struct cl_object *clob; + struct cl_env_nest nest; + bool redirtied = false; + bool unlocked = false; + int result; ENTRY; - LASSERT(PageLocked(page)); - - exp = ll_i2dtexp(inode); - if (exp == NULL) - GOTO(out, rc = -EINVAL); - - llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE); - if (IS_ERR(llap)) - GOTO(out, rc = PTR_ERR(llap)); + LASSERT(PageLocked(vmpage)); + LASSERT(!PageWriteback(vmpage)); + + LASSERT(ll_i2dtexp(inode) != NULL); + + env = cl_env_nested_get(&nest); + if (IS_ERR(env)) + GOTO(out, result = PTR_ERR(env)); + + clob = ll_i2info(inode)->lli_clob; + LASSERT(clob != NULL); + + io = ccc_env_thread_io(env); + io->ci_obj = clob; + io->ci_ignore_layout = 1; + result = cl_io_init(env, io, CIT_MISC, clob); + if (result == 0) { + page = cl_page_find(env, clob, vmpage->index, + vmpage, CPT_CACHEABLE); + if (!IS_ERR(page)) { + lu_ref_add(&page->cp_reference, "writepage", + cfs_current()); + cl_page_assume(env, io, page); + result = cl_page_flush(env, io, page); + if (result != 0) { + /* + * Re-dirty page on error so it retries write, + * but not in case when IO has actually + * occurred and completed with an error. + */ + if (!PageError(vmpage)) { + redirty_page_for_writepage(wbc, vmpage); + result = 0; + redirtied = true; + } + } + cl_page_disown(env, io, page); + unlocked = true; + lu_ref_del(&page->cp_reference, + "writepage", cfs_current()); + cl_page_put(env, page); + } else { + result = PTR_ERR(page); + } + } + cl_io_fini(env, io); + + if (redirtied && wbc->sync_mode == WB_SYNC_ALL) { + loff_t offset = cl_offset(clob, vmpage->index); + + /* Flush page failed because the extent is being written out. + * Wait for the write of extent to be finished to avoid + * breaking kernel which assumes ->writepage should mark + * PageWriteback or clean the page. */ + result = cl_sync_file_range(inode, offset, + offset + CFS_PAGE_SIZE - 1, + CL_FSYNC_LOCAL); + if (result > 0) { + /* actually we may have written more than one page. + * decreasing this page because the caller will count + * it. */ + wbc->nr_to_write -= result - 1; + result = 0; + } + } - LASSERT(!PageWriteback(page)); - set_page_writeback(page); + cl_env_nested_put(&nest, env); + GOTO(out, result); - page_cache_get(page); - if (llap->llap_write_queued) { - LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n"); - rc = obd_set_async_flags(exp, lli->lli_smd, NULL, - llap->llap_cookie, - ASYNC_READY | ASYNC_URGENT); - } else { - 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)) { - end_page_writeback(page); - } - /* resend page only for not started IO*/ - if (!PageError(page)) - ll_redirty_page(page); - unlock_page(page); - } - RETURN(rc); + if (result < 0) { + if (!lli->lli_async_rc) + lli->lli_async_rc = result; + SetPageError(vmpage); + if (!unlocked) + unlock_page(vmpage); + } + return result; } -/* - * for now we do our readpage the same on both 2.4 and 2.5. The kernel's - * read-ahead assumes it is valid to issue readpage all the way up to - * i_size, but our dlm locks make that not the case. We disable the - * kernel's read-ahead and do our own by walking ahead in the page cache - * checking for dlm lock coverage. the main difference between 2.4 and - * 2.6 is how read-ahead gets batched and issued, but we're using our own, - * so they look the same. - */ -int ll_readpage(struct file *filp, struct page *page) +int ll_writepages(struct address_space *mapping, struct writeback_control *wbc) { - struct ll_file_data *fd = LUSTRE_FPRIVATE(filp); - struct inode *inode = page->mapping->host; - struct obd_export *exp; - struct