diff options
Diffstat (limited to 'xlators/performance/io-cache')
| -rw-r--r-- | xlators/performance/io-cache/src/Makefile.am | 8 | ||||
| -rw-r--r-- | xlators/performance/io-cache/src/io-cache.c | 2202 | ||||
| -rw-r--r-- | xlators/performance/io-cache/src/io-cache.h | 308 | ||||
| -rw-r--r-- | xlators/performance/io-cache/src/ioc-inode.c | 288 | ||||
| -rw-r--r-- | xlators/performance/io-cache/src/ioc-mem-types.h | 23 | ||||
| -rw-r--r-- | xlators/performance/io-cache/src/page.c | 1239 |
6 files changed, 2164 insertions, 1904 deletions
diff --git a/xlators/performance/io-cache/src/Makefile.am b/xlators/performance/io-cache/src/Makefile.am index 6dd270e8f..155be9988 100644 --- a/xlators/performance/io-cache/src/Makefile.am +++ b/xlators/performance/io-cache/src/Makefile.am @@ -1,14 +1,16 @@ xlator_LTLIBRARIES = io-cache.la xlatordir = $(libdir)/glusterfs/$(PACKAGE_VERSION)/xlator/performance -io_cache_la_LDFLAGS = -module -avoidversion +io_cache_la_LDFLAGS = -module -avoid-version io_cache_la_SOURCES = io-cache.c page.c ioc-inode.c io_cache_la_LIBADD = $(top_builddir)/libglusterfs/src/libglusterfs.la noinst_HEADERS = io-cache.h ioc-mem-types.h -AM_CFLAGS = -fPIC -D_FILE_OFFSET_BITS=64 -D_GNU_SOURCE -Wall -D$(GF_HOST_OS) \ - -I$(top_srcdir)/libglusterfs/src -I$(CONTRIBDIR)/rbtree -shared -nostartfiles $(GF_CFLAGS) +AM_CPPFLAGS = $(GF_CPPFLAGS) -I$(top_srcdir)/libglusterfs/src \ + -I$(CONTRIBDIR)/rbtree + +AM_CFLAGS = -Wall $(GF_CFLAGS) CLEANFILES = diff --git a/xlators/performance/io-cache/src/io-cache.c b/xlators/performance/io-cache/src/io-cache.c index be7227e99..201777b38 100644 --- a/xlators/performance/io-cache/src/io-cache.c +++ b/xlators/performance/io-cache/src/io-cache.c @@ -1,20 +1,11 @@ /* - Copyright (c) 2007-2010 Gluster, Inc. <http://www.gluster.com> + Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> This file is part of GlusterFS. - GlusterFS is free software; you can redistribute it and/or modify - it under the terms of the GNU Affero General Public License as published - by the Free Software Foundation; either version 3 of the License, - or (at your option) any later version. - - GlusterFS is distributed in the hope that it will be useful, but - WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Affero General Public License for more details. - - You should have received a copy of the GNU Affero General Public License - along with this program. If not, see - <http://www.gnu.org/licenses/>. + This file is licensed to you under your choice of the GNU Lesser + General Public License, version 3 or any later version (LGPLv3 or + later), or the GNU General Public License, version 2 (GPLv2), in all + cases as published by the Free Software Foundation. */ #ifndef _CONFIG_H @@ -37,11 +28,10 @@ int ioc_log2_page_size; uint32_t ioc_get_priority (ioc_table_t *table, const char *path); -uint32_t -ioc_get_priority (ioc_table_t *table, const char *path); +struct volume_options options[]; -inline uint32_t +static inline uint32_t ioc_hashfn (void *data, int len) { off_t offset; @@ -51,53 +41,58 @@ ioc_hashfn (void *data, int len) return (offset >> ioc_log2_page_size); } -inline ioc_inode_t * +static inline ioc_inode_t * ioc_inode_reupdate (ioc_inode_t *ioc_inode) { - ioc_table_t *table = ioc_inode->table; + ioc_table_t *table = NULL; + + table = ioc_inode->table; - list_add_tail (&ioc_inode->inode_lru, - &table->inode_lru[ioc_inode->weight]); + list_add_tail (&ioc_inode->inode_lru, + &table->inode_lru[ioc_inode->weight]); - return ioc_inode; + return ioc_inode; } -inline ioc_inode_t * +static inline ioc_inode_t * ioc_get_inode (dict_t *dict, char *name) { - ioc_inode_t *ioc_inode = NULL; - data_t *ioc_inode_data = dict_get (dict, name); - ioc_table_t *table = NULL; - - if (ioc_inode_data) { - ioc_inode = data_to_ptr (ioc_inode_data); - table = ioc_inode->table; - - ioc_table_lock (table); - { - if (list_empty (&ioc_inode->inode_lru)) { - ioc_inode = ioc_inode_reupdate (ioc_inode); - } - } - ioc_table_unlock (table); - } - - return ioc_inode; + ioc_inode_t *ioc_inode = NULL; + data_t *ioc_inode_data = NULL; + ioc_table_t *table = NULL; + + ioc_inode_data = dict_get (dict, name); + if (ioc_inode_data) { + ioc_inode = data_to_ptr (ioc_inode_data); + table = ioc_inode->table; + + ioc_table_lock (table); + { + if (list_empty (&ioc_inode->inode_lru)) { + ioc_inode = ioc_inode_reupdate (ioc_inode); + } + } + ioc_table_unlock (table); + } + + return ioc_inode; } int32_t ioc_inode_need_revalidate (ioc_inode_t *ioc_inode) { - int8_t need_revalidate = 0; - struct timeval tv = {0,}; - ioc_table_t *table = ioc_inode->table; + int8_t need_revalidate = 0; + struct timeval tv = {0,}; + ioc_table_t *table = NULL; + + table = ioc_inode->table; - gettimeofday (&tv, NULL); + gettimeofday (&tv, NULL); - if (time_elapsed (&tv, &ioc_inode->cache.tv) >= table->cache_timeout) - need_revalidate = 1; + if (time_elapsed (&tv, &ioc_inode->cache.tv) >= table->cache_timeout) + need_revalidate = 1; - return need_revalidate; + return need_revalidate; } /* @@ -110,86 +105,87 @@ ioc_inode_need_revalidate (ioc_inode_t *ioc_inode) int64_t __ioc_inode_flush (ioc_inode_t *ioc_inode) { - ioc_page_t *curr = NULL, *next = NULL; - int64_t destroy_size = 0; - int64_t ret = 0; + ioc_page_t *curr = NULL, *next = NULL; + int64_t destroy_size = 0; + int64_t ret = 0; - list_for_each_entry_safe (curr, next, &ioc_inode->cache.page_lru, + list_for_each_entry_safe (curr, next, &ioc_inode->cache.page_lru, page_lru) { - ret = ioc_page_destroy (curr); + ret = __ioc_page_destroy (curr); - if (ret != -1) - destroy_size += ret; - } + if (ret != -1) + destroy_size += ret; + } - return destroy_size; + return destroy_size; } void ioc_inode_flush (ioc_inode_t *ioc_inode) { - int64_t destroy_size = 0; - - ioc_inode_lock (ioc_inode); - { - destroy_size = __ioc_inode_flush (ioc_inode); - } - ioc_inode_unlock (ioc_inode); - - if (destroy_size) { - ioc_table_lock (ioc_inode->table); - { - ioc_inode->table->cache_used -= destroy_size; - } - ioc_table_unlock (ioc_inode->table); - } - - return; + int64_t destroy_size = 0; + + ioc_inode_lock (ioc_inode); + { + destroy_size = __ioc_inode_flush (ioc_inode); + } + ioc_inode_unlock (ioc_inode); + + if (destroy_size) { + ioc_table_lock (ioc_inode->table); + { + ioc_inode->table->cache_used -= destroy_size; + } + ioc_table_unlock (ioc_inode->table); + } + + return; } int32_t ioc_setattr_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, int32_t op_errno, - struct iatt *preop, struct iatt *postop) + struct iatt *preop, struct iatt *postop, dict_t *xdata) { - STACK_UNWIND_STRICT (setattr, frame, op_ret, op_errno, preop, postop); - return 0; + STACK_UNWIND_STRICT (setattr, frame, op_ret, op_errno, preop, postop, + xdata); + return 0; } int32_t ioc_setattr (call_frame_t *frame, xlator_t *this, loc_t *loc, - struct iatt *stbuf, int32_t valid) + struct iatt *stbuf, int32_t valid, dict_t *xdata) { - uint64_t ioc_inode = 0; + uint64_t ioc_inode = 0; - inode_ctx_get (loc->inode, this, &ioc_inode); + inode_ctx_get (loc->inode, this, &ioc_inode); - if (ioc_inode + if (ioc_inode && ((valid & GF_SET_ATTR_ATIME) || (valid & GF_SET_ATTR_MTIME))) - ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); - STACK_WIND (frame, ioc_setattr_cbk, FIRST_CHILD (this), - FIRST_CHILD (this)->fops->setattr, loc, stbuf, valid); + STACK_WIND (frame, ioc_setattr_cbk, FIRST_CHILD (this), + FIRST_CHILD (this)->fops->setattr, loc, stbuf, valid, xdata); - return 0; + return 0; } int32_t ioc_lookup_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, inode_t *inode, - struct iatt *stbuf, dict_t *dict, struct iatt *postparent) + int32_t op_ret, int32_t op_errno, inode_t *inode, + struct iatt *stbuf, dict_t *xdata, struct iatt *postparent) { - ioc_inode_t *ioc_inode = NULL; - ioc_table_t *table = NULL; - uint8_t cache_still_valid = 0; - uint64_t tmp_ioc_inode = 0; - uint32_t weight = 0xffffffff; - const char *path = NULL; - ioc_local_t *local = NULL; - - if (op_ret != 0) - goto out; + ioc_inode_t *ioc_inode = NULL; + ioc_table_t *table = NULL; + uint8_t cache_still_valid = 0; + uint64_t tmp_ioc_inode = 0; + uint32_t weight = 0xffffffff; + const char *path = NULL; + ioc_local_t *local = NULL; + + if (op_ret != 0) + goto out; local = frame->local; if (local == NULL) { @@ -256,20 +252,19 @@ out: loc_wipe (&local->file_loc); } - STACK_UNWIND_STRICT (lookup, frame, op_ret, op_errno, inode, stbuf, - dict, postparent); - return 0; + STACK_UNWIND_STRICT (lookup, frame, op_ret, op_errno, inode, stbuf, + xdata, postparent); + return 0; } int32_t ioc_lookup (call_frame_t *frame, xlator_t *this, loc_t *loc, - dict_t *xattr_req) + dict_t *xdata) { - ioc_local_t *local = NULL; + ioc_local_t *local = NULL; int32_t op_errno = -1, ret = -1; - local = GF_CALLOC (1, sizeof (*local), - gf_ioc_mt_ioc_local_t); + local = mem_get0 (this->local_pool); if (local == NULL) { op_errno = ENOMEM; gf_log (this->name, GF_LOG_ERROR, "out of memory"); @@ -285,16 +280,16 @@ ioc_lookup (call_frame_t *frame, xlator_t *this, loc_t *loc, frame->local = local; - STACK_WIND (frame, ioc_lookup_cbk, FIRST_CHILD (this), - FIRST_CHILD (this)->fops->lookup, loc, xattr_req); + STACK_WIND (frame, ioc_lookup_cbk, FIRST_CHILD (this), + FIRST_CHILD (this)->fops->lookup, loc, xdata); return 0; unwind: - STACK_UNWIND_STRICT (lookup, frame, -1, op_errno, NULL, NULL, + STACK_UNWIND_STRICT (lookup, frame, -1, op_errno, NULL, NULL, NULL, NULL); - return 0; + return 0; } /* @@ -308,17 +303,31 @@ unwind: int32_t ioc_forget (xlator_t *this, inode_t *inode) { - uint64_t ioc_inode = 0; + uint64_t ioc_inode = 0; + + inode_ctx_get (inode, this, &ioc_inode); - inode_ctx_get (inode, this, &ioc_inode); + if (ioc_inode) + ioc_inode_destroy ((ioc_inode_t *)(long)ioc_inode); + + return 0; +} + +static int32_t +ioc_invalidate(xlator_t *this, inode_t *inode) +{ + uint64_t ioc_addr = 0; + ioc_inode_t *ioc_inode = NULL; + + inode_ctx_get(inode, this, (uint64_t *) &ioc_addr); + ioc_inode = (void *) ioc_addr; if (ioc_inode) - ioc_inode_destroy ((ioc_inode_t *)(long)ioc_inode); + ioc_inode_flush(ioc_inode); return 0; } - /* * ioc_cache_validate_cbk - * @@ -332,86 +341,88 @@ ioc_forget (xlator_t *this, inode_t *inode) */ int32_t ioc_cache_validate_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct iatt *stbuf) + int32_t op_ret, int32_t op_errno, struct iatt *stbuf, + dict_t *xdata) { - ioc_local_t *local = NULL; - ioc_inode_t *ioc_inode = NULL; - size_t destroy_size = 0; - struct iatt *local_stbuf = NULL; + ioc_local_t *local = NULL; + ioc_inode_t *ioc_inode = NULL; + size_t destroy_size = 0; + struct iatt *local_stbuf = NULL; local = frame->local; - ioc_inode = local->inode; + ioc_inode = local->inode; local_stbuf = stbuf; - if ((op_ret == -1) || - ((op_ret >= 0) && !ioc_cache_still_valid(ioc_inode, stbuf))) { - gf_log (ioc_inode->table->xl->name, GF_LOG_DEBUG, - "cache for inode(%p) is invalid. flushing all pages", - ioc_inode); - /* NOTE: only pages with no waiting frames are flushed by - * ioc_inode_flush. page_fault will be generated for all - * the pages which have waiting frames by ioc_inode_wakeup() - */ - ioc_inode_lock (ioc_inode); - { - destroy_size = __ioc_inode_flush (ioc_inode); - if (op_ret >= 0) { - ioc_inode->cache.mtime = stbuf->ia_mtime; - ioc_inode->cache.mtime_nsec = stbuf->ia_mtime_nsec; + if ((op_ret == -1) || + ((op_ret >= 0) && !ioc_cache_still_valid(ioc_inode, stbuf))) { + gf_log (ioc_inode->table->xl->name, GF_LOG_DEBUG, + "cache for inode(%p) is invalid. flushing all pages", + ioc_inode); + /* NOTE: only pages with no waiting frames are flushed by + * ioc_inode_flush. page_fault will be generated for all + * the pages which have waiting frames by ioc_inode_wakeup() + */ + ioc_inode_lock (ioc_inode); + { + destroy_size = __ioc_inode_flush (ioc_inode); + if (op_ret >= 0) { + ioc_inode->cache.mtime = stbuf->ia_mtime; + ioc_inode->cache.mtime_nsec + = stbuf->ia_mtime_nsec; } - } - ioc_inode_unlock (ioc_inode); - local_stbuf = NULL; - } + } + ioc_inode_unlock (ioc_inode); + local_stbuf = NULL; + } - if (destroy_size) { - ioc_table_lock (ioc_inode->table); - { - ioc_inode->table->cache_used -= destroy_size; - } - ioc_table_unlock (ioc_inode->table); - } + if (destroy_size) { + ioc_table_lock (ioc_inode->table); + { + ioc_inode->table->cache_used -= destroy_size; + } + ioc_table_unlock (ioc_inode->table); + } - if (op_ret < 0) - local_stbuf = NULL; + if (op_ret < 0) + local_stbuf = NULL; - ioc_inode_lock (ioc_inode); - { - gettimeofday (&ioc_inode->cache.tv, NULL); - } - ioc_inode_unlock (ioc_inode); + ioc_inode_lock (ioc_inode); + { + gettimeofday (&ioc_inode->cache.tv, NULL); + } + ioc_inode_unlock (ioc_inode); - ioc_inode_wakeup (frame, ioc_inode, local_stbuf); + ioc_inode_wakeup (frame, ioc_inode, local_stbuf); - /* any page-fault initiated by ioc_inode_wakeup() will have its own - * fd_ref on fd, safe to unref validate frame's private copy - */ - fd_unref (local->fd); + /* any page-fault initiated by ioc_inode_wakeup() will have its own + * fd_ref on fd, safe to unref validate frame's private copy + */ + fd_unref (local->fd); - STACK_DESTROY (frame->root); + STACK_DESTROY (frame->root); - return 0; + return 0; } int32_t ioc_wait_on_inode (ioc_inode_t *ioc_inode, ioc_page_t *page) { - ioc_waitq_t *waiter = NULL, *trav = NULL; - uint32_t page_found = 0; - int32_t ret = 0; - - trav = ioc_inode->waitq; - - while (trav) { - if (trav->data == page) { - page_found = 1; - break; - } - trav = trav->next; - } - - if (!page_found) { - waiter = GF_CALLOC (1, sizeof (ioc_waitq_t), + ioc_waitq_t *waiter = NULL, *trav = NULL; + uint32_t page_found = 0; + int32_t ret = 0; + + trav = ioc_inode->waitq; + + while (trav) { + if (trav->data == page) { + page_found = 1; + break; + } + trav = trav->next; + } + + if (!page_found) { + waiter = GF_CALLOC (1, sizeof (ioc_waitq_t), gf_ioc_mt_ioc_waitq_t); if (waiter == NULL) { gf_log (ioc_inode->table->xl->name, GF_LOG_ERROR, @@ -420,13 +431,13 @@ ioc_wait_on_inode (ioc_inode_t *ioc_inode, ioc_page_t *page) goto out; } - waiter->data = page; - waiter->next = ioc_inode->waitq; - ioc_inode->waitq = waiter; - } + waiter->data = page; + waiter->next = ioc_inode->waitq; + ioc_inode->waitq = waiter; + } out: - return ret; + return ret; } /* @@ -439,16 +450,15 @@ out: */ int32_t ioc_cache_validate (call_frame_t *frame, ioc_inode_t *ioc_inode, fd_t *fd, - ioc_page_t *page) + ioc_page_t *page) { - call_frame_t *validate_frame = NULL; - ioc_local_t *validate_local = NULL; - ioc_local_t *local = NULL; - int32_t ret = 0; + call_frame_t *validate_frame = NULL; + ioc_local_t *validate_local = NULL; + ioc_local_t *local = NULL; + int32_t ret = 0; local = frame->local; - validate_local = GF_CALLOC (1, sizeof (ioc_local_t), - gf_ioc_mt_ioc_local_t); + validate_local = mem_get0 (THIS->local_pool); if (validate_local == NULL) { ret = -1; local->op_ret = -1; @@ -458,55 +468,55 @@ ioc_cache_validate (call_frame_t *frame, ioc_inode_t *ioc_inode, fd_t *fd, goto out; } - validate_frame = copy_frame (frame); + validate_frame = copy_frame (frame); if (validate_frame == NULL) { ret = -1; local->op_ret = -1; local->op_errno = ENOMEM; - GF_FREE (validate_local); + mem_put (validate_local); gf_log (ioc_inode->table->xl->name, GF_LOG_ERROR, "out of memory"); goto out; } - validate_local->fd = fd_ref (fd); - validate_local->inode = ioc_inode; - validate_frame->local = validate_local; + validate_local->fd = fd_ref (fd); + validate_local->inode = ioc_inode; + validate_frame->local = validate_local; - STACK_WIND (validate_frame, ioc_cache_validate_cbk, + STACK_WIND (validate_frame, ioc_cache_validate_cbk, FIRST_CHILD (frame->this), - FIRST_CHILD (frame->this)->fops->fstat, fd); + FIRST_CHILD (frame->this)->fops->fstat, fd, NULL); out: - return ret; + return ret; } -inline uint32_t +static inline uint32_t is_match (const char *path, const char *pattern) { - int32_t ret = 0; + int32_t ret = 0; - ret = fnmatch (pattern, path, FNM_NOESCAPE); + ret = fnmatch (pattern, path, FNM_NOESCAPE); - return (ret == 0); + return (ret == 0); } uint32_t ioc_get_priority (ioc_table_t *table, const char *path) { - uint32_t priority = 1; - struct ioc_priority *curr = NULL; + uint32_t priority = 1; + struct ioc_priority *curr = NULL; - if (list_empty(&table->priority_list)) - return priority; + if (list_empty(&table->priority_list)) + return priority; - priority = 0; - list_for_each_entry (curr, &table->priority_list, list) { - if (is_match (path, curr->pattern)) - priority = curr->priority; - } + priority = 0; + list_for_each_entry (curr, &table->priority_list, list) { + if (is_match (path, curr->pattern)) + priority = curr->priority; + } - return priority; + return priority; } /* @@ -522,13 +532,13 @@ ioc_get_priority (ioc_table_t *table, const char *path) */ int32_t ioc_open_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno, fd_t *fd) + int32_t op_errno, fd_t *fd, dict_t *xdata) { - uint64_t tmp_ioc_inode = 0; - ioc_local_t *local = NULL; - ioc_table_t *table = NULL; - ioc_inode_t *ioc_inode = NULL; - uint32_t weight = 0xffffffff; + uint64_t tmp_ioc_inode = 0; + ioc_local_t *local = NULL; + ioc_table_t *table = NULL; + ioc_inode_t *ioc_inode = NULL; + uint32_t weight = 0xffffffff; local = frame->local; if (!this || !this->private) { @@ -543,6 +553,13 @@ ioc_open_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, inode_ctx_get (fd->inode, this, &tmp_ioc_inode); ioc_inode = (ioc_inode_t *)(long)tmp_ioc_inode; + //TODO: see why inode context is NULL and handle it. + if (!ioc_inode) { + gf_log (this->name, GF_LOG_ERROR, "inode context is " + "NULL (%s)", uuid_utoa (fd->inode->gfid)); + goto out; + } + ioc_table_lock (ioc_inode->table); { list_move_tail (&ioc_inode->inode_lru, @@ -553,7 +570,7 @@ ioc_open_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, ioc_inode_lock (ioc_inode); { if ((table->min_file_size > ioc_inode->ia_size) - || ((table->max_file_size >= 0) + || ((table->max_file_size > 0) && (table->max_file_size < ioc_inode->ia_size))) { fd_ctx_set (fd, this, 1); } @@ -565,28 +582,24 @@ ioc_open_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, /* O_DIRECT is only for one fd, not the inode * as a whole */ - fd_ctx_set (fd, this, 1); - } - if ((local->wbflags & GF_OPEN_NOWB) != 0) { - /* disable caching as asked by NFS */ fd_ctx_set (fd, this, 1); } /* weight = 0, we disable caching on it */ if (weight == 0) { /* we allow a pattern-matched cache disable this way - */ + */ fd_ctx_set (fd, this, 1); } } out: - GF_FREE (local); + mem_put (local); frame->local = NULL; - STACK_UNWIND_STRICT (open, frame, op_ret, op_errno, fd); + STACK_UNWIND_STRICT (open, frame, op_ret, op_errno, fd, xdata); - return 0; + return 0; } /* @@ -605,15 +618,15 @@ out: int32_t ioc_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, int32_t op_errno, fd_t *fd, - inode_t *inode, struct iatt *buf, struct iatt *preparent, - struct iatt *postparent) + inode_t *inode, struct iatt *buf, struct iatt *preparent, + struct iatt *postparent, dict_t *xdata) { - ioc_local_t *local = NULL; - ioc_table_t *table = NULL; - ioc_inode_t *ioc_inode = NULL; - uint32_t weight = 0xffffffff; - const char *path = NULL; - int ret = -1; + ioc_local_t *local = NULL; + ioc_table_t *table = NULL; + ioc_inode_t *ioc_inode = NULL; + uint32_t weight = 0xffffffff; + const char *path = NULL; + int ret = -1; local = frame->local; if (!this || !this->private) { @@ -625,7 +638,7 @@ ioc_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this, table = this->private; path = local->file_loc.path; - if (op_ret != -1) { + if (op_ret != -1) { /* assign weight */ weight = ioc_get_priority (table, path); @@ -638,9 +651,13 @@ ioc_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this, ioc_inode->ia_size = buf->ia_size; if ((table->min_file_size > ioc_inode->ia_size) - || ((table->max_file_size >= 0) + || ((table->max_file_size > 0) && (table->max_file_size < ioc_inode->ia_size))) { ret = fd_ctx_set (fd, this, 1); + if (ret) + gf_log (this->name, GF_LOG_WARNING, + "%s: failed to set fd ctx", + local->file_loc.path); } } ioc_inode_unlock (ioc_inode); @@ -648,28 +665,38 @@ ioc_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this, inode_ctx_put (fd->inode, this, (uint64_t)(long)ioc_inode); - /* If O_DIRECT open, we disable caching on it */ - if (local->flags & O_DIRECT) - /* + /* If O_DIRECT open, we disable caching on it */ + if (local->flags & O_DIRECT) { + /* * O_DIRECT is only for one fd, not the inode - * as a whole */ - ret = fd_ctx_set (fd, this, 1); + * as a whole */ + ret = fd_ctx_set (fd, this, 1); + if (ret) + gf_log (this->name, GF_LOG_WARNING, + "%s: failed to set fd ctx", + local->file_loc.path); + } /* if weight == 0, we disable caching on it */ - if (!weight) - /* we allow a pattern-matched cache disable this way */ - ret = fd_ctx_set (fd, this, 1); + if (!weight) { + /* we allow a pattern-matched cache disable this way */ + ret = fd_ctx_set (fd, this, 1); + if (ret) + gf_log (this->name, GF_LOG_WARNING, + "%s: failed to set fd ctx", + local->file_loc.path); + } - } + } out: - frame->local = NULL; - GF_FREE (local); + frame->local = NULL; + mem_put (local); STACK_UNWIND_STRICT (create, frame, op_ret, op_errno, fd, inode, buf, - preparent, postparent); + preparent, postparent, xdata); - return 0; + return 0; } @@ -677,13 +704,13 @@ int32_t ioc_mknod_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, int32_t op_errno, inode_t *inode, struct iatt *buf, struct iatt *preparent, - struct iatt *postparent) + struct iatt *postparent, dict_t *xdata) { ioc_local_t *local = NULL; - ioc_table_t *table = NULL; - ioc_inode_t *ioc_inode = NULL; - uint32_t weight = 0xffffffff; - const char *path = NULL; + ioc_table_t *table = NULL; + ioc_inode_t *ioc_inode = NULL; + uint32_t weight = 0xffffffff; + const char *path = NULL; local = frame->local; if (!this || !this->private) { @@ -695,7 +722,7 @@ ioc_mknod_cbk (call_frame_t *frame, void *cookie, xlator_t *this, table = this->private; path = local->file_loc.path; - if (op_ret != -1) { + if (op_ret != -1) { /* assign weight */ weight = ioc_get_priority (table, path); @@ -711,29 +738,28 @@ ioc_mknod_cbk (call_frame_t *frame, void *cookie, xlator_t *this, inode_ctx_put (inode, this, (uint64_t)(long)ioc_inode); - } + } out: - frame->local = NULL; + frame->local = NULL; loc_wipe (&local->file_loc); - GF_FREE (local); + mem_put (local); - STACK_UNWIND_STRICT (mknod, frame, op_ret, op_errno, inode, buf, - preparent, postparent); - return 0; + STACK_UNWIND_STRICT (mknod, frame, op_ret, op_errno, inode, buf, + preparent, postparent, xdata); + return 0; } int ioc_mknod (call_frame_t *frame, xlator_t *this, loc_t *loc, mode_t mode, - dev_t rdev, dict_t *params) + dev_t rdev, mode_t umask, dict_t *xdata) { - ioc_local_t *local = NULL; + ioc_local_t *local = NULL; int32_t op_errno = -1, ret = -1; - local = GF_CALLOC (1, sizeof (*local), - gf_ioc_mt_ioc_local_t); + local = mem_get0 (this->local_pool); if (local == NULL) { op_errno = ENOMEM; gf_log (this->name, GF_LOG_ERROR, "out of memory"); @@ -749,22 +775,22 @@ ioc_mknod (call_frame_t *frame, xlator_t *this, loc_t *loc, mode_t mode, frame->local = local; - STACK_WIND (frame, ioc_mknod_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->mknod, - loc, mode, rdev, params); + STACK_WIND (frame, ioc_mknod_cbk, + FIRST_CHILD(this), + FIRST_CHILD(this)->fops->mknod, + loc, mode, rdev, umask, xdata); return 0; unwind: if (local != NULL) { loc_wipe (&local->file_loc); - GF_FREE (local); + mem_put (local); } - STACK_UNWIND_STRICT (mknod, frame, -1, op_errno, NULL, NULL, - NULL, NULL); + STACK_UNWIND_STRICT (mknod, frame, -1, op_errno, NULL, NULL, + NULL, NULL, NULL); - return 0; + return 0; } @@ -778,29 +804,29 @@ unwind: */ int32_t ioc_open (call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, - fd_t *fd, int32_t wbflags) + fd_t *fd, dict_t *xdata) { - ioc_local_t *local = NULL; + ioc_local_t *local = NULL; - local = GF_CALLOC (1, sizeof (ioc_local_t), gf_ioc_mt_ioc_local_t); + local = mem_get0 (this->local_pool); if (local == NULL) { gf_log (this->name, GF_LOG_ERROR, "out of memory"); - STACK_UNWIND_STRICT (open, frame, -1, ENOMEM, NULL); + STACK_UNWIND_STRICT (open, frame, -1, ENOMEM, NULL, NULL); return 0; } - local->flags = flags; - local->file_loc.path = loc->path; - local->file_loc.inode = loc->inode; - local->wbflags = wbflags; + local->flags = flags; + local->file_loc.path = loc->path; + local->file_loc.inode = loc->inode; - frame->local = local; + frame->local = local; - STACK_WIND (frame, ioc_open_cbk, FIRST_CHILD(this), - FIRST_CHILD(this)->fops->open, loc, flags, fd, wbflags); + STACK_WIND (frame, ioc_open_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->open, loc, flags, fd, + xdata); - return 0; + return 0; } /* @@ -815,27 +841,27 @@ ioc_open (call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, */ int32_t ioc_create (call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, - mode_t mode, fd_t *fd, dict_t *params) + mode_t mode, mode_t umask, fd_t *fd, dict_t *xdata) { - ioc_local_t *local = NULL; + ioc_local_t *local = NULL; - local = GF_CALLOC (1, sizeof (ioc_local_t), gf_ioc_mt_ioc_local_t); + local = mem_get0 (this->local_pool); if (local == NULL) { gf_log (this->name, GF_LOG_ERROR, "out of memory"); STACK_UNWIND_STRICT (create, frame, -1, ENOMEM, NULL, NULL, - NULL, NULL, NULL); + NULL, NULL, NULL, NULL); return 0; } - local->flags = flags; - local->file_loc.path = loc->path; - frame->local = local; + local->flags = flags; + local->file_loc.path = loc->path; + frame->local = local; - STACK_WIND (frame, ioc_create_cbk, FIRST_CHILD(this), - FIRST_CHILD(this)->fops->create, loc, flags, mode, - fd, params); + STACK_WIND (frame, ioc_create_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->create, loc, flags, mode, + umask, fd, xdata); - return 0; + return 0; } @@ -852,7 +878,7 @@ ioc_create (call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, int32_t ioc_release (xlator_t *this, fd_t *fd) { - return 0; + return 0; } /* @@ -869,30 +895,30 @@ ioc_release (xlator_t *this, fd_t *fd) int32_t ioc_readv_disabled_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, int32_t op_errno, struct iovec *vector, - int32_t count, struct iatt *stbuf, - struct iobref *iobref) + int32_t count, struct iatt *stbuf, + struct iobref *iobref, dict_t *xdata) { - STACK_UNWIND_STRICT (readv, frame, op_ret, op_errno, vector, count, - stbuf, iobref); - return 0; + STACK_UNWIND_STRICT (readv, frame, op_ret, op_errno, vector, count, + stbuf, iobref, xdata); + return 0; } int32_t ioc_need_prune (ioc_table_t *table) { - int64_t cache_difference = 0; - - ioc_table_lock (table); - { - cache_difference = table->cache_used - table->cache_size; - } - ioc_table_unlock (table); - - if (cache_difference > 0) - return 1; - else - return 0; + int64_t cache_difference = 0; + + ioc_table_lock (table); + { + cache_difference = table->cache_used - table->cache_size; + } + ioc_table_unlock (table); + + if (cache_difference > 0) + return 1; + else + return 0; } /* @@ -907,122 +933,132 @@ void ioc_dispatch_requests (call_frame_t *frame, ioc_inode_t *ioc_inode, fd_t *fd, off_t offset, size_t size) { - ioc_local_t *local = NULL; - ioc_table_t *table = NULL; - ioc_page_t *trav = NULL; - ioc_waitq_t *waitq = NULL; - off_t rounded_offset = 0; - off_t rounded_end = 0; - off_t trav_offset = 0; - int32_t fault = 0; - size_t trav_size = 0; - off_t local_offset = 0; - int32_t ret = -1; - int8_t need_validate = 0; - int8_t might_need_validate = 0; /* - * if a page exists, do we need - * to validate it? - */ + ioc_local_t *local = NULL; + ioc_table_t *table = NULL; + ioc_page_t *trav = NULL; + ioc_waitq_t *waitq = NULL; + off_t rounded_offset = 0; + off_t rounded_end = 0; + off_t trav_offset = 0; + int32_t fault = 0; + size_t trav_size = 0; + off_t local_offset = 0; + int32_t ret = -1; + int8_t need_validate = 0; + int8_t might_need_validate = 0; /* + * if a page exists, do we need + * to validate it? + */ local = frame->local; table = ioc_inode->table; - rounded_offset = floor (offset, table->page_size); - rounded_end = roof (offset + size, table->page_size); - trav_offset = rounded_offset; - - /* once a frame does read, it should be waiting on something */ - local->wait_count++; - - /* Requested region can fall in three different pages, - * 1. Ready - region is already in cache, we just have to serve it. - * 2. In-transit - page fault has been generated on this page, we need - * to wait till the page is ready - * 3. Fault - page is not in cache, we have to generate a page fault - */ - - might_need_validate = ioc_inode_need_revalidate (ioc_inode); - - while (trav_offset < rounded_end) { - ioc_inode_lock (ioc_inode); - //{ - - /* look for requested region in the cache */ - trav = ioc_page_get (ioc_inode, trav_offset); - - local_offset = max (trav_offset, offset); - trav_size = min (((offset+size) - local_offset), - table->page_size); - - if (!trav) { - /* page not in cache, we need to generate page fault */ - trav = ioc_page_create (ioc_inode, trav_offset); - fault = 1; - if (!trav) { - gf_log (frame->this->name, GF_LOG_CRITICAL, - "out of memory"); - local->op_ret = -1; - local->op_errno = ENOMEM; - goto out; - } - } - - ioc_wait_on_page (trav, frame, local_offset, trav_size); - - if (trav->ready) { - /* page found in cache */ - if (!might_need_validate && !ioc_inode->waitq) { - /* fresh enough */ - gf_log (frame->this->name, GF_LOG_TRACE, - "cache hit for trav_offset=%"PRId64"" - "/local_offset=%"PRId64"", - trav_offset, local_offset); - waitq = ioc_page_wakeup (trav); - } else { - /* if waitq already exists, fstat revalidate is - already on the way */ - if (!ioc_inode->waitq) { - need_validate = 1; - } - - ret = ioc_wait_on_inode (ioc_inode, trav); - if (ret < 0) { - local->op_ret = -1; - local->op_errno = -ret; - need_validate = 0; + rounded_offset = floor (offset, table->page_size); + rounded_end = roof (offset + size, table->page_size); + trav_offset = rounded_offset; + + /* once a frame does read, it should be waiting on something */ + local->wait_count++; + + /* Requested region can fall in three different pages, + * 1. Ready - region is already in cache, we just have to serve it. + * 2. In-transit - page fault has been generated on this page, we need + * to wait till the page is ready + * 3. Fault - page is not in cache, we have to generate a page fault + */ - waitq = ioc_page_wakeup (trav); - ioc_inode_unlock (ioc_inode); + might_need_validate = ioc_inode_need_revalidate (ioc_inode); - ioc_waitq_return (waitq); - waitq = NULL; + while (trav_offset < rounded_end) { + ioc_inode_lock (ioc_inode); + { + /* look for requested region in the cache */ + trav = __ioc_page_get (ioc_inode, trav_offset); + + local_offset = max (trav_offset, offset); + trav_size = min (((offset+size) - local_offset), + table->page_size); + + if (!trav) { + /* page not in cache, we need to generate page + * fault + */ + trav = __ioc_page_create (ioc_inode, + trav_offset); + fault = 1; + if (!trav) { + gf_log (frame->this->name, + GF_LOG_CRITICAL, + "out of memory"); + local->op_ret = -1; + local->op_errno = ENOMEM; goto out; } - } - } - - //} - ioc_inode_unlock (ioc_inode); - - ioc_waitq_return (waitq); - waitq = NULL; - - if (fault) { - fault = 0; - /* new page created, increase the table->cache_used */ - ioc_page_fault (ioc_inode, frame, fd, trav_offset); - } - - if (need_validate) { - need_validate = 0; - gf_log (frame->this->name, GF_LOG_TRACE, - "sending validate request for " - "inode(%"PRId64") at offset=%"PRId64"", - fd->inode->ino, trav_offset); - ret = ioc_cache_validate (frame, ioc_inode, fd, trav); + } + + __ioc_wait_on_page (trav, frame, local_offset, + trav_size); + + if (trav->ready) { + /* page found in cache */ + if (!might_need_validate && !ioc_inode->waitq) { + /* fresh enough */ + gf_log (frame->this->name, GF_LOG_TRACE, + "cache hit for trav_offset=%" + PRId64"/local_offset=%"PRId64"", + trav_offset, local_offset); + waitq = __ioc_page_wakeup (trav, + trav->op_errno); + } else { + /* if waitq already exists, fstat + * revalidate is + * already on the way + */ + if (!ioc_inode->waitq) { + need_validate = 1; + } + + ret = ioc_wait_on_inode (ioc_inode, + trav); + if (ret < 0) { + local->op_ret = -1; + local->op_errno = -ret; + need_validate = 0; + + waitq = __ioc_page_wakeup (trav, + trav->op_errno); + ioc_inode_unlock (ioc_inode); + + ioc_waitq_return (waitq); + waitq = NULL; + goto out; + } + } + } + + } + ioc_inode_unlock (ioc_inode); + + ioc_waitq_return (waitq); + waitq = NULL; + + if (fault) { + fault = 0; + /* new page created, increase the table->cache_used */ + ioc_page_fault (ioc_inode, frame, fd, trav_offset); + } + + if (need_validate) { + need_validate = 0; + gf_log (frame->this->name, GF_LOG_TRACE, + "sending validate request for " + "inode(%s) at offset=%"PRId64"", + uuid_utoa (fd->inode->gfid), trav_offset); + ret = ioc_cache_validate (frame, ioc_inode, fd, trav); if (ret == -1) { ioc_inode_lock (ioc_inode); { - waitq = ioc_page_wakeup (trav); + waitq = __ioc_page_wakeup (trav, + trav->op_errno); } ioc_inode_unlock (ioc_inode); @@ -1030,19 +1066,19 @@ ioc_dispatch_requests (call_frame_t *frame, ioc_inode_t *ioc_inode, fd_t *fd, waitq = NULL; goto out; } - } + } - trav_offset += table->page_size; - } + trav_offset += table->page_size; + } out: - ioc_frame_return (frame); + ioc_frame_return (frame); - if (ioc_need_prune (ioc_inode->table)) { - ioc_prune (ioc_inode->table); - } + if (ioc_need_prune (ioc_inode->table)) { + ioc_prune (ioc_inode->table); + } - return; + return; } @@ -1058,30 +1094,29 @@ out: */ int32_t ioc_readv (call_frame_t *frame, xlator_t *this, fd_t *fd, - size_t size, off_t offset) + size_t size, off_t offset, uint32_t flags, dict_t *xdata) { - uint64_t tmp_ioc_inode = 0; - ioc_inode_t *ioc_inode = NULL; - ioc_local_t *local = NULL; - uint32_t weight = 0; - ioc_table_t *table = NULL; - uint32_t num_pages = 0; - int32_t op_errno = -1; + uint64_t tmp_ioc_inode = 0; + ioc_inode_t *ioc_inode = NULL; + ioc_local_t *local = NULL; + uint32_t weight = 0; + ioc_table_t *table = NULL; + int32_t op_errno = -1; if (!this) { goto out; } - inode_ctx_get (fd->inode, this, &tmp_ioc_inode); - ioc_inode = (ioc_inode_t *)(long)tmp_ioc_inode; - if (!ioc_inode) { - /* caching disabled, go ahead with normal readv */ - STACK_WIND (frame, ioc_readv_disabled_cbk, + inode_ctx_get (fd->inode, this, &tmp_ioc_inode); + ioc_inode = (ioc_inode_t *)(long)tmp_ioc_inode; + if (!ioc_inode) { + /* caching disabled, go ahead with normal readv */ + STACK_WIND (frame, ioc_readv_disabled_cbk, FIRST_CHILD (frame->this), - FIRST_CHILD (frame->this)->fops->readv, fd, size, - offset); - return 0; - } + FIRST_CHILD (frame->this)->fops->readv, fd, size, + offset, flags, xdata); + return 0; + } table = this->private; @@ -1092,37 +1127,14 @@ ioc_readv (call_frame_t *frame, xlator_t *this, fd_t *fd, goto out; } - - ioc_table_lock (table); - { - if (!table->mem_pool) { - - num_pages = (table->cache_size / table->page_size) - + ((table->cache_size % table->page_size) - ? 