summaryrefslogtreecommitdiffstats
path: root/xlators/cluster/afr/src/afr-inode-write.c
diff options
context:
space:
mode:
Diffstat (limited to 'xlators/cluster/afr/src/afr-inode-write.c')
-rw-r--r--xlators/cluster/afr/src/afr-inode-write.c1474
1 files changed, 1097 insertions, 377 deletions
diff --git a/xlators/cluster/afr/src/afr-inode-write.c b/xlators/cluster/afr/src/afr-inode-write.c
index 4c5502972..c1ec69a55 100644
--- a/xlators/cluster/afr/src/afr-inode-write.c
+++ b/xlators/cluster/afr/src/afr-inode-write.c
@@ -39,34 +39,133 @@
#include "afr-transaction.h"
#include "afr-self-heal-common.h"
+void
+__inode_write_fop_cbk (call_frame_t *frame, int child_index, int read_child,
+ xlator_t *this, int32_t *op_ret, int32_t *op_errno,
+ struct iatt *prebuf, struct iatt *postbuf, dict_t *xdata)
+{
+ afr_local_t *local = NULL;
+
+ local = frame->local;
+
+ if (afr_fop_failed (*op_ret, *op_errno)) {
+ local->child_errno[child_index] = *op_errno;
+
+ switch (local->op) {
+ case GF_FOP_TRUNCATE:
+ case GF_FOP_FTRUNCATE:
+ if (*op_errno != EFBIG)
+ afr_transaction_fop_failed (frame, this,
+ child_index);
+ break;
+ default:
+ afr_transaction_fop_failed (frame, this, child_index);
+ break;
+ }
+ local->op_errno = *op_errno;
+ goto out;
+ }
+
+ if ((local->success_count == 0) || (read_child == child_index)) {
+ local->op_ret = *op_ret;
+ if (prebuf)
+ local->cont.inode_wfop.prebuf = *prebuf;
+ if (postbuf)
+ local->cont.inode_wfop.postbuf = *postbuf;
+ }
+
+ local->success_count++;
+out:
+ return;
+}
+
/* {{{ writev */
-int
+void
+afr_writev_copy_outvars (call_frame_t *src_frame, call_frame_t *dst_frame)
+{
+ afr_local_t *src_local = NULL;
+ afr_local_t *dst_local = NULL;
+
+ src_local = src_frame->local;
+ dst_local = dst_frame->local;
+
+ dst_local->op_ret = src_local->op_ret;
+ dst_local->op_errno = src_local->op_errno;
+ dst_local->cont.inode_wfop.prebuf = src_local->cont.inode_wfop.prebuf;
+ dst_local->cont.inode_wfop.postbuf = src_local->cont.inode_wfop.postbuf;
+}
+
+void
afr_writev_unwind (call_frame_t *frame, xlator_t *this)
{
afr_local_t * local = NULL;
- call_frame_t *main_frame = NULL;
+ local = frame->local;
+
+ AFR_STACK_UNWIND (writev, frame,
+ local->op_ret, local->op_errno,
+ &local->cont.inode_wfop.prebuf,
+ &local->cont.inode_wfop.postbuf,
+ NULL);
+}
+
+call_frame_t*
+afr_transaction_detach_fop_frame (call_frame_t *frame)
+{
+ afr_local_t * local = NULL;
+ call_frame_t *fop_frame = NULL;
local = frame->local;
LOCK (&frame->lock);
{
- if (local->transaction.main_frame)
- main_frame = local->transaction.main_frame;
+ fop_frame = local->transaction.main_frame;
local->transaction.main_frame = NULL;
}
UNLOCK (&frame->lock);
- if (main_frame) {
- AFR_STACK_UNWIND (writev, main_frame,
- local->op_ret, local->op_errno,
- &local->cont.writev.prebuf,
- &local->cont.writev.postbuf,
- NULL);
+ return fop_frame;
+}
+
+int
+afr_transaction_writev_unwind (call_frame_t *frame, xlator_t *this)
+{
+ call_frame_t *fop_frame = NULL;
+
+ fop_frame = afr_transaction_detach_fop_frame (frame);
+
+ if (fop_frame) {
+ afr_writev_copy_outvars (frame, fop_frame);
+ afr_writev_unwind (fop_frame, this);
}
return 0;
}
+static void
+afr_writev_handle_short_writes (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int i = 0;
+
+ local = frame->local;
+ priv = this->private;
+ /*
+ * We already have the best case result of the writev calls staged
+ * as the return value. Any writev that returns some value less
+ * than the best case is now out of sync, so mark the fop as
+ * failed. Note that fops that have returned with errors have
+ * already been marked as failed.
+ */
+ for (i = 0; i < priv->child_count; i++) {
+ if ((!local->replies[i].valid) ||
+ (local->replies[i].op_ret == -1))
+ continue;
+
+ if (local->replies[i].op_ret < local->op_ret)
+ afr_transaction_fop_failed(frame, this, i);
+ }
+}
int
afr_writev_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
@@ -74,11 +173,17 @@ afr_writev_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
struct iatt *postbuf, dict_t *xdata)
{
afr_local_t * local = NULL;
+ afr_private_t *priv = NULL;
+ call_frame_t *fop_frame = NULL;
int child_index = (long) cookie;
int call_count = -1;
int read_child = 0;
+ int ret = 0;
+ uint32_t open_fd_count = 0;
+ uint32_t write_is_append = 0;
local = frame->local;
+ priv = this->private;
read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
@@ -88,32 +193,81 @@ afr_writev_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
local->read_child_returned = _gf_true;
}
- if (afr_fop_failed (op_ret, op_errno))
- afr_transaction_fop_failed (frame, this, child_index);
-
- if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- local->cont.writev.prebuf = *prebuf;
- local->cont.writev.postbuf = *postbuf;
- }
-
- if (child_index == read_child) {
- local->cont.writev.prebuf = *prebuf;
- local->cont.writev.postbuf = *postbuf;
+ __inode_write_fop_cbk (frame, child_index, read_child, this,
+ &op_ret, &op_errno, prebuf, postbuf,
+ xdata);
+
+ local->replies[child_index].valid = 1;
+ local->replies[child_index].op_ret = op_ret;
+ local->replies[child_index].op_errno = op_errno;
+
+
+ /* stage the best case return value for unwind */
+ if ((local->success_count == 0) || (op_ret > local->op_ret)) {
+ local->op_ret = op_ret;
+ local->op_errno = op_errno;
+ }
+
+ if (op_ret != -1) {
+ if (xdata) {
+ ret = dict_get_uint32 (xdata,
+ GLUSTERFS_OPEN_FD_COUNT,
+ &open_fd_count);
+ if ((ret == 0) &&
+ (open_fd_count > local->open_fd_count)) {
+ local->open_fd_count = open_fd_count;
+ local->update_open_fd_count = _gf_true;
+ }
+
+ write_is_append = 0;
+ ret = dict_get_uint32 (xdata,
+ GLUSTERFS_WRITE_IS_APPEND,
+ &write_is_append);
+ if (ret || !write_is_append)
+ local->append_write = _gf_false;
}
- }
- local->op_errno = op_errno;
+ }
}
UNLOCK (&frame->lock);
call_count = afr_frame_return (frame);
if (call_count == 0) {
- local->transaction.unwind (frame, this);
- local->transaction.resume (frame, this);
+ if (local->update_open_fd_count)
+ afr_handle_open_fd_count (frame, this);
+
+ if (!local->stable_write && !local->append_write)
+ /* An appended write removes the necessity to
+ fsync() the file. This is because self-heal
+ has the logic to check for larger file when
+ the xattrs are not reliably pointing at
+ a stale file.
