X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fllite%2Frw.c;h=2243aab4c1a3623f8ecf85464911ecaebe5119ac;hb=c5361360e51de22a59d4427327bddf9fd398f352;hp=d3f59a69a8f26d66caaaa0c9feb28321c91321ec;hpb=b918d20c1073fbb29de95dccec018b1b992cf81d;p=fs%2Flustre-release.git diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c index d3f59a6..2243aab 100644 --- a/lustre/llite/rw.c +++ b/lustre/llite/rw.c @@ -26,7 +26,7 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. */ /* @@ -38,7 +38,6 @@ * Lustre Lite I/O page cache routines shared by different kernel revs */ -#include #include #include #include @@ -61,7 +60,6 @@ #define DEBUG_SUBSYSTEM S_LLITE -//#include #include #include #include "llite_internal.h" @@ -76,186 +74,165 @@ void ll_truncate(struct inode *inode) { struct ll_inode_info *lli = ll_i2info(inode); - loff_t new_size; ENTRY; - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino, - inode->i_generation, inode, i_size_read(inode), - i_size_read(inode)); - ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1); - if (lli->lli_size_sem_owner != cfs_current()) { - EXIT; - return; - } + CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino, + inode->i_generation, inode, i_size_read(inode)); - if (!lli->lli_smd) { - CDEBUG(D_INODE, "truncate on inode %lu with no objects\n", - inode->i_ino); - GOTO(out_unlock, 0); - } - LASSERT_SEM_LOCKED(&lli->lli_size_sem); - - if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) && - (i_size_read(inode) & ~CFS_PAGE_MASK))) { - /* If the truncate leaves a partial page, update its checksum */ - struct page *page = find_get_page(inode->i_mapping, - i_size_read(inode) >> - CFS_PAGE_SHIFT); - if (page != NULL) { -#if 0 /* XXX */ - struct ll_async_page *llap = llap_cast_private(page); - if (llap != NULL) { - char *kaddr = kmap_atomic(page, KM_USER0); - llap->llap_checksum = - init_checksum(OSC_DEFAULT_CKSUM); - llap->llap_checksum = - compute_checksum(llap->llap_checksum, - kaddr, CFS_PAGE_SIZE, - OSC_DEFAULT_CKSUM); - kunmap_atomic(kaddr, KM_USER0); - } - page_cache_release(page); -#endif - } + 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); } - new_size = i_size_read(inode); - ll_inode_size_unlock(inode, 0); - EXIT; return; - - out_unlock: - ll_inode_size_unlock(inode, 0); } /* 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) +{ + struct lu_env *env = lcc->lcc_env; + struct cl_io *io = lcc->lcc_io; + struct cl_page *page = lcc->lcc_page; + + 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); + } + + 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 int ll_cl_init(struct file *file, struct page *vmpage, - struct lu_env **env, - struct cl_io **io, struct cl_page **page, int *refcheck) +static struct ll_cl_context *ll_cl_init(struct file *file, + struct page *vmpage, int create) { - struct lu_env *_env; - struct cl_io *_io; - struct cl_page *_page; + struct ll_cl_context *lcc; + struct lu_env *env; + struct cl_io *io; struct cl_object *clob; + struct ccc_io *cio; - int result; - - *env = NULL; - *io = NULL; - *page = NULL; + 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)) { - struct ccc_io *cio = ccc_env_io(_env); - - *env = _env; - *io = _io = cio->cui_cl.cis_io; - if (_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)) { - *page = _page; - lu_ref_add(&_page->cp_reference, "cl_io", _io); - result = 0; - } else - result = PTR_ERR(_page); - } else - /* - * This is for a case where operation can be called - * either with or without cl_io created by the upper - * layer (e.g., ->prepare_write() called directly from - * loop-back driver). - */ - result = -EALREADY; - } else - result = PTR_ERR(_env); - CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %i %p %p %p\n", - vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result, - *env, *io, *page); - return result; -} + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + return ERR_PTR(PTR_ERR(env)); -/** - * Finalizes cl-data before exiting typical address_space operation. Dual to - * ll_cl_init(). - */ -static void ll_cl_fini(struct lu_env *env, - struct cl_io *io, struct cl_page *page, int *refcheck) -{ - if (page != NULL) { - lu_ref_del(&page->cp_reference, "cl_io", io); - cl_page_put(env, page); - } - if (env != NULL) { + 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_thread_io(env); vio = vvp_env_io(env); - LASSERT(vio->cui_oneshot >= 0); - if (vio->cui_oneshot > 0) { - if (--vio->cui_oneshot == 0) { - cl_io_end(env, io); - cl_io_unlock(env, io); - cl_io_iter_fini(env, io); - cl_io_fini(env, io); - /* to trigger assertion above, if ll_cl_fini() - * is called against freed io. */ - vio->cui_oneshot = -1; + 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); } - /* additional reference on env was acquired by io, - * disable refcheck */ - refcheck = NULL; - } - cl_env_put(env, refcheck); - } else - LASSERT(io == NULL); + } else + result = io->ci_result; + lcc->lcc_created = 1; + } + + 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); + } + + CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n", + vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result, + env, io); + return lcc; } -/** - * Initializes one-shot cl_io for the case when loop driver calls - * ->{prepare,commit}_write() methods directly. - */ -static int ll_prepare_loop(struct lu_env *env, struct cl_io *io, - struct file *file, struct page *vmpage, - unsigned from, unsigned to) +static struct ll_cl_context *ll_cl_get(void) { - struct vvp_io *vio; - struct ccc_io *cio; - int result; - loff_t pos; + struct ll_cl_context *lcc; + struct lu_env *env; + int refcheck; - vio = vvp_env_io(env); - cio = ccc_env_io(env); - ll_io_init(io, file, 1); - pos = (vmpage->index << CFS_PAGE_SHIFT) + from; - /* - * Create IO and quickly drive it through CIS_{INIT,IT_STARTED,LOCKED} - * states. DLM locks are not taken for vio->cui_oneshot IO---we cannot - * take DLM locks here, because page is already locked. With new - * ->write_{being,end}() address_space operations lustre might be - * luckier. - */ - result = cl_io_rw_init(env, io, CIT_WRITE, pos, from - to); - if (result == 0) { - cio->cui_fd = LUSTRE_FPRIVATE(file); - vio->cui_oneshot = 1; - 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; - return result; + 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; } /** @@ -265,34 +242,30 @@ static int ll_prepare_loop(struct lu_env *env, struct cl_io *io, int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from, unsigned to) { - struct lu_env *env; - struct cl_io *io; - struct cl_page *page; + struct ll_cl_context *lcc; int result; - int refcheck; ENTRY; - result = ll_cl_init(file, vmpage, &env, &io, &page, &refcheck); - /* - * Loop-back driver calls ->prepare_write() and ->sendfile() methods - * directly, bypassing file system ->write() operation, so cl_io has - * to be created here. - */ - if (result == -EALREADY) { - io = &ccc_env_info(env)->cti_io; - result = ll_prepare_loop(env, io, file, vmpage, from, to); - if (result == 0) { - result = ll_cl_init(file, vmpage, - &env, &io, &page, &refcheck); - cl_env_put(env, NULL); - } - } - if (result == 0) { + 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) { - struct vvp_io *vio; - /* * Add a reference, so that page is not evicted from * the cache until ->commit_write() is called. @@ -300,40 +273,46 @@ int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from, cl_page_get(page); lu_ref_add(&page->cp_reference, "prepare_write", cfs_current()); - vio = vvp_env_io(env); - if (vio->cui_oneshot > 0) - vio->cui_oneshot++; - } else + } 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); } - ll_cl_fini(env, io, page, &refcheck); 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; - int refcheck; + int result = 0; ENTRY; - result = ll_cl_init(file, vmpage, &env, &io, &page, &refcheck); - LASSERT(result != -EALREADY); - if (result == 0) { - LASSERT(cl_page_is_owned(page, io)); + 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 cl_io_prepare_write(). - */ - lu_ref_del(&page->cp_reference, "prepare_write", cfs_current()); - cl_page_put(env, page); - } - ll_cl_fini(env, io, page, &refcheck); + 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); } @@ -357,20 +336,35 @@ static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which); * 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) +static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, struct ra_io_arg *ria, + unsigned long len) { struct ll_ra_info *ra = &sbi->ll_ra_info; unsigned long ret; ENTRY; - ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), len); - if ((int)ret < 0) + /** + * If read-ahead pages left are less than 1M, do not do read-ahead, + * otherwise it will form small read RPC(< 1M), which hurt server + * performance a lot. + */ + if (ra->ra_max_pages < atomic_read(&ra->ra_cur_pages)) GOTO(out, ret = 0); - if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) { - atomic_sub(ret, &ra->ra_cur_pages); + ret = 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 (ria->ria_pages == 0) + /* it needs 1M align again after trimed by ra_max_pages*/ + if (ret >= ((ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES)) + ret -= (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES; + + 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); } @@ -378,7 +372,7 @@ out: void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len) { struct ll_ra_info *ra = &sbi->ll_ra_info; - atomic_sub(len, &ra->ra_cur_pages); + cfs_atomic_sub(len, &ra->ra_cur_pages); } static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which) @@ -396,11 +390,11 @@ void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which) #define RAS_CDEBUG(ras) \ CDEBUG(D_READA, \ "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \ - "csr %lu sf %lu sp %lu sl %lu \n", \ + "csr %lu sf %lu sp %lu sl %lu \n", \ ras->ras_last_readpage, ras->ras_consecutive_requests, \ ras->ras_consecutive_pages, ras->ras_window_start, \ ras->ras_window_len, ras->ras_next_readahead, \ - ras->ras_requests, ras->ras_request_index, \ + ras->ras_requests, ras->ras_request_index, \ ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \ ras->ras_stride_pages, ras->ras_stride_length) @@ -431,14 +425,14 @@ void ll_ra_read_in(struct file *f, struct ll_ra_read *rar) ras = ll_ras_get(f); - spin_lock(&ras->ras_lock); + cfs_spin_lock(&ras->ras_lock); ras->ras_requests++; ras->ras_request_index = 0; ras->ras_consecutive_requests++; rar->lrr_reader = current; - list_add(&rar->lrr_linkage, &ras->ras_read_beads); - spin_unlock(&ras->ras_lock); + cfs_list_add(&rar->lrr_linkage, &ras->ras_read_beads); + cfs_spin_unlock(&ras->ras_lock); } void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar) @@ -447,16 +441,16 @@ void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar) ras = ll_ras_get(f); - spin_lock(&ras->ras_lock); - list_del_init(&rar->lrr_linkage); - spin_unlock(&ras->ras_lock); + cfs_spin_lock(&ras->ras_lock); + cfs_list_del_init(&rar->lrr_linkage); + cfs_spin_unlock(&ras->ras_lock); } static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras) { struct ll_ra_read *scan; - list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) { + cfs_list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) { if (scan->lrr_reader == current) return scan; } @@ -470,9 +464,9 @@ struct ll_ra_read *ll_ra_read_get(struct file *f) ras = ll_ras_get(f); - spin_lock(&ras->ras_lock); + cfs_spin_lock(&ras->ras_lock); bead = ll_ra_read_get_locked(ras); - spin_unlock(&ras->ras_lock); + cfs_spin_unlock(&ras->ras_lock); return bead; } @@ -520,7 +514,7 @@ static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io, */ static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io, struct cl_page_list *queue, - int index, struct address_space *mapping) + pgoff_t index, struct address_space *mapping) { struct page *vmpage; struct cl_object *clob = ll_i2info(mapping->host)->lli_clob; @@ -576,38 +570,62 @@ 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 (1024 * 1024 >> CFS_PAGE_SHIFT) +#define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES static inline int stride_io_mode(struct ll_readahead_state *ras) { return ras->ras_consecutive_stride_requests > 1; } - /* The function calculates how much pages will be read in - * [off, off + length], which will be read by