diff options
Diffstat (limited to 'xlators/performance/write-behind')
4 files changed, 2840 insertions, 2159 deletions
diff --git a/xlators/performance/write-behind/src/Makefile.am b/xlators/performance/write-behind/src/Makefile.am index f800abad50d..a6a16fcc080 100644 --- a/xlators/performance/write-behind/src/Makefile.am +++ b/xlators/performance/write-behind/src/Makefile.am @@ -1,12 +1,16 @@ xlator_LTLIBRARIES = write-behind.la xlatordir = $(libdir)/glusterfs/$(PACKAGE_VERSION)/xlator/performance -write_behind_la_LDFLAGS = -module -avoidversion +write_behind_la_LDFLAGS = -module $(GF_XLATOR_DEFAULT_LDFLAGS) write_behind_la_SOURCES = write-behind.c write_behind_la_LIBADD = $(top_builddir)/libglusterfs/src/libglusterfs.la -AM_CFLAGS = -fPIC -D_FILE_OFFSET_BITS=64 -D_GNU_SOURCE -Wall -D$(GF_HOST_OS)\ - -I$(top_srcdir)/libglusterfs/src -shared -nostartfiles $(GF_CFLAGS) +noinst_HEADERS = write-behind-mem-types.h write-behind-messages.h + +AM_CPPFLAGS = $(GF_CPPFLAGS) -I$(top_srcdir)/libglusterfs/src \ + -I$(top_srcdir)/rpc/xdr/src -I$(top_builddir)/rpc/xdr/src + +AM_CFLAGS = -Wall $(GF_CFLAGS) CLEANFILES = diff --git a/xlators/performance/write-behind/src/write-behind-mem-types.h b/xlators/performance/write-behind/src/write-behind-mem-types.h new file mode 100644 index 00000000000..a0647299150 --- /dev/null +++ b/xlators/performance/write-behind/src/write-behind-mem-types.h @@ -0,0 +1,24 @@ +/* + Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> + This file is part of GlusterFS. + + 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 __WB_MEM_TYPES_H__ +#define __WB_MEM_TYPES_H__ + +#include <glusterfs/mem-types.h> + +enum gf_wb_mem_types_ { + gf_wb_mt_wb_file_t = gf_common_mt_end + 1, + gf_wb_mt_wb_request_t, + gf_wb_mt_iovec, + gf_wb_mt_wb_conf_t, + gf_wb_mt_wb_inode_t, + gf_wb_mt_end +}; +#endif diff --git a/xlators/performance/write-behind/src/write-behind-messages.h b/xlators/performance/write-behind/src/write-behind-messages.h new file mode 100644 index 00000000000..e9ea474879b --- /dev/null +++ b/xlators/performance/write-behind/src/write-behind-messages.h @@ -0,0 +1,31 @@ +/*Copyright (c) 2015 Red Hat, Inc. <http://www.redhat.com> + This file is part of GlusterFS. + + 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 _WRITE_BEHIND_MESSAGES_H_ +#define _WRITE_BEHIND_MESSAGES_H_ + +#include <glusterfs/glfs-message-id.h> + +/* To add new message IDs, append new identifiers at the end of the list. + * + * Never remove a message ID. If it's not used anymore, you can rename it or + * leave it as it is, but not delete it. This is to prevent reutilization of + * IDs by other messages. + * + * The component name must match one of the entries defined in + * glfs-message-id.h. + */ + +GLFS_MSGID(WRITE_BEHIND, WRITE_BEHIND_MSG_EXCEEDED_MAX_SIZE, + WRITE_BEHIND_MSG_INIT_FAILED, WRITE_BEHIND_MSG_INVALID_ARGUMENT, + WRITE_BEHIND_MSG_NO_MEMORY, WRITE_BEHIND_MSG_SIZE_NOT_SET, + WRITE_BEHIND_MSG_VOL_MISCONFIGURED, + WRITE_BEHIND_MSG_RES_UNAVAILABLE); + +#endif /* _WRITE_BEHIND_MESSAGES_H_ */ diff --git a/xlators/performance/write-behind/src/write-behind.c b/xlators/performance/write-behind/src/write-behind.c index 883d25e9666..00cfca016e6 100644 --- a/xlators/performance/write-behind/src/write-behind.c +++ b/xlators/performance/write-behind/src/write-behind.c @@ -1,2656 +1,3278 @@ /* - Copyright (c) 2006-2009 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 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/>. + 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. */ -/*TODO: check for non null wb_file_data before getting wb_file */ - - -#ifndef _CONFIG_H -#define _CONFIG_H -#include "config.h" -#endif - -#include "glusterfs.h" -#include "logging.h" -#include "dict.h" -#include "xlator.h" -#include "list.h" -#include "compat.h" -#include "compat-errno.h" -#include "common-utils.h" -#include "call-stub.h" -#include "statedump.h" +#include <glusterfs/glusterfs.h> +#include <glusterfs/logging.h> +#include <glusterfs/dict.h> +#include <glusterfs/xlator.h> +#include <glusterfs/list.h> +#include <glusterfs/compat.h> +#include <glusterfs/compat-errno.h> +#include <glusterfs/common-utils.h> +#include <glusterfs/call-stub.h> +#include <glusterfs/statedump.h> +#include <glusterfs/defaults.h> +#include "write-behind-mem-types.h" +#include "write-behind-messages.h" #define MAX_VECTOR_COUNT 8 #define WB_AGGREGATE_SIZE 131072 /* 128 KB */ -#define WB_WINDOW_SIZE 1048576 /* 1MB */ - +#define WB_WINDOW_SIZE 1048576 /* 1MB */ + typedef struct list_head list_head_t; struct wb_conf; -struct wb_page; -struct wb_file; - - -typedef struct wb_file { - int disabled; - uint64_t disable_till; - size_t window_conf; - size_t window_current; - size_t aggregate_current; - int32_t refcount; - int32_t op_ret; - int32_t op_errno; - list_head_t request; - list_head_t passive_requests; - fd_t *fd; - gf_lock_t lock; - xlator_t *this; -}wb_file_t; - +struct wb_inode; + +typedef struct wb_inode { + ssize_t window_conf; + ssize_t window_current; + ssize_t transit; /* size of data stack_wound, and yet + to be fulfilled (wb_fulfill_cbk). + used for trickling_writes + */ + + list_head_t all; /* All requests, from enqueue() till destroy(). + Used only for resetting generation + number when empty. + */ + list_head_t todo; /* Work to do (i.e, STACK_WIND to server). + Once we STACK_WIND, the entry is taken + off the list. If it is non-sync write, + then we continue to track it via @liability + or @temptation depending on the status + of its writeback. + */ + list_head_t liability; /* Non-sync writes which are lied + (STACK_UNWIND'ed to caller) but ack + from server not yet complete. This + is the "liability" which we hold, and + must guarantee that dependent operations + which arrive later (which overlap, etc.) + are issued only after their dependencies + in this list are "fulfilled". + + Server acks for entries in this list + shrinks the window. + + The sum total of all req->write_size + of entries in this list must be kept less + than the permitted window size. + */ + list_head_t temptation; /* Operations for which we are tempted + to 'lie' (write-behind), but temporarily + holding off (because of insufficient + window capacity, etc.) + + This is the list to look at to grow + the window (in __wb_pick_unwinds()). + + Entries typically get chosen from + write-behind from this list, and therefore + get "upgraded" to the "liability" list. + */ + list_head_t wip; /* List of write calls in progress, SYNC or non-SYNC + which are currently STACK_WIND'ed towards the server. + This is for guaranteeing that no two overlapping + writes are in progress at the same time. Modules + like eager-lock in AFR depend on this behavior. + */ + list_head_t invalidate_list; /* list of wb_inodes that were marked for + * iatt invalidation due to requests in + * liability queue fulfilled while there + * was a readdirp session on parent + * directory. For a directory inode, this + * list points to list of children. + */ + uint64_t gen; /* Liability generation number. Represents + the current 'state' of liability. Every + new addition to the liability list bumps + the generation number. + + a newly arrived request is only required + to perform causal checks against the entries + in the liability list which were present + at the time of its addition. the generation + number at the time of its addition is stored + in the request and used during checks. + + the liability list can grow while the request + waits in the todo list waiting for its + dependent operations to complete. however + it is not of the request's concern to depend + itself on those new entries which arrived + after it arrived (i.e, those that have a + liability generation higher than itself) + */ + size_t size; /* Size of the file to catch write after EOF. */ + gf_lock_t lock; + xlator_t *this; + inode_t *inode; + int dontsync; /* If positive, don't pick lies for + * winding. This is needed to break infinite + * recursion during invocation of + * wb_process_queue from + * wb_fulfill_cbk in case of an + * error during fulfill. + */ + gf_atomic_int32_t readdirps; + gf_atomic_int8_t invalidate; + +} wb_inode_t; typedef struct wb_request { - list_head_t list; - list_head_t winds; - list_head_t unwinds; - list_head_t other_requests; - call_stub_t *stub; - size_t write_size; - int32_t refcount; - wb_file_t *file; - union { - struct { - char write_behind; - char stack_wound; - char got_reply; - char virgin; - }write_request; - - struct { - char marked_for_resume; - }other_requests; - }flags; + list_head_t all; + list_head_t todo; + list_head_t lie; /* either in @liability or @temptation */ + list_head_t winds; + list_head_t unwinds; + list_head_t wip; + + call_stub_t *stub; + + ssize_t write_size; /* currently held size + (after collapsing) */ + size_t orig_size; /* size which arrived with the request. + This is the size by which we grow + the window when unwinding the frame. + */ + size_t total_size; /* valid only in @head in wb_fulfill(). + This is the size with which we perform + STACK_WIND to server and therefore the + amount by which we shrink the window. + */ + + int op_ret; + int op_errno; + + int32_t refcount; + wb_inode_t *wb_inode; + glusterfs_fop_t fop; + gf_lkowner_t lk_owner; + pid_t client_pid; + struct iobref *iobref; + uint64_t gen; /* inode liability state at the time of + request arrival */ + + fd_t *fd; + int wind_count; /* number of sync-attempts. Only + for debug purposes */ + struct { + size_t size; /* 0 size == till infinity */ + off_t off; + int append : 1; /* offset is invalid. only one + outstanding append at a time */ + int tempted : 1; /* true only for non-sync writes */ + int lied : 1; /* sin committed */ + int fulfilled : 1; /* got server acknowledgement */ + int go : 1; /* enough aggregating, good to go */ + } ordering; + + /* for debug purposes. A request might outlive the fop it is + * representing. So, preserve essential info for logging. + */ + uint64_t unique; + uuid_t gfid; } wb_request_t; +typedef struct wb_conf { + uint64_t aggregate_size; + uint64_t page_size; + uint64_t window_size; + gf_boolean_t flush_behind; + gf_boolean_t trickling_writes; + gf_boolean_t strict_write_ordering; + gf_boolean_t strict_O_DIRECT; + gf_boolean_t resync_after_fsync; +} wb_conf_t; + +wb_inode_t * +__wb_inode_ctx_get(xlator_t *this, inode_t *inode) +{ + uint64_t value = 0; + wb_inode_t *wb_inode = NULL; + int ret = 0; -struct wb_conf { - uint64_t aggregate_size; - uint64_t window_size; - uint64_t disable_till; - gf_boolean_t enable_O_SYNC; - gf_boolean_t flush_behind; - gf_boolean_t enable_trickling_writes; -}; + ret = __inode_ctx_get(inode, this, &value); + if (ret) + return NULL; + wb_inode = (wb_inode_t *)(unsigned long)value; -typedef struct wb_local { - list_head_t winds; - int32_t flags; - int32_t wbflags; - struct wb_file *file; - wb_request_t *request; - int op_ret; - int op_errno; - call_frame_t *frame; - int32_t reply_count; -} wb_local_t; + return wb_inode; +} + +wb_inode_t * +wb_inode_ctx_get(xlator_t *this, inode_t *inode) +{ + wb_inode_t *wb_inode = NULL; + GF_VALIDATE_OR_GOTO("write-behind", this, out); + GF_VALIDATE_OR_GOTO(this->name, inode, out); -typedef struct wb_conf wb_conf_t; -typedef struct wb_page wb_page_t; + LOCK(&inode->lock); + { + wb_inode = __wb_inode_ctx_get(this, inode); + } + UNLOCK(&inode->lock); +out: + return wb_inode; +} +static void +wb_set_invalidate(wb_inode_t *wb_inode) +{ + int readdirps = 0; + inode_t *parent_inode = NULL; + wb_inode_t *wb_parent_inode = NULL; -int32_t -wb_process_queue (call_frame_t *frame, wb_file_t *file, char flush_all); + parent_inode = inode_parent(wb_inode->inode, NULL, NULL); + if (parent_inode) + wb_parent_inode = wb_inode_ctx_get(wb_inode->this, parent_inode); -ssize_t -wb_sync (call_frame_t *frame, wb_file_t *file, list_head_t *winds); + if (wb_parent_inode) { + LOCK(&wb_parent_inode->lock); + { + readdirps = GF_ATOMIC_GET(wb_parent_inode->readdirps); + if (readdirps && list_empty(&wb_inode->invalidate_list)) { + inode_ref(wb_inode->inode); + GF_ATOMIC_INIT(wb_inode->invalidate, 1); + list_add(&wb_inode->invalidate_list, + &wb_parent_inode->invalidate_list); + } + } + UNLOCK(&wb_parent_inode->lock); + } else { + GF_ATOMIC_INIT(wb_inode->invalidate, 0); + } + + if (parent_inode) + inode_unref(parent_inode); + + return; +} -ssize_t -__wb_mark_winds (list_head_t *list, list_head_t *winds, size_t aggregate_size, - char wind_all, char enable_trickling_writes); +void +wb_process_queue(wb_inode_t *wb_inode); +/* + Below is a succinct explanation of the code deciding whether two regions + overlap, from Pavan <tcp@gluster.com>. -static void -__wb_request_unref (wb_request_t *this) + For any two ranges to be non-overlapping, either the end of the first + range is lesser than the start of the second, or vice versa. Example - + + <---------> <--------------> + p q x y + + ( q < x ) or (y < p) = > No overlap. + + To check for *overlap*, we can negate this (using de morgan's laws), and + it becomes - + + (q >= x ) and (y >= p) + + Either that, or you write the negation using - + + if (! ((q < x) or (y < p)) ) { + "Overlap" + } +*/ + +gf_boolean_t +wb_requests_overlap(wb_request_t *req1, wb_request_t *req2) { - if (this->refcount <= 0) { - gf_log ("wb-request", GF_LOG_DEBUG, - "refcount(%d) is <= 0", this->refcount); - return; - } + uint64_t r1_start = 0; + uint64_t r1_end = 0; + uint64_t r2_start = 0; + uint64_t r2_end = 0; + gf_boolean_t do_overlap = _gf_false; + + r1_start = req1->ordering.off; + if (req1->ordering.size) + r1_end = r1_start + req1->ordering.size - 1; + else + r1_end = ULLONG_MAX; + + r2_start = req2->ordering.off; + if (req2->ordering.size) + r2_end = r2_start + req2->ordering.size - 1; + else + r2_end = ULLONG_MAX; + + do_overlap = ((r1_end >= r2_start) && (r2_end >= r1_start)); + + return do_overlap; +} - this->refcount--; - if (this->refcount == 0) { - list_del_init (&this->list); - if (this->stub && this->stub->fop == GF_FOP_WRITE) { - call_stub_destroy (this->stub); - } +gf_boolean_t +wb_requests_conflict(wb_request_t *lie, wb_request_t *req) +{ + wb_conf_t *conf = NULL; - FREE (this); - } + conf = req->wb_inode->this->private; + + if (lie == req) + /* request cannot conflict with itself */ + return _gf_false; + + if (lie->gen >= req->gen) + /* this liability entry was behind + us in the todo list */ + return _gf_false; + + if (lie->ordering.append) + /* all modifications wait for the completion + of outstanding append */ + return _gf_true; + + if (conf->strict_write_ordering) + /* We are sure (lie->gen < req->gen) by now. So + skip overlap check if strict write ordering is + requested and always return "conflict" against a + lower generation lie. */ + return _gf_true; + + return wb_requests_overlap(lie, req); } +wb_request_t * +wb_liability_has_conflict(wb_inode_t *wb_inode, wb_request_t *req) +{ + wb_request_t *each = NULL; + + list_for_each_entry(each, &wb_inode->liability, lie) + { + if (wb_requests_conflict(each, req) && (!each->ordering.fulfilled)) + /* A fulfilled request shouldn't block another + * request (even a dependent one) from winding. + */ + return each; + } + + return NULL; +} -static void -wb_request_unref (wb_request_t *this) +wb_request_t * +wb_wip_has_conflict(wb_inode_t *wb_inode, wb_request_t *req) { - wb_file_t *file = NULL; - if (this == NULL) { - gf_log ("wb-request", GF_LOG_DEBUG, - "request is NULL"); - return; + wb_request_t *each = NULL; + + if (req->stub->fop != GF_FOP_WRITE) + /* non-writes fundamentally never conflict with WIP requests */ + return NULL; + + list_for_each_entry(each, &wb_inode->wip, wip) + { + if (each == req) + /* request never conflicts with itself, + though this condition should never occur. + */ + continue; + + if (wb_requests_overlap(each, req)) + return each; + } + + return NULL; +} + +static int +__wb_request_unref(wb_request_t *req) +{ + int ret = -1; + wb_inode_t *wb_inode = NULL; + char gfid[64] = { + 0, + }; + + wb_inode = req->wb_inode; + + if (req->refcount <= 0) { + uuid_utoa_r(req->gfid, gfid); + + gf_msg( + "wb-request", GF_LOG_WARNING, 0, WRITE_BEHIND_MSG_RES_UNAVAILABLE, + "(unique=%" PRIu64 ", fop=%s, gfid=%s, gen=%" PRIu64 + "): " + "refcount(%d) is <= 0 ", + req->unique, gf_fop_list[req->fop], gfid, req->gen, req->refcount); + goto out; + } + + ret = --req->refcount; + if (req->refcount == 0) { + uuid_utoa_r(req->gfid, gfid); + + gf_log_callingfn(wb_inode->this->name, GF_LOG_DEBUG, + "(unique = %" PRIu64 + ", fop=%s, gfid=%s, " + "gen=%" PRIu64 + "): destroying request, " + "removing from all queues", + req->unique, gf_fop_list[req->fop], gfid, req->gen); + + list_del_init(&req->todo); + list_del_init(&req->lie); + list_del_init(&req->wip); + + list_del_init(&req->all); + if (list_empty(&wb_inode->all)) { + wb_inode->gen = 0; + /* in case of accounting errors? */ + wb_inode->window_current = 0; } - - file = this->file; - LOCK (&file->lock); - { - __wb_request_unref (this); + + list_del_init(&req->winds); + list_del_init(&req->unwinds); + + if (req->stub) { + call_stub_destroy(req->stub); + req->stub = NULL; } - UNLOCK (&file->lock); + + if (req->iobref) + iobref_unref(req->iobref); + + if (req->fd) + fd_unref(req->fd); + + GF_FREE(req); + } +out: + return ret; } +static int +wb_request_unref(wb_request_t *req) +{ + wb_inode_t *wb_inode = NULL; + int ret = -1; + + GF_VALIDATE_OR_GOTO("write-behind", req, out); + + wb_inode = req->wb_inode; + + LOCK(&wb_inode->lock); + { + ret = __wb_request_unref(req); + } + UNLOCK(&wb_inode->lock); + +out: + return ret; +} static wb_request_t * -__wb_request_ref (wb_request_t *this) +__wb_request_ref(wb_request_t *req) { - if (this->refcount < 0) { - gf_log ("wb-request", GF_LOG_DEBUG, - "refcount(%d) is < 0", this->refcount); - return NULL; - } + GF_VALIDATE_OR_GOTO("write-behind", req, out); - this->refcount++; - return this; -} + if (req->refcount < 0) { + gf_msg("wb-request", GF_LOG_WARNING, 0, + WRITE_BEHIND_MSG_RES_UNAVAILABLE, "refcount(%d) is < 0", + req->refcount); + req = NULL; + goto out; + } + + req->refcount++; +out: + return req; +} wb_request_t * -wb_request_ref (wb_request_t *this) +wb_request_ref(wb_request_t *req) { - wb_file_t *file = NULL; - if (this == NULL) { - gf_log ("wb-request", GF_LOG_DEBUG, - "request is NULL"); - return NULL; - } + wb_inode_t *wb_inode = NULL; - file = this->file; - LOCK (&file->lock); - { - this = __wb_request_ref (this); - } - UNLOCK (&file->lock); + GF_VALIDATE_OR_GOTO("write-behind", req, out); + + wb_inode = req->wb_inode; + LOCK(&wb_inode->lock); + { + req = __wb_request_ref(req); + } + UNLOCK(&wb_inode->lock); - return this; +out: + return req; } +gf_boolean_t +wb_enqueue_common(wb_inode_t *wb_inode, call_stub_t *stub, int tempted) +{ + wb_request_t *req = NULL; + inode_t *inode = NULL; + + GF_VALIDATE_OR_GOTO("write-behind", wb_inode, out); + GF_VALIDATE_OR_GOTO(wb_inode->this->name, stub, out); + + req = GF_CALLOC(1, sizeof(*req), gf_wb_mt_wb_request_t); + if (!req) + goto out; + + INIT_LIST_HEAD(&req->all); + INIT_LIST_HEAD(&req->todo); + INIT_LIST_HEAD(&req->lie); + INIT_LIST_HEAD(&req->winds); + INIT_LIST_HEAD(&req->unwinds); + INIT_LIST_HEAD(&req->wip); + + req->stub = stub; + req->wb_inode = wb_inode; + req->fop = stub->fop; + req->ordering.tempted = tempted; + req->unique = stub->frame->root->unique; + + inode = ((stub->args.fd != NULL) ? stub->args.fd->inode + : stub->args.loc.inode); + + if (inode) + gf_uuid_copy(req->gfid, inode->gfid); + + if (stub->fop == GF_FOP_WRITE) { + req->write_size = iov_length(stub->args.vector, stub->args.count); + + /* req->write_size can change as we collapse + small writes. But the window needs to grow + only by how much we acknowledge the app. so + copy the original size in orig_size for the + purpose of accounting. + */ + req->orig_size = req->write_size; -wb_request_t * -wb_enqueue (wb_file_t *file, call_stub_t *stub) -{ - wb_request_t *request = NULL; - call_frame_t *frame = NULL; - wb_local_t *local = NULL; - struct iovec *vector = NULL; - int32_t count = 0; - - request = CALLOC (1, sizeof (*request)); - if (request == NULL) { - goto out; - } + /* Let's be optimistic that we can + lie about it + */ + req->op_ret = req->write_size; + req->op_errno = 0; + + if (stub->args.fd && (stub->args.fd->flags & O_APPEND)) + req->ordering.append = 1; + } + + req->lk_owner = stub->frame->root->lk_owner; + req->client_pid = stub->frame->root->pid; + + switch (stub->fop) { + case GF_FOP_WRITE: + LOCK(&wb_inode->lock); + { + if (wb_inode->size < stub->args.offset) { + req->ordering.off = wb_inode->size; + req->ordering.size = stub->args.offset + req->write_size - + wb_inode->size; + } else { + req->ordering.off = stub->args.offset; + req->ordering.size = req->write_size; + } - INIT_LIST_HEAD (&request->list); - INIT_LIST_HEAD (&request->winds); - INIT_LIST_HEAD (&request->unwinds); - INIT_LIST_HEAD (&request->other_requests); + if (wb_inode->size < stub->args.offset + req->write_size) + wb_inode->size = stub->args.offset + req->write_size; + } + UNLOCK(&wb_inode->lock); - request->stub = stub; - request->file = file; + req->fd = fd_ref(stub->args.fd); - frame = stub->frame; - local = frame->local; - if (local) { - local->request = request; - } + break; + case GF_FOP_READ: + req->ordering.off = stub->args.offset; + req->ordering.size = stub->args.size; - if (stub->fop == GF_FOP_WRITE) { - vector = stub->args.writev.vector; - count = stub->args.writev.count; + req->fd = fd_ref(stub->args.fd); - frame = stub->frame; - local = frame->local; - request->write_size = iov_length (vector, count); - local->op_ret = request->write_size; - local->op_errno = 0; + break; + case GF_FOP_TRUNCATE: + req->ordering.off = stub->args.offset; + req->ordering.size = 0; /* till infinity */ + LOCK(&wb_inode->lock); + { + wb_inode->size = req->ordering.off; + } + UNLOCK(&wb_inode->lock); + break; + case GF_FOP_FTRUNCATE: + req->ordering.off = stub->args.offset; + req->ordering.size = 0; /* till infinity */ + LOCK(&wb_inode->lock); + { + wb_inode->size = req->ordering.off; + } + UNLOCK(&wb_inode->lock); - request->flags.write_request.virgin = 1; - } + req->fd = fd_ref(stub->args.fd); - LOCK (&file->lock); - { - list_add_tail (&request->list, &file->request); - if (stub->fop == GF_FOP_WRITE) { - /* reference for stack winding */ - __wb_request_ref (request); + break; + default: + if (stub && stub->args.fd) + req->fd = fd_ref(stub->args.fd); - /* reference for stack unwinding */ - __wb_request_ref (request); + break; + } - file->aggregate_current += request->write_size; - } else { - /*reference for resuming */ - __wb_request_ref (request); - } + LOCK(&wb_inode->lock); + { + list_add_tail(&req->all, &wb_inode->all); + + req->gen = wb_inode->gen; + + list_add_tail(&req->todo, &wb_inode->todo); + __wb_request_ref(req); /* for wind */ + + if (req->ordering.tempted) { + list_add_tail(&req->lie, &wb_inode->temptation); + __wb_request_ref(req); /* for unwind */ } - UNLOCK (&file->lock); + } + UNLOCK(&wb_inode->lock); out: - return request; + if (!req) + return _gf_false; + + return _gf_true; } +gf_boolean_t +wb_enqueue(wb_inode_t *wb_inode, call_stub_t *stub) +{ + return wb_enqueue_common(wb_inode, stub, 0); +} -wb_file_t * -wb_file_create (xlator_t *this, fd_t *fd) +gf_boolean_t +wb_enqueue_tempted(wb_inode_t *wb_inode, call_stub_t *stub) { - wb_file_t *file = NULL; - wb_conf_t *conf = this->private; + return wb_enqueue_common(wb_inode, stub, 1); +} - file = CALLOC (1, sizeof (*file)); - if (file == NULL) { - goto out; - } +wb_inode_t * +__wb_inode_create(xlator_t *this, inode_t *inode) +{ + wb_inode_t *wb_inode = NULL; + wb_conf_t *conf = NULL; + int ret = 0; - INIT_LIST_HEAD (&file->request); - INIT_LIST_HEAD (&file->passive_requests); + GF_VALIDATE_OR_GOTO(this->name, inode, out); - /* - fd_ref() not required, file should never decide the existance of - an fd - */ - file->fd= fd; - file->disable_till = conf->disable_till; - file->this = this; - file->refcount = 1; - file->window_conf = conf->window_size; + conf = this->private; + + wb_inode = GF_CALLOC(1, sizeof(*wb_inode), gf_wb_mt_wb_inode_t); + if (!wb_inode) + goto out; + + INIT_LIST_HEAD(&wb_inode->all); + INIT_LIST_HEAD(&wb_inode->todo); + INIT_LIST_HEAD(&wb_inode->liability); + INIT_LIST_HEAD(&wb_inode->temptation); + INIT_LIST_HEAD(&wb_inode->wip); + INIT_LIST_HEAD(&wb_inode->invalidate_list); + + wb_inode->this = this; + + wb_inode->window_conf = conf->window_size; + wb_inode->inode = inode; - fd_ctx_set (fd, this, (uint64_t)(long)file); + LOCK_INIT(&wb_inode->lock); + GF_ATOMIC_INIT(wb_inode->invalidate, 0); + GF_ATOMIC_INIT(wb_inode->readdirps, 0); + + ret = __inode_ctx_put(inode, this, (uint64_t)(unsigned long)wb_inode); + if (ret) { + GF_FREE(wb_inode); + wb_inode = NULL; + } out: - return file; + return wb_inode; } -void -wb_file_destroy (wb_file_t *file) +wb_inode_t * +wb_inode_create(xlator_t *this, inode_t *inode) { - int32_t refcount = 0; + wb_inode_t *wb_inode = NULL; - LOCK (&file->lock); - { - refcount = --file->refcount; - } - UNLOCK (&file->lock); + GF_VALIDATE_OR_GOTO(this->name, inode, out); - if (!refcount){ - LOCK_DESTROY (&file->lock); - FREE (file); - } + LOCK(&inode->lock); + { + wb_inode = __wb_inode_ctx_get(this, inode); + if (!wb_inode) + wb_inode = __wb_inode_create(this, inode); + } + UNLOCK(&inode->lock); - return; +out: + return wb_inode; } +void +wb_inode_destroy(wb_inode_t *wb_inode) +{ + GF_VALIDATE_OR_GOTO("write-behind", wb_inode, out); + + GF_ASSERT(list_empty(&wb_inode->todo)); + GF_ASSERT(list_empty(&wb_inode->liability)); + GF_ASSERT(list_empty(&wb_inode->temptation)); -int32_t -wb_sync_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno, struct stat *prebuf, struct stat *postbuf) + LOCK_DESTROY(&wb_inode->lock); + GF_FREE(wb_inode); +out: + return; +} + +void +__wb_fulfill_request(wb_request_t *req) { - wb_local_t *local = NULL; - list_head_t *winds = NULL; - wb_file_t *file = NULL; - wb_request_t *request = NULL, *dummy = NULL; - wb_local_t *per_request_local = NULL; - int32_t ret = -1; - fd_t *fd = NULL; + wb_inode_t *wb_inode = NULL; + char gfid[64] = { + 0, + }; + + wb_inode = req->wb_inode; + + req->ordering.fulfilled = 1; + wb_inode->window_current -= req->total_size; + wb_inode->transit -= req->total_size; + + uuid_utoa_r(req->gfid, gfid); + + gf_log_callingfn(wb_inode->this->name, GF_LOG_DEBUG, + "(unique=%" PRIu64 + ", fop=%s, gfid=%s, " + "gen=%" PRIu64 + "): request fulfilled. " + "removing the request from liability queue? = %s", + req->unique, gf_fop_list[req->fop], gfid, req->gen, + req->ordering.lied ? "yes" : "no"); + + if (req->ordering.lied) { + /* 1. If yes, request is in liability queue and hence can be + safely removed from list. + 2. If no, request is in temptation queue and hence should be + left in the queue so that wb_pick_unwinds picks it up + */ + list_del_init(&req->lie); + } else { + /* TODO: fail the req->frame with error if + necessary + */ + } + list_del_init(&req->wip); + __wb_request_unref(req); +} - local = frame->local; - winds = &local->winds; - file = local->file; +/* get a flush/fsync waiting on req */ +wb_request_t * +__wb_request_waiting_on(wb_request_t *req) +{ + wb_inode_t *wb_inode = NULL; + wb_request_t *trav = NULL; - LOCK (&file->lock); - { - list_for_each_entry_safe (request, dummy, winds, winds) { - request->flags.write_request.got_reply = 1; + wb_inode = req->wb_inode; - if (!request->flags.write_request.write_behind - && (op_ret == -1)) { - per_request_local = request->stub->frame->local; - per_request_local->op_ret = op_ret; - per_request_local->op_errno = op_errno; - } + list_for_each_entry(trav, &wb_inode->todo, todo) + { + if (((trav->stub->fop == GF_FOP_FLUSH) || + (trav->stub->fop == GF_FOP_FSYNC)) && + (trav->gen >= req->gen)) + return trav; + } - if (request->flags.write_request.write_behind) { - file->window_current -= request->write_size; - } + return NULL; +} - __wb_request_unref (request); - } - - if (op_ret == -1) { - file->op_ret = op_ret; - file->op_errno = op_errno; - } - fd = file->fd; - } - UNLOCK (&file->lock); - - ret = wb_process_queue (frame, file, 0); - if ((ret == -1) && (errno == ENOMEM)) { - LOCK (&file->lock); - { - file->op_ret = -1; - file->op_errno = ENOMEM; - } - UNLOCK (&file->lock); - } +void +__wb_add_request_for_retry(wb_request_t *req) +{ + wb_inode_t *wb_inode = NULL; - /* safe place to do fd_unref */ - fd_unref (fd); + if (!req) + goto out; - STACK_DESTROY (frame->root); + wb_inode = req->wb_inode; - return 0; + /* response was unwound and no waiter waiting on this request, retry + till a flush or fsync (subject to conf->resync_after_fsync). + */ + wb_inode->transit -= req->total_size; + + req->total_size = 0; + + list_del_init(&req->winds); + list_del_init(&req->todo); + list_del_init(&req->wip); + + /* sanitize ordering flags to retry */ + req->ordering.go = 0; + + /* Add back to todo list to retry */ + list_add(&req->todo, &wb_inode->todo); + +out: + return; } +void +__wb_add_head_for_retry(wb_request_t *head) +{ + wb_request_t *req = NULL, *tmp = NULL; -ssize_t -wb_sync (call_frame_t *frame, wb_file_t *file, list_head_t *winds) -{ - wb_request_t *dummy = NULL, *request = NULL; - wb_request_t *first_request = NULL, *next = NULL; - size_t total_count = 0, count = 0; - size_t copied = 0; - call_frame_t *sync_frame = NULL; - struct iobref *iobref = NULL; - wb_local_t *local = NULL; - struct iovec *vector = NULL; - ssize_t current_size = 0, bytes = 0; - size_t bytecount = 0; - wb_conf_t *conf = NULL; - fd_t *fd = NULL; - - conf = file->this->private; - list_for_each_entry (request, winds, winds) { - total_count += request->stub->args.writev.count; - if (total_count > 0) { - break; - } - } + if (!head) + goto out; - if (total_count == 0) { - goto out; - } - - list_for_each_entry_safe (request, dummy, winds, winds) { - if (!vector) { - vector = MALLOC (VECTORSIZE (MAX_VECTOR_COUNT)); - if (vector == NULL) { - bytes = -1; - goto out; - } - - iobref = iobref_new (); - if (iobref == NULL) { - bytes = -1; - goto out; - } - - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - bytes = -1; - goto out; - } - - INIT_LIST_HEAD (&local->winds); - - first_request = request; - current_size = 0; - } + list_for_each_entry_safe_reverse(req, tmp, &head->winds, winds) + { + __wb_add_request_for_retry(req); + } - count += request->stub->args.writev.count; - bytecount = VECTORSIZE (request->stub->args.writev.count); - memcpy (((char *)vector)+copied, - request->stub->args.writev.vector, - bytecount); - copied += bytecount; - - current_size += request->write_size; - - if (request->stub->args.writev.iobref) { - iobref_merge (iobref, - request->stub->args.writev.iobref); - } + __wb_add_request_for_retry(head); - next = NULL; - if (request->winds.next != winds) { - next = list_entry (request->winds.next, - wb_request_t, winds); - } +out: + return; +} - list_del_init (&request->winds); - list_add_tail (&request->winds, &local->winds); - - if ((!next) - || ((count + next->stub->args.writev.count) - > MAX_VECTOR_COUNT) - || ((current_size + next->write_size) - > conf->aggregate_size)) - { - sync_frame = copy_frame (frame); - if (sync_frame == NULL) { - bytes = -1; - goto out; - } - - sync_frame->local = local; - local->file = file; - - LOCK (&file->lock); - { - fd = file->fd; - } - UNLOCK (&file->lock); - - fd_ref (fd); - - bytes += current_size; - STACK_WIND (sync_frame, - wb_sync_cbk, - FIRST_CHILD(sync_frame->this), - FIRST_CHILD(sync_frame->this)->fops->writev, - fd, vector, - count, - first_request->stub->args.writev.off, - iobref); - - iobref_unref (iobref); - FREE (vector); - first_request = NULL; - iobref = NULL; - vector = NULL; - sync_frame = NULL; - local = NULL; - copied = count = 0; - } - } +void +wb_add_head_for_retry(wb_request_t *head) +{ + if (!head) + goto out; + + LOCK(&head->wb_inode->lock); + { + __wb_add_head_for_retry(head); + } + UNLOCK(&head->wb_inode->lock); out: - if (sync_frame != NULL) { - sync_frame->local = NULL; - STACK_DESTROY (sync_frame->root); - } + return; +} - if (local != NULL) { - FREE (local); - } +void +__wb_fulfill_request_err(wb_request_t *req, int32_t op_errno) +{ + wb_inode_t *wb_inode = NULL; + wb_request_t *waiter = NULL; + wb_conf_t *conf = NULL; + + wb_inode = req->wb_inode; + + conf = wb_inode->this->private; - if (iobref != NULL) { - iobref_unref (iobref); + req->op_ret = -1; + req->op_errno = op_errno; + + if (req->ordering.lied) + waiter = __wb_request_waiting_on(req); + + if (!req->ordering.lied || waiter) { + if (!req->ordering.lied) { + /* response to app is still pending, send failure in + * response. + */ + } else { + /* response was sent, store the error in a + * waiter (either an fsync or flush). + */ + waiter->op_ret = -1; + waiter->op_errno = op_errno; } - if (vector != NULL) { - FREE (vector); + if (!req->ordering.lied || (waiter->stub->fop == GF_FOP_FLUSH) || + ((waiter->stub->fop == GF_FOP_FSYNC) && + !conf->resync_after_fsync)) { + /* No retry needed, forget the request */ + __wb_fulfill_request(req); + return; } + } - return bytes; -} + __wb_add_request_for_retry(req); + return; +} -int32_t -wb_stat_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno, struct stat *buf) +void +wb_head_done(wb_request_t *head) { - wb_local_t *local = NULL; - wb_request_t *request = NULL; - call_frame_t *process_frame = NULL; - wb_file_t *file = NULL; - int32_t ret = -1; - fd_t *fd = NULL; - - local = frame->local; - file = local->file; + wb_request_t *req = NULL; + wb_request_t *tmp = NULL; + wb_inode_t *wb_inode = NULL; - request = local->request; - if (request) { - process_frame = copy_frame (frame); - if (process_frame == NULL) { - op_ret = -1; - op_errno = ENOMEM; - } + wb_inode = head->wb_inode; + + LOCK(&wb_inode->lock); + { + list_for_each_entry_safe(req, tmp, &head->winds, winds) + { + __wb_fulfill_request(req); } - STACK_UNWIND_STRICT (stat, frame, op_ret, op_errno, buf); + __wb_fulfill_request(head); + } + UNLOCK(&wb_inode->lock); +} - if (request != NULL) { - wb_request_unref (request); - } +void +__wb_fulfill_err(wb_request_t *head, int op_errno) +{ + wb_request_t *req = NULL, *tmp = NULL; - if (process_frame != NULL) { - ret = wb_process_queue (process_frame, file, 0); - if ((ret == -1) && (errno == ENOMEM) && (file != NULL)) { - LOCK (&file->lock); - { - file->op_ret = -1; - file->op_errno = ENOMEM; - } - UNLOCK (&file->lock); - } + if (!head) + goto out; - STACK_DESTROY (process_frame->root); - } + head->wb_inode->dontsync++; - if (file) { - LOCK (&file->lock); - { - fd = file->fd; - } - UNLOCK (&file->lock); + list_for_each_entry_safe_reverse(req, tmp, &head->winds, winds) + { + __wb_fulfill_request_err(req, op_errno); + } - fd_unref (fd); - } + __wb_fulfill_request_err(head, op_errno); - return 0; +out: + return; } +void +wb_fulfill_err(wb_request_t *head, int op_errno) +{ + wb_inode_t *wb_inode = NULL; + + wb_inode = head->wb_inode; + + LOCK(&wb_inode->lock); + { + __wb_fulfill_err(head, op_errno); + } + UNLOCK(&wb_inode->lock); +} -static int32_t -wb_stat_helper (call_frame_t *frame, xlator_t *this, loc_t *loc) +void +__wb_modify_write_request(wb_request_t *req, int synced_size) { - STACK_WIND (frame, wb_stat_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->stat, - loc); - return 0; + struct iovec *vector = NULL; + int count = 0; + + if (!req || synced_size == 0) + goto out; + + req->write_size -= synced_size; + req->stub->args.offset += synced_size; + + vector = req->stub->args.vector; + count = req->stub->args.count; + + req->stub->args.count = iov_skip(vector, count, synced_size); + +out: + return; } +int +__wb_fulfill_short_write(wb_request_t *req, int size, gf_boolean_t *fulfilled) +{ + int accounted_size = 0; -int32_t -wb_stat (call_frame_t *frame, xlator_t *this, loc_t *loc) -{ - wb_file_t *file = NULL; - fd_t *iter_fd = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - wb_request_t *request = NULL; - int32_t ret = -1, op_errno = EINVAL; - - if (loc->inode) { - /* FIXME: fd_lookup extends life of fd till stat returns */ - iter_fd = fd_lookup (loc->inode, frame->root->pid); - if (iter_fd) { - if (!fd_ctx_get (iter_fd, this, &tmp_file)) { - file = (wb_file_t *)(long)tmp_file; - } else { - fd_unref (iter_fd); - iter_fd = NULL; - } - } - } + if (req == NULL) + goto out; - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - op_errno = ENOMEM; - goto unwind; - } + if (req->write_size <= size) { + accounted_size = req->write_size; + __wb_fulfill_request(req); + *fulfilled = 1; + } else { + accounted_size = size; + __wb_modify_write_request(req, size); + *fulfilled = 0; + } - local->file = file; +out: + return accounted_size; +} - frame->local = local; +void +wb_fulfill_short_write(wb_request_t *head, int size) +{ + wb_inode_t *wb_inode = NULL; + wb_request_t *req = NULL, *next = NULL; + int accounted_size = 0; + gf_boolean_t fulfilled = _gf_false; - if (file) { - stub = fop_stat_stub (frame, wb_stat_helper, loc); - if (stub == NULL) { - op_errno = ENOMEM; - goto unwind; - } + if (!head) + goto out; - request = wb_enqueue (file, stub); - if (request == NULL) { - op_errno = ENOMEM; - goto unwind; - } + wb_inode = head->wb_inode; - ret = wb_process_queue (frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - op_errno = ENOMEM; - goto unwind; - } + req = head; - } else { - STACK_WIND (frame, wb_stat_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->stat, - loc); - } - return 0; + LOCK(&wb_inode->lock); + { + /* hold a reference to head so that __wb_fulfill_short_write + * won't free it. We need head for a cleaner list traversal as + * list_for_each_entry_safe doesn't iterate over "head" member. + * So, if we pass "next->winds" as head to list_for_each_entry, + * "next" is skipped. For a simpler logic we need to traverse + * the list in the order. So, we start traversal from + * "head->winds" and hence we want head to be alive. + */ + __wb_request_ref(head); -unwind: - STACK_UNWIND_STRICT (stat, frame, -1, op_errno, NULL); + next = list_entry(head->winds.next, wb_request_t, winds); - if (stub) { - call_stub_destroy (stub); - } - - if (iter_fd != NULL) { - fd_unref (iter_fd); - } + accounted_size = __wb_fulfill_short_write(head, size, &fulfilled); - return 0; -} + size -= accounted_size; + if (size == 0) { + if (fulfilled && (next != head)) + req = next; -int32_t -wb_fstat_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno, struct stat *buf) -{ - wb_local_t *local = NULL; - wb_request_t *request = NULL; - wb_file_t *file = NULL; - int32_t ret = -1; - - local = frame->local; - file = local->file; + goto done; + } - request = local->request; - if ((file != NULL) && (request != NULL)) { - wb_request_unref (request); - ret = wb_process_queue (frame, file, 0); - if ((ret == -1) && (errno == ENOMEM)) { - op_ret = -1; - op_errno = ENOMEM; - } + list_for_each_entry_safe(req, next, &head->winds, winds) + { + accounted_size = __wb_fulfill_short_write(req, size, &fulfilled); + size -= accounted_size; + + if (size == 0) { + if (fulfilled && (next != head)) + req = next; + break; + } } + done: + __wb_request_unref(head); + } + UNLOCK(&wb_inode->lock); - STACK_UNWIND_STRICT (fstat, frame, op_ret, op_errno, buf); + wb_add_head_for_retry(req); +out: + return; +} - return 0; +int +wb_fulfill_cbk(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, struct iatt *prebuf, + struct iatt *postbuf, dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + wb_request_t *head = NULL; + + head = frame->local; + frame->local = NULL; + + wb_inode = head->wb_inode; + + /* There could be a readdirp session in progress. Since wb_fulfill_cbk + * can potentially remove a request from liability queue, + * wb_readdirp_cbk will miss writes on this inode (as it invalidates + * stats only if liability queue is not empty) and hence mark inode + * for invalidation of stats in readdirp response. Specifically this + * code fixes the following race mentioned in wb_readdirp_cbk: + */ + + /* <removed comment from wb_readdirp_cbk> + * We cannot guarantee integrity of entry->d_stat as there are cached + * writes. The stat is most likely stale as it doesn't account the + * cached writes. However, checking for non-empty liability list here is + * not a fool-proof solution as there can be races like, + * 1. readdirp is successful on posix + * 2. sync of cached write is successful on posix + * 3. write-behind received sync response and removed the request from + * liability queue + * 4. readdirp response is processed at write-behind + * + * In the above scenario, stat for the file is sent back in readdirp + * response but it is stale. + * </comment> */ + wb_set_invalidate(wb_inode); + + if (op_ret == -1) { + wb_fulfill_err(head, op_errno); + } else if (op_ret < head->total_size) { + wb_fulfill_short_write(head, op_ret); + } else { + wb_head_done(head); + } + + wb_process_queue(wb_inode); + + STACK_DESTROY(frame->root); + + return 0; } +#define WB_IOV_LOAD(vec, cnt, req, head) \ + do { \ + memcpy(&vec[cnt], req->stub->args.vector, \ + (req->stub->args.count * sizeof(vec[0]))); \ + cnt += req->stub->args.count; \ + head->total_size += req->write_size; \ + } while (0) -int32_t -wb_fstat_helper (call_frame_t *frame, xlator_t *this, fd_t *fd) +int +wb_fulfill_head(wb_inode_t *wb_inode, wb_request_t *head) { - STACK_WIND (frame, - wb_fstat_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->fstat, - fd); - return 0; + struct iovec vector[MAX_VECTOR_COUNT]; + int count = 0; + wb_request_t *req = NULL; + call_frame_t *frame = NULL; + + /* make sure head->total_size is updated before we run into any + * errors + */ + + WB_IOV_LOAD(vector, count, head, head); + + list_for_each_entry(req, &head->winds, winds) + { + WB_IOV_LOAD(vector, count, req, head); + + if (iobref_merge(head->stub->args.iobref, req->stub->args.iobref)) + goto err; + } + + frame = create_frame(wb_inode->this, wb_inode->this->ctx->pool); + if (!frame) + goto err; + + frame->root->lk_owner = head->lk_owner; + frame->root->pid = head->client_pid; + frame->local = head; + + LOCK(&wb_inode->lock); + { + wb_inode->transit += head->total_size; + } + UNLOCK(&wb_inode->lock); + + STACK_WIND(frame, wb_fulfill_cbk, FIRST_CHILD(frame->this), + FIRST_CHILD(frame->this)->fops->writev, head->fd, vector, count, + head->stub->args.offset, head->stub->args.flags, + head->stub->args.iobref, NULL); + + return 0; +err: + /* frame creation failure */ + wb_fulfill_err(head, ENOMEM); + + return ENOMEM; } +#define NEXT_HEAD(head, req) \ + do { \ + if (head) \ + ret |= wb_fulfill_head(wb_inode, head); \ + head = req; \ + expected_offset = req->stub->args.offset + req->write_size; \ + curr_aggregate = 0; \ + vector_count = 0; \ + } while (0) -int32_t -wb_fstat (call_frame_t *frame, xlator_t *this, fd_t *fd) +int +wb_fulfill(wb_inode_t *wb_inode, list_head_t *liabilities) { - wb_file_t *file = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - wb_request_t *request = NULL; - int32_t ret = -1; - int op_errno = EINVAL; + wb_request_t *req = NULL; + wb_request_t *head = NULL; + wb_request_t *tmp = NULL; + wb_conf_t *conf = NULL; + off_t expected_offset = 0; + size_t curr_aggregate = 0; + size_t vector_count = 0; + int ret = 0; + + conf = wb_inode->this->private; - if ((!S_ISDIR (fd->inode->st_mode)) - && fd_ctx_get (fd, this, &tmp_file)) { - gf_log (this->name, GF_LOG_DEBUG, "write behind file pointer is" - " not stored in context of fd(%p), returning EBADFD", - fd); + list_for_each_entry_safe(req, tmp, liabilities, winds) + { + list_del_init(&req->winds); - STACK_UNWIND_STRICT (fstat, frame, -1, EBADFD, NULL); - return 0; + if (!head) { + NEXT_HEAD(head, req); + continue; } - file = (wb_file_t *)(long)tmp_file; - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - STACK_UNWIND_STRICT (fstat, frame, -1, ENOMEM, NULL); - return 0; + if (req->fd != head->fd) { + NEXT_HEAD(head, req); + continue; } - local->file = file; + if (!is_same_lkowner(&req->lk_owner, &head->lk_owner)) { + NEXT_HEAD(head, req); + continue; + } - frame->local = local; + if (expected_offset != req->stub->args.offset) { + NEXT_HEAD(head, req); + continue; + } - if (file) { - stub = fop_fstat_stub (frame, wb_fstat_helper, fd); - if (stub == NULL) { - op_errno = ENOMEM; - goto unwind; - } - - request = wb_enqueue (file, stub); - if (request == NULL) { - op_errno = ENOMEM; - goto unwind; - } + if ((curr_aggregate + req->write_size) > conf->aggregate_size) { + NEXT_HEAD(head, req); + continue; + } - /* - FIXME:should the request queue be emptied in case of error? - */ - ret = wb_process_queue (frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - op_errno = ENOMEM; - goto unwind; - } - } else { - STACK_WIND (frame, - wb_fstat_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->fstat, - fd); + if (vector_count + req->stub->args.count > MAX_VECTOR_COUNT) { + NEXT_HEAD(head, req); + continue; } - return 0; -unwind: - STACK_UNWIND_STRICT (fstat, frame, -1, op_errno, NULL); + list_add_tail(&req->winds, &head->winds); + curr_aggregate += req->write_size; + vector_count += req->stub->args.count; + } - if (stub) { - call_stub_destroy (stub); - } + if (head) + ret |= wb_fulfill_head(wb_inode, head); - return 0; + return ret; } +void +wb_do_unwinds(wb_inode_t *wb_inode, list_head_t *lies) +{ + wb_request_t *req = NULL; + wb_request_t *tmp = NULL; + call_frame_t *frame = NULL; + struct iatt buf = { + 0, + }; + + list_for_each_entry_safe(req, tmp, lies, unwinds) + { + frame = req->stub->frame; + + STACK_UNWIND_STRICT(writev, frame, req->op_ret, req->op_errno, &buf, + &buf, NULL); /* :O */ + req->stub->frame = NULL; + + list_del_init(&req->unwinds); + wb_request_unref(req); + } + + return; +} -int32_t -wb_truncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct stat *prebuf, - struct stat *postbuf) -{ - wb_local_t *local = NULL; - wb_request_t *request = NULL; - wb_file_t *file = NULL; - call_frame_t *process_frame = NULL; - int32_t ret = -1; - fd_t *fd = NULL; - - local = frame->local; - file = local->file; - request = local->request; - - if ((request != NULL) && (file != NULL)) { - process_frame = copy_frame (frame); - if (process_frame == NULL) { - op_ret = -1; - op_errno = ENOMEM; - } - } +void +__wb_pick_unwinds(wb_inode_t *wb_inode, list_head_t *lies) +{ + wb_request_t *req = NULL; + wb_request_t *tmp = NULL; + char gfid[64] = { + 0, + }; + + list_for_each_entry_safe(req, tmp, &wb_inode->temptation, lie) + { + if (!req->ordering.fulfilled && + wb_inode->window_current > wb_inode->window_conf) + continue; - STACK_UNWIND_STRICT (truncate, frame, op_ret, op_errno, prebuf, postbuf); + list_del_init(&req->lie); + list_move_tail(&req->unwinds, lies); - if (request) { - wb_request_unref (request); + wb_inode->window_current += req->orig_size; + + wb_inode->gen++; + + if (!req->ordering.fulfilled) { + /* burden increased */ + list_add_tail(&req->lie, &wb_inode->liability); + + req->ordering.lied = 1; + + uuid_utoa_r(req->gfid, gfid); + gf_msg_debug(wb_inode->this->name, 0, + "(unique=%" PRIu64 + ", fop=%s, gfid=%s, " + "gen=%" PRIu64 + "): added req to liability " + "queue. inode-generation-number=%" PRIu64, + req->stub->frame->root->unique, gf_fop_list[req->fop], + gfid, req->gen, wb_inode->gen); } + } - if (process_frame != NULL) { - ret = wb_process_queue (process_frame, file, 0); - if ((ret == -1) && (errno == ENOMEM) && (file != NULL)) { - LOCK (&file->lock); - { - file->op_ret = -1; - file->op_errno = ENOMEM; - } - UNLOCK (&file->lock); - } + return; +} - STACK_DESTROY (process_frame->root); +int +__wb_collapse_small_writes(wb_conf_t *conf, wb_request_t *holder, + wb_request_t *req) +{ + char *ptr = NULL; + struct iobuf *iobuf = NULL; + struct iobref *iobref = NULL; + int ret = -1; + ssize_t required_size = 0; + size_t holder_len = 0; + size_t req_len = 0; + + if (!holder->iobref) { + holder_len = iov_length(holder->stub->args.vector, + holder->stub->args.count); + req_len = iov_length(req->stub->args.vector, req->stub->args.count); + + required_size = max((conf->page_size), (holder_len + req_len)); + iobuf = iobuf_get2(req->wb_inode->this->ctx->iobuf_pool, required_size); + if (iobuf == NULL) { + goto out; } - if (file) { - LOCK (&file->lock); - { - fd = file->fd; - } - UNLOCK (&file->lock); + iobref = iobref_new(); + if (iobref == NULL) { + iobuf_unref(iobuf); + goto out; + } - fd_unref (fd); + ret = iobref_add(iobref, iobuf); + if (ret != 0) { + gf_msg(req->wb_inode->this->name, GF_LOG_WARNING, -ret, + WRITE_BEHIND_MSG_INVALID_ARGUMENT, + "cannot add iobuf (%p) into iobref (%p)", iobuf, iobref); + iobuf_unref(iobuf); + iobref_unref(iobref); + goto out; } - return 0; -} + iov_unload(iobuf->ptr, holder->stub->args.vector, + holder->stub->args.count); + holder->stub->args.vector[0].iov_base = iobuf->ptr; + holder->stub->args.count = 1; + iobref_unref(holder->stub->args.iobref); + holder->stub->args.iobref = iobref; -static int32_t -wb_truncate_helper (call_frame_t *frame, xlator_t *this, loc_t *loc, - off_t offset) -{ - STACK_WIND (frame, - wb_truncate_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->truncate, - loc, - offset); + iobuf_unref(iobuf); - return 0; + holder->iobref = iobref_ref(iobref); + } + + ptr = holder->stub->args.vector[0].iov_base + holder->write_size; + + iov_unload(ptr, req->stub->args.vector, req->stub->args.count); + + holder->stub->args.vector[0].iov_len += req->write_size; + holder->write_size += req->write_size; + holder->ordering.size += req->write_size; + + ret = 0; +out: + return ret; } +void +__wb_preprocess_winds(wb_inode_t *wb_inode) +{ + off_t offset_expected = 0; + ssize_t space_left = 0; + wb_request_t *req = NULL; + wb_request_t *tmp = NULL; + wb_request_t *holder = NULL; + wb_conf_t *conf = NULL; + int ret = 0; + ssize_t page_size = 0; + char gfid[64] = { + 0, + }; + + /* With asynchronous IO from a VM guest (as a file), there + can be two sequential writes happening in two regions + of the file. But individual (broken down) IO requests + can arrive interleaved. + + TODO: cycle for each such sequence sifting + through the interleaved ops + */ + + conf = wb_inode->this->private; + page_size = conf->page_size; + + list_for_each_entry_safe(req, tmp, &wb_inode->todo, todo) + { + if (wb_inode->dontsync && req->ordering.lied) { + /* sync has failed. Don't pick lies _again_ for winding + * as winding these lies again will trigger an infinite + * recursion of wb_process_queue being called from a + * failed fulfill. However, pick non-lied requests for + * winding so that application won't block indefinitely + * waiting for write result. + */ + + uuid_utoa_r(req->gfid, gfid); + gf_msg_debug(wb_inode->this->name, 0, + "(unique=%" PRIu64 + ", fop=%s, gfid=%s, " + "gen=%" PRIu64 + "): not setting ordering.go" + "as dontsync is set", + req->unique, gf_fop_list[req->fop], gfid, req->gen); + + continue; + } + + if (!req->ordering.tempted) { + if (holder) { + if (wb_requests_conflict(holder, req)) + /* do not hold on write if a + dependent write is in queue */ + holder->ordering.go = 1; + } + /* collapse only non-sync writes */ + continue; + } else if (!holder) { + /* holder is always a non-sync write */ + holder = req; + continue; + } + + offset_expected = holder->stub->args.offset + holder->write_size; + + if (req->stub->args.offset != offset_expected) { + holder->ordering.go = 1; + holder = req; + continue; + } + + if (!is_same_lkowner(&req->lk_owner, &holder->lk_owner)) { + holder->ordering.go = 1; + holder = req; + continue; + } + + if (req->fd != holder->fd) { + holder->ordering.go = 1; + holder = req; + continue; + } + + space_left = page_size - holder->write_size; + + if (space_left < req->write_size) { + holder->ordering.go = 1; + holder = req; + continue; + } + + ret = __wb_collapse_small_writes(conf, holder, req); + if (ret) + continue; + + /* collapsed request is as good as wound + (from its p.o.v) + */ + list_del_init(&req->todo); + __wb_fulfill_request(req); + + /* Only the last @holder in queue which + + - does not have any non-buffered-writes following it + - has not yet filled its capacity + + does not get its 'go' set, in anticipation of the arrival + of consecutive smaller writes. + */ + } + + /* but if trickling writes are enabled, then do not hold back + writes if there are no outstanding requests + */ -int32_t -wb_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset) + if (conf->trickling_writes && !wb_inode->transit && holder) + holder->ordering.go = 1; + + if (wb_inode->dontsync > 0) + wb_inode->dontsync--; + + return; +} + +int +__wb_handle_failed_conflict(wb_request_t *req, wb_request_t *conflict, + list_head_t *tasks) { - wb_file_t *file = NULL; - fd_t *iter_fd = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - wb_request_t *request = NULL; - int32_t ret = -1, op_errno = ENOMEM; + wb_conf_t *conf = NULL; + char gfid[64] = { + 0, + }; + + conf = req->wb_inode->this->private; + + uuid_utoa_r(req->gfid, gfid); + + if ((req->stub->fop != GF_FOP_FLUSH) && + ((req->stub->fop != GF_FOP_FSYNC) || conf->resync_after_fsync)) { + if (!req->ordering.lied && list_empty(&conflict->wip)) { + /* If request itself is in liability queue, + * 1. We cannot unwind as the response has already been + * sent. + * 2. We cannot wind till conflict clears up. + * 3. So, skip the request for now. + * 4. Otherwise, resume (unwind) it with error. + */ + req->op_ret = -1; + req->op_errno = conflict->op_errno; + if ((req->stub->fop == GF_FOP_TRUNCATE) || + (req->stub->fop == GF_FOP_FTRUNCATE)) { + req->stub->frame->local = NULL; + } + + list_del_init(&req->todo); + list_add_tail(&req->winds, tasks); + + gf_msg_debug(req->wb_inode->this->name, 0, + "(unique=%" PRIu64 + ", fop=%s, gfid=%s, " + "gen=%" PRIu64 + "): A conflicting write " + "request in liability queue has failed " + "to sync (error = \"%s\"), " + "unwinding this request as a failure", + req->unique, gf_fop_list[req->fop], gfid, req->gen, + strerror(req->op_errno)); + + if (req->ordering.tempted) { + /* make sure that it won't be unwound in + * wb_do_unwinds too. Otherwise there'll be + * a double wind. + */ + list_del_init(&req->lie); + + gf_msg_debug(req->wb_inode->this->name, 0, + "(unique=%" PRIu64 + ", fop=%s, " + "gfid=%s, gen=%" PRIu64 + "): " + "removed from liability queue", + req->unique, gf_fop_list[req->fop], gfid, + req->gen); + + __wb_fulfill_request(req); + } + } + } else { + gf_msg_debug(req->wb_inode->this->name, 0, + "(unique=%" PRIu64 + ", fop=%s, gfid=%s, " + "gen=%" PRIu64 + "): A conflicting write request " + "in liability queue has failed to sync " + "(error = \"%s\"). This is an " + "FSYNC/FLUSH and we need to maintain ordering " + "guarantees with other writes in TODO queue. " + "Hence doing nothing now", + req->unique, gf_fop_list[req->fop], gfid, req->gen, + strerror(conflict->op_errno)); + + /* flush and fsync (without conf->resync_after_fsync) act as + barriers. We cannot unwind them out of + order, when there are earlier generation writes just because + there is a conflicting liability with an error. So, wait for + our turn till there are no conflicting liabilities. + + This situation can arise when there liabilities spread across + multiple generations. For eg., consider two writes with + following characterstics: + + 1. they belong to different generations gen1, gen2 and + (gen1 > gen2). + 2. they overlap. + 3. both are liabilities. + 4. gen1 write was attempted to sync, but the attempt failed. + 5. there was no attempt to sync gen2 write yet. + 6. A flush (as part of close) is issued and gets a gen no + gen3. + + In the above scenario, if flush is unwound without waiting + for gen1 and gen2 writes either to be successfully synced or + purged, we end up with these two writes in wb_inode->todo + list forever as there will be no attempt to process the queue + as flush is the last operation. + */ + } - if (loc->inode) - { - /* - FIXME: fd_lookup extends life of fd till the execution of - truncate_cbk - */ - iter_fd = fd_lookup (loc->inode, frame->root->pid); - if (iter_fd) { - if (!fd_ctx_get (iter_fd, this, &tmp_file)){ - file = (wb_file_t *)(long)tmp_file; - } else { - fd_unref (iter_fd); - } - } - } - - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - op_errno = ENOMEM; - goto unwind; - } + return 0; +} - local->file = file; - - frame->local = local; - if (file) { - stub = fop_truncate_stub (frame, wb_truncate_helper, loc, - offset); - if (stub == NULL) { - op_errno = ENOMEM; - goto unwind; - } +int +__wb_pick_winds(wb_inode_t *wb_inode, list_head_t *tasks, + list_head_t *liabilities) +{ + wb_request_t *req = NULL; + wb_request_t *tmp = NULL; + wb_request_t *conflict = NULL; + char req_gfid[64] = + { + 0, + }, + conflict_gfid[64] = { + 0, + }; + + list_for_each_entry_safe(req, tmp, &wb_inode->todo, todo) + { + uuid_utoa_r(req->gfid, req_gfid); + + conflict = wb_liability_has_conflict(wb_inode, req); + if (conflict) { + uuid_utoa_r(conflict->gfid, conflict_gfid); + + gf_msg_debug(wb_inode->this->name, 0, + "Not winding request due to a " + "conflicting write in liability queue. " + "REQ: unique=%" PRIu64 + ", fop=%s, " + "gen=%" PRIu64 + ", gfid=%s. " + "CONFLICT: unique=%" PRIu64 + ", fop=%s, " + "gen=%" PRIu64 + ", gfid=%s, " + "conflicts-sync-failed?=%s, " + "conflicts-error=%s", + req->unique, gf_fop_list[req->fop], req->gen, req_gfid, + conflict->unique, gf_fop_list[conflict->fop], + conflict->gen, conflict_gfid, + (conflict->op_ret == 1) ? "yes" : "no", + strerror(conflict->op_errno)); + + if (conflict->op_ret == -1) { + /* There is a conflicting liability which failed + * to sync in previous attempts, resume the req + * and fail, unless its an fsync/flush. + */ - request = wb_enqueue (file, stub); - if (request == NULL) { - op_errno = ENOMEM; - goto unwind; - } - - ret = wb_process_queue (frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - op_errno = ENOMEM; - goto unwind; - } + __wb_handle_failed_conflict(req, conflict, tasks); + } else { + /* There is a conflicting liability which was + * not attempted to sync even once. Wait till + * at least one attempt to sync is made. + */ + } + + continue; + } + + if (req->ordering.tempted && !req->ordering.go) { + /* wait some more */ + gf_msg_debug(wb_inode->this->name, 0, + "(unique=%" PRIu64 ", fop=%s, gen=%" PRIu64 + ", gfid=%s): ordering.go is not set, " + "hence not winding", + req->unique, gf_fop_list[req->fop], req->gen, + req_gfid); + continue; + } + + if (req->stub->fop == GF_FOP_WRITE) { + conflict = wb_wip_has_conflict(wb_inode, req); + + if (conflict) { + uuid_utoa_r(conflict->gfid, conflict_gfid); + + gf_msg_debug(wb_inode->this->name, 0, + "Not winding write request as " + "a conflicting write is being " + "synced to backend. " + "REQ: unique=%" PRIu64 + " fop=%s," + " gen=%" PRIu64 + ", gfid=%s. " + "CONFLICT: unique=%" PRIu64 + " " + "fop=%s, gen=%" PRIu64 + ", " + "gfid=%s", + req->unique, gf_fop_list[req->fop], req->gen, + req_gfid, conflict->unique, + gf_fop_list[conflict->fop], conflict->gen, + conflict_gfid); + continue; + } + + list_add_tail(&req->wip, &wb_inode->wip); + req->wind_count++; + + if (!req->ordering.tempted) + /* unrefed in wb_writev_cbk */ + req->stub->frame->local = __wb_request_ref(req); + } + + gf_msg_debug(wb_inode->this->name, 0, + "(unique=%" PRIu64 + ", fop=%s, gfid=%s, " + "gen=%" PRIu64 + "): picking the request for " + "winding", + req->unique, gf_fop_list[req->fop], req_gfid, req->gen); + + list_del_init(&req->todo); + + if (req->ordering.tempted) { + list_add_tail(&req->winds, liabilities); } else { - STACK_WIND (frame, - wb_truncate_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->truncate, - loc, - offset); + list_add_tail(&req->winds, tasks); } + } - return 0; + return 0; +} -unwind: - STACK_UNWIND_STRICT (truncate, frame, -1, op_errno, NULL, NULL); +void +wb_do_winds(wb_inode_t *wb_inode, list_head_t *tasks) +{ + wb_request_t *req = NULL; + wb_request_t *tmp = NULL; + + list_for_each_entry_safe(req, tmp, tasks, winds) + { + list_del_init(&req->winds); - if (stub) { - call_stub_destroy (stub); + if (req->op_ret == -1) { + call_unwind_error_keep_stub(req->stub, req->op_ret, req->op_errno); + } else { + call_resume_keep_stub(req->stub); } - return 0; + wb_request_unref(req); + } } +void +wb_process_queue(wb_inode_t *wb_inode) +{ + list_head_t tasks; + list_head_t lies; + list_head_t liabilities; + int wind_failure = 0; -int32_t -wb_ftruncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct stat *prebuf, - struct stat *postbuf) -{ - wb_local_t *local = NULL; - wb_request_t *request = NULL; - wb_file_t *file = NULL; - int32_t ret = -1; - - local = frame->local; - file = local->file; - request = local->request; - - if ((request != NULL) && (file != NULL)) { - wb_request_unref (request); - ret = wb_process_queue (frame, file, 0); - if ((ret == -1) && (errno == ENOMEM)) { - op_ret = -1; - op_errno = ENOMEM; - } + INIT_LIST_HEAD(&tasks); + INIT_LIST_HEAD(&lies); + INIT_LIST_HEAD(&liabilities); + + do { + gf_log_callingfn(wb_inode->this->name, GF_LOG_DEBUG, + "processing queues"); + + LOCK(&wb_inode->lock); + { + __wb_preprocess_winds(wb_inode); + + __wb_pick_winds(wb_inode, &tasks, &liabilities); + + __wb_pick_unwinds(wb_inode, &lies); } + UNLOCK(&wb_inode->lock); - STACK_UNWIND_STRICT (ftruncate, frame, op_ret, op_errno, prebuf, postbuf); + if (!list_empty(&lies)) + wb_do_unwinds(wb_inode, &lies); - return 0; -} + if (!list_empty(&tasks)) + wb_do_winds(wb_inode, &tasks); + /* If there is an error in wb_fulfill before winding write + * requests, we would miss invocation of wb_process_queue + * from wb_fulfill_cbk. So, retry processing again. + */ + if (!list_empty(&liabilities)) + wind_failure = wb_fulfill(wb_inode, &liabilities); + } while (wind_failure); -static int32_t -wb_ftruncate_helper (call_frame_t *frame, xlator_t *this, fd_t *fd, - off_t offset) -{ - STACK_WIND (frame, - wb_ftruncate_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->ftruncate, - fd, - offset); - return 0; + return; } - -int32_t -wb_ftruncate (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset) -{ - wb_file_t *file = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - wb_request_t *request = NULL; - int32_t ret = -1; - int op_errno = EINVAL; - - if ((!S_ISDIR (fd->inode->st_mode)) - && fd_ctx_get (fd, this, &tmp_file)) { - gf_log (this->name, GF_LOG_DEBUG, "write behind file pointer is" - " not stored in context of fd(%p), returning EBADFD", - fd); - - STACK_UNWIND_STRICT (ftruncate, frame, -1, EBADFD, - NULL, NULL); - return 0; - } +void +wb_set_inode_size(wb_inode_t *wb_inode, struct iatt *postbuf) +{ + GF_ASSERT(wb_inode); + GF_ASSERT(postbuf); + + LOCK(&wb_inode->lock); + { + wb_inode->size = postbuf->ia_size; + } + UNLOCK(&wb_inode->lock); +} - file = (wb_file_t *)(long)tmp_file; +int +wb_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, + dict_t *xdata) +{ + wb_request_t *req = NULL; + wb_inode_t *wb_inode; - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - STACK_UNWIND_STRICT (ftruncate, frame, -1, ENOMEM, - NULL, NULL); - return 0; - } + req = frame->local; + frame->local = NULL; + wb_inode = req->wb_inode; - local->file = file; + LOCK(&req->wb_inode->lock); + { + list_del_init(&req->wip); + } + UNLOCK(&req->wb_inode->lock); - frame->local = local; + wb_request_unref(req); - if (file) { - stub = fop_ftruncate_stub (frame, wb_ftruncate_helper, fd, - offset); - if (stub == NULL) { - op_errno = ENOMEM; - goto unwind; - } + /* requests could be pending while this was in progress */ + wb_process_queue(wb_inode); - request = wb_enqueue (file, stub); - if (request == NULL) { - op_errno = ENOMEM; - goto unwind; - } + STACK_UNWIND_STRICT(writev, frame, op_ret, op_errno, prebuf, postbuf, + xdata); + return 0; +} - ret = wb_process_queue (frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - op_errno = ENOMEM; - goto