+ */
+ afr_fd_report_unstable_write (this, local->fd);
+
+ afr_writev_handle_short_writes (frame, this);
+ if (afr_any_fops_failed (local, priv)) {
+ //Don't unwind until post-op is complete
+ local->transaction.resume (frame, this);
+ } else {
+ /*
+ * Generally inode-write fops do transaction.unwind then
+ * transaction.resume, but writev needs to make sure that
+ * delayed post-op frame is placed in fdctx before unwind
+ * happens. This prevents the race of flush doing the
+ * changelog wakeup first in fuse thread and then this
+ * writev placing its delayed post-op frame in fdctx.
+ * This helps flush make sure all the delayed post-ops are
+ * completed.
+ */
+
+ fop_frame = afr_transaction_detach_fop_frame (frame);
+ afr_writev_copy_outvars (frame, fop_frame);
+ local->transaction.resume (frame, this);
+ afr_writev_unwind (fop_frame, this);
+ }
}
return 0;
}
@@ -125,6 +279,8 @@ afr_writev_wind (call_frame_t *frame, xlator_t *this)
afr_private_t *priv = NULL;
int i = 0;
int call_count = -1;
+ dict_t *xdata = NULL;
+ GF_UNUSED int ret = 0;
local = frame->local;
priv = this->private;
@@ -138,6 +294,28 @@ afr_writev_wind (call_frame_t *frame, xlator_t *this)
}
local->call_count = call_count;
+ local->replies = GF_CALLOC(priv->child_count, sizeof(*local->replies),
+ gf_afr_mt_reply_t);
+ if (!local->replies) {
+ local->op_ret = -1;
+ local->op_errno = ENOMEM;
+ local->transaction.unwind(frame, this);
+ local->transaction.resume(frame, this);
+ return 0;
+ }
+
+ xdata = dict_new ();
+ if (xdata) {
+ ret = dict_set_uint32 (xdata, GLUSTERFS_OPEN_FD_COUNT,
+ sizeof (uint32_t));
+ ret = dict_set_uint32 (xdata, GLUSTERFS_WRITE_IS_APPEND,
+ 0);
+ /* Set append_write to be true speculatively. If on any
+ server it turns not be true, we unset it in the
+ callback.
+ */
+ local->append_write = _gf_true;
+ }
for (i = 0; i < priv->child_count; i++) {
if (local->transaction.pre_op[i]) {
@@ -151,13 +329,16 @@ afr_writev_wind (call_frame_t *frame, xlator_t *this)
local->cont.writev.offset,
local->cont.writev.flags,
local->cont.writev.iobref,
- NULL);
+ xdata);
if (!--call_count)
break;
}
}
+ if (xdata)
+ dict_unref (xdata);
+
return 0;
}
@@ -197,7 +378,7 @@ afr_do_writev (call_frame_t *frame, xlator_t *this)
}
transaction_frame->local = local;
- frame->local = NULL;
+ AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
local->op = GF_FOP_WRITE;
@@ -205,10 +386,17 @@ afr_do_writev (call_frame_t *frame, xlator_t *this)
local->transaction.fop = afr_writev_wind;
local->transaction.done = afr_writev_done;
- local->transaction.unwind = afr_writev_unwind;
+ local->transaction.unwind = afr_transaction_writev_unwind;
local->transaction.main_frame = frame;
if (local->fd->flags & O_APPEND) {
+ /*
+ * Backend vfs ignores the 'offset' for append mode fd so
+ * locking just the region provided for the writev does not
+ * give consistency gurantee. The actual write may happen at a
+ * completely different range than the one provided by the
+ * offset, len in the fop. So lock the entire file.
