X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fllite%2Frw.c;h=355d07e54c55d75648ac4c01780fc69f8551a18f;hb=82d485c9b47a2eb15784ccbfca9d656aff81f518;hp=163c2a8e4dd3d6df2dc9cc8b2e9186fe4108a79a;hpb=f5fb215365791507ec98796ea624abeafa31901c;p=fs%2Flustre-release.git diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c index 163c2a8..355d07e 100644 --- a/lustre/llite/rw.c +++ b/lustre/llite/rw.c @@ -27,7 +27,7 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -44,8 +44,7 @@ #include #include #include -#include -#include +#include #include #include @@ -63,28 +62,11 @@ #include "llite_internal.h" #include -/* 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) -{ - ENTRY; - - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino, - inode->i_generation, inode, i_size_read(inode)); - - EXIT; - return; -} /* ll_truncate */ - /** * Finalizes cl-data before exiting typical address_space operation. Dual to * ll_cl_init(). */ -static void ll_cl_fini(struct ll_cl_context *lcc) +void ll_cl_fini(struct ll_cl_context *lcc) { struct lu_env *env = lcc->lcc_env; struct cl_io *io = lcc->lcc_io; @@ -98,12 +80,6 @@ static void ll_cl_fini(struct ll_cl_context *lcc) cl_page_put(env, page); } - 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); } @@ -111,8 +87,7 @@ static void ll_cl_fini(struct ll_cl_context *lcc) * 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 *ll_cl_init(struct file *file, struct page *vmpage) { struct ll_cl_context *lcc; struct lu_env *env; @@ -123,8 +98,8 @@ static struct ll_cl_context *ll_cl_init(struct file *file, int refcheck; int result = 0; - clob = ll_i2info(vmpage->mapping->host)->lli_clob; - LASSERT(clob != NULL); + clob = ll_i2info(file->f_dentry->d_inode)->lli_clob; + LASSERT(clob != NULL); env = cl_env_get(&refcheck); if (IS_ERR(env)) @@ -138,63 +113,10 @@ static struct ll_cl_context *ll_cl_init(struct file *file, 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 (TRYLOCK_INODE_MUTEX(inode)) { - UNLOCK_INODE_MUTEX(inode); - - /* 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); - } - } else - result = io->ci_result; - lcc->lcc_created = 1; - } - lcc->lcc_io = io; - if (io == NULL) - result = -EIO; - if (result == 0) { + if (io == NULL) + result = -EIO; + if (result == 0 && vmpage != NULL) { struct cl_page *page; LASSERT(io != NULL); @@ -214,110 +136,9 @@ static struct ll_cl_context *ll_cl_init(struct file *file, lcc = ERR_PTR(result); } - CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n", - vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result, - env, io); - return lcc; -} - -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; } -/** - * ->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 ll_cl_context *lcc; - int result; - ENTRY; - - 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); -} - -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 = 0; - ENTRY; - - lcc = ll_cl_get(); - env = lcc->lcc_env; - page = lcc->lcc_page; - io = lcc->lcc_io; - - 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); - - /* - * Release reference acquired by ll_prepare_write(). - */ - lu_ref_del(&page->cp_reference, "prepare_write", cfs_current()); - cl_page_put(env, page); - ll_cl_fini(lcc); - RETURN(result); -} - struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt) { __u64 opc; @@ -348,8 +169,8 @@ static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which); * 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 ra_io_arg *ria, + unsigned long pages, unsigned long min) { struct ll_ra_info *ra = &sbi->ll_ra_info; long ret; @@ -358,7 +179,7 @@ static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, /* 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); + ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), pages); if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages)) GOTO(out, ret = 0); @@ -379,19 +200,24 @@ static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, ret -= beyond_rpc; } - if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) { - cfs_atomic_sub(ret, &ra->ra_cur_pages); - ret = 