unwind; - } - } else { - STACK_WIND (frame, - wb_ftruncate_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->ftruncate, - fd, - offset); - } +int +wb_writev_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, + struct iovec *vector, int32_t count, off_t offset, + uint32_t flags, struct iobref *iobref, dict_t *xdata) +{ + STACK_WIND(frame, wb_writev_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->writev, fd, vector, count, offset, + flags, iobref, xdata); + return 0; +} - return 0; +int +wb_writev(call_frame_t *frame, xlator_t *this, fd_t *fd, struct iovec *vector, + int32_t count, off_t offset, uint32_t flags, struct iobref *iobref, + dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + wb_conf_t *conf = NULL; + gf_boolean_t wb_disabled = 0; + call_stub_t *stub = NULL; + int ret = -1; + int32_t op_errno = EINVAL; + int o_direct = O_DIRECT; + + conf = this->private; + + wb_inode = wb_inode_create(this, fd->inode); + if (!wb_inode) { + op_errno = ENOMEM; + goto unwind; + } + + if (!conf->strict_O_DIRECT) + o_direct = 0; + + if (fd->flags & (O_SYNC | O_DSYNC | o_direct)) + wb_disabled = 1; + + if (flags & (O_SYNC | O_DSYNC | o_direct)) + wb_disabled = 1; + + if (wb_disabled) + stub = fop_writev_stub(frame, wb_writev_helper, fd, vector, count, + offset, flags, iobref, xdata); + else + stub = fop_writev_stub(frame, NULL, fd, vector, count, offset, flags, + iobref, xdata); + if (!stub) { + op_errno = ENOMEM; + goto unwind; + } + + if (wb_disabled) + ret = wb_enqueue(wb_inode, stub); + else + ret = wb_enqueue_tempted(wb_inode, stub); + + if (!ret) { + op_errno = ENOMEM; + goto unwind; + } + + wb_process_queue(wb_inode); + + return 0; unwind: - STACK_UNWIND_STRICT (ftruncate, frame, -1, op_errno, NULL, NULL); + STACK_UNWIND_STRICT(writev, frame, -1, op_errno, NULL, NULL, NULL); - if (stub) { - call_stub_destroy (stub); - } + if (stub) + call_stub_destroy(stub); - return 0; + return 0; } +int +wb_readv_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, size_t size, + off_t offset, uint32_t flags, dict_t *xdata) +{ + STACK_WIND(frame, default_readv_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->readv, fd, size, offset, flags, xdata); + return 0; +} -int32_t -wb_setattr_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct stat *statpre, struct stat *statpost) +int +wb_readv(call_frame_t *frame, xlator_t *this, fd_t *fd, size_t size, + off_t offset, uint32_t flags, dict_t *xdata) { - wb_local_t *local = NULL; - wb_request_t *request = NULL; - call_frame_t *process_frame = NULL; - wb_file_t *file = NULL; - int32_t ret = -1; - fd_t *fd = NULL; - - local = frame->local; - file = local->file; - request = local->request; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - if (request) { - process_frame = copy_frame (frame); - if (process_frame == NULL) { - op_ret = -1; - op_errno = ENOMEM; - } - } + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; - STACK_UNWIND_STRICT (setattr, frame, op_ret, op_errno, statpre, statpost); + stub = fop_readv_stub(frame, wb_readv_helper, fd, size, offset, flags, + xdata); + if (!stub) + goto unwind; - if (request) { - wb_request_unref (request); - } + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - if (request && (process_frame != NULL)) { - ret = wb_process_queue (process_frame, file, 0); - if ((ret == -1) && (errno == ENOMEM) && (file != NULL)) { - LOCK (&file->lock); - { - file->op_ret = -1; - file->op_errno = ENOMEM; - } - UNLOCK (&file->lock); - } + wb_process_queue(wb_inode); - STACK_DESTROY (process_frame->root); - } + return 0; - if (file) { - LOCK (&file->lock); - { - fd = file->fd; - } - UNLOCK (&file->lock); +unwind: + STACK_UNWIND_STRICT(readv, frame, -1, ENOMEM, NULL, 0, NULL, NULL, NULL); - fd_unref (fd); - } + if (stub) + call_stub_destroy(stub); + return 0; - return 0; +noqueue: + STACK_WIND(frame, default_readv_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->readv, fd, size, offset, flags, xdata); + return 0; } +int +wb_flush_bg_cbk(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, dict_t *xdata) +{ + STACK_DESTROY(frame->root); + return 0; +} -static int32_t -wb_setattr_helper (call_frame_t *frame, xlator_t *this, loc_t *loc, - struct stat *stbuf, int32_t valid) +int +wb_flush_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, dict_t *xdata) { - STACK_WIND (frame, - wb_setattr_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->setattr, - loc, - stbuf, - valid); + wb_conf_t *conf = NULL; + wb_inode_t *wb_inode = NULL; + call_frame_t *bg_frame = NULL; + int32_t op_errno = 0; + int op_ret = 0; + + conf = this->private; + + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) { + op_ret = -1; + op_errno = EINVAL; + goto unwind; + } + + if (conf->flush_behind) + goto flushbehind; + + STACK_WIND(frame, default_flush_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->flush, fd, xdata); + return 0; + +flushbehind: + bg_frame = copy_frame(frame); + if (!bg_frame) { + op_ret = -1; + op_errno = ENOMEM; + goto unwind; + } + + STACK_WIND(bg_frame, wb_flush_bg_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->flush, fd, xdata); + /* fall through */ +unwind: + STACK_UNWIND_STRICT(flush, frame, op_ret, op_errno, NULL); - return 0; + return 0; } - -int32_t -wb_setattr (call_frame_t *frame, xlator_t *this, loc_t *loc, - struct stat *stbuf, int32_t valid) +int +wb_flush(call_frame_t *frame, xlator_t *this, fd_t *fd, dict_t *xdata) { - wb_file_t *file = NULL; - fd_t *iter_fd = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - wb_request_t *request = NULL; - int32_t ret = -1, op_errno = EINVAL; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - op_errno = ENOMEM; - goto unwind; - } + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; + stub = fop_flush_stub(frame, wb_flush_helper, fd, xdata); + if (!stub) + goto unwind; - frame->local = local; + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - if (!(valid & (GF_SET_ATTR_ATIME | GF_SET_ATTR_MTIME))) { - STACK_WIND (frame, - wb_setattr_cbk, - FIRST_CHILD (this), - FIRST_CHILD (this)->fops->setattr, - loc, stbuf, valid); - goto out; - } + wb_process_queue(wb_inode); - if (loc->inode) { - /* - FIXME: fd_lookup extends life of fd till the execution - of wb_utimens_cbk - */ - iter_fd = fd_lookup (loc->inode, frame->root->pid); - if (iter_fd) { - if (!fd_ctx_get (iter_fd, this, &tmp_file)) { - file = (wb_file_t *)(long)tmp_file; - } else { - fd_unref (iter_fd); - } - } + return 0; - } +unwind: + STACK_UNWIND_STRICT(flush, frame, -1, ENOMEM, NULL); - local->file = file; + if (stub) + call_stub_destroy(stub); - if (file) { - stub = fop_setattr_stub (frame, wb_setattr_helper, loc, stbuf, valid); - if (stub == NULL) { - op_errno = ENOMEM; - goto unwind; - } + return 0; - request = wb_enqueue (file, stub); - if (request == NULL) { - op_errno = ENOMEM; - goto unwind; - } +noqueue: + STACK_WIND(frame, default_flush_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->flush, fd, xdata); + return 0; +} - ret = wb_process_queue (frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - op_errno = ENOMEM; - goto unwind; - } - } else { - STACK_WIND (frame, - wb_setattr_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->setattr, - loc, - stbuf, valid); - } +int +wb_fsync_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t datasync, + dict_t *xdata) +{ + STACK_WIND(frame, default_fsync_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fsync, fd, datasync, xdata); + return 0; +} + +int +wb_fsync(call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t datasync, + dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; + int32_t op_errno = EINVAL; + + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; + + stub = fop_fsync_stub(frame, wb_fsync_helper, fd, datasync, xdata); + if (!stub) + goto unwind; + + if (!wb_enqueue(wb_inode, stub)) + goto unwind; + + wb_process_queue(wb_inode); + + return 0; - return 0; unwind: - STACK_UNWIND_STRICT (setattr, frame, -1, op_errno, - NULL, NULL); + STACK_UNWIND_STRICT(fsync, frame, -1, op_errno, NULL, NULL, NULL); - if (stub) { - call_stub_destroy (stub); - } -out: - return 0; + if (stub) + call_stub_destroy(stub); + return 0; + +noqueue: + STACK_WIND(frame, default_fsync_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fsync, fd, datasync, xdata); + return 0; } -int32_t -wb_open_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno, fd_t *fd) +int +wb_stat_helper(call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xdata) { - int32_t wbflags = 0, flags = 0; - wb_file_t *file = NULL; - wb_conf_t *conf = NULL; - wb_local_t *local = NULL; + STACK_WIND(frame, default_stat_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->stat, loc, xdata); + return 0; +} - conf = this->private; +int +wb_stat(call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - local = frame->local; - if (local == NULL) { - op_ret = -1; - op_errno = EINVAL; - goto out; - } + wb_inode = wb_inode_ctx_get(this, loc->inode); + if (!wb_inode) + goto noqueue; - flags = local->flags; - wbflags = local->wbflags; + stub = fop_stat_stub(frame, wb_stat_helper, loc, xdata); + if (!stub) + goto unwind; - if (op_ret != -1) { - file = wb_file_create (this, fd); - if (file == NULL) { - op_ret = -1; - op_errno = ENOMEM; - goto out; - } + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - /* - If mandatory locking has been enabled on this file, - we disable caching on it - */ - - if ((fd->inode->st_mode & S_ISGID) - && !(fd->inode->st_mode & S_IXGRP)) - file->disabled = 1; - - /* If O_DIRECT then, we disable chaching */ - if (((flags & O_DIRECT) == O_DIRECT) - || ((flags & O_ACCMODE) == O_RDONLY) - || (((flags & O_SYNC) == O_SYNC) - && conf->enable_O_SYNC == _gf_true)) { - file->window_conf = 0; - } + wb_process_queue(wb_inode); - if (wbflags & GF_OPEN_NOWB) { - file->disabled = 1; - } - - LOCK_INIT (&file->lock); - } - -out: - STACK_UNWIND_STRICT (open, frame, op_ret, op_errno, fd); - return 0; + return 0; + +unwind: + STACK_UNWIND_STRICT(stat, frame, -1, ENOMEM, NULL, NULL); + + if (stub) + call_stub_destroy(stub); + return 0; + +noqueue: + STACK_WIND(frame, default_stat_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->stat, loc, xdata); + return 0; } +int +wb_fstat_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, dict_t *xdata) +{ + STACK_WIND(frame, default_fstat_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fstat, fd, xdata); + return 0; +} -int32_t -wb_open (call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, - fd_t *fd, int32_t wbflags) +int +wb_fstat(call_frame_t *frame, xlator_t *this, fd_t *fd, dict_t *xdata) { - wb_local_t *local = NULL; - int32_t op_errno = EINVAL; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - op_errno = ENOMEM; - goto unwind; - } + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; - local->flags = flags; - local->wbflags = wbflags; - - frame->local = local; + stub = fop_fstat_stub(frame, wb_fstat_helper, fd, xdata); + if (!stub) + goto unwind; - STACK_WIND (frame, - wb_open_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->open, - loc, flags, fd, wbflags); - return 0; + if (!wb_enqueue(wb_inode, stub)) + goto unwind; -unwind: - STACK_UNWIND_STRICT (open, frame, -1, op_errno, NULL); - return 0; -} + wb_process_queue(wb_inode); + return 0; -int32_t -wb_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 stat *buf, struct stat *preparent, - struct stat *postparent) -{ - long flags = 0; - wb_file_t *file = NULL; - wb_conf_t *conf = this->private; - - if (op_ret != -1) { - file = wb_file_create (this, fd); - if (file == NULL) { - op_ret = -1; - op_errno = ENOMEM; - goto out; - } - /* - * If mandatory locking has been enabled on this file, - * we disable caching on it - */ - if ((fd->inode->st_mode & S_ISGID) - && !(fd->inode->st_mode & S_IXGRP)) - file->disabled = 1; - - /* If O_DIRECT then, we disable chaching */ - if (frame->local) { - flags = (long)frame->local; - if (((flags & O_DIRECT) == O_DIRECT) - || ((flags & O_ACCMODE) == O_RDONLY) - || (((flags & O_SYNC) == O_SYNC) - && (conf->enable_O_SYNC == _gf_true))) { - file->window_conf = 0; - } - } +unwind: + STACK_UNWIND_STRICT(fstat, frame, -1, ENOMEM, NULL, NULL); - LOCK_INIT (&file->lock); - } - - frame->local = NULL; + if (stub) + call_stub_destroy(stub); + return 0; -out: - STACK_UNWIND_STRICT (create, frame, op_ret, op_errno, fd, inode, buf, preparent, - postparent); - return 0; +noqueue: + STACK_WIND(frame, default_fstat_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fstat, fd, xdata); + return 0; } - int32_t -wb_create (call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, - mode_t mode, fd_t *fd) +wb_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, dict_t *xdata) { - frame->local = (void *)(long)flags; + GF_ASSERT(frame->local); - STACK_WIND (frame, - wb_create_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->create, - loc, flags, mode, fd); - return 0; + if (op_ret == 0) + wb_set_inode_size(frame->local, postbuf); + + frame->local = NULL; + + STACK_UNWIND_STRICT(truncate, frame, op_ret, op_errno, prebuf, postbuf, + xdata); + return 0; } +int +wb_truncate_helper(call_frame_t *frame, xlator_t *this, loc_t *loc, + off_t offset, dict_t *xdata) +{ + STACK_WIND(frame, wb_truncate_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->truncate, loc, offset, xdata); + return 0; +} -size_t -__wb_mark_wind_all (wb_file_t *file, list_head_t *list, list_head_t *winds) +int +wb_truncate(call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset, + dict_t *xdata) { - wb_request_t *request = NULL; - size_t size = 0; - char first_request = 1; - off_t offset_expected = 0; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - list_for_each_entry (request, list, list) - { - if ((request->stub == NULL) - || (request->stub->fop != GF_FOP_WRITE)) { - break; - } + wb_inode = wb_inode_create(this, loc->inode); + if (!wb_inode) + goto unwind; - if (!request->flags.write_request.stack_wound) { - if (first_request) { - first_request = 0; - offset_expected = request->stub->args.writev.off; - } - - if (request->stub->args.writev.off != offset_expected) { - break; - } - - size += request->write_size; - offset_expected += request->write_size; - file->aggregate_current -= request->write_size; - - request->flags.write_request.stack_wound = 1; - list_add_tail (&request->winds, winds); - } - } - - return size; -} + frame->local = wb_inode; + stub = fop_truncate_stub(frame, wb_truncate_helper, loc, offset, xdata); + if (!stub) + goto unwind; -void -__wb_can_wind (list_head_t *list, char *other_fop_in_queue, - char *non_contiguous_writes, char *incomplete_writes) -{ - wb_request_t *request = NULL; - char first_request = 1; - off_t offset_expected = 0; + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - list_for_each_entry (request, list, list) - { - if ((request->stub == NULL) - || (request->stub->fop != GF_FOP_WRITE)) { - if (request->stub && other_fop_in_queue) { - *other_fop_in_queue = 1; - } - break; - } + wb_process_queue(wb_inode); - if (request->flags.write_request.stack_wound - && !request->flags.write_request.got_reply - && (incomplete_writes != NULL)) { - *incomplete_writes = 1; - break; - } + return 0; - if (!request->flags.write_request.stack_wound) { - if (first_request) { - first_request = 0; - offset_expected = request->stub->args.writev.off; - } +unwind: + STACK_UNWIND_STRICT(truncate, frame, -1, ENOMEM, NULL, NULL, NULL); - if (offset_expected != request->stub->args.writev.off) { - if (non_contiguous_writes) { - *non_contiguous_writes = 1; - } - break; - } + if (stub) + call_stub_destroy(stub); - offset_expected += request->write_size; - } - } + return 0; +} - return; +int32_t +wb_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, dict_t *xdata) +{ + GF_ASSERT(frame->local); + + if (op_ret == 0) + wb_set_inode_size(frame->local, postbuf); + + frame->local = NULL; + + STACK_UNWIND_STRICT(ftruncate, frame, op_ret, op_errno, prebuf, postbuf, + xdata); + return 