1 : 0); - - table->mem_pool - = mem_pool_new (rbthash_entry_t, num_pages); - - if (!table->mem_pool) { - gf_log (this->name, GF_LOG_ERROR, - "Unable to allocate mem_pool"); - op_errno = ENOMEM; - ioc_table_unlock (table); - goto out; - } - } - } - ioc_table_unlock (table); - ioc_inode_lock (ioc_inode); { if (!ioc_inode->cache.page_table) { ioc_inode->cache.page_table = rbthash_table_init - (IOC_PAGE_TABLE_BUCKET_COUNT, - ioc_hashfn, NULL, 0, - table->mem_pool); + (IOC_PAGE_TABLE_BUCKET_COUNT, + ioc_hashfn, NULL, 0, + table->mem_pool); if (ioc_inode->cache.page_table == NULL) { op_errno = ENOMEM; @@ -1133,50 +1145,50 @@ ioc_readv (call_frame_t *frame, xlator_t *this, fd_t *fd, } ioc_inode_unlock (ioc_inode); - if (!fd_ctx_get (fd, this, NULL)) { - /* disable caching for this fd, go ahead with normal readv */ - STACK_WIND (frame, ioc_readv_disabled_cbk, + if (!fd_ctx_get (fd, this, NULL)) { + /* disable caching for this fd, go ahead with normal readv */ + STACK_WIND (frame, ioc_readv_disabled_cbk, FIRST_CHILD (frame->this), - FIRST_CHILD (frame->this)->fops->readv, fd, size, - offset); - return 0; - } + FIRST_CHILD (frame->this)->fops->readv, fd, size, + offset, flags, xdata); + return 0; + } - local = (ioc_local_t *) GF_CALLOC (1, sizeof (ioc_local_t), - gf_ioc_mt_ioc_local_t); + local = mem_get0 (this->local_pool); if (local == NULL) { gf_log (this->name, GF_LOG_ERROR, "out of memory"); op_errno = ENOMEM; goto out; } - INIT_LIST_HEAD (&local->fill_list); + INIT_LIST_HEAD (&local->fill_list); - frame->local = local; - local->pending_offset = offset; - local->pending_size = size; - local->offset = offset; - local->size = size; - local->inode = ioc_inode; + frame->local = local; + local->pending_offset = offset; + local->pending_size = size; + local->offset = offset; + local->size = size; + local->inode = ioc_inode; - gf_log (this->name, GF_LOG_TRACE, - "NEW REQ (%p) offset = %"PRId64" && size = %"GF_PRI_SIZET"", - frame, offset, size); + gf_log (this->name, GF_LOG_TRACE, + "NEW REQ (%p) offset = %"PRId64" && size = %"GF_PRI_SIZET"", + frame, offset, size); - weight = ioc_inode->weight; + weight = ioc_inode->weight; - ioc_table_lock (ioc_inode->table); - { - list_move_tail (&ioc_inode->inode_lru, - &ioc_inode->table->inode_lru[weight]); - } - ioc_table_unlock (ioc_inode->table); + ioc_table_lock (ioc_inode->table); + { + list_move_tail (&ioc_inode->inode_lru, + &ioc_inode->table->inode_lru[weight]); + } + ioc_table_unlock (ioc_inode->table); - ioc_dispatch_requests (frame, ioc_inode, fd, offset, size); - return 0; + ioc_dispatch_requests (frame, ioc_inode, fd, offset, size); + return 0; out: - STACK_UNWIND_STRICT (readv, frame, -1, op_errno, NULL, 0, NULL, NULL); + STACK_UNWIND_STRICT (readv, frame, -1, op_errno, NULL, 0, NULL, NULL, + NULL); return 0; } @@ -1192,20 +1204,21 @@ out: */ int32_t ioc_writev_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct iatt *prebuf, - struct iatt *postbuf) + int32_t op_ret, int32_t op_errno, struct iatt *prebuf, + struct iatt *postbuf, dict_t *xdata) { - ioc_local_t *local = NULL; - uint64_t ioc_inode = 0; + ioc_local_t *local = NULL; + uint64_t ioc_inode = 0; local = frame->local; - inode_ctx_get (local->fd->inode, this, &ioc_inode); + inode_ctx_get (local->fd->inode, this, &ioc_inode); - if (ioc_inode) - ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + if (ioc_inode) + ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); - STACK_UNWIND_STRICT (writev, frame, op_ret, op_errno, prebuf, postbuf); - return 0; + STACK_UNWIND_STRICT (writev, frame, op_ret, op_errno, prebuf, postbuf, + xdata); + return 0; } /* @@ -1221,33 +1234,33 @@ ioc_writev_cbk (call_frame_t *frame, void *cookie, xlator_t *this, */ int32_t ioc_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, - struct iovec *vector, int32_t count, off_t offset, - struct iobref *iobref) + struct iovec *vector, int32_t count, off_t offset, + uint32_t flags, struct iobref *iobref, dict_t *xdata) { - ioc_local_t *local = NULL; - uint64_t ioc_inode = 0; + ioc_local_t *local = NULL; + uint64_t ioc_inode = 0; - local = GF_CALLOC (1, sizeof (ioc_local_t), gf_ioc_mt_ioc_local_t); + local = mem_get0 (this->local_pool); if (local == NULL) { gf_log (this->name, GF_LOG_ERROR, "out of memory"); - STACK_UNWIND_STRICT (writev, frame, -1, ENOMEM, NULL, NULL); + STACK_UNWIND_STRICT (writev, frame, -1, ENOMEM, NULL, NULL, NULL); return 0; } - /* TODO: why is it not fd_ref'ed */ - local->fd = fd; - frame->local = local; + /* TODO: why is it not fd_ref'ed */ + local->fd = fd; + frame->local = local; - inode_ctx_get (fd->inode, this, &ioc_inode); - if (ioc_inode) - ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + inode_ctx_get (fd->inode, this, &ioc_inode); + if (ioc_inode) + ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); - STACK_WIND (frame, ioc_writev_cbk, FIRST_CHILD(this), - FIRST_CHILD(this)->fops->writev, fd, vector, count, offset, - iobref); + STACK_WIND (frame, ioc_writev_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->writev, fd, vector, count, offset, + flags, iobref, xdata); - return 0; + return 0; } /* @@ -1264,12 +1277,12 @@ ioc_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t ioc_truncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, int32_t op_errno, struct iatt *prebuf, - struct iatt *postbuf) + struct iatt *postbuf, dict_t *xdata) { - STACK_UNWIND_STRICT (truncate, frame, op_ret, op_errno, prebuf, - postbuf); - return 0; + STACK_UNWIND_STRICT (truncate, frame, op_ret, op_errno, prebuf, + postbuf, xdata); + return 0; } @@ -1286,13 +1299,13 @@ ioc_truncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this, */ int32_t ioc_ftruncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct iatt *prebuf, - struct iatt *postbuf) + int32_t op_ret, int32_t op_errno, struct iatt *prebuf, + struct iatt *postbuf, dict_t *xdata) { - STACK_UNWIND_STRICT (ftruncate, frame, op_ret, op_errno, prebuf, - postbuf); - return 0; + STACK_UNWIND_STRICT (ftruncate, frame, op_ret, op_errno, prebuf, + postbuf, xdata); + return 0; } @@ -1306,17 +1319,19 @@ ioc_ftruncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this, * */ int32_t -ioc_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset) +ioc_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset, + dict_t *xdata) { - uint64_t ioc_inode = 0; - inode_ctx_get (loc->inode, this, &ioc_inode); + uint64_t ioc_inode = 0; - if (ioc_inode) - ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + inode_ctx_get (loc->inode, this, &ioc_inode); - STACK_WIND (frame, ioc_truncate_cbk, FIRST_CHILD(this), - FIRST_CHILD(this)->fops->truncate, loc, offset); - return 0; + if (ioc_inode) + ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + + STACK_WIND (frame, ioc_truncate_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->truncate, loc, offset, xdata); + return 0; } /* @@ -1329,68 +1344,151 @@ ioc_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset) * */ int32_t -ioc_ftruncate (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset) +ioc_ftruncate (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + dict_t *xdata) { - uint64_t ioc_inode = 0; - inode_ctx_get (fd->inode, this, &ioc_inode); + uint64_t ioc_inode = 0; - if (ioc_inode) - ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + inode_ctx_get (fd->inode, this, &ioc_inode); - STACK_WIND (frame, ioc_ftruncate_cbk, FIRST_CHILD(this), - FIRST_CHILD(this)->fops->ftruncate, fd, offset); - return 0; + if (ioc_inode) + ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + + STACK_WIND (frame, ioc_ftruncate_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->ftruncate, fd, offset, xdata); + return 0; } int32_t ioc_lk_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno, struct gf_flock *lock) + int32_t op_errno, struct gf_flock *lock, dict_t *xdata) { - STACK_UNWIND_STRICT (lk, frame, op_ret, op_errno, lock); - return 0; + STACK_UNWIND_STRICT (lk, frame, op_ret, op_errno, lock, xdata); + return 0; } int32_t ioc_lk (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t cmd, - struct gf_flock *lock) + struct gf_flock *lock, dict_t *xdata) +{ + ioc_inode_t *ioc_inode = NULL; + uint64_t tmp_inode = 0; + + inode_ctx_get (fd->inode, this, &tmp_inode); + ioc_inode = (ioc_inode_t *)(long)tmp_inode; + if (!ioc_inode) { + gf_log (this->name, GF_LOG_DEBUG, + "inode context is NULL: returning EBADFD"); + STACK_UNWIND_STRICT (lk, frame, -1, EBADFD, NULL, NULL); + return 0; + } + + ioc_inode_lock (ioc_inode); + { + gettimeofday (&ioc_inode->cache.tv, NULL); + } + ioc_inode_unlock (ioc_inode); + + STACK_WIND (frame, ioc_lk_cbk, FIRST_CHILD (this), + FIRST_CHILD (this)->fops->lk, fd, cmd, lock, xdata); + + return 0; +} + +int +ioc_readdirp_cbk (call_frame_t *frame, void *cookie, xlator_t *this, + int op_ret, int op_errno, gf_dirent_t *entries, dict_t *xdata) +{ + gf_dirent_t *entry = NULL; + + if (op_ret <= 0) + goto unwind; + + list_for_each_entry (entry, &entries->list, list) { + /* TODO: fill things */ + } + +unwind: + STACK_UNWIND_STRICT (readdirp, frame, op_ret, op_errno, entries, xdata); + + return 0; +} +int +ioc_readdirp (call_frame_t *frame, xlator_t *this, fd_t *fd, size_t size, + off_t offset, dict_t *dict) { - ioc_inode_t *ioc_inode = NULL; - uint64_t tmp_inode = 0; - - inode_ctx_get (fd->inode, this, &tmp_inode); - ioc_inode = (ioc_inode_t *)(long)tmp_inode; - if (!ioc_inode) { - gf_log (this->name, GF_LOG_DEBUG, - "inode context is NULL: returning EBADFD"); - STACK_UNWIND_STRICT (lk, frame, -1, EBADFD, NULL); - return 0; - } - - ioc_inode_lock (ioc_inode); - { - gettimeofday (&ioc_inode->cache.tv, NULL); - } - ioc_inode_unlock (ioc_inode); - - STACK_WIND (frame, ioc_lk_cbk, FIRST_CHILD (this), - FIRST_CHILD (this)->fops->lk, fd, cmd, lock); + STACK_WIND (frame, ioc_readdirp_cbk, + FIRST_CHILD(this), FIRST_CHILD(this)->fops->readdirp, + fd, size, offset, dict); + + return 0; +} +static int32_t +ioc_discard_cbk(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, struct iatt *pre, + struct iatt *post, dict_t *xdata) +{ + STACK_UNWIND_STRICT(discard, frame, op_ret, op_errno, pre, post, xdata); return 0; } +static int32_t +ioc_discard(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + size_t len, dict_t *xdata) +{ + uint64_t ioc_inode = 0; + + inode_ctx_get (fd->inode, this, &ioc_inode); + + if (ioc_inode) + ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + + STACK_WIND(frame, ioc_discard_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->discard, fd, offset, len, xdata); + return 0; +} + +static int32_t +ioc_zerofill_cbk(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, struct iatt *pre, + struct iatt *post, dict_t *xdata) +{ + STACK_UNWIND_STRICT(zerofill, frame, op_ret, + op_errno, pre, post, xdata); + return 0; +} + +static int32_t +ioc_zerofill(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + size_t len, dict_t *xdata) +{ + uint64_t ioc_inode = 0; + + inode_ctx_get (fd->inode, this, &ioc_inode); + + if (ioc_inode) + ioc_inode_flush ((ioc_inode_t *)(long)ioc_inode); + + STACK_WIND(frame, ioc_zerofill_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->zerofill, fd, offset, len, xdata); + return 0; +} + + int32_t ioc_get_priority_list (const char *opt_str, struct list_head *first) { - int32_t max_pri = 1; - char *tmp_str = NULL; - char *tmp_str1 = NULL; - char *tmp_str2 = NULL; - char *dup_str = NULL; - char *stripe_str = NULL; - char *pattern = NULL; - char *priority = NULL; - char *string = NULL; - struct ioc_priority *curr = NULL, *tmp = NULL; + int32_t max_pri = 1; + char *tmp_str = NULL; + char *tmp_str1 = NULL; + char *tmp_str2 = NULL; + char *dup_str = NULL; + char *stripe_str = NULL; + char *pattern = NULL; + char *priority = NULL; + char *string = NULL; + struct ioc_priority *curr = NULL, *tmp = NULL; string = gf_strdup (opt_str); if (string == NULL) { @@ -1398,73 +1496,69 @@ ioc_get_priority_list (const char *opt_str, struct list_head *first) goto out; } - /* Get the pattern for cache priority. - * "option priority *.jpg:1,abc*:2" etc - */ - /* TODO: inode_lru in table is statically hard-coded to 5, - * should be changed to run-time configuration - */ - stripe_str = strtok_r (string, ",", &tmp_str); - while (stripe_str) { - curr = GF_CALLOC (1, sizeof (struct ioc_priority), + /* Get the pattern for cache priority. + * "option priority *.jpg:1,abc*:2" etc + */ + /* TODO: inode_lru in table is statically hard-coded to 5, + * should be changed to run-time configuration + */ + stripe_str = strtok_r (string, ",", &tmp_str); + while (stripe_str) { + curr = GF_CALLOC (1, sizeof (struct ioc_priority), gf_ioc_mt_ioc_priority); if (curr == NULL) { max_pri = -1; goto out; } - list_add_tail (&curr->list, first); + list_add_tail (&curr->list, first); - dup_str = gf_strdup (stripe_str); + dup_str = gf_strdup (stripe_str); if (dup_str == NULL) { max_pri = -1; goto out; } - pattern = strtok_r (dup_str, ":", &tmp_str1); - if (!pattern) { + pattern = strtok_r (dup_str, ":", &tmp_str1); + if (!pattern) { max_pri = -1; goto out; } - priority = strtok_r (NULL, ":", &tmp_str1); - if (!priority) { + priority = strtok_r (NULL, ":", &tmp_str1); + if (!priority) { max_pri = -1; goto out; } - gf_log ("io-cache", GF_LOG_TRACE, - "ioc priority : pattern %s : priority %s", - pattern, - priority); + gf_log ("io-cache", GF_LOG_TRACE, + "ioc priority : pattern %s : priority %s", + pattern, + priority); - curr->pattern = gf_strdup (pattern); + curr->pattern = gf_strdup (pattern); if (curr->pattern == NULL) { max_pri = -1; goto out; } - curr->priority = strtol (priority, &tmp_str2, 0); - if (tmp_str2 && (*tmp_str2)) { + curr->priority = strtol (priority, &tmp_str2, 0); + if (tmp_str2 && (*tmp_str2)) { max_pri = -1; goto out; } else { - max_pri = max (max_pri, curr->priority); + max_pri = max (max_pri, curr->priority); } GF_FREE (dup_str); dup_str = NULL; - stripe_str = strtok_r (NULL, ",", &tmp_str); - } -out: - if (string != NULL) { - GF_FREE (string); + stripe_str = strtok_r (NULL, ",", &tmp_str); } +out: + GF_FREE (string); - if (dup_str != NULL) { - GF_FREE (dup_str); - } + GF_FREE (dup_str); if (max_pri == -1) { list_for_each_entry_safe (curr, tmp, first, list) { @@ -1474,7 +1568,7 @@ out: } } - return max_pri; + return max_pri; } int32_t @@ -1489,329 +1583,119 @@ mem_acct_init (xlator_t *this) if (ret != 0) { gf_log (this->name, GF_LOG_ERROR, "Memory accounting init" - "failed"); + "failed"); return ret; } return ret; } -int -validate_options (xlator_t *this, dict_t *options, char **op_errstr) -{ - int32_t cache_timeout; - int64_t min_file_size = 0; - int64_t max_file_size = 0; - char *tmp = NULL; - uint64_t cache_size; - char *cache_size_string = NULL; - int ret = 0; - - - if (dict_get (options, "cache-timeout")) { - cache_timeout = data_to_uint32 (dict_get (options, - "cache-timeout")); - if (cache_timeout < 0){ - gf_log (this->name, GF_LOG_WARNING, - "cache-timeout %d seconds invalid," - " has to be >=0", cache_timeout); - *op_errstr = gf_strdup ("Error, should be >= 0"); - ret = -1; - goto out; - } - - - if (cache_timeout > 60){ - gf_log (this->name, GF_LOG_WARNING, - "cache-timeout %d seconds invalid," - " has to be <=60", cache_timeout); - *op_errstr = gf_strdup ("Error, should be <= 60"); - ret = -1; - goto out; - } - - - gf_log (this->name, GF_LOG_DEBUG, - "Validated cache-timeout revalidate cache"); - } - - - if (dict_get (options, "cache-size")) - cache_size_string = data_to_str (dict_get (options, - "cache-size")); - if (cache_size_string) { - if (gf_string2bytesize (cache_size_string, - &cache_size) != 0) { - gf_log ("io-cache", GF_LOG_ERROR, - "invalid number format \"%s\" of " - "\"option cache-size\" Defaulting" - "to old value", cache_size_string); - *op_errstr = gf_strdup ("Error, Invalid Format"); - ret = -1; - goto out; - } - - if (cache_size < ( 4 * GF_UNIT_MB)) { - gf_log(this->name, GF_LOG_WARNING, "Reconfiguration" - "'option cache-size %s' failed , Max value" - "can be 4MiB, Defaulting to old value " - "(%"PRIu64")", cache_size_string, - cache_size); - *op_errstr = gf_strdup ("Error, " - "Cannot be less than 4MB"); - ret = -1; - goto out; - } - - if (cache_size > ( 6 * GF_UNIT_GB)) { - gf_log(this->name, GF_LOG_WARNING, "Validation" - "'option cache-size %s' failed , Max value" - "can be 6GiB, Defaulting to old value " - "(%"PRIu64")", cache_size_string, - cache_size); - *op_errstr = gf_strdup ("Error, Cannot be more " - "than 6GB"); - ret = -1; - goto out; - } - - - gf_log (this->name, GF_LOG_DEBUG, "Validated " - " cache-size %"PRIu64"", cache_size); - } - - - tmp = data_to_str (dict_get (options, "min-file-size")); - if (tmp != NULL) { - if (gf_string2bytesize (tmp, - (uint64_t *)&min_file_size) - != 0) { - gf_log ("io-cache", GF_LOG_WARNING, - "invalid number format \"%s\" of " - "\"option min-file-size\"", tmp); - *op_errstr = gf_strdup ("Error, Invalid Format"); - ret = -1; - goto out; - } - - gf_log (this->name, GF_LOG_DEBUG, - "Validated min-file-size %"PRIu64"", - min_file_size); +static gf_boolean_t +check_cache_size_ok (xlator_t *this, uint64_t cache_size) +{ + gf_boolean_t ret = _gf_true; + uint64_t total_mem = 0; + uint64_t max_cache_size = 0; + volume_option_t *opt = NULL; + + GF_ASSERT (this); + opt = xlator_volume_option_get (this, "cache-size"); + if (!opt) { + ret = _gf_false; + gf_log (this->name, GF_LOG_ERROR, + "could not get cache-size option"); + goto out; } + total_mem = get_mem_size (); + if (-1 == total_mem) + max_cache_size = opt->max; + else + max_cache_size = total_mem; - tmp = data_to_str (dict_get (options, "max-file-size")); - if (tmp != NULL) { - if (gf_string2bytesize (tmp, - (uint64_t *)&max_file_size) - != 0) { - gf_log ("io-cache", GF_LOG_WARNING, - "invalid number format \"%s\" of " - "\"option max-file-size\"", tmp); - *op_errstr = gf_strdup ("Error, Invalid Format"); - ret = -1; - goto out; - } + gf_log (this->name, GF_LOG_DEBUG, "Max cache size is %"PRIu64, + max_cache_size); - - gf_log (this->name, GF_LOG_WARNING, - "Validated max-file-size %"PRIu64"", - max_file_size); - } - - if ((max_file_size >= 0) & (min_file_size > max_file_size)) { - gf_log ("io-cache", GF_LOG_WARNING, "minimum size (%" - PRIu64") of a file that can be cached is " - "greater than maximum size (%"PRIu64"). ", - min_file_size, max_file_size); - *op_errstr = gf_strdup ("Error, min-file-size greater" - "than max-file-size"); - ret = -1; + if (cache_size > max_cache_size) { + ret = _gf_false; + gf_log (this->name, GF_LOG_ERROR, "Cache size %"PRIu64 + " is greater than the max size of %"PRIu64, + cache_size, max_cache_size); goto out; } - - out: - return ret; - + return ret; } int reconfigure (xlator_t *this, dict_t *options) { - ioc_table_t *table = NULL; - int32_t cache_timeout; - int64_t min_file_size = 0; - int64_t max_file_size = 0; - char *tmp = NULL; - uint64_t cache_size; - char *cache_size_string = NULL; - int ret = 0; - + data_t *data = NULL; + ioc_table_t *table = NULL; + int ret = -1; + uint64_t cache_size_new = 0; if (!this || !this->private) goto out; - table = this->private; - - ioc_table_lock (table); - { - if (dict_get (options, "cache-timeout")) { - cache_timeout = - data_to_uint32 (dict_get (options, - "cache-timeout")); - if (cache_timeout < 0){ - gf_log (this->name, GF_LOG_WARNING, - "cache-timeout %d seconds invalid," - " has to be >=0", cache_timeout); - ret = -1; - goto out; - } - - - if (cache_timeout > 60){ - gf_log (this->name, GF_LOG_WARNING, - "cache-timeout %d seconds invalid," - " has to be <=60", cache_timeout); - ret = -1; - goto out; - } - - table->cache_timeout = cache_timeout; - - gf_log (this->name, GF_LOG_DEBUG, - "Reconfiguring %d seconds to" - " revalidate cache", table->cache_timeout); - } - else - table->cache_timeout = 1; - - - - if (dict_get (options, "cache-size")) - cache_size_string = data_to_str (dict_get (options, - "cache-size")); - if (cache_size_string) { - if (gf_string2bytesize (cache_size_string, - &cache_size) != 0) { - gf_log ("io-cache", GF_LOG_ERROR, - "invalid number format \"%s\" of " - "\"option cache-size\" Defaulting" - "to old value", cache_size_string); - ret = -1; - goto out; - } - - if (cache_size < (4 * GF_UNIT_MB)) { - gf_log(this->name, GF_LOG_ERROR, - "Reconfiguration" - "'option cache-size %s' failed , " - "Max value can be 4MiB, Defaulting to " - "old value (%"PRIu64")", - cache_size_string, table->cache_size); - ret = -1; - goto out; - } - - if (cache_size > (6 * GF_UNIT_GB)) { - gf_log (this->name, GF_LOG_ERROR, - "Reconfiguration" - "'option cache-size %s' failed , " - "Max value can be 6GiB, Defaulting to " - "old value (%"PRIu64")", - cache_size_string, table->cache_size); - ret = -1; - goto out; - } - - - gf_log (this->name, GF_LOG_DEBUG, "Reconfiguring " - " cache-size %"PRIu64"", cache_size); - table->cache_size = cache_size; - } - else - table->cache_size = IOC_CACHE_SIZE; - - - if (dict_get (options, "priority")) { - char *option_list = data_to_str (dict_get (options, - "priority")); - gf_log (this->name, GF_LOG_TRACE, - "option path %s", option_list); - /* parse the list of pattern:priority */ - table->max_pri = ioc_get_priority_list (option_list, - &table->priority_list); - - if (table->max_pri == -1) { - ret = -1; - goto out; - } - table->max_pri ++; - } - - - - min_file_size = table->min_file_size; - tmp = data_to_str (dict_get (options, "min-file-size")); - if (tmp != NULL) { - if (gf_string2bytesize (tmp, - (uint64_t *)&min_file_size) - != 0) { - gf_log ("io-cache", GF_LOG_ERROR, - "invalid number format \"%s\" of " - "\"option min-file-size\"", tmp); - ret = -1; - goto out; - } - - gf_log (this->name, GF_LOG_DEBUG, - "Reconfiguring min-file-size %"PRIu64"", - table->min_file_size); - } - - max_file_size = table->max_file_size; - tmp = data_to_str (dict_get (options, "max-file-size")); - if (tmp != NULL) { - if (gf_string2bytesize (tmp, - (uint64_t *)&max_file_size) - != 0) { - gf_log ("io-cache", GF_LOG_ERROR, - "invalid number format \"%s\" of " - "\"option max-file-size\"", tmp); - ret = -1; - goto out; - } - - - gf_log (this->name, GF_LOG_DEBUG, - "Reconfiguring max-file-size %"PRIu64"", - table->max_file_size); - } - - if ((max_file_size >= 0) & (min_file_size > max_file_size)) { - gf_log ("io-cache", GF_LOG_ERROR, "minimum size (%" + table = this->private; + + ioc_table_lock (table); + { + GF_OPTION_RECONF ("cache-timeout", table->cache_timeout, + options, int32, unlock); + + data = dict_get (options, "priority"); + if (data) { + char *option_list = data_to_str (data); + + gf_log (this->name, GF_LOG_TRACE, + "option path %s", option_list); + /* parse the list of pattern:priority */ + table->max_pri = ioc_get_priority_list (option_list, + &table->priority_list); + + if (table->max_pri == -1) { + goto unlock; + } + table->max_pri ++; + } + + GF_OPTION_RECONF ("max-file-size", table->max_file_size, + options, size, unlock); + + GF_OPTION_RECONF ("min-file-size", table->min_file_size, + options, size, unlock); + + if ((table->max_file_size >= 0) && + (table->min_file_size > table->max_file_size)) { + gf_log (this->name, GF_LOG_ERROR, "minimum size (%" PRIu64") of a file that can be cached is " "greater than maximum size (%"PRIu64"). " - "Hence Defaulting to old value", + "Hence Defaulting to old value", table->min_file_size, table->max_file_size); + goto unlock; + } + + GF_OPTION_RECONF ("cache-size", cache_size_new, + options, size, unlock); + if (!check_cache_size_ok (this, cache_size_new)) { ret = -1; - goto out; - } - - table->min_file_size = min_file_size; - table->max_file_size = max_file_size; - if (!data_to_str (dict_get (options, "min-file-size"))) - table->min_file_size = 0; - if (data_to_str (dict_get (options, "max-file-size"))) - table->max_file_size = 0; - } - - ioc_table_unlock (table); -out: - return ret; + gf_log (this->name, GF_LOG_ERROR, + "Not reconfiguring cache-size"); + goto unlock; + } + table->cache_size = cache_size_new; + ret = 0; + } +unlock: + ioc_table_unlock (table); +out: + return ret; } + /* * init - * @this: @@ -1820,133 +1704,110 @@ out: int32_t init (xlator_t *this) { - ioc_table_t *table = NULL; - dict_t *options = this->options; - uint32_t index = 0; - char *cache_size_string = NULL, *tmp = NULL; - int32_t ret = -1; - glusterfs_ctx_t *ctx = NULL; - - if (!this->children || this->children->next) { - gf_log (this->name, GF_LOG_ERROR, - "FATAL: io-cache not configured with exactly " - "one child"); + ioc_table_t *table = NULL; + dict_t *xl_options = NULL; + uint32_t index = 0; + int32_t ret = -1; + glusterfs_ctx_t *ctx = NULL; + data_t *data = 0; + uint32_t num_pages = 0; + + xl_options = this->options; + + if (!this->children || this->children->next) { + gf_log (this->name, GF_LOG_ERROR, + "FATAL: io-cache not configured with exactly " + "one child"); goto out; - } + } - if (!this->parents) { - gf_log (this->name, GF_LOG_WARNING, - "dangling volume. check volfile "); - } + if (!this->parents) { + gf_log (this->name, GF_LOG_WARNING, + "dangling volume. check volfile "); + } - table = (void *) GF_CALLOC (1, sizeof (*table), gf_ioc_mt_ioc_table_t); + table = (void *) GF_CALLOC (1, sizeof (*table), gf_ioc_mt_ioc_table_t); if (table == NULL) { gf_log (this->name, GF_LOG_ERROR, "out of memory"); goto out; } - table->xl = this; - table->page_size = this->ctx->page_size; - table->cache_size = IOC_CACHE_SIZE; - - if (dict_get (options, "cache-size")) - cache_size_string = data_to_str (dict_get (options, - "cache-size")); - if (cache_size_string) { - if (gf_string2bytesize (cache_size_string, - &table->cache_size) != 0) { - gf_log ("io-cache", GF_LOG_ERROR, - "invalid number format \"%s\" of " - "\"option cache-size\"", - cache_size_string); - goto out; - } - - gf_log (this->name, GF_LOG_TRACE, - "using cache-size %"PRIu64"", table->cache_size); - } - - table->cache_timeout = 1; - - if (dict_get (options, "cache-timeout")) { - table->cache_timeout = - data_to_uint32 (dict_get (options, - "cache-timeout")); - gf_log (this->name, GF_LOG_TRACE, - "Using %d seconds to revalidate cache", - table->cache_timeout); - } - - INIT_LIST_HEAD (&table->priority_list); - table->max_pri = 1; - if (dict_get (options, "priority")) { - char *option_list = data_to_str (dict_get (options, - "priority")); - gf_log (this->name, GF_LOG_TRACE, - "option path %s", option_list); - /* parse the list of pattern:priority */ - table->max_pri = ioc_get_priority_list (option_list, - &table->priority_list); - - if (table->max_pri == -1) { - goto out; - } - } - table->max_pri ++; - - table->min_file_size = 0; - - tmp = data_to_str (dict_get (options, "min-file-size")); - if (tmp != NULL) { - if (gf_string2bytesize (tmp, - (uint64_t *)&table->min_file_size) != 0) { - gf_log ("io-cache", GF_LOG_ERROR, - "invalid number format \"%s\" of " - "\"option min-file-size\"", tmp); - goto out; - } + table->xl = this; + table->page_size = this->ctx->page_size; + + GF_OPTION_INIT ("cache-size", table->cache_size, size, out); - gf_log (this->name, GF_LOG_TRACE, - "using min-file-size %"PRIu64"", table->min_file_size); + GF_OPTION_INIT ("cache-timeout", table->cache_timeout, int32, out); + + GF_OPTION_INIT ("min-file-size", table->min_file_size, size, out); + + GF_OPTION_INIT ("max-file-size", table->max_file_size, size, out); + + if (!check_cache_size_ok (this, table->cache_size)) { + ret = -1; + goto out; } - table->max_file_size = -1; - tmp = data_to_str (dict_get (options, "max-file-size")); - if (tmp != NULL) { - if (gf_string2bytesize (tmp, - (uint64_t *)&table->max_file_size) != 0) { - gf_log ("io-cache", GF_LOG_ERROR, - "invalid number format \"%s\" of " - "\"option max-file-size\"", tmp); + INIT_LIST_HEAD (&table->priority_list); + table->max_pri = 1; + data = dict_get (xl_options, "priority"); + if (data) { + char *option_list = data_to_str (data); + gf_log (this->name, GF_LOG_TRACE, + "option path %s", option_list); + /* parse the list of pattern:priority */ + table->max_pri = ioc_get_priority_list (option_list, + &table->priority_list); + + if (table->max_pri == -1) { goto out; } - - gf_log (this->name, GF_LOG_TRACE, - "using max-file-size %"PRIu64"", table->max_file_size); } - INIT_LIST_HEAD (&table->inodes); + table->max_pri ++; + + INIT_LIST_HEAD (&table->inodes); if ((table->max_file_size >= 0) && (table->min_file_size > table->max_file_size)) { - gf_log ("io-cache", GF_LOG_ERROR, "minimum size (%" - PRIu64") of a file that can be cached is " - "greater than maximum size (%"PRIu64")", - table->min_file_size, table->max_file_size); - goto out; + gf_log ("io-cache", GF_LOG_ERROR, "minimum size (%" + PRIu64") of a file that can be cached is " + "greater than maximum size (%"PRIu64")", + table->min_file_size, table->max_file_size); + goto out; } - table->inode_lru = GF_CALLOC (table->max_pri, + table->inode_lru = GF_CALLOC (table->max_pri, sizeof (struct list_head), gf_ioc_mt_list_head); if (table->inode_lru == NULL) { goto out; } - for (index = 0; index < (table->max_pri); index++) - INIT_LIST_HEAD (&table->inode_lru[index]); + for (index = 0; index < (table->max_pri); index++) + INIT_LIST_HEAD (&table->inode_lru[index]); + + this->local_pool = mem_pool_new (ioc_local_t, 64); + if (!this->local_pool) { + ret = -1; + gf_log (this->name, GF_LOG_ERROR, + "failed to create local_t's memory pool"); + goto out; + } + + pthread_mutex_init (&table->table_lock, NULL); + this->private = table; + + num_pages = (table->cache_size / table->page_size) + + ((table->cache_size % table->page_size) + ? 1 : 0); + + table->mem_pool = mem_pool_new (rbthash_entry_t, num_pages); + if (!table->mem_pool) { + gf_log (this->name, GF_LOG_ERROR, + "Unable to allocate mem_pool"); + goto out; + } - pthread_mutex_init (&table->table_lock, NULL); - this->private = table; ret = 0; ctx = this->ctx; @@ -1960,34 +1821,227 @@ out: } } - return ret; + return ret; +} + +void +ioc_page_waitq_dump (ioc_page_t *page, char *prefix) +{ + ioc_waitq_t *trav = NULL; + call_frame_t *frame = NULL; + int32_t i = 0; + char key[GF_DUMP_MAX_BUF_LEN] = {0, }; + + trav = page->waitq; + + while (trav) { + frame = trav->data; + sprintf (key, "waitq.frame[%d]", i++); + gf_proc_dump_write (key, "%"PRId64, frame->root->unique); + + trav = trav->next; + } +} + +void +__ioc_inode_waitq_dump (ioc_inode_t *ioc_inode, char *prefix) +{ + ioc_waitq_t *trav = NULL; + ioc_page_t *page = NULL; + int32_t i = 0; + char key[GF_DUMP_MAX_BUF_LEN] = {0, }; + + trav = ioc_inode->waitq; + + while (trav) { + page = trav->data; + + sprintf (key, "cache-validation-waitq.page[%d].offset", i++); + gf_proc_dump_write (key, "%"PRId64, page->offset); + + trav = trav->next; + } +} + +void +__ioc_page_dump (ioc_page_t *page, char *prefix) +{ + + int ret = -1; + + if (!page) + return; + /* ioc_page_lock can be used to hold the mutex. But in statedump + * its better to use trylock to avoid deadlocks. + */ + ret = pthread_mutex_trylock (&page->page_lock); + if (ret) + goto out; + { + gf_proc_dump_write ("offset", "%"PRId64, page->offset); + gf_proc_dump_write ("size", "%"PRId64, page->size); + gf_proc_dump_write ("dirty", "%s", page->dirty ? "yes" : "no"); + gf_proc_dump_write ("ready", "%s", page->ready ? "yes" : "no"); + ioc_page_waitq_dump (page, prefix); + } + pthread_mutex_unlock (&page->page_lock); + +out: + if (ret && page) + gf_proc_dump_write ("Unable to dump the page information", + "(Lock acquisition failed) %p", page); + + return; +} + +void +__ioc_cache_dump (ioc_inode_t *ioc_inode, char *prefix) +{ + off_t offset = 0; + ioc_table_t *table = NULL; + ioc_page_t *page = NULL; + int i = 0; + char key[GF_DUMP_MAX_BUF_LEN] = {0, }; + char timestr[256] = {0, }; + + if ((ioc_inode == NULL) || (prefix == NULL)) { + goto out; + } + + table = ioc_inode->table; + + if (ioc_inode->cache.tv.tv_sec) { + gf_time_fmt (timestr, sizeof timestr, + ioc_inode->cache.tv.tv_sec, gf_timefmt_FT); + snprintf (timestr + strlen (timestr), sizeof timestr - strlen (timestr), + ".