+ */
local->transaction.start = 0;
local->transaction.len = 0;
} else {
@@ -234,153 +422,74 @@ out:
return 0;
}
-static int
-afr_prepare_loc (call_frame_t *frame, fd_t *fd)
+static void
+afr_trigger_open_fd_self_heal (fd_t *fd, xlator_t *this)
{
- afr_local_t *local = NULL;
- char *name = NULL;
- char *path = NULL;
- int ret = 0;
-
- if ((!fd) || (!fd->inode))
- return -1;
-
- local = frame->local;
- ret = inode_path (fd->inode, NULL, (char **)&path);
- if (ret <= 0) {
- gf_log (frame->this->name, GF_LOG_DEBUG,
- "Unable to get path for gfid: %s",
- uuid_utoa (fd->inode->gfid));
- return -1;
- }
-
- if (local->loc.path) {
- if (strcmp (path, local->loc.path))
- gf_log (frame->this->name, GF_LOG_DEBUG,
- "overwriting old loc->path %s with %s",
- local->loc.path, path);
- GF_FREE ((char *)local->loc.path);
- }
- local->loc.path = path;
-
- name = strrchr (local->loc.path, '/');
- if (name)
- name++;
- local->loc.name = name;
-
- if (local->loc.inode) {
- inode_unref (local->loc.inode);
- }
- local->loc.inode = inode_ref (fd->inode);
-
- if (local->loc.parent) {
- inode_unref (local->loc.parent);
+ call_frame_t *frame = NULL;
+ afr_local_t *local = NULL;
+ afr_self_heal_t *sh = NULL;
+ char *reason = NULL;
+ int32_t op_errno = 0;
+ int ret = 0;
+
+ if (!fd || !fd->inode || uuid_is_null (fd->inode->gfid)) {
+ gf_log_callingfn (this->name, GF_LOG_ERROR, "Invalid args: "
+ "fd: %p, inode: %p", fd,
+ fd ? fd->inode : NULL);
+ goto out;
}
- local->loc.parent = inode_parent (local->loc.inode, 0, NULL);
-
- return 0;
-}
-
-afr_fd_paused_call_t*
-afr_paused_call_create (call_frame_t *frame)
-{
- afr_local_t *local = NULL;
- afr_fd_paused_call_t *paused_call = NULL;
+ frame = create_frame (this, this->ctx->pool);
+ if (!frame)
+ goto out;
+ AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
local = frame->local;
- GF_ASSERT (local->fop_call_continue);
-
- paused_call = GF_CALLOC (1, sizeof (*paused_call),
- gf_afr_fd_paused_call_t);
- if (paused_call) {
- INIT_LIST_HEAD (&paused_call->call_list);
- paused_call->frame = frame;
- }
-
- return paused_call;
-}
-
-static int
-afr_pause_fd_fop (call_frame_t *frame, xlator_t *this, afr_fd_ctx_t *fd_ctx)
-{
- afr_fd_paused_call_t *paused_call = NULL;
- int ret = 0;
-
- paused_call = afr_paused_call_create (frame);
- if (paused_call)
- list_add (&paused_call->call_list, &fd_ctx->paused_calls);
- else
- ret = -ENOMEM;
-
- return ret;
-}
+ ret = afr_local_init (local, this->private, &op_errno);
+ if (ret < 0)
+ goto out;
-static void
-afr_trigger_open_fd_self_heal (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- inode_t *inode = NULL;
- char *reason = NULL;
+ local->loc.inode = inode_ref (fd->inode);
+ ret = loc_path (&local->loc, NULL);
+ if (ret < 0)
+ goto out;
- local = frame->local;
sh = &local->self_heal;
- inode = local->fd->inode;
-
- sh->do_missing_entry_self_heal = _gf_true;
- sh->do_gfid_self_heal = _gf_true;
- sh->do_data_self_heal = _gf_true;
+ sh->do_metadata_self_heal = _gf_true;
+ if (fd->inode->ia_type == IA_IFREG)
+ sh->do_data_self_heal = _gf_true;
+ else if (fd->inode->ia_type == IA_IFDIR)
+ sh->do_entry_self_heal = _gf_true;
reason = "subvolume came online";
- afr_launch_self_heal (frame, this, inode, _gf_true, inode->ia_type,
- reason, NULL, NULL);
+ afr_launch_self_heal (frame, this, fd->inode, _gf_true,
+ fd->inode->ia_type, reason, NULL, NULL);
+ return;
+out:
+ AFR_STACK_DESTROY (frame);
}
-int
-afr_open_fd_fix (call_frame_t *frame, xlator_t *this, gf_boolean_t pause_fop)
-{
- int ret = 0;
- int i = 0;
- afr_fd_ctx_t *fd_ctx = NULL;
- inode_t *inode = NULL;
- gf_boolean_t need_self_heal = _gf_false;
- int *need_open = NULL;
- int need_open_count = 0;
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- gf_boolean_t fop_continue = _gf_true;
+void
+afr_open_fd_fix (fd_t *fd, xlator_t *this)
+{
+ int ret = 0;
+ int i = 0;
+ afr_fd_ctx_t *fd_ctx = NULL;
+ gf_boolean_t need_self_heal = _gf_false;
+ int *need_open = NULL;
+ size_t need_open_count = 0;
+ afr_private_t *priv = NULL;
- local = frame->local;
priv = this->private;
- GF_ASSERT (local->fd);
-
- inode = local->fd->inode;
- //gfid is not set in rebalance, that case needs to be handled.
- if (fd_is_anonymous (local->fd) ||
- !inode || uuid_is_null (inode->gfid)) {
- fop_continue = _gf_true;
- goto out;
- }
-
- if (pause_fop)
- GF_ASSERT (local->fop_call_continue);
-
- ret = afr_prepare_loc (frame, local->fd);
- if (ret < 0) {
- //File does not exist we cant open it.
- ret = 0;
+ if (!afr_is_fd_fixable (fd))
goto out;
- }
- fd_ctx = afr_fd_ctx_get (local->fd, this);
- if (!fd_ctx) {
- ret = -EINVAL;
+ fd_ctx = afr_fd_ctx_get (fd, this);
+ if (!fd_ctx)
goto out;
- }
- LOCK (&local->fd->lock);
+ LOCK (&fd->lock);
{
if (fd_ctx->up_count < priv->up_count) {
need_self_heal = _gf_true;
@@ -388,55 +497,34 @@ afr_open_fd_fix (call_frame_t *frame, xlator_t *this, gf_boolean_t pause_fop)
fd_ctx->down_count = priv->down_count;
}
+ need_open = alloca (priv->child_count * sizeof (*need_open));
for (i = 0; i < priv->child_count; i++) {
- if ((fd_ctx->opened_on[i] == AFR_FD_NOT_OPENED) &&
- local->child_up[i]) {
- fd_ctx->opened_on[i] = AFR_FD_OPENING;
- if (!need_open)
- need_open = GF_CALLOC (priv->child_count,
- sizeof (*need_open),
- gf_afr_mt_int32_t);
- need_open[i] = 1;
- need_open_count++;
- } else if (pause_fop && local->child_up[i] &&
- (fd_ctx->opened_on[i] == AFR_FD_OPENING)) {
- local->fop_paused = _gf_true;
- }
- }
+ need_open[i] = 0;
+ if (fd_ctx->opened_on[i] != AFR_FD_NOT_OPENED)