0; } +int +wb_ftruncate_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + dict_t *xdata) +{ + STACK_WIND(frame, wb_ftruncate_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->ftruncate, fd, offset, xdata); + return 0; +} -ssize_t -__wb_mark_winds (list_head_t *list, list_head_t *winds, size_t aggregate_conf, - char wind_all, char enable_trickling_writes) +int +wb_ftruncate(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + dict_t *xdata) { - size_t size = 0; - char other_fop_in_queue = 0; - char incomplete_writes = 0; - char non_contiguous_writes = 0; - wb_request_t *request = NULL; - wb_file_t *file = NULL; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; + int32_t op_errno = 0; - if (list_empty (list)) { - goto out; - } + wb_inode = wb_inode_create(this, fd->inode); + if (!wb_inode) { + op_errno = ENOMEM; + goto unwind; + } - request = list_entry (list->next, typeof (*request), list); - file = request->file; + frame->local = wb_inode; - if (!wind_all && (file->aggregate_current < aggregate_conf)) { - __wb_can_wind (list, &other_fop_in_queue, - &non_contiguous_writes, &incomplete_writes); - } + stub = fop_ftruncate_stub(frame, wb_ftruncate_helper, fd, offset, xdata); + if (!stub) { + op_errno = ENOMEM; + goto unwind; + } - if ((enable_trickling_writes && !incomplete_writes) - || (wind_all) || (non_contiguous_writes) - || (other_fop_in_queue) - || (file->aggregate_current >= aggregate_conf)) { - size = __wb_mark_wind_all (file, list, winds); - } + if (!wb_enqueue(wb_inode, stub)) { + op_errno = ENOMEM; + goto unwind; + } -out: - return size; + wb_process_queue(wb_inode); + + return 0; + +unwind: + frame->local = NULL; + + STACK_UNWIND_STRICT(ftruncate, frame, -1, op_errno, NULL, NULL, NULL); + + if (stub) + call_stub_destroy(stub); + return 0; } +int +wb_setattr_helper(call_frame_t *frame, xlator_t *this, loc_t *loc, + struct iatt *stbuf, int32_t valid, dict_t *xdata) +{ + STACK_WIND(frame, default_setattr_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->setattr, loc, stbuf, valid, xdata); + return 0; +} -size_t -__wb_mark_unwind_till (list_head_t *list, list_head_t *unwinds, size_t size) +int +wb_setattr(call_frame_t *frame, xlator_t *this, loc_t *loc, struct iatt *stbuf, + int32_t valid, dict_t *xdata) { - size_t written_behind = 0; - wb_request_t *request = NULL; - wb_file_t *file = NULL; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - if (list_empty (list)) { - goto out; - } + wb_inode = wb_inode_ctx_get(this, loc->inode); + if (!wb_inode) + goto noqueue; - request = list_entry (list->next, typeof (*request), list); - file = request->file; + stub = fop_setattr_stub(frame, wb_setattr_helper, loc, stbuf, valid, xdata); + if (!stub) + goto unwind; - list_for_each_entry (request, list, list) - { - if ((request->stub == NULL) - || (request->stub->fop != GF_FOP_WRITE)) { - continue; - } + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - if (written_behind <= size) { - if (!request->flags.write_request.write_behind) { - written_behind += request->write_size; - request->flags.write_request.write_behind = 1; - list_add_tail (&request->unwinds, unwinds); - - if (!request->flags.write_request.got_reply) { - file->window_current += request->write_size; - } - } - } else { - break; - } - } + wb_process_queue(wb_inode); -out: - return written_behind; + return 0; +unwind: + STACK_UNWIND_STRICT(setattr, frame, -1, ENOMEM, NULL, NULL, NULL); + + if (stub) + call_stub_destroy(stub); + return 0; + +noqueue: + STACK_WIND(frame, default_setattr_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->setattr, loc, stbuf, valid, xdata); + return 0; } +int +wb_fsetattr_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, + struct iatt *stbuf, int32_t valid, dict_t *xdata) +{ + STACK_WIND(frame, default_fsetattr_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fsetattr, fd, stbuf, valid, xdata); + return 0; +} -void -__wb_mark_unwinds (list_head_t *list, list_head_t *unwinds) +int +wb_fsetattr(call_frame_t *frame, xlator_t *this, fd_t *fd, struct iatt *stbuf, + int32_t valid, dict_t *xdata) { - wb_request_t *request = NULL; - wb_file_t *file = NULL; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - if (list_empty (list)) { - goto out; - } + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; - request = list_entry (list->next, typeof (*request), list); - file = request->file; + stub = fop_fsetattr_stub(frame, wb_fsetattr_helper, fd, stbuf, valid, + xdata); + if (!stub) + goto unwind; - if (file->window_current <= file->window_conf) { - __wb_mark_unwind_till (list, unwinds, - file->window_conf - file->window_current); - } + if (!wb_enqueue(wb_inode, stub)) + goto unwind; -out: - return; -} + wb_process_queue(wb_inode); + return 0; +unwind: + STACK_UNWIND_STRICT(fsetattr, frame, -1, ENOMEM, NULL, NULL, NULL); -uint32_t -__wb_get_other_requests (list_head_t *list, list_head_t *other_requests) -{ - wb_request_t *request = NULL; - uint32_t count = 0; - list_for_each_entry (request, list, list) { - if ((request->stub == NULL) - || (request->stub->fop == GF_FOP_WRITE)) { - break; - } - - if (!request->flags.other_requests.marked_for_resume) { - request->flags.other_requests.marked_for_resume = 1; - list_add_tail (&request->other_requests, - other_requests); - count++; - } - } + if (stub) + call_stub_destroy(stub); + return 0; - return count; +noqueue: + STACK_WIND(frame, default_fsetattr_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fsetattr, fd, stbuf, valid, xdata); + return 0; } - int32_t -wb_stack_unwind (list_head_t *unwinds) +wb_create(call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, + mode_t mode, mode_t umask, fd_t *fd, dict_t *xdata) { - struct stat buf = {0,}; - wb_request_t *request = NULL, *dummy = NULL; - call_frame_t *frame = NULL; - wb_local_t *local = NULL; + wb_inode_t *wb_inode = NULL; - list_for_each_entry_safe (request, dummy, unwinds, unwinds) - { - frame = request->stub->frame; - local = frame->local; + wb_inode = wb_inode_create(this, fd->inode); + if (!wb_inode) + goto unwind; - STACK_UNWIND (frame, local->op_ret, local->op_errno, &buf, - &buf); + if (((flags & O_RDWR) || (flags & O_WRONLY)) && (flags & O_TRUNC)) + wb_inode->size = 0; - wb_request_unref (request); - } + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->create, + loc, flags, mode, umask, fd, xdata); + return 0; - return 0; +unwind: + STACK_UNWIND_STRICT(create, frame, -1, ENOMEM, NULL, NULL, NULL, NULL, NULL, + NULL); + return 0; } - int32_t -wb_resume_other_requests (call_frame_t *frame, wb_file_t *file, - list_head_t *other_requests) +wb_open(call_frame_t *frame, xlator_t *this, loc_t *loc, int32_t flags, + fd_t *fd, dict_t *xdata) { - int32_t ret = 0; - wb_request_t *request = NULL, *dummy = NULL; - int32_t fops_removed = 0; - char wind = 0; - call_stub_t *stub = NULL; + wb_inode_t *wb_inode = NULL; - if (list_empty (other_requests)) { - goto out; - } + wb_inode = wb_inode_create(this, fd->inode); + if (!wb_inode) + goto unwind; - list_for_each_entry_safe (request, dummy, other_requests, - other_requests) { - wind = request->stub->wind; - stub = request->stub; - - LOCK (&file->lock); - { - request->stub = NULL; - } - UNLOCK (&file->lock); - - if (!wind) { - wb_request_unref (request); - fops_removed++; - } - - call_resume (stub); - } + if (((flags & O_RDWR) || (flags & O_WRONLY)) && (flags & O_TRUNC)) + wb_inode->size = 0; - if (fops_removed > 0) { - ret = wb_process_queue (frame, file, 0); - } - -out: - return ret; + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->open, + loc, flags, fd, xdata); + return 0; + +unwind: + STACK_UNWIND_STRICT(open, frame, -1, ENOMEM, NULL, NULL); + return 0; +} + +int32_t +wb_lookup_cbk(call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, + int32_t op_errno, inode_t *inode, struct iatt *buf, dict_t *xdata, + struct iatt *postparent) +{ + if (op_ret == 0) { + wb_inode_t *wb_inode = wb_inode_ctx_get(this, inode); + if (wb_inode) + wb_set_inode_size(wb_inode, buf); + } + + STACK_UNWIND_STRICT(lookup, frame, op_ret, op_errno, inode, buf, xdata, + postparent); + return 0; } +int +wb_lookup_helper(call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xdata) +{ + STACK_WIND(frame, wb_lookup_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->lookup, loc, xdata); + return 0; +} int32_t -wb_do_ops (call_frame_t *frame, wb_file_t *file, list_head_t *winds, - list_head_t *unwinds, list_head_t *other_requests) +wb_lookup(call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xdata) { - int32_t ret = -1; + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - ret = wb_stack_unwind (unwinds); - if (ret == -1) { - goto out; - } + wb_inode = wb_inode_ctx_get(this, loc->inode); + if (!wb_inode) + goto noqueue; - ret = wb_sync (frame, file, winds); - if (ret == -1) { - goto out; - } + stub = fop_lookup_stub(frame, wb_lookup_helper, loc, xdata); + if (!stub) + goto unwind; - ret = wb_resume_other_requests (frame, file, other_requests); + if (!wb_enqueue(wb_inode, stub)) + goto unwind; -out: - return ret; -} + wb_process_queue(wb_inode); + return 0; -inline int -__wb_copy_into_holder (wb_request_t *holder, wb_request_t *request) -{ - char *ptr = NULL; - struct iobuf *iobuf = NULL; - struct iobref *iobref = NULL; - int ret = -1; +unwind: + if (stub) + call_stub_destroy(stub); - if (holder->flags.write_request.virgin) { - iobuf = iobuf_get (request->file->this->ctx->iobuf_pool); - if (iobuf == NULL) { - gf_log (request->file->this->name, GF_LOG_ERROR, - "out of memory"); - goto out; - } + STACK_UNWIND_STRICT(lookup, frame, -1, ENOMEM, NULL, NULL, NULL, NULL); + return 0; - iobref = iobref_new (); - if (iobref == NULL) { - iobuf_unref (iobuf); - gf_log (request->file->this->name, GF_LOG_ERROR, - "out of memory"); - goto out; - } - - ret = iobref_add (iobref, iobuf); - if (ret != 0) { - iobuf_unref (iobuf); - iobref_unref (iobref); - gf_log (request->file->this->name, GF_LOG_DEBUG, - "cannot add iobuf (%p) into iobref (%p)", - iobuf, iobref); - goto out; - } - - iov_unload (iobuf->ptr, holder->stub->args.writev.vector, - holder->stub->args.writev.count); - holder->stub->args.writev.vector[0].iov_base = iobuf->ptr; - - iobref_unref (holder->stub->args.writev.iobref); - holder->stub->args.writev.iobref = iobref; - - iobuf_unref (iobuf); - - holder->flags.write_request.virgin = 0; - } +noqueue: + STACK_WIND(frame, wb_lookup_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->lookup, loc, xdata); + return 0; +} - ptr = holder->stub->args.writev.vector[0].iov_base + holder->write_size; +static void +wb_mark_readdirp_start(xlator_t *this, inode_t *directory) +{ + wb_inode_t *wb_directory_inode = NULL; - iov_unload (ptr, - request->stub->args.writev.vector, - request->stub->args.writev.count); + wb_directory_inode = wb_inode_create(this, directory); - holder->stub->args.writev.vector[0].iov_len += request->write_size; - holder->write_size += request->write_size; + if (!wb_directory_inode) + return; - request->flags.write_request.stack_wound = 1; - list_move_tail (&request->list, &request->file->passive_requests); + LOCK(&wb_directory_inode->lock); + { + GF_ATOMIC_INC(wb_directory_inode->readdirps); + } + UNLOCK(&wb_directory_inode->lock); - ret = 0; -out: - return ret; + return; } +static void +wb_mark_readdirp_end(xlator_t *this, inode_t *directory) +{ + wb_inode_t *wb_directory_inode = NULL, *wb_inode = NULL, *tmp = NULL; + int readdirps = 0; -/* this procedure assumes that write requests have only one vector to write */ -void -__wb_collapse_write_bufs (list_head_t *requests, size_t page_size) -{ - off_t offset_expected = 0; - size_t space_left = 0; - wb_request_t *request = NULL, *tmp = NULL, *holder = NULL; - int ret = 0; - - list_for_each_entry_safe (request, tmp, requests, list) { - if ((request->stub == NULL) - || (request->stub->fop != GF_FOP_WRITE) - || (request->flags.write_request.stack_wound)) { - holder = NULL; - continue; - } + wb_directory_inode = wb_inode_ctx_get(this, directory); - if (request->flags.write_request.write_behind) { - if (holder == NULL) { - holder = request; - continue; - } - - offset_expected = holder->stub->args.writev.off - + holder->write_size; - - if (request->stub->args.writev.off != offset_expected) { - holder = request; - continue; - } - - space_left = page_size - holder->write_size; - - if (space_left >= request->write_size) { - ret = __wb_copy_into_holder (holder, request); - if (ret != 0) { - break; - } - - __wb_request_unref (request); - } else { - holder = request; - } - } else { - break; - } + if (!wb_directory_inode) + return; + + LOCK(&wb_directory_inode->lock); + { + readdirps = GF_ATOMIC_DEC(wb_directory_inode->readdirps); + if (readdirps) + goto unlock; + + list_for_each_entry_safe(wb_inode, tmp, + &wb_directory_inode->invalidate_list, + invalidate_list) + { + list_del_init(&wb_inode->invalidate_list); + GF_ATOMIC_INIT(wb_inode->invalidate, 0); + inode_unref(wb_inode->inode); } + } +unlock: + UNLOCK(&wb_directory_inode->lock); - return; + return; } - -int32_t -wb_process_queue (call_frame_t *frame, wb_file_t *file, char flush_all) +int32_t +wb_readdirp_cbk(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, gf_dirent_t *entries, + dict_t *xdata) { - list_head_t winds, unwinds, other_requests; - size_t size = 0; - wb_conf_t *conf = NULL; - uint32_t count = 0; - int32_t ret = -1; + wb_inode_t *wb_inode = NULL; + gf_dirent_t *entry = NULL; + inode_t *inode = NULL; + fd_t *fd = NULL; - INIT_LIST_HEAD (&winds); - INIT_LIST_HEAD (&unwinds); - INIT_LIST_HEAD (&other_requests); - - if (file == NULL) { - errno = EINVAL; - goto out; - } + fd = frame->local; + frame->local = NULL; - conf = file->this->private; - size = conf->aggregate_size; - LOCK (&file->lock); - { - /* - * make sure requests are marked for unwinding and adjacent - * continguous write buffers (each of size less than that of - * an iobuf) are packed properly so that iobufs are filled to - * their maximum capacity, before calling __wb_mark_winds. - */ - __wb_mark_unwinds (&file->request, &unwinds); + if (op_ret <= 0) + goto unwind; - __wb_collapse_write_bufs (&file->request, - file->this->ctx->page_size); + list_for_each_entry(entry, &entries->list, list) + { + if (!entry->inode || !IA_ISREG(entry->d_stat.ia_type)) + continue; - count = __wb_get_other_requests (&file->request, - &other_requests); + wb_inode = wb_inode_ctx_get(this, entry->inode); + if (!wb_inode) + continue; - if (count == 0) { - __wb_mark_winds (&file->request, &winds, size, - flush_all, - conf->enable_trickling_writes); - } + LOCK(&wb_inode->lock); + { + if (!list_empty(&wb_inode->liability) || + GF_ATOMIC_GET(wb_inode->invalidate)) { + inode = entry->inode; + entry->inode = NULL; + memset(&entry->d_stat, 0, sizeof(entry->d_stat)); + } } - UNLOCK (&file->lock); + UNLOCK(&wb_inode->lock); - ret = wb_do_ops (frame, file, &winds, &unwinds, &other_requests); + if (inode) { + inode_unref(inode); + inode = NULL; + } + } -out: - return ret; -} +unwind: + wb_mark_readdirp_end(this, fd->inode); + frame->local = NULL; + STACK_UNWIND_STRICT(readdirp, frame, op_ret, op_errno, entries, xdata); + return 0; +} int32_t -wb_writev_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, struct stat *prebuf, - struct stat *postbuf) +wb_readdirp(call_frame_t *frame, xlator_t *this, fd_t *fd, size_t size, + off_t off, dict_t *xdata) { - STACK_UNWIND_STRICT (writev, frame, op_ret, op_errno, prebuf, postbuf); - return 0; + wb_mark_readdirp_start(this, fd->inode); + + frame->local = fd; + + STACK_WIND(frame, wb_readdirp_cbk, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->readdirp, fd, size, off, xdata); + + return 0; } +int32_t +wb_link_helper(call_frame_t *frame, xlator_t *this, loc_t *oldloc, + loc_t *newloc, dict_t *xdata) +{ + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->link, + oldloc, newloc, xdata); + return 0; +} int32_t -wb_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, struct