%"GF_PRI_SUSECONDS, ioc_inode->cache.tv.tv_usec); + + gf_proc_dump_write ("last-cache-validation-time", "%s", + timestr); + } + + for (offset = 0; offset < ioc_inode->ia_size; + offset += table->page_size) { + page = __ioc_page_get (ioc_inode, offset); + if (page == NULL) { + continue; + } + + sprintf (key, "inode.cache.page[%d]", i++); + __ioc_page_dump (page, key); + } +out: + return; +} + + +int +ioc_inode_dump (xlator_t *this, inode_t *inode) +{ + + char *path = NULL; + int ret = -1; + char key_prefix[GF_DUMP_MAX_BUF_LEN] = {0, }; + uint64_t tmp_ioc_inode = 0; + ioc_inode_t *ioc_inode = NULL; + gf_boolean_t section_added = _gf_false; + char uuid_str[64] = {0,}; + + if (this == NULL || inode == NULL) + goto out; + + gf_proc_dump_build_key (key_prefix, "io-cache", "inode"); + + inode_ctx_get (inode, this, &tmp_ioc_inode); + ioc_inode = (ioc_inode_t *)(long)tmp_ioc_inode; + if (ioc_inode == NULL) + goto out; + + /* Similar to ioc_page_dump function its better to use + * pthread_mutex_trylock and not to use gf_log in statedump + * to avoid deadlocks. + */ + ret = pthread_mutex_trylock (&ioc_inode->inode_lock); + if (ret) + goto out; + + { + if (uuid_is_null (ioc_inode->inode->gfid)) + goto unlock; + + gf_proc_dump_add_section (key_prefix); + section_added = _gf_true; + + __inode_path (ioc_inode->inode, NULL, &path); + + gf_proc_dump_write ("inode.weight", "%d", ioc_inode->weight); + + if (path) { + gf_proc_dump_write ("path", "%s", path); + GF_FREE (path); + } + + gf_proc_dump_write ("uuid", "%s", uuid_utoa_r + (ioc_inode->inode->gfid, uuid_str)); + __ioc_cache_dump (ioc_inode, key_prefix); + __ioc_inode_waitq_dump (ioc_inode, key_prefix); + } +unlock: + pthread_mutex_unlock (&ioc_inode->inode_lock); + +out: + if (ret && ioc_inode) { + if (section_added == _gf_false) + gf_proc_dump_add_section (key_prefix); + gf_proc_dump_write ("Unable to print the status of ioc_inode", + "(Lock acquisition failed) %s", + uuid_utoa (inode->gfid)); + } + return ret; } int ioc_priv_dump (xlator_t *this) { - ioc_table_t *priv = NULL; - char key_prefix[GF_DUMP_MAX_BUF_LEN]; - char key[GF_DUMP_MAX_BUF_LEN]; + ioc_table_t *priv = NULL; + char key_prefix[GF_DUMP_MAX_BUF_LEN] = {0, }; + int ret = -1; + gf_boolean_t add_section = _gf_false; if (!this || !this->private) goto out; priv = this->private; - gf_proc_dump_build_key (key_prefix, "xlator.performance.io-cache", - "priv"); - gf_proc_dump_add_section (key_prefix); - gf_proc_dump_build_key (key, key_prefix, "page_size"); - gf_proc_dump_write (key, "%ld", priv->page_size); - gf_proc_dump_build_key (key, key_prefix, "cache_size"); - gf_proc_dump_write (key, "%ld", priv->cache_size); - gf_proc_dump_build_key (key, key_prefix, "cache_used"); - gf_proc_dump_write (key, "%ld", priv->cache_used); - gf_proc_dump_build_key (key, key_prefix, "inode_count"); - gf_proc_dump_write (key, "%u", priv->inode_count); + gf_proc_dump_build_key (key_prefix, "io-cache", "priv"); + gf_proc_dump_add_section (key_prefix); + add_section = _gf_true; + ret = pthread_mutex_trylock (&priv->table_lock); + if (ret) + goto out; + { + gf_proc_dump_write ("page_size", "%ld", priv->page_size); + gf_proc_dump_write ("cache_size", "%ld", priv->cache_size); + gf_proc_dump_write ("cache_used", "%ld", priv->cache_used); + gf_proc_dump_write ("inode_count", "%u", priv->inode_count); + gf_proc_dump_write ("cache_timeout", "%u", priv->cache_timeout); + gf_proc_dump_write ("min-file-size", "%u", priv->min_file_size); + gf_proc_dump_write ("max-file-size", "%u", priv->max_file_size); + } + pthread_mutex_unlock (&priv->table_lock); out: + if (ret && priv) { + if (!add_section) { + gf_proc_dump_build_key (key_prefix, "xlator." + "performance.io-cache", "priv"); + gf_proc_dump_add_section (key_prefix); + } + gf_proc_dump_write ("Unable to dump the state of private " + "structure of io-cache xlator", "(Lock " + "acquisition failed) %s", this->name); + } + return 0; } @@ -2000,71 +2054,105 @@ out: void fini (xlator_t *this) { - ioc_table_t *table = NULL; + ioc_table_t *table = NULL; + struct ioc_priority *curr = NULL, *tmp = NULL; + int i = 0; table = this->private; if (table == NULL) return; + this->private = NULL; + if (table->mem_pool != NULL) { mem_pool_destroy (table->mem_pool); table->mem_pool = NULL; } - pthread_mutex_destroy (&table->table_lock); - GF_FREE (table); + list_for_each_entry_safe (curr, tmp, &table->priority_list, list) { + list_del_init (&curr->list); + GF_FREE (curr->pattern); + GF_FREE (curr); + } + + for (i = 0; i < table->max_pri; i++) { + GF_ASSERT (list_empty (&table->inode_lru[i])); + } - this->private = NULL; - return; + GF_ASSERT (list_empty (&table->inodes)); + pthread_mutex_destroy (&table->table_lock); + GF_FREE (table); + + this->private = NULL; + return; } struct xlator_fops fops = { - .open = ioc_open, - .create = ioc_create, - .readv = ioc_readv, - .writev = ioc_writev, - .truncate = ioc_truncate, - .ftruncate = ioc_ftruncate, - .lookup = ioc_lookup, - .lk = ioc_lk, + .open = ioc_open, + .create = ioc_create, + .readv = ioc_readv, + .writev = ioc_writev, + .truncate = ioc_truncate, + .ftruncate = ioc_ftruncate, + .lookup = ioc_lookup, + .lk = ioc_lk, .setattr = ioc_setattr, - .mknod = ioc_mknod + .mknod = ioc_mknod, + + .readdirp = ioc_readdirp, + .discard = ioc_discard, + .zerofill = ioc_zerofill, }; struct xlator_dumpops dumpops = { .priv = ioc_priv_dump, + .inodectx = ioc_inode_dump, }; struct xlator_cbks cbks = { - .forget = ioc_forget, - .release = ioc_release + .forget = ioc_forget, + .release = ioc_release, + .invalidate = ioc_invalidate, }; struct volume_options options[] = { - { .key = {"priority"}, - .type = GF_OPTION_TYPE_ANY - }, - { .key = {"cache-timeout", "force-revalidate-timeout"}, - .type = GF_OPTION_TYPE_INT, - .min = 0, - .max = 60 - }, - { .key = {"cache-size"}, - .type = GF_OPTION_TYPE_SIZET, - .min = 4 * GF_UNIT_MB, - .max = 6 * GF_UNIT_GB - }, + { .key = {"priority"}, + .type = GF_OPTION_TYPE_PRIORITY_LIST, + .default_value = "", + .description = "Assigns priority to filenames with specific " + "patterns so that when a page needs to be ejected " + "out of the cache, the page of a file whose " + "priority is the lowest will be ejected earlier" + }, + { .key = {"cache-timeout", "force-revalidate-timeout"}, + .type = GF_OPTION_TYPE_INT, + .min = 0, + .max = 60, + .default_value = "1", + .description = "The cached data for a file will be retained till " + "'cache-refresh-timeout' seconds, after which data " + "re-validation is performed." + }, + { .key = {"cache-size"}, + .type = GF_OPTION_TYPE_SIZET, + .min = 4 * GF_UNIT_MB, + .max = 32 * GF_UNIT_GB, + .default_value = "32MB", + .description = "Size of the read cache." + }, { .key = {"min-file-size"}, .type = GF_OPTION_TYPE_SIZET, - .min = -1, - .max = -1 + .default_value = "0", + .description = "Minimum file size which would be cached by the " + "io-cache translator." }, { .key = {"max-file-size"}, .type = GF_OPTION_TYPE_SIZET, - .min = -1, - .max = -1 + .default_value = "0", + .description = "Maximum file size which would be cached by the " + "io-cache translator." }, - { .key = {NULL} }, + { .key = {NULL} }, }; diff --git a/xlators/performance/io-cache/src/io-cache.h b/xlators/performance/io-cache/src/io-cache.h index e9cb56e0a..46d758a66 100644 --- a/xlators/performance/io-cache/src/io-cache.h +++ b/xlators/performance/io-cache/src/io-cache.h @@ -1,20 +1,11 @@ /* - Copyright (c) 2007-2010 Gluster, Inc. <http://www.gluster.com> + Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> This file is part of GlusterFS. - GlusterFS is free software; you can redistribute it and/or modify - it under the terms of the GNU Affero General Public License as published - by the Free Software Foundation; either version 3 of the License, - or (at your option) any later version. - - GlusterFS is distributed in the hope that it will be useful, but - WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Affero General Public License for more details. - - You should have received a copy of the GNU Affero General Public License - along with this program. If not, see - <http://www.gnu.org/licenses/>. + This file is licensed to you under your choice of the GNU Lesser + General Public License, version 3 or any later version (LGPLv3 or + later), or the GNU General Public License, version 2 (GPLv2), in all + cases as published by the Free Software Foundation. */ #ifndef __IO_CACHE_H @@ -49,132 +40,134 @@ struct ioc_page; struct ioc_inode; struct ioc_priority { - struct list_head list; - char *pattern; - uint32_t priority; + struct list_head list; + char *pattern; + uint32_t priority; }; /* - * ioc_waitq - this structure is used to represents the waiting + * ioc_waitq - this structure is used to represents the waiting * frames on a page * * @next: pointer to next object in waitq * @data: pointer to the frame which is waiting */ struct ioc_waitq { - struct ioc_waitq *next; - void *data; - off_t pending_offset; - size_t pending_size; + struct ioc_waitq *next; + void *data; + off_t pending_offset; + size_t pending_size; }; /* - * ioc_fill - + * ioc_fill - * */ struct ioc_fill { - struct list_head list; /* list of ioc_fill structures of a frame */ - off_t offset; - size_t size; - struct iovec *vector; - int32_t count; - struct iobref *iobref; + struct list_head list; /* list of ioc_fill structures of a frame */ + off_t offset; + size_t size; + struct iovec *vector; + int32_t count; + struct iobref *iobref; }; struct ioc_local { - mode_t mode; - int32_t flags; - int32_t wbflags; - loc_t file_loc; - off_t offset; - size_t size; - int32_t op_ret; - int32_t op_errno; - struct list_head fill_list; /* list of ioc_fill structures */ - off_t pending_offset; /* + mode_t mode; + int32_t flags; + loc_t file_loc; + off_t offset; + size_t size; + int32_t op_ret; + int32_t op_errno; + struct list_head fill_list; /* list of ioc_fill structures */ + off_t pending_offset; /* * offset from this frame should * continue */ - size_t pending_size; /* + size_t pending_size; /* * size of data this frame is waiting * on */ - struct ioc_inode *inode; - int32_t wait_count; - pthread_mutex_t local_lock; - struct ioc_waitq *waitq; - void *stub; - fd_t *fd; - int32_t need_xattr; - dict_t *xattr_req; + struct ioc_inode *inode; + int32_t wait_count; + pthread_mutex_t local_lock; + struct ioc_waitq *waitq; + void *stub; + fd_t *fd; + int32_t need_xattr; + dict_t *xattr_req; }; /* - * ioc_page - structure to store page of data from file + * ioc_page - structure to store page of data from file * */ struct ioc_page { - struct list_head page_lru; - struct ioc_inode *inode; /* inode this page belongs to */ - struct ioc_priority *priority; - char dirty; - char ready; - struct iovec *vector; - int32_t count; - off_t offset; - size_t size; - struct ioc_waitq *waitq; - struct iobref *iobref; - pthread_mutex_t page_lock; + struct list_head page_lru; + struct ioc_inode *inode; /* inode this page belongs to */ + struct ioc_priority *priority; + char dirty; + char ready; + struct iovec *vector; + int32_t count; + off_t offset; + size_t size; + struct ioc_waitq *waitq; + struct iobref *iobref; + pthread_mutex_t page_lock; + int32_t op_errno; + char stale; }; struct ioc_cache { rbthash_table_t *page_table; struct list_head page_lru; - time_t mtime; /* + time_t mtime; /* * seconds component of file mtime */ time_t mtime_nsec; /* * nanosecond component of file mtime */ - struct timeval tv; /* + struct timeval tv; /* * time-stamp at last re-validate */ }; struct ioc_inode { - struct ioc_table *table; + struct ioc_table *table; off_t ia_size; - struct ioc_cache cache; - struct list_head inode_list; /* + struct ioc_cache cache; + struct list_head inode_list; /* * list of inodes, maintained by * io-cache translator */ - struct list_head inode_lru; - struct ioc_waitq *waitq; - pthread_mutex_t inode_lock; - uint32_t weight; /* + struct list_head inode_lru; + struct ioc_waitq *waitq; + pthread_mutex_t inode_lock; + uint32_t weight; /* * weight of the inode, increases * on each read */ + inode_t *inode; }; struct ioc_table { - uint64_t page_size; - uint64_t cache_size; - uint64_t cache_used; - int64_t min_file_size; - int64_t max_file_size; - struct list_head inodes; /* list of inodes cached */ - struct list_head active; - struct list_head *inode_lru; - struct list_head priority_list; - int32_t readv_count; - pthread_mutex_t table_lock; - xlator_t *xl; - uint32_t inode_count; - int32_t cache_timeout; - int32_t max_pri; + uint64_t page_size; + uint64_t cache_size; + uint64_t cache_used; + uint64_t min_file_size; + uint64_t max_file_size; + struct list_head inodes; /* list of inodes cached */ + struct list_head active; + struct list_head *inode_lru; + struct list_head priority_list; + int32_t readv_count; + pthread_mutex_t table_lock; + xlator_t *xl; + uint32_t inode_count; + int32_t cache_timeout; + int32_t max_pri; struct mem_pool *mem_pool; }; @@ -191,36 +184,33 @@ str_to_ptr (char *string); char * ptr_to_str (void *ptr); -int32_t +int32_t ioc_readv_disabled_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct iovec *vector, - int32_t count, struct iatt *stbuf, - struct iobref *iobref); + int32_t op_ret, int32_t op_errno, struct iovec *vector, + int32_t count, struct iatt *stbuf, + struct iobref *iobref, dict_t *xdata); ioc_page_t * -ioc_page_get (ioc_inode_t *ioc_inode, off_t offset); +__ioc_page_get (ioc_inode_t *ioc_inode, off_t offset); ioc_page_t * -ioc_page_create (ioc_inode_t *ioc_inode, off_t offset); +__ioc_page_create (ioc_inode_t *ioc_inode, off_t offset); void -ioc_page_fault (ioc_inode_t *ioc_inode, call_frame_t *frame, fd_t *fd, - off_t offset); +ioc_page_fault (ioc_inode_t *ioc_inode, call_frame_t *frame, fd_t *fd, + off_t offset); void -ioc_wait_on_page (ioc_page_t *page, call_frame_t *frame, off_t offset, - size_t size); +__ioc_wait_on_page (ioc_page_t *page, call_frame_t *frame, off_t offset, + size_t size); ioc_waitq_t * -ioc_page_wakeup (ioc_page_t *page); +__ioc_page_wakeup (ioc_page_t *page, int32_t op_errno); void ioc_page_flush (ioc_page_t *page); ioc_waitq_t * -ioc_page_error (ioc_page_t *page, int32_t op_ret, int32_t op_errno); - -void -ioc_page_purge (ioc_page_t *page); +__ioc_page_error (ioc_page_t *page, int32_t op_ret, int32_t op_errno); void ioc_frame_return (call_frame_t *frame); @@ -230,95 +220,95 @@ ioc_waitq_return (ioc_waitq_t *waitq); int32_t ioc_frame_fill (ioc_page_t *page, call_frame_t *frame, off_t offset, - size_t size); + size_t size, int32_t op_errno); -#define ioc_inode_lock(ioc_inode) \ - do { \ - gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, \ - "locked inode(%p)", ioc_inode); \ - pthread_mutex_lock (&ioc_inode->inode_lock); \ - } while (0) +#define ioc_inode_lock(ioc_inode) \ + do { \ + gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, \ + "locked inode(%p)", ioc_inode); \ + pthread_mutex_lock (&ioc_inode->inode_lock); \ + } while (0) -#define ioc_inode_unlock(ioc_inode) \ - do { \ - gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, \ - "unlocked inode(%p)", ioc_inode); \ - pthread_mutex_unlock (&ioc_inode->inode_lock); \ - } while (0) +#define ioc_inode_unlock(ioc_inode) \ + do { \ + gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, \ + "unlocked inode(%p)", ioc_inode); \ + pthread_mutex_unlock (&ioc_inode->inode_lock); \ + } while (0) -#define ioc_table_lock(table) \ - do { \ - gf_log (table->xl->name, GF_LOG_TRACE, \ - "locked table(%p)", table); \ - pthread_mutex_lock (&table->table_lock); \ - } while (0) +#define ioc_table_lock(table) \ + do { \ + gf_log (table->xl->name, GF_LOG_TRACE, \ + "locked table(%p)", table); \ + pthread_mutex_lock (&table->table_lock); \ + } while (0) -#define ioc_table_unlock(table) \ - do { \ - gf_log (table->xl->name, GF_LOG_TRACE, \ - "unlocked table(%p)", table); \ - pthread_mutex_unlock (&table->table_lock); \ - } while (0) +#define ioc_table_unlock(table) \ + do { \ + gf_log (table->xl->name, GF_LOG_TRACE, \ + "unlocked table(%p)", table); \ + pthread_mutex_unlock (&table->table_lock); \ + } while (0) -#define