+ continue;
- if (local->fop_paused) {
- GF_ASSERT (pause_fop);
- gf_log (this->name, GF_LOG_INFO, "Pause fd %p",
- local->fd);
- ret = afr_pause_fd_fop (frame, this, fd_ctx);
- if (ret)
- goto unlock;
- fop_continue = _gf_false;
+ if (!priv->child_up[i])
+ continue;
+
+ fd_ctx->opened_on[i] = AFR_FD_OPENING;
+
+ need_open[i] = 1;
+ need_open_count++;
}
}
-unlock:
- UNLOCK (&local->fd->lock);
- if (ret) {
- gf_log (this->name, GF_LOG_ERROR, "Failed to fix fd for %s",
- local->loc.path);
- fop_continue = _gf_false;
+ UNLOCK (&fd->lock);
+ if (ret)
goto out;
- }
if (need_self_heal)
- afr_trigger_open_fd_self_heal (frame, this);
+ afr_trigger_open_fd_self_heal (fd, this);
if (!need_open_count)
goto out;
- gf_log (this->name, GF_LOG_INFO, "Opening fd %p", local->fd);
- afr_fix_open (frame, this, fd_ctx, need_open_count, need_open);
- fop_continue = _gf_false;
+ afr_fix_open (this, fd, need_open_count, need_open);
out:
- GF_FREE (need_open);
- if (fop_continue && local->fop_call_continue)
- local->fop_call_continue (frame, this);
- return ret;
+ return;
}
int
@@ -455,6 +543,11 @@ afr_writev (call_frame_t *frame, xlator_t *this, fd_t *fd,
priv = this->private;
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
+
QUORUM_CHECK(writev,out);
AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
@@ -471,13 +564,15 @@ afr_writev (call_frame_t *frame, xlator_t *this, fd_t *fd,
local->cont.writev.iobref = iobref_ref (iobref);
local->fd = fd_ref (fd);
- local->fop_call_continue = afr_do_writev;
- ret = afr_open_fd_fix (frame, this, _gf_true);
- if (ret) {
- op_errno = -ret;
- goto out;
- }
+ /* detect here, but set it in writev_wind_cbk *after* the unstable
+ write is performed
+ */
+ local->stable_write = !!((fd->flags|flags)&(O_SYNC|O_DSYNC));
+
+ afr_open_fd_fix (fd, this);
+
+ afr_do_writev (frame, this);
ret = 0;
out:
@@ -511,8 +606,8 @@ afr_truncate_unwind (call_frame_t *frame, xlator_t *this)
if (main_frame) {
AFR_STACK_UNWIND (truncate, main_frame, local->op_ret,
local->op_errno,
- &local->cont.truncate.prebuf,
- &local->cont.truncate.postbuf,
+ &local->cont.inode_wfop.prebuf,
+ &local->cont.inode_wfop.postbuf,
NULL);
}
@@ -526,14 +621,11 @@ afr_truncate_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
struct iatt *postbuf, dict_t *xdata)
{
afr_local_t * local = NULL;
- afr_private_t * priv = NULL;
int child_index = (long) cookie;
int read_child = 0;
int call_count = -1;
- int need_unwind = 0;
local = frame->local;
- priv = this->private;
read_child = afr_inode_get_read_ctx (this, local->loc.inode, NULL);
@@ -543,38 +635,22 @@ afr_truncate_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
local->read_child_returned = _gf_true;
}
- if (afr_fop_failed (op_ret, op_errno) && op_errno != EFBIG)
- afr_transaction_fop_failed (frame, this, child_index);
-
if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- local->cont.truncate.prebuf = *prebuf;
- local->cont.truncate.postbuf = *postbuf;
- }
-
- if (child_index == read_child) {
- local->cont.truncate.prebuf = *prebuf;
- local->cont.truncate.postbuf = *postbuf;
- }
-
- local->success_count++;
-
- if ((local->success_count >= priv->wait_count)
- && local->read_child_returned) {
- need_unwind = 1;
- }
+ if (prebuf->ia_size != postbuf->ia_size)
+ local->stable_write = _gf_false;
}
- local->op_errno = op_errno;
+ __inode_write_fop_cbk (frame, child_index, read_child, this,
+ &op_ret, &op_errno, prebuf, postbuf,
+ xdata);
}
UNLOCK (&frame->lock);
- if (need_unwind)
- local->transaction.unwind (frame, this);
-
call_count = afr_frame_return (frame);
if (call_count == 0) {
+ if (local->stable_write && afr_txn_nothing_failed (frame, this))
+ local->transaction.unwind (frame, this);
+
local->transaction.resume (frame, this);
}
@@ -602,6 +678,7 @@ afr_truncate_wind (call_frame_t *frame, xlator_t *this)
}
local->call_count = call_count;
+ local->stable_write = _gf_true;
for (i = 0; i < priv->child_count; i++) {
if (local->transaction.pre_op[i]) {
@@ -722,8 +799,8 @@ afr_ftruncate_unwind (call_frame_t *frame, xlator_t *this)
if (main_frame) {
AFR_STACK_UNWIND (ftruncate, main_frame, local->op_ret,
local->op_errno,
- &local->cont.ftruncate.prebuf,
- &local->cont.ftruncate.postbuf,
+ &local->cont.inode_wfop.prebuf,
+ &local->cont.inode_wfop.postbuf,
NULL);
}
return 0;
@@ -736,14 +813,11 @@ afr_ftruncate_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
struct iatt *postbuf, dict_t *xdata)
{
afr_local_t * local = NULL;
- afr_private_t * priv = NULL;
int child_index = (long) cookie;
int call_count = -1;
- int need_unwind = 0;
int read_child = 0;
local = frame->local;
- priv = this->private;
read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
@@ -753,38 +827,22 @@ afr_ftruncate_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
local->read_child_returned = _gf_true;
}
- if (afr_fop_failed (op_ret, op_errno))
- afr_transaction_fop_failed (frame, this, child_index);
-
if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- local->cont.ftruncate.prebuf = *prebuf;
- local->cont.ftruncate.postbuf = *postbuf;
- }
-
- if (child_index == read_child) {
- local->cont.ftruncate.prebuf = *prebuf;
- local->cont.ftruncate.postbuf = *postbuf;
- }
-
- local->success_count++;
-
- if ((local->success_count >= priv->wait_count)
- && local->read_child_returned) {
- need_unwind = 1;
- }
+ if (prebuf->ia_size != postbuf->ia_size)
+ local->stable_write = _gf_false;
}
- local->op_errno = op_errno;
+ __inode_write_fop_cbk (frame, child_index, read_child, this,