0; - } + if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) { + atomic_sub(ret, &ra->ra_cur_pages); + ret = 0; + } out: - RETURN(ret); + if (ret < min) { + /* override ra limit for maximum performance */ + atomic_add(min - ret, &ra->ra_cur_pages); + ret = min; + } + RETURN(ret); } 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_ra_info *ra = &sbi->ll_ra_info; + atomic_sub(len, &ra->ra_cur_pages); } static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which) @@ -400,10 +226,10 @@ static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which) lprocfs_counter_incr(sbi->ll_ra_stats, which); } -void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which) +void ll_ra_stats_inc(struct inode *inode, enum ra_stat which) { - struct ll_sb_info *sbi = ll_i2sbi(mapping->host); - ll_ra_stats_inc_sbi(sbi, which); + struct ll_sb_info *sbi = ll_i2sbi(inode); + ll_ra_stats_inc_sbi(sbi, which); } #define RAS_CDEBUG(ras) \ @@ -440,36 +266,36 @@ 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); - cfs_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; - cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads); - cfs_spin_unlock(&ras->ras_lock); + 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); - cfs_spin_lock(&ras->ras_lock); - cfs_list_del_init(&rar->lrr_linkage); - cfs_spin_unlock(&ras->ras_lock); + spin_lock(&ras->ras_lock); + 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; - cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) { + list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) { if (scan->lrr_reader == current) return scan; } @@ -478,47 +304,55 @@ 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); + ras = ll_ras_get(f); - cfs_spin_lock(&ras->ras_lock); - bead = ll_ra_read_get_locked(ras); - cfs_spin_unlock(&ras->ras_lock); - return bead; + 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 cl_page_list *queue, struct cl_page *page, + struct cl_object *clob, pgoff_t *max_index) { - struct ccc_page *cp; - int rc; + struct page *vmpage = page->cp_vmpage; + struct ccc_page *cp; + int rc; - ENTRY; + ENTRY; - 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); + rc = 0; + cl_page_assume(env, io, page); + lu_ref_add(&page->cp_reference, "ra", current); + cp = cl2ccc_page(cl_object_page_slice(clob, page)); + if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) { + CDEBUG(D_READA, "page index %lu, max_index: %lu\n", + ccc_index(cp), *max_index); + /* Disable the optimization on prefetching maximum readahead + * index because there is a race with lock cancellation. This + * optimization will be revived later. + * if (*max_index == 0 || ccc_index(cp) > *max_index) */ + rc = cl_page_is_under_lock(env, io, page, max_index); + if (rc == 0) { + cp->cpg_defer_uptodate = 1; + cp->cpg_ra_used = 0; + cl_page_list_add(queue, page); + rc = 1; + } else { + cl_page_discard(env, io, page); + rc = -ENOLCK; + } + } else { + /* skip completed pages */ + cl_page_unassume(env, io, page); + } + lu_ref_del(&page->cp_reference, "ra", current); + cl_page_put(env, page); + RETURN(rc); } /** @@ -532,11 +366,12 @@ static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io, * \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 cl_page_list *queue, + pgoff_t index, pgoff_t *max_index) { + struct cl_object *clob = io->ci_obj; + struct inode *inode = ccc_object_inode(clob); 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; @@ -549,15 +384,15 @@ static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io, #ifdef __GFP_NOWARN gfp_mask |= __GFP_NOWARN; #endif - 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); + vmpage = grab_cache_page_nowait(inode->i_mapping, index); + if (vmpage != NULL) { + /* Check if vmpage was truncated or reclaimed */ + if (vmpage->mapping == inode->i_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, clob, max_index); if (rc == -ENOLCK) { which = RA_STAT_FAILED_MATCH; msg = "lock match failed"; @@ -577,11 +412,11 @@ static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io, 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); - } - RETURN(rc); + if (msg != NULL) { + ll_ra_stats_inc(inode, which); + CDEBUG(D_READA, "%s\n", msg); + } + RETURN(rc); } #define RIA_DEBUG(ria) \ @@ -589,7 +424,17 @@ static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io, ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\ ria->ria_pages) -#define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES +/* 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 - PAGE_CACHE_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 >> PAGE_CACHE_SHIFT) static inline int stride_io_mode(struct ll_readahead_state *ras) { @@ -676,25 +521,26 @@ 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); + int rc, count = 0; + bool stride_ria; + pgoff_t page_idx; + pgoff_t max_index = 0; + + 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, &max_index); if (rc == 1) { (*reserved_pages)--; - count ++; + count++; } else if (rc == -ENOLCK) break; } else if (stride_ria) { @@ -724,42 +570,40 @@ static int ll_read_ahead_pages(const struct lu_env *env, } 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 cl_page_list *queue, struct ll_readahead_state *ras, + bool hit) { - 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; + 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, mlen = 0; + struct inode *inode; + struct ll_ra_read *bead; + struct ra_io_arg *ria = &vti->vti_ria; + struct cl_object *clob; + int ret = 0; + __u64 kms; + ENTRY; - inode = mapping->host; - lli = ll_i2info(inode); - clob = lli->lli_clob; + clob = io->ci_obj; + inode = ccc_object_inode(clob); - memset(ria, 0, sizeof *ria); + memset(ria, 0, sizeof *ria); - cl_object_attr_lock(clob); - ret = cl_object_attr_get(env, clob, attr); - cl_object_attr_unlock(clob); + 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); - } + if (ret != 0) + RETURN(ret); + kms = attr->cat_kms; + if (kms == 0) { + ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN); + RETURN(0); + } - cfs_spin_lock(&ras->ras_lock); + spin_lock(&ras->ras_lock); if (vio->cui_ra_window_set) bead = &vio->cui_bead; else @@ -797,7 +641,7 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, end = rpc_boundary; /* Truncate RA window to end of file */ - end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT)); + end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT)); ras->ras_next_readahead = max(end, end + 1); RAS_CDEBUG(ras); @@ -810,72 +654,93 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, ria->ria_length = ras->ras_stride_length; ria->ria_pages = ras->ras_stride_pages; } - cfs_spin_unlock(&ras->ras_lock); + spin_unlock(&ras->ras_lock); - if (end == 0) { - ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW); - RETURN(0); - } - len = ria_page_count(ria); - if (len == 0) - RETURN(0); - - reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len); - if (reserved < len) - ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT); - - 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) - ll_ra_count_put(ll_i2sbi(inode), reserved); - - if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT)) - ll_ra_stats_inc(mapping, RA_STAT_EOF); - - /* if we didn't get to the end of the region we reserved from - * 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); - } + if (end == 0) { + ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW); + RETURN(0); + } + len = ria_page_count(ria); + if (len == 0) { + ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW); + RETURN(0); + } - RETURN(ret); + CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n", + PFID(lu_object_fid(&clob->co_lu)), + ria->ria_start, ria->ria_end, + bead == NULL ? 0 : bead->lrr_start, + bead == NULL ? 