ioc_local_lock(local) \ - do { \ - gf_log (local->inode->table->xl->name, GF_LOG_TRACE, \ - "locked local(%p)", local); \ - pthread_mutex_lock (&local->local_lock); \ - } while (0) +#define ioc_local_lock(local) \ + do { \ + gf_log (local->inode->table->xl->name, GF_LOG_TRACE, \ + "locked local(%p)", local); \ + pthread_mutex_lock (&local->local_lock); \ + } while (0) -#define ioc_local_unlock(local) \ - do { \ - gf_log (local->inode->table->xl->name, GF_LOG_TRACE, \ - "unlocked local(%p)", local); \ - pthread_mutex_unlock (&local->local_lock); \ - } while (0) +#define ioc_local_unlock(local) \ + do { \ + gf_log (local->inode->table->xl->name, GF_LOG_TRACE, \ + "unlocked local(%p)", local); \ + pthread_mutex_unlock (&local->local_lock); \ + } while (0) -#define ioc_page_lock(page) \ - do { \ - gf_log (page->inode->table->xl->name, GF_LOG_TRACE, \ - "locked page(%p)", page); \ - pthread_mutex_lock (&page->page_lock); \ - } while (0) +#define ioc_page_lock(page) \ + do { \ + gf_log (page->inode->table->xl->name, GF_LOG_TRACE, \ + "locked page(%p)", page); \ + pthread_mutex_lock (&page->page_lock); \ + } while (0) -#define ioc_page_unlock(page) \ - do { \ - gf_log (page->inode->table->xl->name, GF_LOG_TRACE, \ - "unlocked page(%p)", page); \ - pthread_mutex_unlock (&page->page_lock); \ - } while (0) +#define ioc_page_unlock(page) \ + do { \ + gf_log (page->inode->table->xl->name, GF_LOG_TRACE, \ + "unlocked page(%p)", page); \ + pthread_mutex_unlock (&page->page_lock); \ + } while (0) static inline uint64_t time_elapsed (struct timeval *now, - struct timeval *then) + struct timeval *then) { - uint64_t sec = now->tv_sec - then->tv_sec; + uint64_t sec = now->tv_sec - then->tv_sec; + + if (sec) + return sec; - if (sec) - return sec; - - return 0; + return 0; } ioc_inode_t * ioc_inode_search (ioc_table_t *table, inode_t *inode); -void +void ioc_inode_destroy (ioc_inode_t *ioc_inode); ioc_inode_t * ioc_inode_update (ioc_table_t *table, inode_t *inode, uint32_t weight); -int64_t -ioc_page_destroy (ioc_page_t *page); +int64_t +__ioc_page_destroy (ioc_page_t *page); int64_t __ioc_inode_flush (ioc_inode_t *ioc_inode); @@ -339,6 +329,4 @@ ioc_prune (ioc_table_t *table); int32_t ioc_need_prune (ioc_table_t *table); -inline uint32_t -ioc_hashfn (void *data, int len); #endif /* __IO_CACHE_H */ diff --git a/xlators/performance/io-cache/src/ioc-inode.c b/xlators/performance/io-cache/src/ioc-inode.c index bbf46d935..86a54bb14 100644 --- a/xlators/performance/io-cache/src/ioc-inode.c +++ b/xlators/performance/io-cache/src/ioc-inode.c @@ -1,20 +1,11 @@ /* - Copyright (c) 2007-2010 Gluster, Inc. <http://www.gluster.com> + Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> This file is part of GlusterFS. - GlusterFS is free software; you can redistribute it and/or modify - it under the terms of the GNU Affero General Public License as published - by the Free Software Foundation; either version 3 of the License, - or (at your option) any later version. - - GlusterFS is distributed in the hope that it will be useful, but - WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Affero General Public License for more details. - - You should have received a copy of the GNU Affero General Public License - along with this program. If not, see - <http://www.gnu.org/licenses/>. + This file is licensed to you under your choice of the GNU Lesser + General Public License, version 3 or any later version (LGPLv3 or + later), or the GNU General Public License, version 2 (GPLv2), in all + cases as published by the Free Software Foundation. */ #ifndef _CONFIG_H @@ -35,10 +26,14 @@ extern int ioc_log2_page_size; void * str_to_ptr (char *string) { - void *ptr = NULL; + void *ptr = NULL; + + GF_VALIDATE_OR_GOTO ("io-cache", string, out); ptr = (void *)strtoul (string, NULL, 16); - return ptr; + +out: + return ptr; } @@ -51,102 +46,126 @@ char * ptr_to_str (void *ptr) { int ret = 0; - char *str = NULL; - ret = gf_asprintf (&str, "%p", ptr); + char *str = NULL; + + GF_VALIDATE_OR_GOTO ("io-cache", ptr, out); + + ret = gf_asprintf (&str, "%p", ptr); if (-1 == ret) { - gf_log ("ioc", GF_LOG_ERROR, + gf_log ("io-cache", GF_LOG_WARNING, "asprintf failed while converting ptr to str"); - return NULL; + str = NULL; + goto out; } - return str; + +out: + return str; } + void -ioc_inode_wakeup (call_frame_t *frame, ioc_inode_t *ioc_inode, +ioc_inode_wakeup (call_frame_t *frame, ioc_inode_t *ioc_inode, struct iatt *stbuf) { - ioc_waitq_t *waiter = NULL, *waited = NULL; - ioc_waitq_t *page_waitq = NULL; - int8_t cache_still_valid = 1; - ioc_local_t *local = NULL; - int8_t need_fault = 0; - ioc_page_t *waiter_page = NULL; + ioc_waitq_t *waiter = NULL, *waited = NULL; + ioc_waitq_t *page_waitq = NULL; + int8_t cache_still_valid = 1; + ioc_local_t *local = NULL; + int8_t need_fault = 0; + ioc_page_t *waiter_page = NULL; + + GF_VALIDATE_OR_GOTO ("io-cache", frame, out); local = frame->local; - ioc_inode_lock (ioc_inode); - { - waiter = ioc_inode->waitq; - ioc_inode->waitq = NULL; - } - ioc_inode_unlock (ioc_inode); - - if (stbuf) - cache_still_valid = ioc_cache_still_valid (ioc_inode, stbuf); - else - cache_still_valid = 0; - - if (!waiter) { - gf_log (frame->this->name, GF_LOG_DEBUG, - "cache validate called without any " - "page waiting to be validated"); - } - - while (waiter) { - waiter_page = waiter->data; - page_waitq = NULL; - - if (waiter_page) { - if (cache_still_valid) { - /* cache valid, wake up page */ - ioc_inode_lock (ioc_inode); - { - page_waitq = - ioc_page_wakeup (waiter_page); - } - ioc_inode_unlock (ioc_inode); - if (page_waitq) - ioc_waitq_return (page_waitq); - } else { - /* cache invalid, generate page fault and set - * page->ready = 0, to avoid double faults - */ - ioc_inode_lock (ioc_inode); - - if (waiter_page->ready) { - waiter_page->ready = 0; - need_fault = 1; - } else { - gf_log (frame->this->name, - GF_LOG_TRACE, - "validate frame(%p) is waiting" - "for in-transit page = %p", - frame, waiter_page); - } - - ioc_inode_unlock (ioc_inode); - - if (need_fault) { - need_fault = 0; - ioc_page_fault (ioc_inode, frame, - local->fd, - waiter_page->offset); - } - } - } - - waited = waiter; - waiter = waiter->next; - - waited->data = NULL; - GF_FREE (waited); - } + GF_VALIDATE_OR_GOTO (frame->this->name, local, out); + + if (ioc_inode == NULL) { + local->op_ret = -1; + local->op_errno = EINVAL; + gf_log (frame->this->name, GF_LOG_WARNING, "ioc_inode is NULL"); + goto out; + } + + ioc_inode_lock (ioc_inode); + { + waiter = ioc_inode->waitq; + ioc_inode->waitq = NULL; + } + ioc_inode_unlock (ioc_inode); + + if (stbuf) + cache_still_valid = ioc_cache_still_valid (ioc_inode, stbuf); + else + cache_still_valid = 0; + + if (!waiter) { + gf_log (frame->this->name, GF_LOG_WARNING, + "cache validate called without any " + "page waiting to be validated"); + } + + while (waiter) { + waiter_page = waiter->data; + page_waitq = NULL; + + if (waiter_page) { + if (cache_still_valid) { + /* cache valid, wake up page */ + ioc_inode_lock (ioc_inode); + { + page_waitq = + __ioc_page_wakeup (waiter_page, + waiter_page->op_errno); + } + ioc_inode_unlock (ioc_inode); + if (page_waitq) + ioc_waitq_return (page_waitq); + } else { + /* cache invalid, generate page fault and set + * page->ready = 0, to avoid double faults + */ + ioc_inode_lock (ioc_inode); + { + if (waiter_page->ready) { + waiter_page->ready = 0; + need_fault = 1; + } else { + gf_log (frame->this->name, + GF_LOG_TRACE, + "validate frame(%p) is " + "waiting for in-transit" + " page = %p", frame, + waiter_page); + } + } + ioc_inode_unlock (ioc_inode); + + if (need_fault) { + need_fault = 0; + ioc_page_fault (ioc_inode, frame, + local->fd, + waiter_page->offset); + } + } + } + + waited = waiter; + waiter = waiter->next; + + waited->data = NULL; + GF_FREE (waited); + } + +out: + return; } -/* - * ioc_inode_update - create a new ioc_inode_t structure and add it to - * the table table. fill in the fields which are derived + +/* + * ioc_inode_update - create a new ioc_inode_t structure and add it to + * the table table. fill in the fields which are derived * from inode_t corresponding to the file - * + * * @table: io-table structure * @inode: inode structure * @@ -155,60 +174,67 @@ ioc_inode_wakeup (call_frame_t *frame, ioc_inode_t *ioc_inode, ioc_inode_t * ioc_inode_update (ioc_table_t *table, inode_t *inode, uint32_t weight) { - ioc_inode_t *ioc_inode = NULL; + ioc_inode_t *ioc_inode = NULL; - ioc_inode = GF_CALLOC (1, sizeof (ioc_inode_t), - gf_ioc_mt_ioc_inode_t); + GF_VALIDATE_OR_GOTO ("io-cache", table, out); + + ioc_inode = GF_CALLOC (1, sizeof (ioc_inode_t), gf_ioc_mt_ioc_inode_t); if (ioc_inode == NULL) { goto out; } - - ioc_inode->table = table; - INIT_LIST_HEAD (&ioc_inode->cache.page_lru); - - ioc_table_lock (table); - - table->inode_count++; - list_add (&ioc_inode->inode_list, &table->inodes); - list_add_tail (&ioc_inode->inode_lru, &table->inode_lru[weight]); - gf_log (table->xl->name, - GF_LOG_TRACE, - "adding to inode_lru[%d]", weight); + ioc_inode->inode = inode; + ioc_inode->table = table; + INIT_LIST_HEAD (&ioc_inode->cache.page_lru); + pthread_mutex_init (&ioc_inode->inode_lock, NULL); + ioc_inode->weight = weight; + + ioc_table_lock (table); + { + table->inode_count++; + list_add (&ioc_inode->inode_list, &table->inodes); + list_add_tail (&ioc_inode->inode_lru, + &table->inode_lru[weight]); + } + ioc_table_unlock (table); - ioc_table_unlock (table); + gf_log (table->xl->name, GF_LOG_TRACE, + "adding to inode_lru[%d]", weight); - pthread_mutex_init (&ioc_inode->inode_lock, NULL); - ioc_inode->weight = weight; - out: - return ioc_inode; + return ioc_inode; } -/* +/* * ioc_inode_destroy - destroy an ioc_inode_t object. * * @inode: inode to destroy * - * to be called only from ioc_forget. + * to be called only from ioc_forget. */ void ioc_inode_destroy (ioc_inode_t *ioc_inode) { - ioc_table_t *table = NULL; + ioc_table_t *table = NULL; + + GF_VALIDATE_OR_GOTO ("io-cache", ioc_inode, out); table = ioc_inode->table; - ioc_table_lock (table); - table->inode_count--; - list_del (&ioc_inode->inode_list); - list_del (&ioc_inode->inode_lru); - ioc_table_unlock (table); - - ioc_inode_flush (ioc_inode); + ioc_table_lock (table); + { + table->inode_count--; + list_del (&ioc_inode->inode_list); + list_del (&ioc_inode->inode_lru); + } + ioc_table_unlock (table); + + ioc_inode_flush (ioc_inode); rbthash_table_destroy (ioc_inode->cache.page_table); - pthread_mutex_destroy (&ioc_inode->inode_lock); - GF_FREE (ioc_inode); + pthread_mutex_destroy (&ioc_inode->inode_lock); + GF_FREE (ioc_inode); +out: + return; } diff --git a/xlators/performance/io-cache/src/ioc-mem-types.h b/xlators/performance/io-cache/src/ioc-mem-types.h index 2cb5ea878..9b68f9fce 100644 --- a/xlators/performance/io-cache/src/ioc-mem-types.h +++ b/xlators/performance/io-cache/src/ioc-mem-types.h @@ -1,20 +1,11 @@ /* - Copyright (c) 2007-2010 Gluster, Inc. <http://www.gluster.com> - This file is part of GlusterFS. + Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> + This file is part of GlusterFS. - GlusterFS is free software; you can redistribute it and/or modify - it under the terms of the GNU Affero General Public License as published - by the Free Software Foundation; either version 3 of the License, - or (at your option) any later version. - - GlusterFS is distributed in the hope that it will be useful, but - WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Affero General Public License for more details. - - You should have received a copy of the GNU Affero General Public License - along with this program. If not, see - <http://www.gnu.org/licenses/>. + This file is licensed to you under your choice of the GNU Lesser + General Public License, version 3 or any later version (LGPLv3 or + later), or the GNU General Public License, version 2 (GPLv2), in all + cases as published by the Free Software Foundation. */ #ifndef __IOC_MT_H__ @@ -26,7 +17,6 @@ enum gf_ioc_mem_types_ { gf_ioc_mt_iovec = gf_common_mt_end + 1, gf_ioc_mt_ioc_table_t, gf_ioc_mt_char, - gf_ioc_mt_ioc_local_t, gf_ioc_mt_ioc_waitq_t, gf_ioc_mt_ioc_priority, gf_ioc_mt_list_head, @@ -37,4 +27,3 @@ enum gf_ioc_mem_types_ { gf_ioc_mt_end }; #endif - diff --git a/xlators/performance/io-cache/src/page.c b/xlators/performance/io-cache/src/page.c index 47a8fbb66..b2e20ba65 100644 --- a/xlators/performance/io-cache/src/page.c +++ b/xlators/performance/io-cache/src/page.c @@ -1,20 +1,11 @@ /* - Copyright (c) 2007-2010 Gluster, Inc. <http://www.gluster.com> + Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> This file is part of GlusterFS. - GlusterFS is free software; you can redistribute it and/or modify - it under the terms of the GNU Affero General Public License as published - by the Free Software Foundation; either version 3 of the License, - or (at your option) any later version. - - GlusterFS is distributed in the hope that it will be useful, but - WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Affero General Public License for more details. - - You should have received a copy of the GNU Affero General Public License - along with this program. If not, see - <http://www.gnu.org/licenses/>. + This file is licensed to you under your choice of the GNU Lesser + General Public License, version 3 or any later version (LGPLv3 or + later), or the GNU General Public License, version 2 (GPLv2), in all + cases as published by the Free Software Foundation. */ #ifndef _CONFIG_H @@ -34,74 +25,169 @@ char ioc_empty (struct ioc_cache *cache) { - return list_empty (&cache->page_lru); + char is_empty = -1; + + GF_VALIDATE_OR_GOTO ("io-cache", cache, out); + + is_empty = list_empty (&cache->page_lru); + +out: + return is_empty; } + ioc_page_t * -ioc_page_get (ioc_inode_t *ioc_inode, off_t offset) +__ioc_page_get (ioc_inode_t *ioc_inode, off_t offset) { - ioc_page_t *page = NULL; - ioc_table_t *table = NULL; - off_t rounded_offset = 0; + ioc_page_t *page = NULL; + ioc_table_t *table = NULL; + off_t rounded_offset = 0; + + GF_VALIDATE_OR_GOTO ("io-cache", ioc_inode, out); table = ioc_inode->table; + GF_VALIDATE_OR_GOTO ("io-cache", ioc_inode, out); + rounded_offset = floor (offset, table->page_size); - + page = rbthash_get (ioc_inode->cache.page_table, &rounded_offset, sizeof (rounded_offset)); if (page != NULL) { - /* push the page to the end of the lru list */ - list_move_tail (&page->page_lru, &ioc_inode->cache.page_lru); - } + /* push the page to the end of the lru list */ + list_move_tail (&page->page_lru, &ioc_inode->cache.page_lru); + } - return page; +out: + return page; +} + + +ioc_page_t * +ioc_page_get (ioc_inode_t *ioc_inode, off_t offset) +{ + ioc_page_t *page = NULL; + + if (ioc_inode == NULL) { + goto out; + } + + ioc_inode_lock (ioc_inode); + { + page = __ioc_page_get (ioc_inode, offset); + } + ioc_inode_unlock (ioc_inode); + +out: + return page; } /* - * ioc_page_destroy - + * __ioc_page_destroy - * * @page: * */ int64_t -ioc_page_destroy (ioc_page_t *page) +__ioc_page_destroy (ioc_page_t *page) { - int64_t page_size = 0; + int64_t page_size = 0; + + GF_VALIDATE_OR_GOTO ("io-cache", page, out); - page_size = iobref_size (page->iobref); + if (page->iobref) + page_size = iobref_size (page->iobref); - if (page->waitq) { - /* frames waiting on this page, do not destroy this page */ - page_size = -1; - } else { + if (page->waitq) { + /* frames waiting on this page, do not destroy this page */ + page_size = -1; + page->stale = 1; + } else { rbthash_remove (page->inode->cache.page_table, &page->offset, sizeof (page->offset)); - list_del (&page->page_lru); - - gf_log (page->inode->table->xl->name, GF_LOG_TRACE, - "destroying page = %p, offset = %"PRId64" " - "&& inode = %p", - page, page->offset, page->inode); - - if (page->vector){ - iobref_unref (page->iobref); - GF_FREE (page->vector); - page->vector = NULL; - } - - page->inode = NULL; - } - - if (page_size != -1) { - pthread_mutex_destroy (&page->page_lock); - GF_FREE (page); - } - - return page_size; + list_del (&page->page_lru); + + gf_log (page->inode->table->xl->name, GF_LOG_TRACE, + "destroying page = %p, offset = %"PRId64" " + "&& inode = %p", + page, page->offset, page->inode); + + if (page->vector){ + iobref_unref (page->iobref); + GF_FREE (page->vector); + page->vector = NULL; + } + + page->inode = NULL; + } + + if (page_size != -1) { + pthread_mutex_destroy (&page->page_lock); + GF_FREE (page); + } + +out: + return page_size; } + +int64_t +ioc_page_destroy (ioc_page_t *page) +{ + int64_t ret = 0; + + if (page == NULL) { + goto out; + } + + ioc_inode_lock (page->inode); + { + ret = __ioc_page_destroy (page); + } + ioc_inode_unlock (page->inode); + +out: + return ret; +} + +int32_t +__ioc_inode_prune (ioc_inode_t *curr, uint64_t *size_pruned, + uint64_t size_to_prune, uint32_t index) +{ + ioc_page_t *page = NULL, *next = NULL; + int32_t ret = 0; + ioc_table_t *table = NULL; + + if (curr == NULL) { + goto out; + } + + table = curr->table; + + list_for_each_entry_safe (page, next, &curr->cache.page_lru, page_lru) { + *size_pruned += page->size; + ret = __ioc_page_destroy (page); + + if (ret != -1) + table->cache_used -= ret; + + gf_log (table->xl->name, GF_LOG_TRACE, + "index = %d && table->cache_used = %"PRIu64" && table->" + "cache_size = %"PRIu64, index, table->cache_used, + table->cache_size); + + if ((*size_pruned) >= size_to_prune) + break; + } + + if (ioc_empty (&curr->cache)) { + list_del_init (&curr->inode_lru); + } + +out: + return 0; +} /* * ioc_prune - prune the cache. we have a limit to the number of pages we * can have in-memory. @@ -112,160 +198,149 @@ ioc_page_destroy (ioc_page_t *page) int32_t ioc_prune (ioc_table_t *table) { - ioc_inode_t *curr = NULL, *next_ioc_inode = NULL; - ioc_page_t *page = NULL, *next = NULL; - int32_t ret = -1; - int32_t index = 0; - uint64_t size_to_prune = 0; - uint64_t size_pruned = 0; - - ioc_table_lock (table); - { - size_to_prune = table->cache_used - table->cache_size; - /* take out the least recently used inode */ - for (index=0; index < table->max_pri; index++) { - list_for_each_entry_safe (curr, next_ioc_inode, - &table->inode_lru[index], - inode_lru) { - /* prune page-by-page for this inode, till - * we reach the equilibrium */ - ioc_inode_lock (curr); - /* { */ - - list_for_each_entry_safe (page, next, - &curr->cache.page_lru, - page_lru) { - /* done with all pages, and not - * reached equilibrium yet?? - * continue with next inode in - * lru_list */ - size_pruned += page->size; - ret = ioc_page_destroy (page); - - if (ret != -1) - table->cache_used -= ret; - - gf_log (table->xl->name, - GF_LOG_TRACE, - "index = %d && table->cache_" - "used = %"PRIu64" && table->" - "cache_size = %"PRIu64, - index, table->cache_used, - table->cache_size); - - if (size_pruned >= size_to_prune) - break; - } /* list_for_each_entry_safe(page...) */ - if (ioc_empty (&curr->cache)) { - list_del_init (&curr->inode_lru); - } - - /* } */ - ioc_inode_unlock (curr); - - if (size_pruned >= size_to_prune) - break; - } /* list_for_each_entry_safe (curr...) */ - - if (size_pruned >= size_to_prune) - break; - } /* for(index=0;...) */ - - } /* ioc_inode_table locked region end */ - ioc_table_unlock (table); - - return 0; + ioc_inode_t *curr = NULL, *next_ioc_inode = NULL; + int32_t index = 0; + uint64_t size_to_prune = 0; + uint64_t size_pruned = 0; + + GF_VALIDATE_OR_GOTO ("io-cache", table, out); + + ioc_table_lock (table); + { + size_to_prune = table->cache_used - table->cache_size; + /* take out the least recently used inode */ + for (index=0; index < table->max_pri; index++) { + list_for_each_entry_safe (curr, next_ioc_inode, + &table->inode_lru[index], + inode_lru) { + /* prune page-by-page for this inode, till + * we reach the equilibrium */ + ioc_inode_lock (curr); + { + __ioc_inode_prune (curr, &size_pruned, + size_to_prune, + index); + } + ioc_inode_unlock (curr); + + if (size_pruned >= size_to_prune) + break; + } /* list_for_each_entry_safe (curr...) */ + + if (size_pruned >= size_to_prune) + break; + } /* for(index=0;...) */ + + } /* ioc_inode_table locked region end */ + ioc_table_unlock (table); + +out: + return 0; } /* - * ioc_page_create - create a new page. + * __ioc_page_create - create a new page. * * @ioc_inode: * @offset: * */ ioc_page_t * -ioc_page_create (ioc_inode_t *ioc_inode, off_t offset) +__ioc_page_create (ioc_inode_t *ioc_inode, off_t offset) { - ioc_table_t *table = NULL; - ioc_page_t *page = NULL; - off_t rounded_offset = 0; - ioc_page_t *newpage = NULL; + ioc_table_t *table = NULL; + ioc_page_t *page = NULL; + off_t rounded_offset = 0; + ioc_page_t *newpage = NULL; + + GF_VALIDATE_OR_GOTO ("io-cache", ioc_inode, out); table = ioc_inode->table; + GF_VALIDATE_OR_GOTO ("io-cache", table, out); + rounded_offset = floor (offset, table->page_size); - newpage = GF_CALLOC (1, sizeof (*newpage), - gf_ioc_mt_ioc_newpage_t); + newpage = GF_CALLOC (1, sizeof (*newpage), gf_ioc_mt_ioc_newpage_t); if (newpage == NULL) { goto out; } - if (!ioc_inode) { + if (!ioc_inode) { GF_FREE (newpage); newpage = NULL; goto out; - } + } - newpage->offset = rounded_offset; - newpage->inode = ioc_inode; - pthread_mutex_init (&newpage->page_lock, NULL); + newpage->offset = rounded_offset; + newpage->inode = ioc_inode; + pthread_mutex_init (&newpage->page_lock, NULL); rbthash_insert (ioc_inode->cache.page_table, newpage, &rounded_offset, sizeof (rounded_offset)); - list_add_tail (&newpage->page_lru, &ioc_inode->cache.page_lru); + list_add_tail (&newpage->page_lru, &ioc_inode->cache.page_lru); - page = newpage; + page = newpage; - gf_log ("io-cache", GF_LOG_TRACE, - "returning new page %p", page); + gf_log ("io-cache", GF_LOG_TRACE, + "returning new page %p", page); out: - return page; + return page; } -/* - * ioc_wait_on_page - pause a frame to wait till the arrival of a page. - * here we need to handle the case when the frame who calls wait_on_page - * himself has caused page_fault +/* + * ioc_wait_on_page - pause a frame to wait till the arrival of a page. + * here we need to handle the case when the frame who calls wait_on_page + * himself has caused page_fault * * @page: page to wait on * @frame: call frame who is waiting on page * */ void -ioc_wait_on_page (ioc_page_t *page, call_frame_t *frame, off_t offset, - size_t size) +__ioc_wait_on_page (ioc_page_t *page, call_frame_t *frame, off_t offset, + size_t size) { - ioc_waitq_t *waitq = NULL; - ioc_local_t *local = frame->local; + ioc_waitq_t *waitq = NULL; + ioc_local_t *local = NULL; + + GF_VALIDATE_OR_GOTO ("io-cache", frame, out); + local = frame->local; + + GF_VALIDATE_OR_GOTO (frame->this->name, local, out); + + if (page == NULL) { + local->op_ret = -1; + local->op_errno = ENOMEM; + gf_log (frame->this->name, GF_LOG_WARNING, + "asked to wait on a NULL page"); + } - waitq = GF_CALLOC (1, sizeof (*waitq), gf_ioc_mt_ioc_waitq_t); + waitq = GF_CALLOC (1, sizeof (*waitq), gf_ioc_mt_ioc_waitq_t); if (waitq == NULL) { local->op_ret = -1; local->op_errno = ENOMEM; - gf_log (frame->this->name, GF_LOG_ERROR, "out of memory"); goto out; - } - - gf_log (frame->this->name, GF_LOG_TRACE, - "frame(%p) waiting on page = %p, offset=%"PRId64", " - "size=%"GF_PRI_SIZET"", - frame, page, offset, size); - - waitq->data = frame; - waitq->next = page->waitq; - waitq->pending_offset = offset; - waitq->pending_size = size; - page->waitq = waitq; - /* one frame can wait only once on a given page, - * local->wait_count is number of pages a frame is waiting on */ - ioc_local_lock (local); - { - local->wait_count++; - } - ioc_local_unlock (local); + } + + gf_log (frame->this->name, GF_LOG_TRACE, + "frame(%p) waiting on page = %p, offset=%"PRId64", " + "size=%"GF_PRI_SIZET"", + frame, page, offset, size); + + waitq->data = frame; + waitq->next = page->waitq; + waitq->pending_offset = offset; + waitq->pending_size = size; + page->waitq = waitq; + /* one frame can wait only once on a given page, + * local->wait_count is number of pages a frame is waiting on */ + ioc_local_lock (local); + { + local->wait_count++; + } + ioc_local_unlock (local); out: return; @@ -273,7 +348,7 @@ out: /* - * ioc_cache_still_valid - see if cached pages ioc_inode are still valid + * ioc_cache_still_valid - see if cached pages ioc_inode are still valid * against given stbuf * * @ioc_inode: @@ -284,194 +359,210 @@ out: int8_t ioc_cache_still_valid (ioc_inode_t *ioc_inode, struct iatt *stbuf) { - int8_t cache_still_valid = 1; + int8_t cache_still_valid = 1; + + GF_VALIDATE_OR_GOTO ("io-cache", ioc_inode, out); #if 0 - if (!stbuf || (stbuf->ia_mtime != ioc_inode->cache.mtime) || - (stbuf->st_mtim.tv_nsec != ioc_inode->stbuf.st_mtim.tv_nsec)) - cache_still_valid = 0; + if (!stbuf || (stbuf->ia_mtime != ioc_inode->cache.mtime) || + (stbuf->st_mtim.tv_nsec != ioc_inode->stbuf.st_mtim.tv_nsec)) + cache_still_valid = 0; #else - if (!stbuf || (stbuf->ia_mtime != ioc_inode->cache.mtime) + if (!stbuf || (stbuf->ia_mtime != ioc_inode->cache.mtime) || (stbuf->ia_mtime_nsec != ioc_inode->cache.mtime_nsec)) - cache_still_valid = 0; + cache_still_valid = 0; #endif #if 0 - /* talk with avati@gluster.com to enable this section */ - if (!ioc_inode->mtime && stbuf) { - cache_still_valid = 1; - ioc_inode->mtime = stbuf->ia_mtime; - } + /* talk with avati@gluster.com to enable this section */ + if (!ioc_inode->mtime && stbuf) { + cache_still_valid = 1; + ioc_inode->mtime = stbuf->ia_mtime; + } #endif - return cache_still_valid; +out: + return cache_still_valid; } void ioc_waitq_return (ioc_waitq_t *waitq) { - ioc_waitq_t *trav = NULL; - ioc_waitq_t *next = NULL; - call_frame_t *frame = NULL; + ioc_waitq_t *trav = NULL; + ioc_waitq_t *next = NULL; + call_frame_t *frame = NULL; - for (trav = waitq; trav; trav = next) { - next = trav->next; + for (trav = waitq; trav; trav = next) { + next = trav->next; - frame = trav->data; - ioc_frame_return (frame); - GF_FREE (trav); - } + frame = trav->data; + ioc_frame_return (frame); + GF_FREE (trav); + } } int ioc_fault_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, int32_t op_errno, struct iovec *vector, - int32_t count, struct iatt *stbuf, struct iobref *iobref) + int32_t count, struct iatt *stbuf, struct iobref *iobref, + dict_t *xdata) { - ioc_local_t *local = NULL; - off_t offset = 0; - ioc_inode_t *ioc_inode = NULL; - ioc_table_t *table = NULL; - ioc_page_t *page = NULL; - int32_t destroy_size = 0; - size_t page_size = 0; - ioc_waitq_t *waitq = NULL; - size_t iobref_page_size = 0; - char zero_filled = 0; + ioc_local_t *local = NULL; + off_t offset = 0; + ioc_inode_t *ioc_inode = NULL; + ioc_table_t *table = NULL; + ioc_page_t *page = NULL; + int32_t destroy_size = 0; + size_t page_size = 0; + ioc_waitq_t *waitq = NULL; + size_t iobref_page_size = 0; + char zero_filled = 0; + + GF_ASSERT (frame); local = frame->local; + GF_ASSERT (local); + offset = local->pending_offset; ioc_inode = local->inode; + GF_ASSERT (ioc_inode); + table = ioc_inode->table; + GF_ASSERT (table); + + zero_filled = ((op_ret >=0) && (stbuf->ia_mtime == 0)); + + ioc_inode_lock (ioc_inode); + { + if (op_ret == -1 || !(zero_filled || + ioc_cache_still_valid(ioc_inode, + stbuf))) { + gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, + "cache for inode(%p) is invalid. flushing " + "all pages", ioc_inode); + destroy_size = __ioc_inode_flush (ioc_inode); + } - zero_filled = ((op_ret >=0) - && (stbuf->ia_mtime == 0)); - - ioc_inode_lock (ioc_inode); - { - if (op_ret == -1 || - !(zero_filled || - ioc_cache_still_valid(ioc_inode, stbuf))) { - gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, - "cache for inode(%p) is invalid. flushing " - "all pages", ioc_inode); - destroy_size = __ioc_inode_flush (ioc_inode); - } - - if ((op_ret >= 0) && !zero_filled) { - ioc_inode->cache.mtime = stbuf->ia_mtime; + if ((op_ret >= 0) && !zero_filled) { + ioc_inode->cache.mtime = stbuf->ia_mtime; ioc_inode->cache.mtime_nsec = stbuf->ia_mtime_nsec; } - gettimeofday (&ioc_inode->cache.tv, NULL); - - if (op_ret < 0) { - /* error, readv returned -1 */ - page = ioc_page_get (ioc_inode, offset); - if (page) - waitq = ioc_page_error (page, op_ret, - op_errno); - } else { - gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, - "op_ret = %d", op_ret); - page = ioc_page_get (ioc_inode, offset); - if (!page) { - /* page was flushed */ - /* some serious bug ? */ - gf_log (this->name, GF_LOG_DEBUG, - "wasted copy: %"PRId64"[+%"PRId64"] " - "ioc_inode=%p", offset, - table->page_size, ioc_inode); - } else { - if (page->vector) { - iobref_unref (page->iobref); - GF_FREE (page->vector); - page->vector = NULL; - } - - /* keep a copy of the page for our cache */ - page->vector = iov_dup (vector, count); + gettimeofday (&ioc_inode->cache.tv, NULL); + + if (op_ret < 0) { + /* error, readv returned -1 */ + page = __ioc_page_get (ioc_inode, offset); + if (page) + waitq = __ioc_page_error (page, op_ret, + op_errno); + } else { + gf_log (ioc_inode->table->xl->name, GF_LOG_TRACE, + "op_ret = %d", op_ret); + page = __ioc_page_get (ioc_inode, offset); + if (!page) { + /* page was flushed */ + /* some serious bug ? */ + gf_log (frame->this->name, GF_LOG_WARNING, + "wasted copy: %"PRId64"[+%"PRId64"] " + "ioc_inode=%p", offset, + table->page_size, ioc_inode); + } else { + if (page->vector) { + iobref_unref (page->iobref); + GF_FREE (page->vector); + page->vector = NULL; + } + + /* keep a copy of the page for our cache */ + page->vector = iov_dup (vector, count); if (page->vector == NULL) { - page = ioc_page_get (ioc_inode, offset); + page = __ioc_page_get (ioc_inode, + offset); if (page != NULL) - waitq = ioc_page_error (page, - -1, - ENOMEM); + waitq = __ioc_page_error (page, + -1, + ENOMEM); goto unlock; } - page->count = count; - if (iobref) { - page->iobref = iobref_ref (iobref); - } else { - /* TODO: we have got a response to - * our request and no data */ - gf_log (this->name, GF_LOG_CRITICAL, - "frame>root>rsp_refs is null"); - } /* if(frame->root->rsp_refs) */ - - /* page->size should indicate exactly how - * much the readv call to the child - * translator returned. earlier op_ret - * from child translator was used, which - * gave rise to a bug where reads from - * io-cached volume were resulting in 0 - * byte replies */ - page_size = iov_length(vector, count); - page->size = page_size; + page->count = count; + if (iobref) { + page->iobref = iobref_ref (iobref); + } else { + /* TODO: we have got a response to + * our request and no data */ + gf_log (frame->this->name, + GF_LOG_CRITICAL, + "frame>root>rsp_refs is null"); + } /* if(frame->root->rsp_refs) */ + + /* page->size should indicate exactly how + * much the readv call to the child + * translator returned. earlier op_ret + * from child translator was used, which + * gave rise to a bug where reads from + * io-cached volume were resulting in 0 + * byte replies */ + page_size = iov_length(vector, count); + page->size = page_size; + page->op_errno = op_errno; iobref_page_size = iobref_size (page->iobref); - if (page->waitq) { - /* wake up all the frames waiting on - * this page, including - * the frame which triggered fault */ - waitq = ioc_page_wakeup (page); - } /* if(page->waitq) */ - } /* if(!page)...else */ - } /* if(op_ret < 0)...else */ - } /* ioc_inode locked region end */ + if (page->waitq) { + /* wake up all the frames waiting on + * this page, including + * the frame which triggered fault */ + waitq = __ioc_page_wakeup (page, + op_errno); + } /* if(page->waitq) */ + } /* if(!page)...else */ + } /* if(op_ret < 0)...else */ + } /* ioc_inode locked region end */ unlock: - ioc_inode_unlock (ioc_inode); + ioc_inode_unlock (ioc_inode); - ioc_waitq_return (waitq); + ioc_waitq_return (waitq); - if (iobref_page_size) { - ioc_table_lock (table); - { - table->cache_used += iobref_page_size; - } - ioc_table_unlock (table); - } + if (iobref_page_size) { + ioc_table_lock (table); + { + table->cache_used += iobref_page_size; + } + ioc_table_unlock (table); + } - if (destroy_size) { - ioc_table_lock (table); - { - table->cache_used -= destroy_size; - } - ioc_table_unlock (table); - } + if (destroy_size) { + ioc_table_lock (table); + { + table->cache_used -= destroy_size; + } + ioc_table_unlock (table); + } - if (ioc_need_prune (ioc_inode->table)) { - ioc_prune (ioc_inode->table); - } + if (ioc_need_prune (ioc_inode->table)) { + ioc_prune (ioc_inode->table); + } - gf_log (this->name, GF_LOG_TRACE, "fault frame %p returned", frame); - pthread_mutex_destroy (&local->local_lock); + gf_log (frame->this->name, GF_LOG_TRACE, "fault frame %p returned", + frame); + pthread_mutex_destroy (&local->local_lock); - fd_unref (local->fd); + fd_unref (local->fd); - STACK_DESTROY (frame->root); - return 0; + STACK_DESTROY (frame->root); + return 0; } + /* * ioc_page_fault - - * + * * @ioc_inode: * @frame: * @fd: @@ -479,145 +570,169 @@ unlock: * */ void -ioc_page_fault (ioc_inode_t *ioc_inode, call_frame_t *frame, fd_t *fd, - off_t offset) +ioc_page_fault (ioc_inode_t *ioc_inode, call_frame_t *frame, fd_t *fd, + off_t offset) { - ioc_table_t *table = NULL; - call_frame_t *fault_frame = NULL; - ioc_local_t *fault_local = NULL; - int32_t op_ret = -1, op_errno = -1; - ioc_waitq_t *waitq = NULL; - ioc_page_t *page = NULL; + ioc_table_t *table = NULL; + call_frame_t *fault_frame = NULL; + ioc_local_t *fault_local = NULL; + int32_t op_ret = -1, op_errno = -1; + ioc_waitq_t *waitq = NULL; + ioc_page_t *page = NULL; + + GF_ASSERT (ioc_inode); + if (frame == NULL) { + op_ret = -1; + op_errno = EINVAL; + gf_log ("io-cache", GF_LOG_WARNING, + "page fault on a NULL frame"); + goto err; + } table = ioc_inode->table; fault_frame = copy_frame (frame); if (fault_frame == NULL) { op_ret = -1; op_errno = ENOMEM; - gf_log (ioc_inode->table->xl->name, GF_LOG_ERROR, - "out of memory"); goto err; } - fault_local = GF_CALLOC (1, sizeof (ioc_local_t), - gf_ioc_mt_ioc_local_t); + fault_local = mem_get0 (THIS->local_pool); if (fault_local == NULL) { op_ret = -1; op_errno = ENOMEM; STACK_DESTROY (fault_frame->root); - gf_log (ioc_inode->table->xl->name, GF_LOG_ERROR, - "out of memory"); goto err; } - /* NOTE: copy_frame() means, the frame the fop whose fd_ref we - * are using till now won't be valid till we get reply from server. - * we unref this fd, in fault_cbk */ - fault_local->fd = fd_ref (fd); + /* NOTE: copy_frame() means, the frame the fop whose fd_ref we + * are using till now won't be valid till we get reply from server. + * we unref this fd, in fault_cbk */ + fault_local->fd = fd_ref (fd); + + fault_frame->local = fault_local; + pthread_mutex_init (&fault_local->local_lock, NULL); - fault_frame->local = fault_local; - pthread_mutex_init (&fault_local->local_lock, NULL); + INIT_LIST_HEAD (&fault_local->fill_list); + fault_local->pending_offset = offset; + fault_local->pending_size = table->page_size; + fault_local->inode = ioc_inode; - INIT_LIST_HEAD (&fault_local->fill_list); - fault_local->pending_offset = offset; - fault_local->pending_size = table->page_size; - fault_local->inode = ioc_inode; + gf_log (frame->this->name, GF_LOG_TRACE, + "stack winding page fault for offset = %"PRId64" with " + "frame %p", offset, fault_frame); - gf_log (frame->this->name, GF_LOG_TRACE, - "stack winding page fault for offset = %"PRId64" with " - "frame %p", offset, fault_frame); - - STACK_WIND (fault_frame, ioc_fault_cbk, FIRST_CHILD(fault_frame->this), - FIRST_CHILD(fault_frame->this)->fops->readv, fd, - table->page_size, offset); - return; + STACK_WIND (fault_frame, ioc_fault_cbk, FIRST_CHILD(fault_frame->this), + FIRST_CHILD(fault_frame->this)->fops->readv, fd, + table->page_size, offset, 0, NULL); + return; err: - page = ioc_page_get (ioc_inode, offset); - if (page != NULL) { - waitq = ioc_page_error (page, op_ret, op_errno); - if (waitq != NULL) { - ioc_waitq_return (waitq); + ioc_inode_lock (ioc_inode); + { + page = __ioc_page_get (ioc_inode, offset); + if (page != NULL) { + waitq = __ioc_page_error (page, op_ret, op_errno); } } + ioc_inode_unlock (ioc_inode); + + if (waitq != NULL) { + ioc_waitq_return (waitq); + } } + int32_t -ioc_frame_fill (ioc_page_t *page, call_frame_t *frame, off_t offset, - size_t size) +__ioc_frame_fill (ioc_page_t *page, call_frame_t *frame, off_t offset, + size_t size, int32_t op_errno) { - ioc_local_t *local = NULL; - ioc_fill_t *fill = NULL; - off_t src_offset = 0; - off_t dst_offset = 0; - ssize_t copy_size = 0; - ioc_inode_t *ioc_inode = NULL; - ioc_fill_t *new = NULL; - int8_t found = 0; - int32_t ret = 0; + ioc_local_t *local = NULL; + ioc_fill_t *fill = NULL; + off_t src_offset = 0; + off_t dst_offset = 0; + ssize_t copy_size = 0; + ioc_inode_t *ioc_inode = NULL; + ioc_fill_t *new = NULL; + int8_t found = 0; + int32_t ret = -1; + + GF_VALIDATE_OR_GOTO ("io-cache", frame, out); local = frame->local; + GF_VALIDATE_OR_GOTO (frame->this->name, local, out); + + if (page == NULL) { + gf_log (frame->this->name, GF_LOG_WARNING, + "NULL page has been provided to serve read request"); + local->op_ret = -1; + local->op_errno = EINVAL; + goto out; + } + ioc_inode = page->inode; - gf_log (frame->this->name, GF_LOG_TRACE, - "frame (%p) offset = %"PRId64" && size = %"GF_PRI_SIZET" " - "&& page->size = %"GF_PRI_SIZET" && wait_count = %d", - frame, offset, size, page->size, local->wait_count); - - /* immediately move this page to the end of the page_lru list */ - list_move_tail (&page->page_lru, &ioc_inode->cache.page_lru); - /* fill local->pending_size bytes from local->pending_offset */ - if (local->op_ret != -1 && page->size) { - if (offset > page->offset) - /* offset is offset in file, convert it to offset in - * page */ - src_offset = offset - page->offset; - /*FIXME: since offset is the offset within page is the - * else case valid? */ - else - /* local->pending_offset is in previous page. do not - * fill until we have filled all previous pages */ - dst_offset = page->offset - offset; - - /* we have to copy from offset to either end of this page - * or till the requested size */ - copy_size = min (page->size - src_offset, - size - dst_offset); - - if (copy_size < 0) { - /* if page contains fewer bytes and the required offset - is beyond the page size in the page */ - copy_size = src_offset = 0; - } - - gf_log (page->inode->table->xl->name, GF_LOG_TRACE, - "copy_size = %"GF_PRI_SIZET" && src_offset = " - "%"PRId64" && dst_offset = %"PRId64"", - copy_size, src_offset, dst_offset); - - { + gf_log (frame->this->name, GF_LOG_TRACE, + "frame (%p) offset = %"PRId64" && size = %"GF_PRI_SIZET" " + "&& page->size = %"GF_PRI_SIZET" && wait_count = %d", + frame, offset, size, page->size, local->wait_count); + + /* immediately move this page to the end of the page_lru list */ + list_move_tail (&page->page_lru, &ioc_inode->cache.page_lru); + /* fill local->pending_size bytes from local->pending_offset */ + if (local->op_ret != -1) { + local->op_errno = op_errno; + + if (page->size == 0) { + goto done; + } + + if (offset > page->offset) + /* offset is offset in file, convert it to offset in + * page */ + src_offset = offset - page->offset; + /*FIXME: since offset is the offset within page is the + * else case valid? */ + else + /* local->pending_offset is in previous page. do not + * fill until we have filled all previous pages */ + dst_offset = page->offset - offset; + + /* we have to copy from offset to either end of this page + * or till the requested size */ + copy_size = min (page->size - src_offset, + size - dst_offset); + + if (copy_size < 0) { + /* if page contains fewer bytes and the required offset + is beyond the page size in the page */ + copy_size = src_offset = 0; + } + + gf_log (page->inode->table->xl->name, GF_LOG_TRACE, + "copy_size = %"GF_PRI_SIZET" && src_offset = " + "%"PRId64" && dst_offset = %"PRId64"", + copy_size, src_offset, dst_offset); + + { new = GF_CALLOC (1, sizeof (*new), gf_ioc_mt_ioc_fill_t); if (new == NULL) { local->op_ret = -1; local->op_errno = ENOMEM; - ret = -1; - gf_log (page->inode->table->xl->name, - GF_LOG_ERROR, "out of memory"); goto out; } - new->offset = page->offset; - new->size = copy_size; - new->iobref = iobref_ref (page->iobref); - new->count = iov_subset (page->vector, - page->count, - src_offset, - src_offset + copy_size, - NULL); - - new->vector = GF_CALLOC (new->count, - sizeof (struct iovec), + new->offset = page->offset; + new->size = copy_size; + new->iobref = iobref_ref (page->iobref); + new->count = iov_subset (page->vector, page->count, + src_offset, + src_offset + copy_size, + NULL); + + new->vector = GF_CALLOC (new->count, + sizeof (struct iovec), gf_ioc_mt_iovec); if (new->vector == NULL) { local->op_ret = -1; @@ -625,57 +740,53 @@ ioc_frame_fill (ioc_page_t *page, call_frame_t *frame, off_t offset, iobref_unref (new->iobref); GF_FREE (new); - - ret = -1; - gf_log (page->inode->table->xl->name, - GF_LOG_ERROR, "out of memory"); goto out; } - new->count = iov_subset (page->vector, - page->count, - src_offset, - src_offset + copy_size, - new->vector); - + new->count = iov_subset (page->vector, page->count, + src_offset, + src_offset + copy_size, + new->vector); + + /* add the ioc_fill to fill_list for this frame */ + if (list_empty (&local->fill_list)) { + /* if list is empty, then this is the first + * time we are filling frame, add the + * ioc_fill_t to the end of list */ + list_add_tail (&new->list, &local->fill_list); + } else { + found = 0; + /* list is not empty, we need to look for + * where this offset fits in list */ + list_for_each_entry (fill, &local->fill_list, + list) { + if (fill->offset > new->offset) { + found = 1; + break; + } + } + if (found) { + list_add_tail (&new->list, + &fill->list); + } else { + list_add_tail (&new->list, + &local->fill_list); + } + } + } - /* add the ioc_fill to fill_list for this frame */ - if (list_empty (&local->fill_list)) { - /* if list is empty, then this is the first - * time we are filling frame, add the - * ioc_fill_t to the end of list */ - list_add_tail (&new->list, &local->fill_list); - } else { - found = 0; - /* list is not empty, we need to look for - * where this offset fits in list */ - list_for_each_entry (fill, &local->fill_list, - list) { - if (fill->offset > new->offset) { - found = 1; - break; - } - } - - if (found) { - list_add_tail (&new->list, - &fill->list); - } else { - list_add_tail (&new->list, - &local->fill_list); - } - } - } - local->op_ret += copy_size; - } + local->op_ret += copy_size; + } +done: + ret = 0; out: return ret; } /* - * ioc_frame_unwind - frame unwinds only from here + * ioc_frame_unwind - frame unwinds only from here * * @frame: call frame to unwind * @@ -686,43 +797,57 @@ out: static void ioc_frame_unwind (call_frame_t *frame) { - ioc_local_t *local = NULL; - ioc_fill_t *fill = NULL, *next = NULL; - int32_t count = 0; - struct iovec *vector = NULL; - int32_t copied = 0; - struct iobref *iobref = NULL; - struct iatt stbuf = {0,}; - int32_t op_ret = 0; + ioc_local_t *local = NULL; + ioc_fill_t *fill = NULL, *next = NULL; + int32_t count = 0; + struct iovec *vector = NULL; + int32_t copied = 0; + struct iobref *iobref = NULL; + struct iatt stbuf = {0,}; + int32_t op_ret = 0, op_errno = 0; + + GF_ASSERT (frame); local = frame->local; - // ioc_local_lock (local); - frame->local = NULL; - iobref = iobref_new (); + if (local == NULL) { + gf_log (frame->this->name, GF_LOG_WARNING, + "local is NULL"); + op_ret = -1; + op_errno = ENOMEM; + goto unwind; + } + + if (local->op_ret < 0) { + op_ret = local->op_ret; + op_errno = local->op_errno; + goto unwind; + } + + // ioc_local_lock (local); + iobref = iobref_new (); if (iobref == NULL) { op_ret = -1; - gf_log (frame->this->name, GF_LOG_ERROR, "out of memory"); + op_errno = ENOMEM; } - if (list_empty (&local->fill_list)) { - gf_log (frame->this->name, GF_LOG_TRACE, - "frame(%p) has 0 entries in local->fill_list " - "(offset = %"PRId64" && size = %"GF_PRI_SIZET")", - frame, local->offset, local->size); - } + if (list_empty (&local->fill_list)) { + gf_log (frame->this->name, GF_LOG_TRACE, + "frame(%p) has 0 entries in local->fill_list " + "(offset = %"PRId64" && size = %"GF_PRI_SIZET")", + frame, local->offset, local->size); + } - list_for_each_entry (fill, &local->fill_list, list) { - count += fill->count; - } + list_for_each_entry (fill, &local->fill_list, list) { + count += fill->count; + } - vector = GF_CALLOC (count, sizeof (*vector), gf_ioc_mt_iovec); + vector = GF_CALLOC (count, sizeof (*vector), gf_ioc_mt_iovec); if (vector == NULL) { op_ret = -1; - - gf_log (frame->this->name, GF_LOG_ERROR, "out of memory"); + op_errno = ENOMEM; } - list_for_each_entry_safe (fill, next, &local->fill_list, list) { + list_for_each_entry_safe (fill, next, &local->fill_list, list) { if ((vector != NULL) && (iobref != NULL)) { memcpy (((char *)vector) + copied, fill->vector, @@ -733,23 +858,25 @@ ioc_frame_unwind (call_frame_t *frame) iobref_merge (iobref, fill->iobref); } - list_del (&fill->list); - iobref_unref (fill->iobref); - GF_FREE (fill->vector); - GF_FREE (fill); - } + list_del (&fill->list); + iobref_unref (fill->iobref); + GF_FREE (fill->vector); + GF_FREE (fill); + } if (op_ret != -1) { op_ret = iov_length (vector, count); } - gf_log (frame->this->name, GF_LOG_TRACE, - "frame(%p) unwinding with op_ret=%d", frame, op_ret); +unwind: + gf_log (frame->this->name, GF_LOG_TRACE, + "frame(%p) unwinding with op_ret=%d", frame, op_ret); - // ioc_local_unlock (local); + // ioc_local_unlock (local); - STACK_UNWIND_STRICT (readv, frame, op_ret, local->op_errno, vector, - count, &stbuf, iobref); + frame->local = NULL; + STACK_UNWIND_STRICT (readv, frame, op_ret, op_errno, vector, + count, &stbuf, iobref, NULL); if (iobref != NULL) { iobref_unref (iobref); @@ -760,8 +887,9 @@ ioc_frame_unwind (call_frame_t *frame) vector = NULL; } - pthread_mutex_destroy (&local->local_lock); - GF_FREE (local); + pthread_mutex_destroy (&local->local_lock); + if (local) + mem_put (local); return; } @@ -775,23 +903,25 @@ ioc_frame_unwind (call_frame_t *frame) void ioc_frame_return (call_frame_t *frame) { - ioc_local_t *local = NULL; - int32_t wait_count = 0; + ioc_local_t *local = NULL; + int32_t wait_count = 0; + + GF_ASSERT (frame); local = frame->local; GF_ASSERT (local->wait_count > 0); - ioc_local_lock (local); - { - wait_count = --local->wait_count; - } - ioc_local_unlock (local); + ioc_local_lock (local); + { + wait_count = --local->wait_count; + } + ioc_local_unlock (local); - if (!wait_count) { - ioc_frame_unwind (frame); - } + if (!wait_count) { + ioc_frame_unwind (frame); + } - return; + return; } /* @@ -801,33 +931,91 @@ ioc_frame_return (call_frame_t *frame) * to be called only when a frame is waiting on an in-transit page */ ioc_waitq_t * -ioc_page_wakeup (ioc_page_t *page) +__ioc_page_wakeup (ioc_page_t *page, int32_t op_errno) { - ioc_waitq_t *waitq = NULL, *trav = NULL; - call_frame_t *frame = NULL; - int32_t ret = -1; + ioc_waitq_t *waitq = NULL, *trav = NULL; + call_frame_t *frame = NULL; + int32_t ret = -1; + + GF_VALIDATE_OR_GOTO ("io-cache", page, out); - waitq = page->waitq; - page->waitq = NULL; + waitq = page->waitq; + page->waitq = NULL; - page->ready = 1; + page->ready = 1; - gf_log (page->inode->table->xl->name, GF_LOG_TRACE, - "page is %p && waitq = %p", page, waitq); + gf_log (page->inode->table->xl->name, GF_LOG_TRACE, + "page is %p && waitq = %p", page, waitq); - for (trav = waitq; trav; trav = trav->next) { - frame = trav->data; - ret = ioc_frame_fill (page, frame, trav->pending_offset, - trav->pending_size); + for (trav = waitq; trav; trav = trav->next) { + frame = trav->data; + ret = __ioc_frame_fill (page, frame, trav->pending_offset, + trav->pending_size, op_errno); if (ret == -1) { break; } - } - - return waitq; + } + + if (page->stale) { + __ioc_page_destroy (page); + } + +out: + return waitq; } + +/* + * ioc_page_error - + * @page: + * @op_ret: + * @op_errno: + * + */ +ioc_waitq_t * +__ioc_page_error (ioc_page_t *page, int32_t op_ret, int32_t op_errno) +{ + ioc_waitq_t *waitq = NULL, *trav = NULL; + call_frame_t *frame = NULL; + int64_t ret = 0; + ioc_table_t *table = NULL; + ioc_local_t *local = NULL; + + GF_VALIDATE_OR_GOTO ("io-cache", page, out); + + waitq = page->waitq; + page->waitq = NULL; + + gf_log (page->inode->table->xl->name, GF_LOG_WARNING, + "page error for page = %p & waitq = %p", page, waitq); + + for (trav = waitq; trav; trav = trav->next) { + + frame = trav->data; + + local = frame->local; + ioc_local_lock (local); + { + if (local->op_ret != -1) { + local->op_ret = op_ret; + local->op_errno = op_errno; + } + } + ioc_local_unlock (local); + } + + table = page->inode->table; + ret = __ioc_page_destroy (page); + + if (ret != -1) { + table->cache_used -= ret; + } + +out: + return waitq; +} + /* * ioc_page_error - * @page: @@ -838,39 +1026,18 @@ ioc_page_wakeup (ioc_page_t *page) ioc_waitq_t * ioc_page_error (ioc_page_t *page, int32_t op_ret, int32_t op_errno) { - ioc_waitq_t *waitq = NULL, *trav = NULL; - call_frame_t *frame = NULL; - int64_t ret = 0; - ioc_table_t *table = NULL; - ioc_local_t *local = NULL; - - waitq = page->waitq; - page->waitq = NULL; - - gf_log (page->inode->table->xl->name, GF_LOG_DEBUG, - "page error for page = %p & waitq = %p", page, waitq); - - for (trav = waitq; trav; trav = trav->next) { - - frame = trav->data; - - local = frame->local; - ioc_local_lock (local); - { - if (local->op_ret != -1) { - local->op_ret = op_ret; - local->op_errno = op_errno; - } - } - ioc_local_unlock (local); - } - - table = page->inode->table; - ret = ioc_page_destroy (page); - - if (ret != -1) { - table->cache_used -= ret; - } - - return waitq; + ioc_waitq_t *waitq = NULL; + + if (page == NULL) { + goto out; + } + + ioc_inode_lock (page->inode); + { + waitq = __ioc_page_error (page, op_ret, op_errno); + } + ioc_inode_unlock (page->inode); + +out: + return waitq; } |