+ &op_ret, &op_errno, prebuf, postbuf,
+ xdata);
}
UNLOCK (&frame->lock);
- if (need_unwind)
- local->transaction.unwind (frame, this);
-
call_count = afr_frame_return (frame);
if (call_count == 0) {
+ if (local->stable_write && afr_txn_nothing_failed (frame, this))
+ local->transaction.unwind (frame, this);
+
local->transaction.resume (frame, this);
}
@@ -812,6 +870,7 @@ afr_ftruncate_wind (call_frame_t *frame, xlator_t *this)
}
local->call_count = call_count;
+ local->stable_write = _gf_true;
for (i = 0; i < priv->child_count; i++) {
if (local->transaction.pre_op[i]) {
@@ -911,6 +970,10 @@ afr_ftruncate (call_frame_t *frame, xlator_t *this,
priv = this->private;
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
QUORUM_CHECK(ftruncate,out);
AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
@@ -923,13 +986,10 @@ afr_ftruncate (call_frame_t *frame, xlator_t *this,
local->cont.ftruncate.offset = offset;
local->fd = fd_ref (fd);
- local->fop_call_continue = afr_do_ftruncate;
- ret = afr_open_fd_fix (frame, this, _gf_true);
- if (ret) {
- op_errno = -ret;
- goto out;
- }
+ afr_open_fd_fix (fd, this);
+
+ afr_do_ftruncate (frame, this);
ret = 0;
out:
@@ -965,8 +1025,8 @@ afr_setattr_unwind (call_frame_t *frame, xlator_t *this)
if (main_frame) {
AFR_STACK_UNWIND (setattr, main_frame, local->op_ret,
local->op_errno,
- &local->cont.setattr.preop_buf,
- &local->cont.setattr.postop_buf,
+ &local->cont.inode_wfop.prebuf,
+ &local->cont.inode_wfop.postbuf,
NULL);
}
@@ -997,29 +1057,14 @@ afr_setattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
local->read_child_returned = _gf_true;
}
- if (afr_fop_failed (op_ret, op_errno))
- afr_transaction_fop_failed (frame, this, child_index);
+ __inode_write_fop_cbk (frame, child_index, read_child, this,
+ &op_ret, &op_errno, preop, postop,
+ xdata);
- if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- local->cont.setattr.preop_buf = *preop;
- local->cont.setattr.postop_buf = *postop;
- }
-
- if (child_index == read_child) {
- local->cont.setattr.preop_buf = *preop;
- local->cont.setattr.postop_buf = *postop;
- }
-
- local->success_count++;
-
- if ((local->success_count >= priv->wait_count)
- && local->read_child_returned) {
- need_unwind = 1;
- }
+ if ((local->success_count >= priv->wait_count)
+ && local->read_child_returned) {
+ need_unwind = 1;
}
- local->op_errno = op_errno;
}
UNLOCK (&frame->lock);
@@ -1174,8 +1219,8 @@ afr_fsetattr_unwind (call_frame_t *frame, xlator_t *this)
if (main_frame) {
AFR_STACK_UNWIND (fsetattr, main_frame, local->op_ret,
local->op_errno,
- &local->cont.fsetattr.preop_buf,
- &local->cont.fsetattr.postop_buf,
+ &local->cont.inode_wfop.prebuf,
+ &local->cont.inode_wfop.postbuf,
NULL);
}
@@ -1206,29 +1251,14 @@ afr_fsetattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
local->read_child_returned = _gf_true;
}
- if (afr_fop_failed (op_ret, op_errno))
- afr_transaction_fop_failed (frame, this, child_index);
-
- if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- local->cont.fsetattr.preop_buf = *preop;
- local->cont.fsetattr.postop_buf = *postop;
- }
+ __inode_write_fop_cbk (frame, child_index, read_child, this,
+ &op_ret, &op_errno, preop, postop,
+ xdata);
- if (child_index == read_child) {
- local->cont.fsetattr.preop_buf = *preop;
- local->cont.fsetattr.postop_buf = *postop;
- }
-
- local->success_count++;
-
- if ((local->success_count >= priv->wait_count)
- && local->read_child_returned) {
- need_unwind = 1;
- }
+ if ((local->success_count >= priv->wait_count)
+ && local->read_child_returned) {
+ need_unwind = 1;
}
- local->op_errno = op_errno;
}
UNLOCK (&frame->lock);
@@ -1316,6 +1346,11 @@ afr_fsetattr (call_frame_t *frame, xlator_t *this,
priv = this->private;
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
+
QUORUM_CHECK(fsetattr,out);
transaction_frame = copy_frame (frame);
@@ -1340,11 +1375,7 @@ afr_fsetattr (call_frame_t *frame, xlator_t *this,
local->fd = fd_ref (fd);
- ret = afr_open_fd_fix (transaction_frame, this, _gf_false);
- if (ret) {
- op_errno = -ret;
- goto out;
- }
+ afr_open_fd_fix (fd, this);
local->transaction.main_frame = frame;
local->transaction.start = LLONG_MAX - 1;
@@ -1400,28 +1431,23 @@ int
afr_setxattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
int32_t op_ret, int32_t op_errno, dict_t *xdata)
{
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- int call_count = -1;
- int need_unwind = 0;
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int call_count = -1;
+ int need_unwind = 0;
+ int child_index = (long) cookie;
local = frame->local;
priv = this->private;
LOCK (&frame->lock);
{
- if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- }
- local->success_count++;
-
- if (local->success_count == priv->child_count) {
- need_unwind = 1;
- }
+ __inode_write_fop_cbk (frame, child_index, -1, this,
+ &op_ret, &op_errno, NULL, NULL,
+ xdata);
+ if (local->success_count == priv->child_count) {
+ need_unwind = 1;
}
-
- local->op_errno = op_errno;
}
UNLOCK (&frame->lock);
@@ -1590,28 +1616,24 @@ int
afr_fsetxattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
int32_t op_ret, int32_t op_errno, dict_t *xdata)
{
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- int call_count = -1;
- int need_unwind = 0;
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int call_count = -1;
+ int need_unwind = 0;
+ int child_index = (long) cookie;
local = frame->local;
priv = this->private;
LOCK (&frame->lock);
{
- if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- }
