diff options
author | Vikas Gorur <vikas@zresearch.com> | 2009-02-18 17:36:07 +0530 |
---|---|---|
committer | Vikas Gorur <vikas@zresearch.com> | 2009-02-18 17:36:07 +0530 |
commit | 77adf4cd648dce41f89469dd185deec6b6b53a0b (patch) | |
tree | 02e155a5753b398ee572b45793f889b538efab6b /xlators/performance/io-cache/src/page.c | |
parent | f3b2e6580e5663292ee113c741343c8a43ee133f (diff) |
Added all files
Diffstat (limited to 'xlators/performance/io-cache/src/page.c')
-rw-r--r-- | xlators/performance/io-cache/src/page.c | 778 |
1 files changed, 778 insertions, 0 deletions
diff --git a/xlators/performance/io-cache/src/page.c b/xlators/performance/io-cache/src/page.c new file mode 100644 index 000000000..e549f0bb5 --- /dev/null +++ b/xlators/performance/io-cache/src/page.c @@ -0,0 +1,778 @@ +/* + Copyright (c) 2007, 2008, 2009 Z RESEARCH, Inc. <http://www.zresearch.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 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 + General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see + <http://www.gnu.org/licenses/>. +*/ + +#ifndef _CONFIG_H +#define _CONFIG_H +#include "config.h" +#endif + +#include "glusterfs.h" +#include "logging.h" +#include "dict.h" +#include "xlator.h" +#include "io-cache.h" +#include <assert.h> +#include <sys/time.h> + +ioc_page_t * +ioc_page_get (ioc_inode_t *ioc_inode, + off_t offset) +{ + int8_t found = 0; + ioc_page_t *page = NULL; + ioc_table_t *table = ioc_inode->table; + off_t rounded_offset = floor (offset, table->page_size); + + if (list_empty (&ioc_inode->pages)) { + return NULL; + } + + list_for_each_entry (page, &ioc_inode->pages, pages) { + if (page->offset == rounded_offset) { + found = 1; + break; + } + } + + /* was previously returning ioc_inode itself.., + * 1st of its type and found one more downstairs :O */ + if (!found){ + page = NULL; + } else { + /* push the page to the end of the lru list */ + list_move_tail (&page->page_lru, &ioc_inode->page_lru); + } + + return page; +} + + +/* + * ioc_page_destroy - + * + * @page: + * + */ +int64_t +ioc_page_destroy (ioc_page_t *page) +{ + int64_t page_size = 0; + + page_size = page->size; + + if (page->waitq) { + /* frames waiting on this page, do not destroy this page */ + page_size = -1; + } else { + + list_del (&page->pages); + list_del (&page->page_lru); + + gf_log (page->inode->table->xl->name, GF_LOG_DEBUG, + "destroying page = %p, offset = %"PRId64" " + "&& inode = %p", + page, page->offset, page->inode); + + if (page->vector){ + dict_unref (page->ref); + free (page->vector); + page->vector = NULL; + } + + page->inode = NULL; + + } + + if (page_size != -1) { + pthread_mutex_destroy (&page->page_lock); + free (page); + } + + return page_size; +} + +/* + * ioc_prune - prune the cache. we have a limit to the number of pages we + * can have in-memory. + * + * @table: ioc_table_t of this translator + * + */ +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->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_DEBUG, + "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 (list_empty (&curr->pages)) { + 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_page_create - create a new page. + * + * @ioc_inode: + * @offset: + * + */ +ioc_page_t * +ioc_page_create (ioc_inode_t *ioc_inode, + off_t offset) +{ + ioc_table_t *table = ioc_inode->table; + ioc_page_t *page = NULL; + off_t rounded_offset = floor (offset, table->page_size); + ioc_page_t *newpage = CALLOC (1, sizeof (*newpage)); + ERR_ABORT (newpage); + + if (ioc_inode) + table = ioc_inode->table; + else { + return NULL; + } + + newpage->offset = rounded_offset; + newpage->inode = ioc_inode; + pthread_mutex_init (&newpage->page_lock, NULL); + + list_add_tail (&newpage->page_lru, &ioc_inode->page_lru); + list_add_tail (&newpage->pages, &ioc_inode->pages); + + page = newpage; + + gf_log ("io-cache", GF_LOG_DEBUG, + "returning new page %p", 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 + * + * @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_waitq_t *waitq = NULL; + ioc_local_t *local = frame->local; + + waitq = CALLOC (1, sizeof (*waitq)); + ERR_ABORT (waitq); + + gf_log (frame->this->name, GF_LOG_DEBUG, + "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); +} + + +/* + * ioc_cache_still_valid - see if cached pages ioc_inode are still valid + * against given stbuf + * + * @ioc_inode: + * @stbuf: + * + * assumes ioc_inode is locked + */ +int8_t +ioc_cache_still_valid (ioc_inode_t *ioc_inode, + struct stat *stbuf) +{ + int8_t cache_still_valid = 1; + +#if 0 + if (!stbuf || (stbuf->st_mtime != ioc_inode->mtime) || + (stbuf->st_mtim.tv_nsec != ioc_inode->stbuf.st_mtim.tv_nsec)) + cache_still_valid = 0; + +#else + if (!stbuf || (stbuf->st_mtime != ioc_inode->mtime)) + cache_still_valid = 0; + +#endif + +#if 0 + /* talk with avati@zresearch.com to enable this section */ + if (!ioc_inode->mtime && stbuf) { + cache_still_valid = 1; + ioc_inode->mtime = stbuf->st_mtime; + } +#endif + + 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; + + for (trav = waitq; trav; trav = next) { + next = trav->next; + + frame = trav->data; + ioc_frame_return (frame); + 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 stat *stbuf) +{ + ioc_local_t *local = frame->local; + off_t offset = local->pending_offset; + ioc_inode_t *ioc_inode = local->inode; + ioc_table_t *table = ioc_inode->table; + ioc_page_t *page = NULL; + off_t trav_offset = 0; + size_t payload_size = 0; + int32_t destroy_size = 0; + size_t page_size = 0; + ioc_waitq_t *waitq = NULL; + + trav_offset = offset; + payload_size = op_ret; + + ioc_inode_lock (ioc_inode); + { + 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); + destroy_size = __ioc_inode_flush (ioc_inode); + } + + if (op_ret >= 0) + ioc_inode->mtime = stbuf->st_mtime; + + gettimeofday (&ioc_inode->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_DEBUG, + "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) { + dict_unref (page->ref); + free (page->vector); + page->vector = NULL; + } + + /* keep a copy of the page for our cache */ + page->vector = iov_dup (vector, count); + page->count = count; + if (frame->root->rsp_refs) { + dict_ref (frame->root->rsp_refs); + page->ref = frame->root->rsp_refs; + } 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; + + 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 */ + ioc_inode_unlock (ioc_inode); + + ioc_waitq_return (waitq); + + if (page_size) { + ioc_table_lock (table); + { + table->cache_used += page_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); + } + + gf_log (this->name, GF_LOG_DEBUG, "fault frame %p returned", frame); + pthread_mutex_destroy (&local->local_lock); + + fd_unref (local->fd); + + STACK_DESTROY (frame->root); + return 0; +} + +/* + * ioc_page_fault - + * + * @ioc_inode: + * @frame: + * @fd: + * @offset: + * + */ +void +ioc_page_fault (ioc_inode_t *ioc_inode, + call_frame_t *frame, + fd_t *fd, + off_t offset) +{ + ioc_table_t *table = ioc_inode->table; + call_frame_t *fault_frame = copy_frame (frame); + ioc_local_t *fault_local = CALLOC (1, sizeof (ioc_local_t)); + ERR_ABORT (fault_local); + + /* 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); + + 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_DEBUG, + "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; +} + +void +ioc_frame_fill (ioc_page_t *page, + call_frame_t *frame, + off_t offset, + size_t size) +{ + ioc_local_t *local = frame->local; + 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 = page->inode; + + gf_log (frame->this->name, GF_LOG_DEBUG, + "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->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_DEBUG, + "copy_size = %"GF_PRI_SIZET" && src_offset = " + "%"PRId64" && dst_offset = %"PRId64"", + copy_size, src_offset, dst_offset); + + { + ioc_fill_t *new = CALLOC (1, sizeof (*new)); + ERR_ABORT (new); + new->offset = page->offset; + new->size = copy_size; + new->refs = dict_ref (page->ref); + new->count = iov_subset (page->vector, + page->count, + src_offset, + src_offset + copy_size, + NULL); + new->vector = CALLOC (new->count, + sizeof (struct iovec)); + ERR_ABORT (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 { + int8_t 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) { + found = 0; + list_add_tail (&new->list, + &fill->list); + } else { + list_add_tail (&new->list, + &local->fill_list); + } + } + } + local->op_ret += copy_size; + } +} + +/* + * ioc_frame_unwind - frame unwinds only from here + * + * @frame: call frame to unwind + * + * to be used only by ioc_frame_return(), when a frame has + * finished waiting on all pages, required + * + */ +static void +ioc_frame_unwind (call_frame_t *frame) +{ + ioc_local_t *local = frame->local; + ioc_fill_t *fill = NULL, *next = NULL; + int32_t count = 0; + struct iovec *vector = NULL; + int32_t copied = 0; + dict_t *refs = NULL; + struct stat stbuf = {0,}; + int32_t op_ret = 0; + + // ioc_local_lock (local); + refs = get_new_dict (); + + frame->local = NULL; + + if (list_empty (&local->fill_list)) { + gf_log (frame->this->name, GF_LOG_DEBUG, + "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; + } + + vector = CALLOC (count, sizeof (*vector)); + ERR_ABORT (vector); + + list_for_each_entry_safe (fill, next, &local->fill_list, list) { + memcpy (((char *)vector) + copied, + fill->vector, + fill->count * sizeof (*vector)); + + copied += (fill->count * sizeof (*vector)); + + dict_copy (fill->refs, refs); + + list_del (&fill->list); + dict_unref (fill->refs); + free (fill->vector); + free (fill); + } + + frame->root->rsp_refs = dict_ref (refs); + + op_ret = iov_length (vector, count); + gf_log (frame->this->name, GF_LOG_DEBUG, + "frame(%p) unwinding with op_ret=%d", frame, op_ret); + + // ioc_local_unlock (local); + + STACK_UNWIND (frame, + op_ret, + local->op_errno, + vector, + count, + &stbuf); + + dict_unref (refs); + + pthread_mutex_destroy (&local->local_lock); + free (local); + free (vector); + + return; +} + +/* + * ioc_frame_return - + * @frame: + * + * to be called only when a frame is waiting on an in-transit page + */ +void +ioc_frame_return (call_frame_t *frame) +{ + ioc_local_t *local = frame->local; + int32_t wait_count; + assert (local->wait_count > 0); + + ioc_local_lock (local); + { + wait_count = --local->wait_count; + } + ioc_local_unlock (local); + + if (!wait_count) { + ioc_frame_unwind (frame); + } + + return; +} + +/* + * ioc_page_wakeup - + * @page: + * + * 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_waitq_t *waitq = NULL, *trav = NULL; + call_frame_t *frame = NULL; + + waitq = page->waitq; + page->waitq = NULL; + + trav = waitq; + page->ready = 1; + + gf_log (page->inode->table->xl->name, GF_LOG_DEBUG, + "page is %p && waitq = %p", page, waitq); + + for (trav = waitq; trav; trav = trav->next) { + frame = trav->data; + ioc_frame_fill (page, frame, trav->pending_offset, + trav->pending_size); + } + + 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; + + 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; +} |