ll_async_page *llap; - struct obd_io_group *oig = NULL; - int rc; - ENTRY; - - LASSERT(PageLocked(page)); - LASSERT(!PageUptodate(page)); - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n", - inode->i_ino, inode->i_generation, inode, - (((loff_t)page->index) << CFS_PAGE_SHIFT), - (((loff_t)page->index) << CFS_PAGE_SHIFT)); - LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0); - - if (!ll_i2info(inode)->lli_smd) { - /* File with no objects - one big hole */ - /* We use this just for remove_from_page_cache that is not - * exported, we'd make page back up to date. */ - ll_truncate_complete_page(page); - clear_page(kmap(page)); - kunmap(page); - SetPageUptodate(page); - unlock_page(page); - RETURN(0); - } - - rc = oig_init(&oig); - if (rc < 0) - GOTO(out, rc); - - exp = ll_i2dtexp(inode); - if (exp == NULL) - GOTO(out, rc = -EINVAL); - - llap = llap_from_page(page, LLAP_ORIGIN_READPAGE); - if (IS_ERR(llap)) - 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, - llap->llap_defer_uptodate); - - - if (llap->llap_defer_uptodate) { - /* This is the callpath if we got the page from a readahead */ - llap->llap_ra_used = 1; - rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig, - fd->fd_flags); - if (rc > 0) - obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, - NULL, oig); - LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n"); - SetPageUptodate(page); - unlock_page(page); - GOTO(out_oig, rc = 0); - } + struct inode *inode = mapping->host; + loff_t start; + loff_t end; + enum cl_fsync_mode mode; + int range_whole = 0; + int result; + ENTRY; + + if (wbc->range_cyclic) { + start = mapping->writeback_index << CFS_PAGE_SHIFT; + end = OBD_OBJECT_EOF; + } else { + start = wbc->range_start; + end = wbc->range_end; + if (end == LLONG_MAX) { + end = OBD_OBJECT_EOF; + range_whole = start == 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); - } + mode = CL_FSYNC_NONE; + if (wbc->sync_mode == WB_SYNC_ALL) + mode = CL_FSYNC_LOCAL; - 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); - } - } + result = cl_sync_file_range(inode, start, end, mode); + if (result > 0) { + wbc->nr_to_write -= result; + result = 0; + } - rc = ll_issue_page_read(exp, llap, oig, 0); - if (rc) - GOTO(out, rc); + if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) { + if (end == OBD_OBJECT_EOF) + end = i_size_read(inode); + mapping->writeback_index = (end >> CFS_PAGE_SHIFT) + 1; + } + RETURN(result); +} - LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n"); - /* We have just requested the actual page we want, see if we can tack - * on some readahead to that page's RPC before it is sent. */ - if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages) - ll_readahead(&fd->fd_ras, exp, page->mapping, oig, - fd->fd_flags); +int ll_readpage(struct file *file, struct page *vmpage) +{ + struct ll_cl_context *lcc; + int result; + ENTRY; - rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig); + lcc = ll_cl_init(file, vmpage, 0); + if (!IS_ERR(lcc)) { + struct lu_env *env = lcc->lcc_env; + struct cl_io *io = lcc->lcc_io; + struct cl_page *page = lcc->lcc_page; -out: - if (rc) - unlock_page(page); -out_oig: - if (oig != NULL) - oig_release(oig); - RETURN(rc); + LASSERT(page->cp_type == CPT_CACHEABLE); + if (likely(!PageUptodate(vmpage))) { + cl_page_assume(env, io, page); + result = cl_io_read_page(env, io, page); + } else { + /* Page from a non-object file. */ + unlock_page(vmpage); + result = 0; + } + ll_cl_fini(lcc); + } else { + unlock_page(vmpage); + result = PTR_ERR(lcc); + } + RETURN(result); } +