Whamcloud - gitweb
LU-613 clio: Client dead-lock during binary exec
[fs/lustre-release.git] / lustre / llite / rw.c
index 0247349..6bc1525 100644 (file)
@@ -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 <linux/autoconf.h>
 #include <linux/kernel.h>
 #include <linux/mm.h>
 #include <linux/string.h>
@@ -61,7 +60,6 @@
 
 #define DEBUG_SUBSYSTEM S_LLITE
 
-//#include <lustre_mdc.h>
 #include <lustre_lite.h>
 #include <obd_cksum.h>
 #include "llite_internal.h"
  * must be called under ->lli_size_sem */
 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
-                }
-        }
-
-        new_size = i_size_read(inode);
-        ll_inode_size_unlock(inode, 0);
+        ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 
         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);
+        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) {
+                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);
+                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
-                        /*
-                         * 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;
-}
+                        result = io->ci_result;
+                lcc->lcc_created = 1;
+        }
 
-/**
- * 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);
+        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 (env != NULL) {
-                struct vvp_io *vio;
-
-                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;
-                        }
-                        /* additional reference on env was acquired by io,
-                         * disable refcheck */
-                        refcheck = NULL;
-                }
-                cl_env_put(env, refcheck);
-        } else
-                LASSERT(io == NULL);
+        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;
 }
 
 /**
@@ -265,29 +234,16 @@ 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;
@@ -302,8 +258,6 @@ int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
                 }
                 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.
@@ -311,40 +265,46 @@ 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);
 }
 
@@ -368,7 +328,8 @@ 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;
@@ -379,14 +340,23 @@ static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
          * otherwise it will form small read RPC(< 1M), which hurt server
          * performance a lot.
          */
-        ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), len);
+        if (ra->ra_max_pages < atomic_read(&ra->ra_cur_pages))
+                GOTO(out, ret = 0);
+
+        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 (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
-                atomic_sub(ret, &ra->ra_cur_pages);
+        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);
 }
@@ -394,7 +364,7 @@ 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)
@@ -447,14 +417,14 @@ 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)
@@ -463,16 +433,16 @@ 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;
         }
@@ -486,9 +456,9 @@ 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;
 }
 
@@ -536,7 +506,7 @@ 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;
@@ -620,8 +590,8 @@ static unsigned long
 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
                 unsigned long off, unsigned long length)
 {
-        unsigned long start = off > st_off ? off - st_off : 0;
-        unsigned long end = off + length > st_off ? off + length - st_off : 0;
+        __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;
@@ -639,7 +609,7 @@ stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
         if (end_left > st_pgs)
                 end_left = st_pgs;
 
-        CDEBUG(D_READA, "start %lu, end %lu start_left %lu end_left %lu \n",
+        CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
                start, end, start_left, end_left);
 
         if (start == end)
@@ -671,7 +641,8 @@ 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,
@@ -707,10 +678,10 @@ 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) {
@@ -761,7 +732,7 @@ 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
@@ -779,8 +750,22 @@ 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);
         }
@@ -792,7 +777,7 @@ 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);
@@ -802,8 +787,7 @@ 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);
 
@@ -828,14 +812,14 @@ 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);
@@ -870,10 +854,10 @@ 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);
 }
 
 /*
@@ -995,8 +979,7 @@ 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);
 
@@ -1095,8 +1078,16 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode,
         ras->ras_consecutive_pages++;
         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
@@ -1127,12 +1118,11 @@ void ras_update(struct ll_sb_info *sbi, struct inode *inode,
 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 *unused)
+int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
 {
         struct inode           *inode = vmpage->mapping->host;
         struct lu_env          *env;
@@ -1154,11 +1144,11 @@ int ll_writepage(struct page *vmpage, struct writeback_control *unused)
         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) {
@@ -1179,21 +1169,18 @@ int ll_writepage(struct page *vmpage, struct writeback_control *unused)
                         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());
@@ -1208,15 +1195,16 @@ int ll_writepage(struct page *vmpage, struct writeback_control *unused)
 
 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);
@@ -1227,9 +1215,11 @@ 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);
 }