X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Frw.c;h=4c6da36d9665f8e04815f6601b9f07f310748fca;hp=01ad4a11a39fc7da8986fa5cc6ee66876a0ec012;hb=a221fb78b48c82df803e1f6f1548a2f75e0ad13f;hpb=eef0dc3356976bad84f7aefb6595f10443a5980d diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c index 01ad4a1..4c6da36 100644 --- a/lustre/llite/rw.c +++ b/lustre/llite/rw.c @@ -1,27 +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. + * + * 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 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). * - * Copyright (c) 2001-2003 Cluster File Systems, Inc. + * 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 * - * This file is part of Lustre, http://www.lustre.org. + * 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. * - * 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. + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. 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 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. + * lustre/llite/rw.c * - * 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. + * Lustre Lite I/O page cache routines shared by different kernel revs */ -#include +#include #include #include #include @@ -36,881 +53,1173 @@ #include #include #include -#include #include #include #include +/* current_is_kswapd() */ +#include #define DEBUG_SUBSYSTEM S_LLITE -#include -#include +//#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 - -/* 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) +/* 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) { struct ll_inode_info *lli = ll_i2info(inode); - struct lov_stripe_md *lsm = lli->lli_smd; - struct brw_page pg; - int rc; ENTRY; - pg.pg = page; - pg.off = ((obd_off)page->index) << PAGE_SHIFT; + CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu\n",inode->i_ino, + inode->i_generation, inode, i_size_read(inode)); - if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size)) - pg.count = inode->i_size % PAGE_SIZE; - else - pg.count = PAGE_SIZE; - - CDEBUG(D_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, inode->i_size, page->mapping->host, - page->mapping->host->i_size, page->index, pg.off); + ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1); + if (lli->lli_size_sem_owner == cfs_current()) { + LASSERT_SEM_LOCKED(&lli->lli_size_sem); + ll_inode_size_unlock(inode, 0); } - pg.flag = flags; - - if (cmd == OBD_BRW_WRITE) - lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, - LPROC_LL_BRW_WRITE, pg.count); - else - lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, - LPROC_LL_BRW_READ, pg.count); - rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL); - 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); -} + EXIT; + return; +} /* ll_truncate */ -/* this isn't where truncate starts. roughly: - * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate - * we grab the lock back in setattr_raw to avoid races. */ -void ll_truncate(struct inode *inode) +/** + * 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 lov_stripe_md *lsm = ll_i2info(inode)->lli_smd; - struct obdo oa; - int rc; - ENTRY; - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino, - inode->i_generation, inode); + struct lu_env *env = lcc->lcc_env; + struct cl_io *io = lcc->lcc_io; + struct cl_page *page = lcc->lcc_page; - if (!lsm) { - CDEBUG(D_INODE, "truncate on inode %lu with no objects\n", - inode->i_ino); - EXIT; - return; + LASSERT(lcc->lcc_cookie == current); + LASSERT(env != NULL); + + if (page != NULL) { + lu_ref_del(&page->cp_reference, "cl_io", io); + cl_page_put(env, page); } - oa.o_id = lsm->lsm_object_id; - oa.o_valid = OBD_MD_FLID; - obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME| - OBD_MD_FLMTIME | OBD_MD_FLCTIME); + 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); +} + +/** + * 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_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 vvp_io *vio; + loff_t pos; + + /* + * 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_info(env)->cti_io; + vio = vvp_env_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); + } + } else + result = io->ci_result; + lcc->lcc_created = 1; + } - CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n", - oa.o_id, inode->i_size); + 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); + } + if (result) { + ll_cl_fini(lcc); + lcc = ERR_PTR(result); + } - /* truncate == punch from new size to absolute end of file */ - /* NB: obd_punch must be called with i_sem held! It updates the kms! */ - rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size, - OBD_OBJECT_EOF, NULL); - 