0 : bead->lrr_count, + hit); + + /* at least to extend the readahead window to cover current read */ + if (!hit && bead != NULL && + bead->lrr_start + bead->lrr_count > ria->ria_start) { + /* to the end of current read window. */ + mlen = bead->lrr_start + bead->lrr_count - ria->ria_start; + /* trim to RPC boundary */ + start = ria->ria_start & (PTLRPC_MAX_BRW_PAGES - 1); + mlen = min(mlen, PTLRPC_MAX_BRW_PAGES - start); + } + + reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len, mlen); + if (reserved < len) + ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT); + + CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n", + reserved, len, mlen, + 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, &ra_end); + + if (reserved != 0) + ll_ra_count_put(ll_i2sbi(inode), reserved); + + if (ra_end == end + 1 && ra_end == (kms >> PAGE_CACHE_SHIFT)) + ll_ra_stats_inc(inode, RA_STAT_EOF); + + /* 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) { + ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END); + 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); } -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 */ @@ -889,33 +754,34 @@ static void ras_stride_reset(struct ll_readahead_state *ras) void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) { - cfs_spin_lock_init(&ras->ras_lock); - ras_reset(ras, 0); - ras->ras_requests = 0; - CFS_INIT_LIST_HEAD(&ras->ras_read_beads); + spin_lock_init(&ras->ras_lock); + ras_reset(inode, ras, 0); + ras->ras_requests = 0; + INIT_LIST_HEAD(&ras->ras_read_beads); } /* * Check whether the read request is in the stride window. * If it is in the stride window, return 1, otherwise return 0. */ -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) { - unsigned long stride_gap = index - ras->ras_last_readpage - 1; + 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; - 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; + /* 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; + /* 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, @@ -991,30 +857,31 @@ static void ras_stride_increase_window(struct ll_readahead_state *ras, RAS_CDEBUG(ras); } -static void ras_increase_window(struct ll_readahead_state *ras, - struct ll_ra_info *ra, struct inode *inode) +static void ras_increase_window(struct inode *inode, + struct ll_readahead_state *ras, + struct ll_ra_info *ra) { - /* 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); + /* 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); } void ras_update(struct ll_sb_info *sbi, struct inode *inode, - struct ll_readahead_state *ras, unsigned long index, - unsigned hit) + 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; + struct ll_ra_info *ra = &sbi->ll_ra_info; + int zero = 0, stride_detect = 0, ra_miss = 0; + ENTRY; - cfs_spin_lock(&ras->ras_lock); + spin_lock(&ras->ras_lock); ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS); @@ -1043,8 +910,8 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, 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; + kms_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> + PAGE_CACHE_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); @@ -1059,125 +926,128 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, GOTO(out_unlock, 0); } } - 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; - } - } - } - 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); + 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 { + if (ras->ras_next_readahead < ras->ras_window_start) + ras->ras_next_readahead = ras->ras_window_start; + if (!hit) + ras->ras_next_readahead = index + 1; + } + 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); - } + /* 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); + } - /* 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; - } + /* 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); + } - /* 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; + /* 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++; - cfs_spin_unlock(&ras->ras_lock); - return; + RAS_CDEBUG(ras); + ras->ras_request_index++; + spin_unlock(&ras->ras_lock); + return; } int ll_writepage(struct page *vmpage, struct writeback_control *wbc) { - struct inode *inode = vmpage->mapping->host; + 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; - int redirtied = 0; + bool redirtied = false; + bool unlocked = false; int result; ENTRY; LASSERT(PageLocked(vmpage)); LASSERT(!PageWriteback(vmpage)); - if (ll_i2dtexp(inode) == NULL) - RETURN(-EINVAL); + LASSERT(ll_i2dtexp(inode) != NULL); - env = cl_env_nested_get(&nest); - if (IS_ERR(env)) - RETURN(PTR_ERR(env)); + 