- local->success_count++;
- if (local->success_count == priv->child_count) {
- need_unwind = 1;
- }
+ __inode_write_fop_cbk (frame, child_index, -1, this,
+ &op_ret, &op_errno, NULL, NULL,
+ xdata);
+ if (local->success_count == priv->child_count) {
+ need_unwind = 1;
}
-
- local->op_errno = op_errno;
}
UNLOCK (&frame->lock);
@@ -1701,11 +1723,13 @@ afr_fsetxattr (call_frame_t *frame, xlator_t *this,
GF_IF_INTERNAL_XATTR_GOTO ("trusted.glusterfs.afr.*", dict,
op_errno, out);
- if (ret)
- goto out;
-
priv = this->private;
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
+
QUORUM_CHECK(fsetxattr,out);
AFR_LOCAL_ALLOC_OR_GOTO (local, out);
@@ -1788,28 +1812,23 @@ int
afr_removexattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
int32_t op_ret, int32_t op_errno, dict_t *xdata)
{
- afr_local_t * local = NULL;
- afr_private_t * priv = NULL;
- int call_count = -1;
- int need_unwind = 0;
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int call_count = -1;
+ int need_unwind = 0;
+ int child_index = (long) cookie;
local = frame->local;
priv = this->private;
LOCK (&frame->lock);
{
- if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- }
- local->success_count++;
-
- if (local->success_count == priv->wait_count) {
- need_unwind = 1;
- }
+ __inode_write_fop_cbk (frame, child_index, -1, this,
+ &op_ret, &op_errno, NULL, NULL,
+ xdata);
+ if (local->success_count == priv->wait_count) {
+ need_unwind = 1;
}
-
- local->op_errno = op_errno;
}
UNLOCK (&frame->lock);
@@ -1977,28 +1996,24 @@ int
afr_fremovexattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
int32_t op_ret, int32_t op_errno, dict_t *xdata)
{
- afr_local_t * local = NULL;
- afr_private_t * priv = NULL;
- int call_count = -1;
- int need_unwind = 0;
+ afr_local_t * local = NULL;
+ afr_private_t * priv = NULL;
+ int call_count = -1;
+ int need_unwind = 0;
+ int child_index = (long) cookie;
local = frame->local;
priv = this->private;
LOCK (&frame->lock);
{
- if (op_ret != -1) {
- if (local->success_count == 0) {
- local->op_ret = op_ret;
- }
- local->success_count++;
+ __inode_write_fop_cbk (frame, child_index, -1, this,
+ &op_ret, &op_errno, NULL, NULL,
+ xdata);
- if (local->success_count == priv->wait_count) {
- need_unwind = 1;
- }
+ if (local->success_count == priv->wait_count) {
+ need_unwind = 1;
}
-
- local->op_errno = op_errno;
}
UNLOCK (&frame->lock);
@@ -2091,6 +2106,10 @@ afr_fremovexattr (call_frame_t *frame, xlator_t *this,
VALIDATE_OR_GOTO (this->private, out);
priv = this->private;
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
QUORUM_CHECK(fremovexattr, out);
@@ -2139,3 +2158,704 @@ out:
return 0;
}
+
+static int
+afr_fallocate_unwind (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t * local = NULL;
+ call_frame_t *main_frame = NULL;
+
+ local = frame->local;
+
+ LOCK (&frame->lock);
+ {
+ if (local->transaction.main_frame)
+ main_frame = local->transaction.main_frame;
+ local->transaction.main_frame = NULL;
+ }
+ UNLOCK (&frame->lock);
+
+ if (main_frame) {
+ AFR_STACK_UNWIND (fallocate, main_frame, local->op_ret,
+ local->op_errno,
+ &local->cont.inode_wfop.prebuf,
+ &local->cont.inode_wfop.postbuf,
+ NULL);
+ }
+ return 0;
+}
+
+static int
+afr_fallocate_wind_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)
+{
+ afr_local_t * local = NULL;
+ afr_private_t * priv = NULL;
+ int child_index = (long) cookie;
+ int call_count = -1;
+ int need_unwind = 0;
+ int read_child = 0;
+
+ local = frame->local;
+ priv = this->private;
+
+ read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
+
+ LOCK (&frame->lock);
+ {
+ if (child_index == read_child) {
+ local->read_child_returned = _gf_true;
+ }
+
+ __inode_write_fop_cbk (frame, child_index, read_child, this,
+ &op_ret, &op_errno, prebuf, postbuf,
+ xdata);
+
+ if ((local->success_count >= priv->wait_count)
+ && local->read_child_returned) {
+ need_unwind = 1;
+ }
+ }
+ UNLOCK (&frame->lock);
+
+ if (need_unwind)
+ local->transaction.unwind (frame, this);
+
+ call_count = afr_frame_return (frame);
+
+ if (call_count == 0) {
+ local->transaction.resume (frame, this);
+ }
+
+ return 0;
+}
+
+static int
+afr_fallocate_wind (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int call_count = -1;
+ int i = 0;
+
+ local = frame->local;
+ priv = this->private;
+
+ call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
+ priv->child_count);
+
+ if (call_count == 0) {
+ local->transaction.resume (frame, this);
+ return 0;
+ }
+
+ local->call_count = call_count;
+
+ for (i = 0; i < priv->child_count; i++) {
+ if (local->transaction.pre_op[i]) {
+ STACK_WIND_COOKIE (frame, afr_fallocate_wind_cbk,
+ (void *) (long) i,
+ priv->children[i],
+ priv->children[i]->fops->fallocate,
+ local->fd,
+ local->cont.fallocate.mode,
+ local->cont.fallocate.offset,
+ local->cont.fallocate.len,
+ NULL);
+
+ if (!--call_count)
+ break;
+ }
+ }
+
+ return 0;
+}
+
+static int
+afr_fallocate_done (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+
+ local = frame->local;
+
+ local->transaction.unwind (frame, this);
+
+ AFR_STACK_DESTROY (frame);
+
+ return 0;
+}
+
+static int
+afr_do_fallocate (call_frame_t *frame, xlator_t *this)
+{
+ call_frame_t * transaction_frame = NULL;
+ afr_local_t * local = NULL;
+ int op_ret = -1;
+ int op_errno = 0;
+
+ local = frame->local;
+
+ transaction_frame = copy_frame (frame);
+ if (!transaction_frame) {
+ goto out;
+ }
+
+ transaction_frame->local = local;
+ frame->local = NULL;
+
+ local->op = GF_FOP_FALLOCATE;
+