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); + CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %i %p %p\n", + vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result, + env, io); + return lcc; +} - EXIT; - return; -} /* ll_truncate */ +static struct ll_cl_context *ll_cl_get(void) +{ + 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; +} -__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms); -int ll_prepare_write(struct file *file, struct page *page, unsigned from, +/** + * ->prepare_write() address space operation called by generic_file_write() + * for every page during write. + */ +int ll_prepare_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; - obd_off offset = ((obd_off)page->index) << PAGE_SHIFT; - struct brw_page pga; - struct obdo oa; - __u64 kms; - int rc = 0; + struct ll_cl_context *lcc; + int result; ENTRY; - if (!PageLocked(page)) - LBUG(); + 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); + if (cl_io_is_append(io)) { + struct cl_object *obj = io->ci_obj; + struct inode *inode = ccc_object_inode(obj); + /** + * In VFS file->page write loop, for appending, the + * write offset might be reset according to the new + * file size before holding i_mutex. So crw_pos should + * be reset here. BUG:17711. + */ + io->u.ci_wr.wr.crw_pos = i_size_read(inode); + } + 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 { + 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); + } + RETURN(result); +} - /* Check to see if we should return -EIO right away */ - pga.pg = page; - pga.off = offset; - pga.count = PAGE_SIZE; - pga.flag = 0; +int ll_commit_write(struct file *file, struct page *vmpage, unsigned from, + unsigned to) +{ + struct ll_cl_context *lcc; + struct lu_env *env; + struct cl_io *io; + struct cl_page *page; + int result; + ENTRY; - oa.o_id = lsm->lsm_object_id; - oa.o_mode = inode->i_mode; - oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE; + lcc = ll_cl_get(); + env = lcc->lcc_env; + page = lcc->lcc_page; + io = lcc->lcc_io; + + LASSERT(cl_page_is_owned(page, io)); + result = cl_io_commit_write(env, io, page, from, to); + if (cl_page_is_owned(page, io)) + cl_page_unassume(env, io, page); + /* + * Release reference acquired by cl_io_prepare_write(). + */ + lu_ref_del(&page->cp_reference, "prepare_write", cfs_current()); + cl_page_put(env, page); + ll_cl_fini(lcc); + RETURN(result); +} - rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga, - NULL); - if (rc) - RETURN(rc); +struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt) +{ + __u64 opc; - if (PageUptodate(page)) - RETURN(0); + opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW; + return ll_osscapa_get(inode, opc); +} - /* We're completely overwriting an existing page, so _don't_ set it up - * to date until commit_write */ - if (from == 0 && to == PAGE_SIZE) { - POISON_PAGE(page, 0x11); - RETURN(0); - } +static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which); - /* 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. */ - kms = lov_merge_size(lsm, 1); - if (kms <= offset) { - memset(kmap(page), 0, PAGE_SIZE); - kunmap(page); - GOTO(prepare_done, rc = 0); - } +/* 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; - /* 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 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), len); + if ((int)ret < 0 || ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len)) + GOTO(out, ret = 0); + + if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) { + cfs_atomic_sub(ret, &ra->ra_cur_pages); + ret = 0; } +out: + RETURN(ret); +} - EXIT; - prepare_done: - if (rc == 0) - SetPageUptodate(page); - - return rc; +void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len) +{ + struct ll_ra_info *ra = &sbi->ll_ra_info; + cfs_atomic_sub(len, &ra->ra_cur_pages); } -struct ll_async_page *llap_from_cookie(void *cookie) +static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which) { - struct ll_async_page *llap = cookie; - if (llap->llap_magic != LLAP_MAGIC) - return ERR_PTR(-EINVAL); - return llap; -}; + LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which); + lprocfs_counter_incr(sbi->ll_ra_stats, which); +} -static int ll_ap_make_ready(void *data, int cmd) +void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which) { - struct ll_async_page *llap; - struct page *page; - ENTRY; + struct ll_sb_info *sbi = ll_i2sbi(mapping->host); + ll_ra_stats_inc_sbi(sbi, which); +} - llap = llap_from_cookie(data); - if (IS_ERR(llap)) - RETURN(-EINVAL); +#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", \ + 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_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) +{ + unsigned