iovec *vector, - int32_t count, off_t offset, struct iobref *iobref) -{ - wb_file_t *file = NULL; - char wb_disabled = 0; - call_frame_t *process_frame = NULL; - size_t size = 0; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - wb_local_t *local = NULL; - wb_request_t *request = NULL; - int32_t ret = -1; - int32_t op_ret = -1, op_errno = EINVAL; - - if (vector != NULL) - size = iov_length (vector, count); - - if ((!S_ISDIR (fd->inode->st_mode)) - && fd_ctx_get (fd, this, &tmp_file)) { - gf_log (this->name, GF_LOG_DEBUG, "write behind file pointer is" - " not stored in context of fd(%p), returning EBADFD", - fd); - - op_errno = EBADFD; - goto unwind; - } +wb_link(call_frame_t *frame, xlator_t *this, loc_t *oldloc, loc_t *newloc, + dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - file = (wb_file_t *)(long)tmp_file; - if ((!S_ISDIR (fd->inode->st_mode)) && (file == NULL)) { - gf_log (this->name, GF_LOG_DEBUG, - "wb_file not found for fd %p", fd); - op_errno = EBADFD; - goto unwind; - } + wb_inode = wb_inode_ctx_get(this, oldloc->inode); + if (!wb_inode) + goto noqueue; - if (file != NULL) { - LOCK (&file->lock); - { - op_ret = file->op_ret; - op_errno = file->op_errno; - - file->op_ret = 0; - - if ((op_ret == 0) - && (file->disabled || file->disable_till)) { - if (size > file->disable_till) { - file->disable_till = 0; - } else { - file->disable_till -= size; - } - wb_disabled = 1; - } - } - UNLOCK (&file->lock); - } else { - wb_disabled = 1; - } + stub = fop_link_stub(frame, wb_link_helper, oldloc, newloc, xdata); + if (!stub) + goto unwind; - if (op_ret == -1) { - STACK_UNWIND_STRICT (writev, frame, op_ret, op_errno, - NULL, NULL); - return 0; - } + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - if (wb_disabled) { - STACK_WIND (frame, wb_writev_cbk, - FIRST_CHILD (frame->this), - FIRST_CHILD (frame->this)->fops->writev, - fd, vector, count, offset, iobref); - return 0; - } + wb_process_queue(wb_inode); - process_frame = copy_frame (frame); - if (process_frame == NULL) { - op_errno = ENOMEM; - goto unwind; - } + return 0; - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - op_errno = ENOMEM; - goto unwind; - return 0; - } +unwind: + STACK_UNWIND_STRICT(link, frame, -1, ENOMEM, NULL, NULL, NULL, NULL, NULL); - frame->local = local; - local->file = file; + if (stub) + call_stub_destroy(stub); - stub = fop_writev_stub (frame, NULL, fd, vector, count, offset, - iobref); - if (stub == NULL) { - op_errno = ENOMEM; - goto unwind; - } + return 0; - request = wb_enqueue (file, stub); - if (request == NULL) { - op_errno = ENOMEM; - goto unwind; - } - - ret = wb_process_queue (process_frame, file, 0); - if ((ret == -1) && (errno == ENOMEM)) { - op_errno = ENOMEM; - goto unwind; - } +noqueue: + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->link, + oldloc, newloc, xdata); + return 0; +} - STACK_DESTROY (process_frame->root); +int32_t +wb_fallocate_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, + int32_t keep_size, off_t offset, size_t len, dict_t *xdata) +{ + STACK_WIND_TAIL(frame, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fallocate, fd, keep_size, offset, + len, xdata); + return 0; +} - return 0; +int32_t +wb_fallocate(call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t keep_size, + off_t offset, size_t len, dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; -unwind: - STACK_UNWIND_STRICT (writev, frame, -1, op_errno, NULL, NULL); + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; - if (process_frame) { - STACK_DESTROY (process_frame->root); - } + stub = fop_fallocate_stub(frame, wb_fallocate_helper, fd, keep_size, offset, + len, xdata); + if (!stub) + goto unwind; - if (stub) { - call_stub_destroy (stub); - } + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - return 0; -} + wb_process_queue(wb_inode); + return 0; -int32_t -wb_readv_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, struct iobref *iobref) -{ - wb_local_t *local = NULL; - wb_file_t *file = NULL; - wb_request_t *request = NULL; - int32_t ret = 0; - - local = frame->local; - file = local->file; - request = local->request; - - if ((request != NULL) && (file != NULL)) { - wb_request_unref (request); - - ret = wb_process_queue (frame, file, 0); - if ((ret == -1) && (errno == ENOMEM)) { - op_ret = -1; - op_errno = ENOMEM; - } - } +unwind: + STACK_UNWIND_STRICT(fallocate, frame, -1, ENOMEM, NULL, NULL, NULL); - STACK_UNWIND_STRICT (readv, frame, op_ret, op_errno, vector, count, stbuf, iobref); + if (stub) + call_stub_destroy(stub); - return 0; -} + return 0; +noqueue: + STACK_WIND_TAIL(frame, FIRST_CHILD(this), + FIRST_CHILD(this)->fops->fallocate, fd, keep_size, offset, + len, xdata); + return 0; +} -static int32_t -wb_readv_helper (call_frame_t *frame, xlator_t *this, fd_t *fd, size_t size, - off_t offset) +int32_t +wb_discard_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + size_t len, dict_t *xdata) { - STACK_WIND (frame, - wb_readv_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->readv, - fd, size, offset); - - return 0; + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->discard, + fd, offset, len, xdata); + return 0; } - int32_t -wb_readv (call_frame_t *frame, xlator_t *this, fd_t *fd, size_t size, - off_t offset) -{ - wb_file_t *file = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - int32_t ret = -1; - wb_request_t *request = NULL; - - if ((!S_ISDIR (fd->inode->st_mode)) - && fd_ctx_get (fd, this, &tmp_file)) { - gf_log (this->name, GF_LOG_DEBUG, "write behind file pointer is" - " not stored in context of fd(%p), returning EBADFD", - fd); - - STACK_UNWIND_STRICT (readv, frame, -1, EBADFD, - NULL, 0, NULL, NULL); - return 0; - } +wb_discard(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + size_t len, dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - file = (wb_file_t *)(long)tmp_file; + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - STACK_UNWIND_STRICT (readv, frame, -1, ENOMEM, - NULL, 0, NULL, NULL); - return 0; - } + stub = fop_discard_stub(frame, wb_discard_helper, fd, offset, len, xdata); + if (!stub) + goto unwind; - local->file = file; + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - frame->local = local; - if (file) { - stub = fop_readv_stub (frame, wb_readv_helper, fd, size, - offset); - if (stub == NULL) { - STACK_UNWIND_STRICT (readv, frame, -1, ENOMEM, - NULL, 0, NULL, NULL); - return 0; - } + wb_process_queue(wb_inode); - request = wb_enqueue (file, stub); - if (request == NULL) { - STACK_UNWIND_STRICT (readv, frame, -1, ENOMEM, - NULL, 0, NULL, NULL); - call_stub_destroy (stub); - return 0; - } + return 0; - ret = wb_process_queue (frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - STACK_UNWIND_STRICT (readv, frame, -1, ENOMEM, - NULL, 0, NULL, NULL); - call_stub_destroy (stub); - return 0; - } +unwind: + STACK_UNWIND_STRICT(discard, frame, -1, ENOMEM, NULL, NULL, NULL); - } else { - STACK_WIND (frame, - wb_readv_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->readv, - fd, size, offset); - } + if (stub) + call_stub_destroy(stub); + return 0; - return 0; -} +noqueue: + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->discard, + fd, offset, len, xdata); + return 0; +} int32_t -wb_ffr_bg_cbk (call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno) +wb_zerofill_helper(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + off_t len, dict_t *xdata) { - STACK_DESTROY (frame->root); - return 0; + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->zerofill, + fd, offset, len, xdata); + return 0; } - int32_t -wb_ffr_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno) -{ - wb_local_t *local = NULL; - wb_file_t *file = NULL; - wb_conf_t *conf = NULL; - char unwind = 0; - int32_t ret = -1; - int disabled = 0; - int64_t disable_till = 0; - - conf = this->private; - local = frame->local; - - if ((local != NULL) && (local->file != NULL)) { - file = local->file; - - LOCK (&file->lock); - { - disabled = file->disabled; - disable_till = file->disable_till; - } - UNLOCK (&file->lock); +wb_zerofill(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + off_t len, dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - if (conf->flush_behind - && (!disabled) && (disable_till == 0)) { - unwind = 1; - } else { - local->reply_count++; - /* - * without flush-behind, unwind should wait for replies - * of writes queued before and the flush - */ - if (local->reply_count == 2) { - unwind = 1; - } - } - } else { - unwind = 1; - } + wb_inode = wb_inode_ctx_get(this, fd->inode); + if (!wb_inode) + goto noqueue; - if (unwind) { - if (file != NULL) { - LOCK (&file->lock); - { - if (file->op_ret == -1) { - op_ret = file->op_ret; - op_errno = file->op_errno; - - file->op_ret = 0; - } - } - UNLOCK (&file->lock); - - ret = wb_process_queue (frame, file, 0); - if ((ret == -1) && (errno == ENOMEM)) { - op_ret = -1; - op_errno = ENOMEM; - } - } - - STACK_UNWIND_STRICT (flush, frame, op_ret, op_errno); - } + stub = fop_zerofill_stub(frame, wb_zerofill_helper, fd, offset, len, xdata); + if (!stub) + goto unwind; - return 0; -} + if (!wb_enqueue(wb_inode, stub)) + goto unwind; + + wb_process_queue(wb_inode); + + return 0; + +unwind: + STACK_UNWIND_STRICT(zerofill, frame, -1, ENOMEM, NULL, NULL, NULL); + + if (stub) + call_stub_destroy(stub); +noqueue: + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->zerofill, + fd, offset, len, xdata); + return 0; +} int32_t -wb_flush (call_frame_t *frame, xlator_t *this, fd_t *fd) -{ - wb_conf_t *conf = NULL; - wb_file_t *file = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - call_frame_t *process_frame = NULL; - wb_local_t *tmp_local = NULL; - wb_request_t *request = NULL; - int32_t ret = 0; - int disabled = 0; - int64_t disable_till = 0; - - conf = this->private; - - if ((!S_ISDIR (fd->inode->st_mode)) - && fd_ctx_get (fd, this, &tmp_file)) { - gf_log (this->name, GF_LOG_DEBUG, "write behind file pointer is" - " not stored in context of fd(%p), returning EBADFD", - fd); - - STACK_UNWIND_STRICT (flush, frame, -1, EBADFD); - return 0; - } +wb_rename(call_frame_t *frame, xlator_t *this, loc_t *oldloc, loc_t *newloc, + dict_t *xdata) +{ + wb_inode_t *wb_inode = NULL; + call_stub_t *stub = NULL; - file = (wb_file_t *)(long)tmp_file; + wb_inode = wb_inode_ctx_get(this, oldloc->inode); + if (!wb_inode) + goto noqueue; - if (file != NULL) { - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - STACK_UNWIND (frame, -1, ENOMEM, NULL); - return 0; - } + stub = fop_rename_stub(frame, default_rename_resume, oldloc, newloc, xdata); + if (!stub) + goto unwind; - local->file = file; + if (!wb_enqueue(wb_inode, stub)) + goto unwind; - frame->local = local; - stub = fop_flush_cbk_stub (frame, wb_ffr_cbk, 0, 0); - if (stub == NULL) { - STACK_UNWIND_STRICT (flush, frame, -1, ENOMEM); - return 0; - } + wb_process_queue(wb_inode); - process_frame = copy_frame (frame); - if (process_frame == NULL) { - STACK_UNWIND_STRICT (flush, frame, -1, ENOMEM); - call_stub_destroy (stub); - return 0; - } + return 0; - LOCK (&file->lock); - { - disabled = file->disabled; - disable_till = file->disable_till; - } - UNLOCK (&file->lock); - - if (conf->flush_behind - && (!disabled) && (disable_till == 0)) { - tmp_local = CALLOC (1, sizeof (*local)); - if (tmp_local == NULL) { - STACK_UNWIND_STRICT (flush, frame, -1, ENOMEM); - - STACK_DESTROY (process_frame->root); - call_stub_destroy (stub); - return 0; - } - tmp_local->file = file; - - process_frame->local = tmp_local; - } +unwind: + if (stub) + call_stub_destroy(stub); - fd_ref (fd); + STACK_UNWIND_STRICT(rename, frame, -1, ENOMEM, NULL, NULL, NULL, NULL, NULL, + NULL); - request = wb_enqueue (file, stub); - if (request == NULL) { - STACK_UNWIND_STRICT (flush, frame, -1, ENOMEM); + return 0; - fd_unref (fd); - call_stub_destroy (stub); - STACK_DESTROY (process_frame->root); - return 0; - } +noqueue: + STACK_WIND_TAIL(frame, FIRST_CHILD(this), FIRST_CHILD(this)->fops->rename, + oldloc, newloc, xdata); + return 0; +} - ret = wb_process_queue (process_frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - STACK_UNWIND_STRICT (flush, frame, -1, ENOMEM); +int +wb_forget(xlator_t *this, inode_t *inode) +{ + uint64_t tmp = 0; + wb_inode_t *wb_inode = NULL; - fd_unref (fd); - call_stub_destroy (stub); - STACK_DESTROY (process_frame->root); - return 0; - } - } - - if ((file != NULL) && conf->flush_behind - && (!disabled) && (disable_till == 0)) { - STACK_WIND (process_frame, - wb_ffr_bg_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->flush, - fd); - } else { - STACK_WIND (frame, - wb_ffr_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->flush, - fd); - - if (process_frame != NULL) { - STACK_DESTROY (process_frame->root); - } - } + inode_ctx_del(inode, this, &tmp); - - if (file != NULL) { - fd_unref (fd); - } + wb_inode = (wb_inode_t *)(long)tmp; + if (!wb_inode) return 0; + + wb_inode_destroy(wb_inode); + + return 0; } +int +wb_release(xlator_t *this, fd_t *fd) +{ + uint64_t tmp = 0; + + (void)fd_ctx_del(fd, this, &tmp); -static int32_t -wb_fsync_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, - int32_t op_errno, struct stat *prebuf, struct stat *postbuf) + return 0; +} + +int +wb_priv_dump(xlator_t *this) { - wb_local_t *local = NULL; - wb_file_t *file = NULL; - wb_request_t *request = NULL; - int32_t ret = -1; + wb_conf_t *conf = NULL; + char key_prefix[GF_DUMP_MAX_BUF_LEN] = { + 0, + }; + int ret = -1; - local = frame->local; - file = local->file; - request = local->request; + GF_VALIDATE_OR_GOTO("write-behind", this, out); - if (file != NULL) { - LOCK (&file->lock); - { - if (file->op_ret == -1) { - op_ret = file->op_ret; - op_errno = file->op_errno; + conf = this->private; + GF_VALIDATE_OR_GOTO(this->name, conf, out); - file->op_ret = 0; - } - } - UNLOCK (&file->lock); - - if (request) { - wb_request_unref (request); - ret = wb_process_queue (frame, file, 0); - if ((ret == -1) && (errno == ENOMEM)) { - op_ret = -1; - op_errno = ENOMEM; - } - } + gf_proc_dump_build_key(key_prefix, "xlator.performance.write-behind", + "priv"); - } + gf_proc_dump_add_section("%s", key_prefix); - STACK_UNWIND_STRICT (fsync, frame, op_ret, op_errno, prebuf, postbuf); - - return 0; -} + gf_proc_dump_write("aggregate_size", "%" PRIu64, conf->aggregate_size); + gf_proc_dump_write("window_size", "%" PRIu64, conf->window_size); + gf_proc_dump_write("flush_behind", "%d", conf->flush_behind); + gf_proc_dump_write("trickling_writes", "%d", conf->trickling_writes); + ret = 0; +out: + return ret; +} -static int32_t -wb_fsync_helper (call_frame_t *frame, xlator_t *this, fd_t *fd, - int32_t datasync) +void +__wb_dump_requests(struct list_head *head, char *prefix) { - STACK_WIND (frame, - wb_fsync_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->fsync, - fd, datasync); - return 0; -} + char key[GF_DUMP_MAX_BUF_LEN] = { + 0, + }; + char key_prefix[GF_DUMP_MAX_BUF_LEN] = + { + 0, + }, + flag = 0; + wb_request_t *req = NULL; + list_for_each_entry(req, head, all) + { + gf_proc_dump_build_key(key_prefix, key, "%s", + (char *)gf_fop_list[req->fop]); -int32_t -wb_fsync (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t datasync) -{ - wb_file_t *file = NULL; - wb_local_t *local = NULL; - uint64_t tmp_file = 0; - call_stub_t *stub = NULL; - wb_request_t *request = NULL; - int32_t ret = -1; - - if ((!S_ISDIR (fd->inode->st_mode)) - && fd_ctx_get (fd, this, &tmp_file)) { - gf_log (this->name, GF_LOG_DEBUG, "write behind file pointer is" - " not stored in context of fd(%p), returning EBADFD", - fd); - - STACK_UNWIND_STRICT (fsync, frame, -1, EBADFD, NULL, NULL); - return 0; - } + gf_proc_dump_add_section("%s", key_prefix); - file = (wb_file_t *)(long)tmp_file; + gf_proc_dump_write("unique", "%" PRIu64, req->unique); - local = CALLOC (1, sizeof (*local)); - if (local == NULL) { - STACK_UNWIND_STRICT (fsync, frame, -1, ENOMEM, NULL, NULL); - return 0; - } + gf_proc_dump_write("refcount", "%d", req->refcount); - local->file = file; + if (list_empty(&req->todo)) + gf_proc_dump_write("wound", "yes"); + else + gf_proc_dump_write("wound", "no"); - frame->local = local; + gf_proc_dump_write("generation-number", "%" PRIu64, req->gen); - if (file) { - stub = fop_fsync_stub (frame, wb_fsync_helper, fd, datasync); - if (stub == NULL) { - STACK_UNWIND_STRICT (fsync, frame, -1, ENOMEM, - NULL, NULL); - return 0; - } - - request = wb_enqueue (file, stub); - if (request == NULL) { - STACK_UNWIND_STRICT (fsync, frame, -1, ENOMEM, - NULL, NULL); - call_stub_destroy (stub); - return 0; - } + gf_proc_dump_write("req->op_ret", "%d", req->op_ret); + gf_proc_dump_write("req->op_errno", "%d", req->op_errno); + gf_proc_dump_write("sync-attempts", "%d", req->wind_count); - ret = wb_process_queue (frame, file, 1); - if ((ret == -1) && (errno == ENOMEM)) { - STACK_UNWIND_STRICT (fsync, frame, -1, ENOMEM, - NULL, NULL); - call_stub_destroy (stub); - return 0; - } - - } else { - STACK_WIND (frame, - wb_fsync_cbk, - FIRST_CHILD(this), - FIRST_CHILD(this)->fops->fsync, - fd, datasync); - } + if (req->fop == GF_FOP_WRITE) { + if (list_empty(&req->wip)) + gf_proc_dump_write("sync-in-progress", "no"); + else + gf_proc_dump_write("sync-in-progress", "yes"); - return 0; -} + gf_proc_dump_write("size", "%" GF_PRI_SIZET, req->write_size); + if (req->stub) + gf_proc_dump_write("offset", "%" PRId64, + req->stub->args.offset); -int32_t -wb_release (xlator_t *this, fd_t *fd) -{ - uint64_t file_ptr = 0; - wb_file_t *file = NULL; + flag = req->ordering.lied; + gf_proc_dump_write("lied", "%d", flag); - fd_ctx_get (fd, this, &file_ptr); - file = (wb_file_t *) (long) file_ptr; + flag = req->ordering.append; + gf_proc_dump_write("append", "%d", flag); - if (file != NULL) { - LOCK (&file->lock); - { - assert (list_empty (&file->request)); - } - UNLOCK (&file->lock); + flag = req->ordering.fulfilled; + gf_proc_dump_write("fulfilled", "%d", flag); - wb_file_destroy (file); + flag = req->ordering.go; + gf_proc_dump_write("go", "%d", flag); } - - return 0; + } } int -wb_priv_dump (xlator_t *this) +wb_inode_dump(xlator_t *this, inode_t *inode) { - wb_conf_t *conf = NULL; - char key[GF_DUMP_MAX_BUF_LEN]; - char key_prefix[GF_DUMP_MAX_BUF_LEN]; + wb_inode_t *wb_inode = NULL; + int32_t ret = -1; + char *path = NULL; + char key_prefix[GF_DUMP_MAX_BUF_LEN] = { + 0, + }; + char uuid_str[64] = { + 0, + }; + + if ((inode == NULL) || (this == NULL)) { + ret = 0; + goto out; + } - if (!this) - return -1; + wb_inode = wb_inode_ctx_get(this, inode); + if (wb_inode == NULL) { + ret = 0; + goto out; + } - conf = this->private; - if (!conf) { - gf_log (this->name, GF_LOG_WARNING, - "conf null in xlator"); - return -1; + uuid_utoa_r(inode->gfid, uuid_str); + + gf_proc_dump_build_key(key_prefix, "xlator.performance.write-behind", + "wb_inode"); + + gf_proc_dump_add_section("%s", key_prefix); + + __inode_path(inode, NULL, &path); + if (path != NULL) { + gf_proc_dump_write("path", "%s", path); + GF_FREE(path); + } + + gf_proc_dump_write("inode", "%p", inode); + + gf_proc_dump_write("gfid", "%s", uuid_str); + + gf_proc_dump_write("window_conf", "%" GF_PRI_SIZET, wb_inode->window_conf); + + gf_proc_dump_write("window_current", "%" GF_PRI_SIZET, + wb_inode->window_current); + + gf_proc_dump_write("transit-size", "%" GF_PRI_SIZET, wb_inode->transit); + + gf_proc_dump_write("dontsync", "%d", wb_inode->dontsync); + + ret = TRY_LOCK(&wb_inode->lock); + if (!ret) { + if (!list_empty(&wb_inode->all)) { + __wb_dump_requests(&wb_inode->all, key_prefix); } + UNLOCK(&wb_inode->lock); + } - gf_proc_dump_build_key (key_prefix, - "xlator.performance.write-behind", - "priv"); - - gf_proc_dump_add_section (key_prefix); - - gf_proc_dump_build_key (key, key_prefix, "aggregate_size"); - gf_proc_dump_write (key, "%d", conf->aggregate_size); - gf_proc_dump_build_key (key, key_prefix, "window_size"); - gf_proc_dump_write (key, "%d", conf->window_size); - gf_proc_dump_build_key (key, key_prefix, "disable_till"); - gf_proc_dump_write (key, "%d", conf->disable_till); - gf_proc_dump_build_key (key, key_prefix, "enable_O_SYNC"); - gf_proc_dump_write (key, "%d", conf->enable_O_SYNC); - gf_proc_dump_build_key (key, key_prefix, "flush_behind"); - gf_proc_dump_write (key, "%d", conf->flush_behind); - gf_proc_dump_build_key (key, key_prefix, "enable_trickling_writes"); - gf_proc_dump_write (key, "%d", conf->enable_trickling_writes); + if (ret && wb_inode) + gf_proc_dump_write("Unable to dump the inode information", + "(Lock acquisition failed) %p (gfid: %s)", wb_inode, + uuid_str); - return 0; + ret = 0; +out: + return ret; } -int32_t -init (xlator_t *this) +int +mem_acct_init(xlator_t *this) { - dict_t *options = NULL; - wb_conf_t *conf = NULL; - char *str = NULL; - int32_t ret = -1; + int ret = -1; - if ((this->children == NULL) - || this->children->next) { - gf_log (this->name, GF_LOG_ERROR, - "FATAL: write-behind (%s) not configured with exactly " - "one child", - this->name); - return -1; - } + if (!this) { + goto out; + } - if (this->parents == NULL) { - gf_log (this->name, GF_LOG_WARNING, - "dangling volume. check volfile"); - } - - options = this->options; - - conf = CALLOC (1, sizeof (*conf)); - if (conf == NULL) { - gf_log (this->name, GF_LOG_ERROR, - "FATAL: Out of memory"); - return -1; - } - - conf->enable_O_SYNC = _gf_false; - ret = dict_get_str (options, "enable-O_SYNC", - &str); - if (ret == 0) { - ret = gf_string2boolean (str, - &conf->enable_O_SYNC); - if (ret == -1) { - gf_log (this->name, GF_LOG_ERROR, - "'enable-O_SYNC' takes only boolean arguments"); - return -1; - } - } + ret = xlator_mem_acct_init(this, gf_wb_mt_end + 1); - /* configure 'options aggregate-size <size>' */ - conf->aggregate_size = WB_AGGREGATE_SIZE; - conf->disable_till = 0; - ret = dict_get_str (options, "disable-for-first-nbytes", - &str); - if (ret == 0) { - ret = gf_string2bytesize (str, - &conf->disable_till); - if (ret != 0) { - gf_log (this->name, GF_LOG_ERROR, - "invalid number format \"%s\" of \"option " - "disable-for-first-nbytes\"", - str); - return -1; - } - } + if (ret != 0) { + gf_msg(this->name, GF_LOG_ERROR, ENOMEM, WRITE_BEHIND_MSG_NO_MEMORY, + "Memory accounting init" + "failed"); + } - gf_log (this->name, GF_LOG_DEBUG, - "disabling write-behind for first %"PRIu64" bytes", - conf->disable_till); - - /* configure 'option window-size <size>' */ - conf->window_size = WB_WINDOW_SIZE; - ret = dict_get_str (options, "cache-size", - &str); - if (ret == 0) { - ret = gf_string2bytesize (str, - &conf->window_size); - if (ret != 0) { - gf_log (this->name, GF_LOG_ERROR, - "invalid number format \"%s\" of \"option " - "window-size\"", - str); - FREE (conf); - return -1; - } - } +out: + return ret; +} - if (!conf->window_size && conf->aggregate_size) { - gf_log (this->name, GF_LOG_WARNING, - "setting window-size to be equal to " - "aggregate-size(%"PRIu64")", - conf->aggregate_size); - conf->window_size = conf->aggregate_size; - } +int +reconfigure(xlator_t *this, dict_t *options) +{ + wb_conf_t *conf = NULL; + int ret = -1; - if (conf->window_size < conf->aggregate_size) { - gf_log (this->name, GF_LOG_ERROR, - "aggregate-size(%"PRIu64") cannot be more than " - "window-size" - "(%"PRIu64")", conf->window_size, conf->aggregate_size); - FREE (conf); - return -1; - } + conf = this->private; - /* configure 'option flush-behind <on/off>' */ - conf->flush_behind = 1; - ret = dict_get_str (options, "flush-behind", - &str); - if (ret == 0) { - ret = gf_string2boolean (str, - &conf->flush_behind); - if (ret == -1) { - gf_log (this->name, GF_LOG_ERROR, - "'flush-behind' takes only boolean arguments"); - return -1; - } + GF_OPTION_RECONF("cache-size", conf->window_size, options, size_uint64, + out); - if (conf->flush_behind) { - gf_log (this->name, GF_LOG_DEBUG, - "enabling flush-behind"); - } - } + GF_OPTION_RECONF("flush-behind", conf->flush_behind, options, bool, out); - conf->enable_trickling_writes = _gf_true; - ret = dict_get_str (options, "enable-trickling-writes", - &str); - if (ret == 0) { - ret = gf_string2boolean (str, - &conf->enable_trickling_writes); - if (ret == -1) { - gf_log (this->name, GF_LOG_ERROR, - "'enable-trickling_writes' takes only boolean" - " arguments"); - return -1; - } - } + GF_OPTION_RECONF("trickling-writes", conf->trickling_writes, options, bool, + out); - this->private = conf; - return 0; + GF_OPTION_RECONF("strict-O_DIRECT", conf->strict_O_DIRECT, options, bool, + out); + + GF_OPTION_RECONF("strict-write-ordering", conf->strict_write_ordering, + options, bool, out); + GF_OPTION_RECONF("resync-failed-syncs-after-fsync", + conf->resync_after_fsync, options, bool, out); + + ret = 0; +out: + return ret; } +int32_t +init(xlator_t *this) +{ + wb_conf_t *conf = NULL; + int32_t ret = -1; + + if ((this->children == NULL) || this->children->next) { + gf_msg(this->name, GF_LOG_ERROR, 0, WRITE_BEHIND_MSG_INIT_FAILED, + "FATAL: write-behind (%s) not configured with exactly " + "one child", + this->name); + goto out; + } + + if (this->parents == NULL) { + gf_msg(this->name, GF_LOG_WARNING, 0, + WRITE_BEHIND_MSG_VOL_MISCONFIGURED, + "dangling volume. check volfilex"); + } + + conf = GF_CALLOC(1, sizeof(*conf), gf_wb_mt_wb_conf_t); + if (conf == NULL) { + goto out; + } + + /* configure 'options aggregate-size <size>' */ + GF_OPTION_INIT("aggregate-size", conf->aggregate_size, size_uint64, out); + conf->page_size = conf->aggregate_size; + + /* configure 'option window-size <size>' */ + GF_OPTION_INIT("cache-size", conf->window_size, size_uint64, out); + + if (!conf->window_size && conf->aggregate_size) { + gf_msg(this->name, GF_LOG_WARNING, 0, WRITE_BEHIND_MSG_SIZE_NOT_SET, + "setting window-size to be equal to " + "aggregate-size(%" PRIu64 ")", + conf->aggregate_size); + conf->window_size = conf->aggregate_size; + } + + if (conf->window_size < conf->aggregate_size) { + gf_msg(this->name, GF_LOG_ERROR, 0, WRITE_BEHIND_MSG_EXCEEDED_MAX_SIZE, + "aggregate-size(%" PRIu64 + ") cannot be more than " + "window-size(%" PRIu64 ")", + conf->aggregate_size, conf->window_size); + goto out; + } + + /* configure 'option flush-behind <on/off>' */ + GF_OPTION_INIT("flush-behind", conf->flush_behind, bool, out); + + GF_OPTION_INIT("trickling-writes", conf->trickling_writes, bool, out); + + GF_OPTION_INIT("strict-O_DIRECT", conf->strict_O_DIRECT, bool, out); + + GF_OPTION_INIT("strict-write-ordering", conf->strict_write_ordering, bool, + out); + + GF_OPTION_INIT("resync-failed-syncs-after-fsync", conf->resync_after_fsync, + bool, out); + + this->private = conf; + ret = 0; + +out: + if (ret) { + GF_FREE(conf); + } + return ret; +} void -fini (xlator_t *this) +fini(xlator_t *this) { - wb_conf_t *conf = this->private; + wb_conf_t *conf = NULL; - FREE (conf); - return; -} + GF_VALIDATE_OR_GOTO("write-behind", this, out); + conf = this->private; + if (!conf) { + goto out; + } -struct xlator_fops fops = { - .writev = wb_writev, - .open = wb_open, - .create = wb_create, - .readv = wb_readv, - .flush = wb_flush, - .fsync = wb_fsync, - .stat = wb_stat, - .fstat = wb_fstat, - .truncate = wb_truncate, - .ftruncate = wb_ftruncate, - .setattr = wb_setattr, -}; + this->private = NULL; + GF_FREE(conf); -struct xlator_mops mops = { -}; +out: + return; +} -struct xlator_cbks cbks = { - .release = wb_release +struct xlator_fops fops = { + .writev = wb_writev, + .readv = wb_readv, + .flush = wb_flush, + .fsync = wb_fsync, + .stat = wb_stat, + .fstat = wb_fstat, + .truncate = wb_truncate, + .ftruncate = wb_ftruncate, + .setattr = wb_setattr, + .fsetattr = wb_fsetattr, + .lookup = wb_lookup, + .readdirp = wb_readdirp, + .link = wb_link, + .fallocate = wb_fallocate, + .discard = wb_discard, + .zerofill = wb_zerofill, + .rename = wb_rename, }; +struct xlator_cbks cbks = {.forget = wb_forget, .release = wb_release}; + struct xlator_dumpops dumpops = { - .priv = wb_priv_dump, + .priv = wb_priv_dump, + .inodectx = wb_inode_dump, }; struct volume_options options[] = { - { .key = {"flush-behind"}, - .type = GF_OPTION_TYPE_BOOL - }, - { .key = {"cache-size", "window-size"}, - .type = GF_OPTION_TYPE_SIZET, - .min = 512 * GF_UNIT_KB, - .max = 1 * GF_UNIT_GB - }, - { .key = {"disable-for-first-nbytes"}, - .type = GF_OPTION_TYPE_SIZET, - .min = 1, - .max = 1 * GF_UNIT_MB, - }, - { .key = {"enable-O_SYNC"}, - .type = GF_OPTION_TYPE_BOOL, - }, - { .key = {"enable-trickling-writes"}, - .type = GF_OPTION_TYPE_BOOL, - }, - { .key = {NULL} }, + { + .key = {"write-behind"}, + .type = GF_OPTION_TYPE_BOOL, + .default_value = "off", + .description = "enable/disable write-behind", + .op_version = {GD_OP_VERSION_6_0}, + .flags = OPT_FLAG_SETTABLE, + }, + {.key = {"flush-behind"}, + .type = GF_OPTION_TYPE_BOOL, + .default_value = "on", + .op_version = {1}, + .flags = OPT_FLAG_SETTABLE | OPT_FLAG_DOC | OPT_FLAG_CLIENT_OPT, + .tags = {"write-behind"}, + .description = "If this option is set ON, instructs write-behind " + "translator to perform flush in background, by " + "returning success (or any errors, if any of " + "previous writes were failed) to application even " + "before flush FOP is sent to backend filesystem. "}, + {.key = {"cache-size", "window-size"}, + .type = GF_OPTION_TYPE_SIZET, + .min = 512 * GF_UNIT_KB, + .max = 1 * GF_UNIT_GB, + .default_value = "1MB", + .op_version = {1}, + .flags = OPT_FLAG_SETTABLE | OPT_FLAG_DOC | OPT_FLAG_CLIENT_OPT, + .tags = {"write-behind"}, + .description = "Size of the write-behind buffer for a single file " + "(inode)."}, + { + .key = {"trickling-writes"}, + .type = GF_OPTION_TYPE_BOOL, + .op_version = {GD_OP_VERSION_3_13_1}, + .flags = OPT_FLAG_SETTABLE | OPT_FLAG_DOC | OPT_FLAG_CLIENT_OPT, + .tags = {"write-behind"}, + .default_value = "on", + }, + {.key = {"strict-O_DIRECT"}, + .type = GF_OPTION_TYPE_BOOL, + .default_value = "off", + .op_version = {2}, + .flags = OPT_FLAG_SETTABLE | OPT_FLAG_DOC | OPT_FLAG_CLIENT_OPT, + .tags = {"write-behind"}, + .description = "This option when set to off, ignores the " + "O_DIRECT flag."}, + { + .key = {"strict-write-ordering"}, + .type = GF_OPTION_TYPE_BOOL, + .default_value = "off", + .op_version = {2}, + .flags = OPT_FLAG_SETTABLE | OPT_FLAG_DOC | OPT_FLAG_CLIENT_OPT, + .tags = {"write-behind"}, + .description = "Do not let later writes overtake earlier writes even " + "if they do not overlap", + }, + { + .key = {"resync-failed-syncs-after-fsync"}, + .type = GF_OPTION_TYPE_BOOL, + .default_value = "off", + .op_version = {GD_OP_VERSION_3_7_7}, + .flags = OPT_FLAG_SETTABLE | OPT_FLAG_DOC | OPT_FLAG_CLIENT_OPT, + .tags = {"write-behind"}, + .description = "If sync of \"cached-writes issued before fsync\" " + "(to backend) fails, this option configures whether " + "to retry syncing them after fsync or forget them. " + "If set to on, cached-writes are retried " + "till a \"flush\" fop (or a successful sync) on sync " + "failures. " + "fsync itself is failed irrespective of the value of " + "this option. ", + }, + { + .key = {"aggregate-size"}, + .type = GF_OPTION_TYPE_SIZET, + .default_value = "128KB", + .op_version = {GD_OP_VERSION_4_1_0}, + .flags = OPT_FLAG_SETTABLE | OPT_FLAG_DOC | OPT_FLAG_CLIENT_OPT, + .description = "Will aggregate writes until data of specified " + "size is fully filled for a single file provided " + "there are no dependent fops on cached writes. This " + "option just sets the aggregate size. Note that " + "aggregation won't happen if " + "performance.write-behind-trickling-writes" + " is turned on. Hence turn off " + "performance.write-behind.trickling-writes" + " so that writes are aggregated till a max of " + "\"aggregate-size\" bytes", + }, + {.key = {NULL}}, +}; + +xlator_api_t xlator_api = { + .init = init, + .fini = fini, + .reconfigure = reconfigure, + .mem_acct_init = mem_acct_init, + .op_version = {1}, /* Present from the initial version */ + .dumpops = &dumpops, + .fops = &fops, + .cbks = &cbks, + .options = options, + .identifier = "write-behind", + .category = GF_MAINTAINED, }; |