+ local->transaction.fop = afr_fallocate_wind;
+ local->transaction.done = afr_fallocate_done;
+ local->transaction.unwind = afr_fallocate_unwind;
+
+ local->transaction.main_frame = frame;
+
+ local->transaction.start = local->cont.fallocate.offset;
+ local->transaction.len = 0;
+
+ /* fallocate can modify the file size */
+ op_ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
+ if (op_ret < 0) {
+ op_errno = -op_ret;
+ goto out;
+ }
+
+ op_ret = 0;
+out:
+ if (op_ret < 0) {
+ if (transaction_frame)
+ AFR_STACK_DESTROY (transaction_frame);
+ AFR_STACK_UNWIND (fallocate, frame, op_ret, op_errno, NULL,
+ NULL, NULL);
+ }
+
+ return 0;
+}
+
+int
+afr_fallocate (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t mode,
+ off_t offset, size_t len, dict_t *xdata)
+{
+ afr_private_t * priv = NULL;
+ afr_local_t * local = NULL;
+ call_frame_t *transaction_frame = NULL;
+ int ret = -1;
+ int op_errno = 0;
+
+ VALIDATE_OR_GOTO (frame, out);
+ VALIDATE_OR_GOTO (this, out);
+ VALIDATE_OR_GOTO (this->private, out);
+
+ priv = this->private;
+
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
+ QUORUM_CHECK(fallocate,out);
+
+ AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
+ local = frame->local;
+
+ ret = afr_local_init (local, priv, &op_errno);
+ if (ret < 0)
+ goto out;
+
+ local->cont.fallocate.mode = mode;
+ local->cont.fallocate.offset = offset;
+ local->cont.fallocate.len = len;
+
+ local->fd = fd_ref (fd);
+
+ afr_open_fd_fix (fd, this);
+
+ afr_do_fallocate (frame, this);
+
+ ret = 0;
+out:
+ if (ret < 0) {
+ if (transaction_frame)
+ AFR_STACK_DESTROY (transaction_frame);
+ AFR_STACK_UNWIND (fallocate, frame, -1, op_errno, NULL, NULL, NULL);
+ }
+
+ return 0;
+}
+
+/* }}} */
+
+/* {{{ discard */
+
+static int
+afr_discard_unwind (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t * local = NULL;
+ call_frame_t *main_frame = NULL;
+
+ local = frame->local;
+
+ LOCK (&frame->lock);
+ {
+ if (local->transaction.main_frame)
+ main_frame = local->transaction.main_frame;
+ local->transaction.main_frame = NULL;
+ }
+ UNLOCK (&frame->lock);
+
+ if (main_frame) {
+ AFR_STACK_UNWIND (discard, main_frame, local->op_ret,
+ local->op_errno,
+ &local->cont.inode_wfop.prebuf,
+ &local->cont.inode_wfop.postbuf,
+ NULL);
+ }
+ return 0;
+}
+
+static int
+afr_discard_wind_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)
+{
+ afr_local_t * local = NULL;
+ afr_private_t * priv = NULL;
+ int child_index = (long) cookie;
+ int call_count = -1;
+ int need_unwind = 0;
+ int read_child = 0;
+
+ local = frame->local;
+ priv = this->private;
+
+ read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
+
+ LOCK (&frame->lock);
+ {
+ if (child_index == read_child) {
+ local->read_child_returned = _gf_true;
+ }
+
+ __inode_write_fop_cbk (frame, child_index, read_child, this,
+ &op_ret, &op_errno, prebuf, postbuf,
+ xdata);
+
+ if ((local->success_count >= priv->wait_count)
+ && local->read_child_returned) {
+ need_unwind = 1;
+ }
+ }
+ UNLOCK (&frame->lock);
+
+ if (need_unwind)
+ local->transaction.unwind (frame, this);
+
+ call_count = afr_frame_return (frame);
+
+ if (call_count == 0) {
+ local->transaction.resume (frame, this);
+ }
+
+ return 0;
+}
+
+static int
+afr_discard_wind (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int call_count = -1;
+ int i = 0;
+
+ local = frame->local;
+ priv = this->private;
+
+ call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
+ priv->child_count);
+
+ if (call_count == 0) {
+ local->transaction.resume (frame, this);
+ return 0;
+ }
+
+ local->call_count = call_count;
+
+ for (i = 0; i < priv->child_count; i++) {
+ if (local->transaction.pre_op[i]) {
+ STACK_WIND_COOKIE (frame, afr_discard_wind_cbk,
+ (void *) (long) i,
+ priv->children[i],
+ priv->children[i]->fops->discard,
+ local->fd,
+ local->cont.discard.offset,
+ local->cont.discard.len,
+ NULL);
+
+ if (!--call_count)
+ break;
+ }
+ }
+
+ return 0;
+}
+
+static int
+afr_discard_done (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+
+ local = frame->local;
+
+ local->transaction.unwind (frame, this);
+
+ AFR_STACK_DESTROY (frame);
+
+ return 0;
+}
+
+static int
+afr_do_discard (call_frame_t *frame, xlator_t *this)
+{
+ call_frame_t * transaction_frame = NULL;
+ afr_local_t * local = NULL;
+ int op_ret = -1;
+ int op_errno = 0;
+
+ local = frame->local;
+
+ transaction_frame = copy_frame (frame);
+ if (!transaction_frame) {
+ goto out;
+ }
+
+ transaction_frame->local = local;
+ frame->local = NULL;
+
+ local->op = GF_FOP_DISCARD;
+
+ local->transaction.fop = afr_discard_wind;
+ local->transaction.done = afr_discard_done;
+ local->transaction.unwind = afr_discard_unwind;
+
+ local->transaction.main_frame = frame;
+
+ local->transaction.start = local->cont.discard.offset;
+ local->transaction.len = 0;
+
+ op_ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
+ if (op_ret < 0) {
+ op_errno = -op_ret;
+ goto out;
+ }
+
+ op_ret = 0;
+out:
+ if (op_ret < 0) {
+ if (transaction_frame)
+ AFR_STACK_DESTROY (transaction_frame);
+ AFR_STACK_UNWIND (discard, frame, op_ret, op_errno, NULL,
+ NULL, NULL);
+ }
+
+ return 0;
+}
+
+int
+afr_discard (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
+ size_t len, dict_t *xdata)
+{
+ afr_private_t * priv = NULL;
+ afr_local_t * local = NULL;
+ call_frame_t *transaction_frame = NULL;
+ int ret = -1;
+ int op_errno = 0;
+
+ VALIDATE_OR_GOTO (frame, out);
+ VALIDATE_OR_GOTO (this, out);
+ VALIDATE_OR_GOTO (this->private, out);
+
+ priv = this->private;
+