long start = point - before, end = point + after; - page = llap->llap_page; + if (start > point) + start = 0; + if (end < point) + end = ~0; - if (cmd == OBD_BRW_READ) - RETURN(0); - - /* we're trying to write, but the page is locked.. come back later */ - if (TryLockPage(page)) - RETURN(-EAGAIN); + return start <= index && index <= end; +} - LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n"); - page_cache_get(page); +static struct ll_readahead_state *ll_ras_get(struct file *f) +{ + struct ll_file_data *fd; - /* 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 */ - clear_page_dirty(page); - RETURN(0); + fd = LUSTRE_FPRIVATE(f); + return &fd->fd_ras; } -/* 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. - * 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. - */ -static int ll_ap_refresh_count(void *data, int cmd) +void ll_ra_read_in(struct file *f, struct ll_ra_read *rar) { - struct ll_async_page *llap; - struct lov_stripe_md *lsm; - struct page *page; - __u64 kms; - ENTRY; + struct ll_readahead_state *ras; - /* readpage queues with _COUNT_STABLE, shouldn't get here. */ - LASSERT(cmd != OBD_BRW_READ); + ras = ll_ras_get(f); - llap = llap_from_cookie(data); - if (IS_ERR(llap)) - RETURN(PTR_ERR(llap)); + cfs_spin_lock(&ras->ras_lock); + ras->ras_requests++; + ras->ras_request_index = 0; + ras->ras_consecutive_requests++; + rar->lrr_reader = current; - page = llap->llap_page; - lsm = ll_i2info(page->mapping->host)->lli_smd; - kms = lov_merge_size(lsm, 1); + cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads); + cfs_spin_unlock(&ras->ras_lock); +} - /* catch race with truncate */ - if (((__u64)page->index << PAGE_SHIFT) >= kms) - return 0; +void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar) +{ + struct ll_readahead_state *ras; - /* catch sub-page write at end of file */ - if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms) - return kms % PAGE_SIZE; + ras = ll_ras_get(f); - return PAGE_SIZE; + cfs_spin_lock(&ras->ras_lock); + cfs_list_del_init(&rar->lrr_linkage); + cfs_spin_unlock(&ras->ras_lock); } -void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa) +static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras) { - struct lov_stripe_md *lsm; - obd_flag valid_flags; + struct ll_ra_read *scan; - lsm = ll_i2info(inode)->lli_smd; + cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) { + if (scan->lrr_reader == current) + return scan; + } + return NULL; +} - oa->o_id = lsm->lsm_object_id; - oa->o_valid = OBD_MD_FLID; - valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME; - if (cmd == OBD_BRW_WRITE) { - oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH; - mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode); - oa->o_easize = ll_i2info(inode)->lli_io_epoch; +struct ll_ra_read *ll_ra_read_get(struct file *f) +{ + struct ll_readahead_state *ras; + struct ll_ra_read *bead; - valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME; - } + ras = ll_ras_get(f); - obdo_from_inode(oa, inode, valid_flags); + cfs_spin_lock(&ras->ras_lock); + bead = ll_ra_read_get_locked(ras); + cfs_spin_unlock(&ras->ras_lock); + return bead; } -static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa) +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 ll_async_page *llap; - ENTRY; + struct ccc_page *cp; + int rc; - llap = llap_from_cookie(data); - if (IS_ERR(llap)) { - EXIT; - return; - } + ENTRY; - ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa); - EXIT; + 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 && !Page_Uptodate(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 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_completion = ll_ap_completion, -}; - -/* XXX have the exp be an argument? */ -struct ll_async_page *llap_from_page(struct page *page) +/** + * 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 obd_export *exp; - struct inode *inode = page->mapping->host; - struct ll_sb_info *sbi = ll_i2sbi(inode); - int rc; + 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; - llap = (struct ll_async_page *)page->private; - if (llap != NULL) { - if (llap->llap_magic != LLAP_MAGIC) - RETURN(ERR_PTR(-EINVAL)); - RETURN(llap); + gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT; +#ifdef __GFP_NOWARN + gfp_mask |= __GFP_NOWARN; +#endif + vmpage = grab_cache_page_nowait_gfp(mapping, index, gfp_mask); + 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"; + } + if (rc != 1) + unlock_page(vmpage); + page_cache_release(vmpage); + } else { + which = RA_STAT_FAILED_GRAB_PAGE; + msg = "g_c_p_n failed"; } - - exp = ll_i2obdexp(page->mapping->host); - if (exp == NULL) - RETURN(ERR_PTR(-EINVAL)); - - OBD_ALLOC(llap, sizeof(*llap)); - if (llap == NULL) - RETURN(ERR_PTR(-ENOMEM)); - llap->llap_magic = LLAP_MAGIC; - rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page, - (obd_off)page->index << PAGE_SHIFT, - &ll_async_page_ops, llap, &llap->llap_cookie); - if (rc) { - OBD_FREE(llap, sizeof(*llap)); - RETURN(ERR_PTR(rc)); + if (msg != NULL) { + ll_ra_stats_inc(mapping, which); + CDEBUG(D_READA, "%s\n", msg); } + RETURN(rc); +} - CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap, - page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT); - /* also zeroing the PRIVBITS low order bitflags */ - page->private = (unsigned long)llap; - llap->llap_page = page; +#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) - spin_lock(&sbi->ll_pglist_lock); - sbi->ll_pglist_gen++; - list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist); - spin_unlock(&sbi->ll_pglist_lock); +#define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES - RETURN(llap); +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], 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 long length) +{ + __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 || end == 0) + return length; + + start_left = do_div(start, st_len); + if (start_left < st_pgs) + start_left = st_pgs - start_left; + else + start_left = 0; -void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm, - obd_off size); -/* update our write count to account for i_size increases that may have - * happened since we've queued the page for io. */ + end_left = do_div(end, st_len); + if (end_left > st_pgs) + end_left = st_pgs; -/* 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, - unsigned to) + CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n", + start, end, start_left, end_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 %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) { - 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 = NULL; - struct ll_async_page *llap; - loff_t size; - int rc = 0; - ENTRY; + __u64 length = ria->ria_end >= ria->ria_start ? + ria->ria_end - ria->ria_start + 1 : 0; - SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */ - LASSERT(inode == file->f_dentry->d_inode); - LASSERT(PageLocked(page)); - - 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); - if (IS_ERR(llap)) - RETURN(PTR_ERR(llap)); - - /* queue a write for some time in the future the first time we - * dirty the page */ - if (!PageDirty(page)) { - lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, - LPROC_LL_DIRTY_MISSES); - - exp = ll_i2obdexp(inode); - if (exp == NULL) - RETURN(-EINVAL); - - /* _make_ready only sees llap once we've unlocked the page */ - llap->llap_write_queued = 1; - rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie, - OBD_BRW_WRITE, 0, 0, 0, 0); - if (rc != 0) { /* async failed, try sync.. */ - struct obd_io_group *oig; - rc = oig_init(&oig); - if (rc) - GOTO(out, rc); - - llap->llap_write_queued = 0; - rc = obd_queue_group_io(exp, lsm, NULL, oig, - llap->llap_cookie, - OBD_BRW_WRITE, 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, lsm, NULL, oig); - if (rc) - GOTO(free_oig, rc); - - rc = oig_wait(oig); -free_oig: - oig_release(oig); - GOTO(out, rc); - } - LL_CDEBUG_PAGE(D_PAGE, page, "write queued\n"); - //llap_write_pending(inode, llap); - } else { - lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, - LPROC_LL_DIRTY_HITS); - } + return stride_pg_count(ria->ria_stoff, ria->ria_length, + ria->ria_pages, ria->ria_start, + length); +} - /* put the page in the page cache, from now on ll_removepage is - * responsible for cleaning up the llap */ - set_page_dirty(page); +/*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 && (idx - ria->ria_stoff) % + ria->ria_length < ria->ria_pages); +} -out: - if (rc == 0) { - size = (((obd_off)page->index) << PAGE_SHIFT) + to; - lov_increase_kms(exp, lsm, size); - if (size > inode->i_size) - inode->i_size = size; - SetPageUptodate(page); +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) +{ + 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(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 + * 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, "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, + ria->ria_length - offset); + continue; + } + } } - RETURN(rc); + *ra_end = page_idx; + return count; } -/* called for each page in a completed rpc.