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); @@ -1189,10 +1059,10 @@ int ll_writepage(struct page *vmpage, struct writeback_control *wbc) 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); + if (!IS_ERR(page)) { + lu_ref_add(&page->cp_reference, "writepage", + current); + cl_page_assume(env, io, page); result = cl_page_flush(env, io, page); if (result != 0) { /* @@ -1203,14 +1073,17 @@ int ll_writepage(struct page *vmpage, struct writeback_control *wbc) if (!PageError(vmpage)) { redirty_page_for_writepage(wbc, vmpage); result = 0; - redirtied = 1; + redirtied = true; } } cl_page_disown(env, io, page); - lu_ref_del(&page->cp_reference, - "writepage", cfs_current()); - cl_page_put(env, page); - } + unlocked = true; + lu_ref_del(&page->cp_reference, + "writepage", current); + cl_page_put(env, page); + } else { + result = PTR_ERR(page); + } } cl_io_fini(env, io); @@ -1222,8 +1095,8 @@ int ll_writepage(struct page *vmpage, struct writeback_control *wbc) * 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); + offset + PAGE_CACHE_SIZE - 1, + CL_FSYNC_LOCAL, 1); if (result > 0) { /* actually we may have written more than one page. * decreasing this page because the caller will count @@ -1234,21 +1107,33 @@ int ll_writepage(struct page *vmpage, struct writeback_control *wbc) } cl_env_nested_put(&nest, env); - RETURN(result); + GOTO(out, result); + +out: + if (result < 0) { + if (!lli->lli_async_rc) + lli->lli_async_rc = result; + SetPageError(vmpage); + if (!unlocked) + unlock_page(vmpage); + } + return result; } int ll_writepages(struct address_space *mapping, struct writeback_control *wbc) { struct inode *inode = mapping->host; + struct ll_sb_info *sbi = ll_i2sbi(inode); loff_t start; loff_t end; enum cl_fsync_mode mode; int range_whole = 0; int result; + int ignore_layout = 0; ENTRY; if (wbc->range_cyclic) { - start = mapping->writeback_index << CFS_PAGE_SHIFT; + start = mapping->writeback_index << PAGE_CACHE_SHIFT; end = OBD_OBJECT_EOF; } else { start = wbc->range_start; @@ -1263,7 +1148,16 @@ int ll_writepages(struct address_space *mapping, struct writeback_control *wbc) if (wbc->sync_mode == WB_SYNC_ALL) mode = CL_FSYNC_LOCAL; - result = cl_sync_file_range(inode, start, end, mode); + if (sbi->ll_umounting) + /* if the mountpoint is being umounted, all pages have to be + * evicted to avoid hitting LBUG when truncate_inode_pages() + * is called later on. */ + ignore_layout = 1; + + if (cl_i2info(inode)->lli_clob == NULL) + RETURN(0); + + result = cl_sync_file_range(inode, start, end, mode, ignore_layout); if (result > 0) { wbc->nr_to_write -= result; result = 0; @@ -1271,8 +1165,9 @@ int ll_writepages(struct address_space *mapping, struct writeback_control *wbc) 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; + mapping->writeback_index = 0; + else + mapping->writeback_index = (end >> PAGE_CACHE_SHIFT) +1; } RETURN(result); } @@ -1283,7 +1178,7 @@ int ll_readpage(struct file *file, struct page *vmpage) int result; ENTRY; - lcc = ll_cl_init(file, vmpage, 0); + lcc = ll_cl_init(file, vmpage); if (!IS_ERR(lcc)) { struct lu_env *env = lcc->lcc_env; struct cl_io *io = lcc->lcc_io; @@ -1295,7 +1190,6 @@ int ll_readpage(struct file *file, struct page *vmpage) result = cl_io_read_page(env, io, page); } else { /* Page from a non-object file. */ - LASSERT(!ll_i2info(vmpage->mapping->host)->lli_has_smd); unlock_page(vmpage); result = 0; } @@ -1307,3 +1201,27 @@ int ll_readpage(struct file *file, struct page *vmpage) RETURN(result); } +int ll_page_sync_io(const struct lu_env *env, struct cl_io *io, + struct cl_page *page, enum cl_req_type crt) +{ + struct cl_2queue *queue; + int result; + + LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); + + queue = &io->ci_queue; + cl_2queue_init_page(queue, page); + + result = cl_io_submit_sync(env, io, crt, queue, 0); + LASSERT(cl_page_is_owned(page, io)); + + if (crt == CRT_READ) + /* + * in CRT_WRITE case page is left locked even in case of + * error. + */ + cl_page_list_disown(env, io, &queue->c2_qin); + cl_2queue_fini(env, queue); + + return result; +}