stride I/O mode, + * [off, off + length], in such stride IO area, * stride_offset = st_off, stride_lengh = st_len, * stride_pages = st_pgs + * + * |------------------|*****|------------------|*****|------------|*****|.... + * st_off + * |--- st_pgs ---| + * |----- st_len -----| + * + * How many pages it should read in such pattern + * |-------------------------------------------------------------| + * off + * |<------ length ------->| + * + * = |<----->| + |-------------------------------------| + |---| + * start_left st_pgs * i end_left */ static unsigned long stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs, - unsigned long off, unsigned length) + unsigned long off, unsigned long length) { - unsigned long cont_len = st_off > off ? st_off - off : 0; - __u64 stride_len = length + off > st_off ? - length + off + 1 - st_off : 0; - unsigned long left, pg_count; + __u64 start = off > st_off ? off - st_off : 0; + __u64 end = off + length > st_off ? off + length - st_off : 0; + unsigned long start_left = 0; + unsigned long end_left = 0; + unsigned long pg_count; - if (st_len == 0 || length == 0) + if (st_len == 0 || length == 0 || end == 0) return length; - left = do_div(stride_len, st_len); - left = min(left, st_pgs); + start_left = do_div(start, st_len); + if (start_left < st_pgs) + start_left = st_pgs - start_left; + else + start_left = 0; - pg_count = left + stride_len * st_pgs + cont_len; + end_left = do_div(end, st_len); + if (end_left > st_pgs) + end_left = st_pgs; - LASSERT(pg_count >= left); + 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 %u" + 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; @@ -631,7 +649,8 @@ static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria) * For stride I/O mode, just check whether the idx is inside * the ria_pages. */ return ria->ria_length == 0 || ria->ria_length == ria->ria_pages || - (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages; + (idx >= ria->ria_stoff && (idx - ria->ria_stoff) % + ria->ria_length < ria->ria_pages); } static int ll_read_ahead_pages(const struct lu_env *env, @@ -667,10 +686,10 @@ static int ll_read_ahead_pages(const struct lu_env *env, /* FIXME: This assertion only is valid when it is for * forward read-ahead, it will be fixed when backward * read-ahead is implemented */ - LASSERTF(page_idx > ria->ria_stoff, "since %lu in the" - " gap of ra window,it should bigger than stride" - " offset %lu \n", page_idx, ria->ria_stoff); - + LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu" + "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx, + ria->ria_start, ria->ria_end, ria->ria_stoff, + ria->ria_length, ria->ria_pages); offset = page_idx - ria->ria_stoff; offset = offset % (ria->ria_length); if (offset > ria->ria_pages) { @@ -691,7 +710,7 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, { struct vvp_io *vio = vvp_env_io(env); struct vvp_thread_info *vti = vvp_env_info(env); - struct ccc_thread_info *cti = ccc_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; @@ -699,7 +718,6 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, struct ra_io_arg *ria = &vti->vti_ria; struct ll_inode_info *lli; struct cl_object *clob; - struct cl_attr *attr = &cti->cti_attr; int ret = 0; __u64 kms; ENTRY; @@ -722,7 +740,7 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, RETURN(0); } - spin_lock(&ras->ras_lock); + cfs_spin_lock(&ras->ras_lock); if (vio->cui_ra_window_set) bead = &vio->cui_bead; else @@ -740,8 +758,22 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, 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); } @@ -753,7 +785,7 @@ 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; } - spin_unlock(&ras->ras_lock); + cfs_spin_unlock(&ras->ras_lock); if (end == 0) { ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW); @@ -763,8 +795,7 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, if (len == 0) RETURN(0); - reserved = ll_ra_count_get(ll_i2sbi(inode), len); - + reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len); if (reserved < len) ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT); @@ -789,14 +820,14 @@ int ll_readahead(const struct lu_env *env, struct cl_io *io, ra_end, end, ria->ria_end); if (ra_end != end + 1) { - spin_lock(&ras->ras_lock); + 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); } - spin_unlock(&ras->ras_lock); + cfs_spin_unlock(&ras->ras_lock); } RETURN(ret); @@ -831,10 +862,10 @@ static void ras_stride_reset(struct ll_readahead_state *ras) void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) { - spin_lock_init(&ras->ras_lock); + cfs_spin_lock_init(&ras->ras_lock); ras_reset(ras, 0); ras->ras_requests = 0; - INIT_LIST_HEAD(&ras->ras_read_beads); + CFS_INIT_LIST_HEAD(&ras->ras_read_beads); } /* @@ -847,16 +878,17 @@ static int index_in_stride_window(unsigned long index, { unsigned long stride_gap = index - ras->ras_last_readpage - 1; - if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0) + if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 || + ras->ras_stride_pages == ras->ras_stride_length) return 0; /* 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 */ + /*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; + ras->ras_consecutive_pages == ras->ras_stride_pages; } static void ras_update_stride_detector(struct ll_readahead_state *ras, @@ -869,7 +901,20 @@ static void ras_update_stride_detector(struct ll_readahead_state *ras, 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; } static unsigned long @@ -890,7 +935,7 @@ static void ras_stride_increase_window(struct ll_readahead_state *ras, unsigned long stride_len; LASSERT(ras->ras_stride_length > 0); - LASSERTF(ras->ras_window_start + ras->ras_window_len + 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); @@ -913,26 +958,25 @@ static void ras_stride_increase_window(struct ll_readahead_state *ras, window_len += step * ras->ras_stride_length + left; - if (stride_page_count(ras, window_len) <= ra->ra_max_pages) + if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file) ras->ras_window_len = window_len; RAS_CDEBUG(ras); } -/* Set stride I/O read-ahead window start offset */ -static void ras_set_stride_offset(struct ll_readahead_state *ras) +static void ras_increase_window(struct ll_readahead_state *ras, + struct ll_ra_info *ra, struct inode *inode) { - unsigned long window_len = ras->ras_next_readahead - - ras->ras_window_start; - unsigned long left; - - LASSERT(ras->ras_stride_length != 0); - - left = window_len % ras->ras_stride_length; - - ras->ras_stride_offset = ras->ras_next_readahead - left; - - RAS_CDEBUG(ras); + /* The stretch of ra-window should be aligned with max rpc_size + * but current clio architecture does not support retrieve such + * information from lower layer. FIXME later + */ + if (stride_io_mode(ras)) + ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP); + 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, @@ -943,8 +987,7 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, int zero = 0, stride_detect = 0, ra_miss = 0; ENTRY; - spin_lock(&sbi->ll_lock); - spin_lock(&ras->ras_lock); + cfs_spin_lock(&ras->ras_lock); ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS); @@ -960,14 +1003,14 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, index < ras->ras_next_readahead && index_in_window(index, ras->ras_window_start, 0, ras->ras_window_len)) { - ra_miss = 1; + ra_miss = 1; ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW); } /* On the second access to a file smaller than the tunable * ra_max_read_ahead_whole_pages trigger RA on all pages in the - * file up to ra_max_pages. This is simply a best effort and - * only occurs once per open file. Normal RA behavior is reverted + * file up to ra_max_pages_per_file. This is simply a best effort + * and only occurs once per open file. Normal RA behavior is reverted * to for subsequent IO. The mmap case does not increment * ras_requests and thus can never trigger this behavior. */ if (ras->ras_requests == 2 && !ras->ras_request_index) { @@ -977,27 +1020,34 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, CFS_PAGE_SHIFT; CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages, - ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages); + ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file); if (kms_pages && kms_pages <= ra->ra_max_read_ahead_whole_pages) { ras->ras_window_start = 0; ras->ras_last_readpage = 0; ras->ras_next_readahead = 0; - ras->ras_window_len = min(ra->ra_max_pages, + ras->ras_window_len = min(ra->ra_max_pages_per_file, ra->ra_max_read_ahead_whole_pages); GOTO(out_unlock, 0); } } if (zero) { - /* check whether it is in stride I/O mode*/ + /* 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++; - ras_stride_reset(ras); GOTO(out_unlock, 0); } else { - ras->ras_consecutive_requests = 0; + ras->ras_consecutive_pages = 0; + ras->ras_consecutive_requests = 0; if (++ras->ras_consecutive_stride_requests > 1) stride_detect = 1; RAS_CDEBUG(ras); @@ -1011,9 +1061,11 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, *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 */ + /* Reset both stride window and normal RA + * window */ ras_reset(ras, index); ras->ras_consecutive_pages++; ras_stride_reset(ras); @@ -1022,7 +1074,7 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, } 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 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); @@ -1032,11 +1084,18 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, } } ras->ras_consecutive_pages++; - ras_update_stride_detector(ras, index); ras->ras_last_readpage = index; ras_set_start(ras, index); - ras->ras_next_readahead = max(ras->ras_window_start, - ras->ras_next_readahead); + + 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 @@ -1047,31 +1106,31 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode, } /* Initially reset the stride window offset to next_readahead*/ - if (ras->ras_consecutive_stride_requests == 2 && stride_detect) - ras_set_stride_offset(ras); + 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; + } /* The initial ras_window_len is set to the request size. To avoid * uselessly reading and discarding pages for random IO the window is * only increased once per consecutive request received. */ - if ((ras->ras_consecutive_requests > 1 && - !ras->ras_request_index) || stride_detect) { - if (stride_io_mode(ras)) - ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP); - else - ras->ras_window_len = min(ras->ras_window_len + - RAS_INCREASE_STEP, - ra->ra_max_pages); - } + 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++; - spin_unlock(&ras->ras_lock); - spin_unlock(&sbi->ll_lock); + cfs_spin_unlock(&ras->ras_lock); return; } -int ll_writepage(struct page *vmpage, struct writeback_control *_) +int ll_writepage(struct page *vmpage, struct writeback_control *wbc) { struct inode *inode = vmpage->mapping->host; struct lu_env *env; @@ -1093,11 +1152,11 @@ int ll_writepage(struct page *vmpage, struct writeback_control *_) 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 = ccc_env_thread_io(env); io->ci_obj = clob; result = cl_io_init(env, io, CIT_MISC, clob); if (result == 0) { @@ -1118,21 +1177,18 @@ int ll_writepage(struct page *vmpage, struct writeback_control *_) 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); + if (!PageError(vmpage)) { + redirty_page_for_writepage(wbc, vmpage); + result = 0; + } } + cl_page_list_disown(env, io, &queue->c2_qin); LASSERT(!cl_page_is_owned(page, io)); lu_ref_del(&page->cp_reference, "writepage", cfs_current()); @@ -1147,15 +1203,16 @@ int ll_writepage(struct page *vmpage, struct writeback_control *_) int ll_readpage(struct file *file, struct page *vmpage) { - struct lu_env *env; - struct cl_io *io; - struct cl_page *page; + struct ll_cl_context *lcc; int result; - int refcheck; ENTRY; - result = ll_cl_init(file, vmpage, &env, &io, &page, &refcheck); - if (result == 0) { + 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); @@ -1166,9 +1223,11 @@ int ll_readpage(struct file *file, struct page *vmpage) unlock_page(vmpage); result = 0; } + ll_cl_fini(lcc); + } else { + unlock_page(vmpage); + result = PTR_ERR(lcc); } - LASSERT(!cl_page_is_owned(page, io)); - ll_cl_fini(env, io, page, &refcheck); RETURN(result); }