*/ -void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc) +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 ll_async_page *llap; - struct page *page; + 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; + struct inode *inode; + struct ll_ra_read *bead; + struct ra_io_arg *ria = &vti->vti_ria; + struct ll_inode_info *lli; + struct cl_object *clob; + int ret = 0; + __u64 kms; ENTRY; - llap = llap_from_cookie(data); - if (IS_ERR(llap)) { - EXIT; - return; - } + inode = mapping->host; + lli = ll_i2info(inode); + clob = lli->lli_clob; - page = llap->llap_page; - LASSERT(PageLocked(page)); + memset(ria, 0, sizeof *ria); - LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc); + cl_object_attr_lock(clob); + ret = cl_object_attr_get(env, clob, attr); + cl_object_attr_unlock(clob); - 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 (ret != 0) + RETURN(ret); + kms = attr->cat_kms; + if (kms == 0) { + ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN); + RETURN(0); } + cfs_spin_lock(&ras->ras_lock); + if (vio->cui_ra_window_set) + bead = &vio->cui_bead; + else + bead = NULL; - unlock_page(page); + /* 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 */ + 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 tmp_end; + /* + * 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. + */ + tmp_end = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1))) - 1; + if (tmp_end > start) + end = tmp_end; + + /* 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_stoff = ras->ras_stride_offset; + ria->ria_length = ras->ras_stride_length; + ria->ria_pages = ras->ras_stride_pages; + } + cfs_spin_unlock(&ras->ras_lock); - if (0 && cmd == OBD_BRW_WRITE) { - llap_write_complete(page->mapping->host, llap); - ll_try_done_writing(page->mapping->host); + if (end == 0) { + ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW); + RETURN(0); } + len = ria_page_count(ria); + if (len == 0) + RETURN(0); - page_cache_release(page); - EXIT; -} + reserved = ll_ra_count_get(ll_i2sbi(inode), len); -/* 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; + if (reserved < len) + ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT); - LASSERT(!in_interrupt()); + CDEBUG(D_READA, "reserved page %lu \n", reserved); - /* 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 == 0) { - EXIT; - return; - } + ret = ll_read_ahead_pages(env, io, queue, + ria, &reserved, mapping, &ra_end); - LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n"); + LASSERTF(reserved >= 0, "reserved %lu\n", reserved); + if (reserved != 0) + ll_ra_count_put(ll_i2sbi(inode), reserved); - exp = ll_i2obdexp(inode); - if (exp == NULL) { - CERROR("page %p ind %lu gave null export\n", page, page->index); - EXIT; - return; - } + if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT)) + ll_ra_stats_inc(mapping, RA_STAT_EOF); - llap = llap_from_page(page); - if (IS_ERR(llap)) { - CERROR("page %p ind %lu couldn't find llap: %ld\n", page, - page->index, PTR_ERR(llap)); - EXIT; - return; + /* if we didn't get to the end of the region we reserved from + * the ras we need to go back and update the ras so that the + * 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 */ + CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n", + ra_end, end, ria->ria_end); + + if (ra_end != end + 1) { + cfs_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); + } + cfs_spin_unlock(&ras->ras_lock); } - //llap_write_complete(inode, llap); - 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..*/ - page->private = 0; - - spin_lock(&sbi->ll_pglist_lock); - if (!list_empty(&llap->llap_proc_item)) - list_del_init(&llap->llap_proc_item); - sbi->ll_pglist_gen++; - spin_unlock(&sbi->ll_pglist_lock); - OBD_FREE(llap, sizeof(*llap)); - EXIT; + RETURN(ret); } -static int ll_page_matches(struct page *page, int fd_flags) +static void ras_set_start(struct ll_readahead_state *ras, unsigned long index) { - struct lustre_handle match_lockh = {0}; - struct inode *inode = page->mapping->host; - ldlm_policy_data_t page_extent; - int flags, matches; - ENTRY; + ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1)); +} - if (fd_flags & LL_FILE_GROUP_LOCKED) - RETURN(1); - - page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT; - page_extent.l_extent.end = - page_extent.l_extent.start + PAGE_CACHE_SIZE - 1; - flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK; - matches = obd_match(ll_i2sbi(inode)->ll_osc_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) -{ - struct page *page = llap->llap_page; - int rc; - - page_cache_get(page); - llap->llap_defer_uptodate = defer; - rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd, - NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0, - PAGE_SIZE, 0, ASYNC_COUNT_STABLE); - if (rc) { - LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc); - page_cache_release(page); - } - RETURN(rc); +/* 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) +{ + 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); } -#define LL_RA_MIN(inode) ((unsigned long)PTLRPC_MAX_BRW_PAGES / 2) -#define LL_RA_MAX(inode) ((ll_i2info(inode)->lli_smd->lsm_xfersize * 3) >> \ - PAGE_CACHE_SHIFT) +/* 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); +} -static void ll_readahead(struct ll_readahead_state *ras, - struct obd_export *exp, struct address_space *mapping, - struct obd_io_group *oig, int flags) +void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) { - unsigned long i, start, end; - struct ll_async_page *llap; - struct page *page; - int rc; + cfs_spin_lock_init(&ras->ras_lock); + ras_reset(ras, 0); + ras->ras_requests = 0; + CFS_INIT_LIST_HEAD(&ras->ras_read_beads); +} - if (mapping->host->i_size == 0) - return; +/* + * Check whether the read request is in the stride window. + * If it is in the stride window, return 1, otherwise return 0. + */ +static int index_in_stride_window(unsigned long index, + struct ll_readahead_state *ras, + struct inode *inode) +{ + unsigned long stride_gap = index - ras->ras_last_readpage - 1; - spin_lock(&ras->ras_lock); - - /* make sure to issue a window's worth of read-ahead pages */ - end = ras->ras_last; - start = end - ras->ras_window; - if (start > end) - start = 0; - - /* but don't iterate over pages that we've already issued. this - * will set start to end + 1 if we've already read-ahead up to - * ras_last sothe for() won't be entered */ - if (ras->ras_next_index > start) - start = ras->ras_next_index; - if (end != ~0UL) - ras->ras_next_index = end + 1; - - CDEBUG(D_READA, "ni %lu last %lu win %lu: reading from %lu to %lu\n", - ras->ras_next_index, ras->ras_last, ras->ras_window, - start, end); - - spin_unlock(&ras->ras_lock); - - /* clamp to filesize */ - i = (mapping->host->i_size - 1) >> PAGE_CACHE_SHIFT; - end = min(end, i); - - for (i = start; i <= end; i++) { - /* grab_cache_page_nowait returns null if this races with - * truncating the page (page->mapping == NULL) */ - page = grab_cache_page_nowait(mapping, i); - if (page == NULL) - break; - - /* the book-keeping above promises that we've tried - * all the indices from start to end, so we don't - * stop if anyone returns an error. This may not be good. */ - if (Page_Uptodate(page)) - goto next_page; - - if ((rc = ll_page_matches(page, flags)) <= 0) { - LL_CDEBUG_PAGE(D_READA | D_PAGE, page, - "lock match failed: rc %d\n", rc); - goto next_page; - } + if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 || + ras->ras_stride_pages == ras->ras_stride_length) + return 0; - llap = llap_from_page(page); - if (IS_ERR(llap) || llap->llap_defer_uptodate) - goto next_page; + /* If it is contiguous read */ + if (stride_gap == 0) + return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages; - rc = ll_issue_page_read(exp, llap, oig, 1); - if (rc == 0) - LL_CDEBUG_PAGE(D_PAGE, page, "started read-ahead\n"); - if (rc) { - next_page: - LL_CDEBUG_PAGE(D_PAGE, page, "skipping read-ahead\n"); + /*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; +} - unlock_page(page); - } - page_cache_release(page); +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; } + 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; + + RAS_CDEBUG(ras); + return; } -/* called with the ras_lock held or from places where it doesn't matter */ -static void ll_readahead_set(struct inode *inode, - struct ll_readahead_state *ras, - unsigned long index) -{ - ras->ras_next_index = index; - if (ras->ras_next_index != ~0UL) - ras->ras_next_index++; - ras->ras_window = LL_RA_MIN(inode); - ras->ras_last = ras->ras_next_index + ras->ras_window; - if (ras->ras_last < ras->ras_next_index) - ras->ras_last = ~0UL; - CDEBUG(D_READA, "ni %lu last %lu win %lu: set %lu\n", - ras->ras_next_index, ras->ras_last, ras->ras_window, - index); +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); } -void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) +/* 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) { - spin_lock_init(&ras->ras_lock); - ll_readahead_set(inode, ras, 0); -} - -static void ll_readahead_update(struct inode *inode, - struct ll_readahead_state *ras, - unsigned long index, int hit) -{ - unsigned long issued_start, new_last; - - spin_lock(&ras->ras_lock); - - /* we're interested in noticing the index's relation to the - * previously issued read-ahead pages */ - issued_start = ras->ras_next_index - ras->ras_window - 1; - if (issued_start > ras->ras_next_index) - issued_start = 0; - - CDEBUG(D_READA, "ni %lu last %lu win %lu: %s ind %lu start %lu\n", - ras->ras_next_index, ras->ras_last, ras->ras_window, - hit ? "hit" : "miss", index, issued_start); - if (!hit && - index == ras->ras_next_index && index == ras->ras_last + 1) { - /* special case the kernel's read-ahead running into the - * page just beyond our read-ahead window as an extension - * of our read-ahead. sigh. wishing it was easier to - * turn off 2.4's read-ahead. */ - ras->ras_window = min(LL_RA_MAX(inode), ras->ras_window + 1); - if (index != ~0UL) - ras->ras_next_index = index + 1; - ras->ras_last = index; - } else if (!hit && - (index > issued_start || ras->ras_next_index >= index)) { - /* deal with a miss way out of the window. we interpret - * this as a seek and restart the window */ - ll_readahead_set(inode, ras, index); - - } else if (!hit && - issued_start <= index && index < ras->ras_next_index) { - /* a miss inside the window? surely its memory pressure - * evicting our read pages before the app can see them. - * we shrink the window aggressively */ - unsigned long old_window = ras->ras_window; - - ras->ras_window = max(ras->ras_window / 2, LL_RA_MIN(inode)); - ras->ras_last -= old_window - ras->ras_window; - if (ras->ras_next_index > ras->ras_last) - ras->ras_next_index = ras->ras_last + 1; - CDEBUG(D_READA, "ni %lu last %lu win %lu: miss inside\n", - ras->ras_next_index, ras->ras_last, ras->ras_window); - - } else if (hit && - issued_start <= index && index < ras->ras_next_index) { - /* a hit inside the window. grow the window by twice the - * number of pages that are satisified within the window. */ - ras->ras_window = min(LL_RA_MAX(inode), ras->ras_window + 2); - - /* we want the next readahead pass to issue a windows worth - * beyond where the app currently is */ - new_last = index + ras->ras_window; - if (new_last > ras->ras_last) - ras->ras_last = new_last; - - CDEBUG(D_READA, "ni %lu last %lu win %lu: extended window/last\n", - ras->ras_next_index, ras->ras_last, ras->ras_window); - } + unsigned long left, step, window_len; + 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; - spin_unlock(&ras->ras_lock); + 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_per_file) + ras->ras_window_len = window_len; + + RAS_CDEBUG(ras); } -/* - * 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) -{ - struct ll_file_data *fd = filp->private_data; - struct inode *inode = page->mapping->host; - struct obd_export *exp; - struct ll_async_page *llap; - struct obd_io_group *oig = NULL; - int rc; +static void ras_increase_window(struct ll_readahead_state *ras, + struct ll_ra_info *ra, struct inode *inode) +{ + /* 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); + else + ras->ras_window_len = min(ras->ras_window_len + + RAS_INCREASE_STEP, + ra->ra_max_pages_per_file); +} + +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_detect = 0, ra_miss = 0; ENTRY; - LASSERT(PageLocked(page)); - LASSERT(!PageUptodate(page)); - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n", - inode->i_ino, inode->i_generation, inode, - (((obd_off)page->index) << PAGE_SHIFT)); - LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0); - - rc = oig_init(&oig); - if (rc < 0) - GOTO(out, rc); - - exp = ll_i2obdexp(inode); - if (exp == NULL) - GOTO(out, rc = -EINVAL); - - llap = llap_from_page(page); - if (IS_ERR(llap)) - GOTO(out, rc = PTR_ERR(llap)); - - if (llap->llap_defer_uptodate) { - ll_readahead_update(inode, &fd->fd_ras, page->index, 1); - ll_readahead(&fd->fd_ras, exp, page->mapping, oig,fd->fd_flags); - 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); + cfs_spin_lock(&ras->ras_lock); + + 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 + * read-ahead miss that we think we've previously issued. This can + * be a symptom of there being so many read-ahead pages that the VM is + * reclaiming it before we get to it. */ + if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) { + 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)) { + ra_miss = 1; + ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW); } - ll_readahead_update(inode, &fd->fd_ras, page->index, 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); + /* 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_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) { + __u64 kms_pages; + + kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >> + 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_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_per_file, + ra->ra_max_read_ahead_whole_pages); + GOTO(out_unlock, 0); + } } - - if (rc == 0) { - static unsigned long next_print; - CDEBUG(D_INODE, "ino %lu page %lu (%llu) didn't match a lock\n", - inode->i_ino, page->index, - (long long)page->index << PAGE_CACHE_SHIFT); - if (time_after(jiffies, next_print)) { - CERROR("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); - ldlm_dump_all_namespaces(); - if (next_print == 0) { - CERROR("%s\n", portals_debug_dumpstack()); - portals_debug_dumplog(); + if (zero) { + /* check whether it is in stride I/O mode*/ + if (!index_in_stride_window(index, ras, inode)) { + 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(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(index, ras, inode) && + 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(ras, index); + RAS_CDEBUG(ras); + } else { + /* Reset both stride window and normal RA + * window */ + ras_reset(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(index, ras, inode)) { + /* Shrink stride read-ahead window to be zero */ + ras_stride_reset(ras); + ras->ras_window_len = 0; + ras->ras_next_readahead = index; } - next_print = jiffies + 30 * HZ; } } + ras->ras_consecutive_pages++; + ras->ras_last_readpage = index; + ras_set_start(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 1M 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; + GOTO(out_unlock, 0); + } - rc = ll_issue_page_read(exp, llap, oig, 0); - if (rc) - GOTO(out, rc); - - LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n"); - if ((ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)) - ll_readahead(&fd->fd_ras, exp, page->mapping, oig,fd->fd_flags); - - rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig); + /* 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; + } -out: - if (rc) - unlock_page(page); -out_oig: - if (oig != NULL) - oig_release(oig); - RETURN(rc); + /* 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(ras, ra, inode); + EXIT; +out_unlock: + RAS_CDEBUG(ras); + ras->ras_request_index++; + cfs_spin_unlock(&ras->ras_lock); + return; } -#if 0 -/* this is for read pages. we issue them as ready but not urgent. when - * someone waits on them we fire them off, hopefully merged with adjacent - * reads that were queued by read-ahead. */ -int ll_sync_page(struct page *page) +int ll_writepage(struct page *vmpage, struct writeback_control *unused) { - struct obd_export *exp; - struct ll_async_page *llap; - int rc; + struct inode *inode = vmpage->mapping->host; + struct lu_env *env; + struct cl_io *io; + struct cl_page *page; + struct cl_object *clob; + struct cl_2queue *queue; + struct cl_env_nest nest; + int result; ENTRY; - /* we're using a low bit flag to signify that a queued read should - * be issued once someone goes to lock it. it is also cleared - * as the page is built into an RPC */ - if (!test_and_clear_bit(LL_PRIVBITS_READ, &page->private)) - RETURN(0); + LASSERT(PageLocked(vmpage)); + LASSERT(!PageWriteback(vmpage)); - /* careful to only deref page->mapping after checking the bit */ - exp = ll_i2obdexp(page->mapping->host); - if (exp == NULL) + if (ll_i2dtexp(inode) == NULL) RETURN(-EINVAL); - llap = llap_from_page(page); - if (IS_ERR(llap)) - RETURN(PTR_ERR(llap)); + env = cl_env_nested_get(&nest); + if (IS_ERR(env)) + RETURN(PTR_ERR(env)); + + io = &ccc_env_info(env)->cti_io; + queue = &vvp_env_info(env)->vti_queue; + clob = ll_i2info(inode)->lli_clob; + LASSERT(clob != NULL); + + io->ci_obj = clob; + 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); + /* + * Mark page dirty, because this is what + * ->vio_submit()->cpo_prep_write() assumes. + * + * XXX better solution is to detect this from within + * cl_io_submit_rw() somehow. + */ + set_page_dirty(vmpage); + cl_2queue_init_page(queue, page); + result = cl_io_submit_rw(env, io, CRT_WRITE, + queue, CRP_NORMAL); + cl_page_list_disown(env, io, &queue->c2_qin); + if (result != 0) { + /* + * There is no need to clear PG_writeback, as + * cl_io_submit_rw() calls completion callback + * on failure. + */ + /* + * 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)) + set_page_dirty(vmpage); + } + LASSERT(!cl_page_is_owned(page, io)); + lu_ref_del(&page->cp_reference, + "writepage", cfs_current()); + cl_page_put(env, page); + cl_2queue_fini(env, queue); + } + } + cl_io_fini(env, io); + cl_env_nested_put(&nest, env); + RETURN(result); +} - LL_CDEBUG_PAGE(D_PAGE, page, "setting ready|urgent\n"); +int ll_readpage(struct file *file, struct page *vmpage) +{ + struct ll_cl_context *lcc; + int result; + ENTRY; - rc = obd_set_async_flags(exp, ll_i2info(page->mapping->host)->lli_smd, - NULL, llap->llap_cookie, - ASYNC_READY|ASYNC_URGENT); - return rc; + 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; + + 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. */ + LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd); + unlock_page(vmpage); + result = 0; + } + ll_cl_fini(lcc); + } else { + result = PTR_ERR(lcc); + } + RETURN(result); } -#endif +