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
+ QUORUM_CHECK(discard, out);
+
+ AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
+ local = frame->local;
+
+ ret = afr_local_init (local, priv, &op_errno);
+ if (ret < 0)
+ goto out;
+
+ local->cont.discard.offset = offset;
+ local->cont.discard.len = len;
+
+ local->fd = fd_ref (fd);
+
+ afr_open_fd_fix (fd, this);
+
+ afr_do_discard(frame, this);
+
+ ret = 0;
+out:
+ if (ret < 0) {
+ if (transaction_frame)
+ AFR_STACK_DESTROY (transaction_frame);
+ AFR_STACK_UNWIND (discard, frame, -1, op_errno, NULL, NULL, NULL);
+ }
+
+ return 0;
+}
+
+
+/* {{{ zerofill */
+
+static int
+afr_zerofill_unwind (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+ call_frame_t *main_frame = NULL;
+
+ local = frame->local;
+
+ LOCK (&frame->lock);
+ {
+ if (local->transaction.main_frame) {
+ main_frame = local->transaction.main_frame;
+ }
+ local->transaction.main_frame = NULL;
+ }
+ UNLOCK (&frame->lock);
+
+ if (main_frame) {
+ AFR_STACK_UNWIND (zerofill, main_frame, local->op_ret,
+ local->op_errno,
+ &local->cont.zerofill.prebuf,
+ &local->cont.zerofill.postbuf,
+ NULL);
+ }
+ return 0;
+}
+
+static int
+afr_zerofill_wind_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)
+{
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int child_index = (long) cookie;
+ int call_count = -1;
+ int need_unwind = 0;
+ int read_child = 0;
+
+ local = frame->local;
+ priv = this->private;
+
+ read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
+
+ LOCK (&frame->lock);
+ {
+ if (child_index == read_child) {
+ local->read_child_returned = _gf_true;
+ }
+
+ if (afr_fop_failed (op_ret, op_errno)) {
+ afr_transaction_fop_failed (frame, this, child_index);
+ }
+
+ if (op_ret != -1) {
+ if (local->success_count == 0) {
+ local->op_ret = op_ret;
+ local->cont.zerofill.prebuf = *prebuf;
+ local->cont.zerofill.postbuf = *postbuf;
+ }
+
+ if (child_index == read_child) {
+ local->cont.zerofill.prebuf = *prebuf;
+ local->cont.zerofill.postbuf = *postbuf;
+ }
+
+ local->success_count++;
+
+ if ((local->success_count >= priv->wait_count)
+ && local->read_child_returned) {
+ need_unwind = 1;
+ }
+ }
+ local->op_errno = op_errno;
+ }
+ UNLOCK (&frame->lock);
+
+ if (need_unwind) {
+ local->transaction.unwind (frame, this);
+ }
+ call_count = afr_frame_return (frame);
+
+ if (call_count == 0) {
+ local->transaction.resume (frame, this);
+ }
+
+ return 0;
+}
+
+static int
+afr_zerofill_wind (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int call_count = -1;
+ int i = 0;
+
+ local = frame->local;
+ priv = this->private;
+
+ call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
+ priv->child_count);
+
+ if (call_count == 0) {
+ local->transaction.resume (frame, this);
+ return 0;
+ }
+
+ local->call_count = call_count;
+
+ for (i = 0; i < priv->child_count; i++) {
+ if (local->transaction.pre_op[i]) {
+ STACK_WIND_COOKIE (frame, afr_zerofill_wind_cbk,
+ (void *) (long) i,
+ priv->children[i],
+ priv->children[i]->fops->zerofill,
+ local->fd,
+ local->cont.zerofill.offset,
+ local->cont.zerofill.len,
+ NULL);
+
+ if (!--call_count)
+ break;
+ }
+ }
+
+ return 0;
+}
+
+static int
+afr_zerofill_done (call_frame_t *frame, xlator_t *this)
+{
+ afr_local_t *local = NULL;
+
+ local = frame->local;
+
+ local->transaction.unwind (frame, this);
+
+ AFR_STACK_DESTROY (frame);
+
+ return 0;
+}
+
+static int
+afr_do_zerofill(call_frame_t *frame, xlator_t *this)
+{
+ call_frame_t *transaction_frame = NULL;
+ afr_local_t *local = NULL;
+ int op_ret = -1;
+ int op_errno = 0;
+
+ local = frame->local;
+
+ transaction_frame = copy_frame (frame);
+ if (!transaction_frame) {
+ goto out;
+ }
+
+ transaction_frame->local = local;
+ frame->local = NULL;
+
+ local->op = GF_FOP_ZEROFILL;
+
+ local->transaction.fop = afr_zerofill_wind;
+ local->transaction.done = afr_zerofill_done;
+ local->transaction.unwind = afr_zerofill_unwind;
+
+ local->transaction.main_frame = frame;
+
+ local->transaction.start = local->cont.zerofill.offset;
+ local->transaction.len = 0;
+
+ op_ret = afr_transaction (transaction_frame, this,
+ AFR_DATA_TRANSACTION);
+ if (op_ret < 0) {
+ op_errno = -op_ret;
+ goto out;
+ }
+
+ op_ret = 0;
+out:
+ if (op_ret < 0) {
+ if (transaction_frame) {
+ AFR_STACK_DESTROY (transaction_frame);
+ }
+ AFR_STACK_UNWIND (zerofill, frame, op_ret, op_errno, NULL,
+ NULL, NULL);
+ }
+
+ return 0;
+}
+
+int
+afr_zerofill (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
+ size_t len, dict_t *xdata)
+{
+ afr_private_t *priv = NULL;
+ afr_local_t *local = NULL;
+ call_frame_t *transaction_frame = NULL;
+ int ret = -1;
+ int op_errno = 0;
+
+ VALIDATE_OR_GOTO (frame, out);
+ VALIDATE_OR_GOTO (this, out);
+ VALIDATE_OR_GOTO (this->private, out);
+
+ priv = this->private;
+
+ if (afr_is_split_brain (this, fd->inode)) {
+ op_errno = EIO;
+ goto out;
+ }
+ QUORUM_CHECK(zerofill, out);
+
+ AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
+ local = frame->local;
+
+ ret = afr_local_init (local, priv, &op_errno);
+ if (ret < 0) {
+ goto out;
+ }
+ local->cont.zerofill.offset = offset;
+ local->cont.zerofill.len = len;
+
+ local->fd = fd_ref (fd);
+
+ afr_open_fd_fix (fd, this);
+
+ afr_do_zerofill(frame, this);
+
+ ret = 0;
+out:
+ if (ret < 0) {
+ if (transaction_frame) {
+ AFR_STACK_DESTROY (transaction_frame);
+ }
+ AFR_STACK_UNWIND (zerofill, frame, -1, op_errno, NULL,
+ NULL, NULL);
+ }
+
+ return 0;
+}
+
+/* }}} */
+
+
generated by cgit (git 2.34.1) at 2026-01-05 20:42:30 +0000