summaryrefslogtreecommitdiffstats
path: root/xlators/cluster/afr/src/afr-self-heal-data.c
diff options
context:
space:
mode:
Diffstat (limited to 'xlators/cluster/afr/src/afr-self-heal-data.c')
-rw-r--r--xlators/cluster/afr/src/afr-self-heal-data.c2022
1 files changed, 742 insertions, 1280 deletions
diff --git a/xlators/cluster/afr/src/afr-self-heal-data.c b/xlators/cluster/afr/src/afr-self-heal-data.c
index cfc00a3728e..37bcc2b3f9e 100644
--- a/xlators/cluster/afr/src/afr-self-heal-data.c
+++ b/xlators/cluster/afr/src/afr-self-heal-data.c
@@ -1,5 +1,5 @@
/*
- Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com>
+ Copyright (c) 2013 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
@@ -8,1422 +8,884 @@
cases as published by the Free Software Foundation.
*/
-#include <libgen.h>
-#include <unistd.h>
-#include <fnmatch.h>
-#include <sys/time.h>
-#include <stdlib.h>
-#include <signal.h>
-
-#ifndef _CONFIG_H
-#define _CONFIG_H
-#include "config.h"
-#endif
-
-#include "glusterfs.h"
#include "afr.h"
-#include "dict.h"
-#include "xlator.h"
-#include "hashfn.h"
-#include "logging.h"
-#include "stack.h"
-#include "list.h"
-#include "call-stub.h"
-#include "defaults.h"
-#include "common-utils.h"
-#include "compat-errno.h"
-#include "compat.h"
-#include "byte-order.h"
-
-#include "afr-transaction.h"
#include "afr-self-heal.h"
-#include "afr-self-heal-common.h"
-#include "afr-self-heal-algorithm.h"
+#include <glusterfs/byte-order.h>
+#include "protocol-common.h"
+#include "afr-messages.h"
+#include <glusterfs/events.h>
-int
-afr_sh_data_fail (call_frame_t *frame, xlator_t *this);
+#define HAS_HOLES(i) ((i->ia_blocks * 512) < (i->ia_size))
+static int
+__checksum_cbk(call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
+ int op_errno, uint32_t weak, uint8_t *strong, dict_t *xdata)
+{
+ afr_local_t *local = NULL;
+ struct afr_reply *replies = NULL;
+ int i = (long)cookie;
+
+ local = frame->local;
+ replies = local->replies;
+
+ replies[i].valid = 1;
+ replies[i].op_ret = op_ret;
+ replies[i].op_errno = op_errno;
+ if (xdata) {
+ replies[i].buf_has_zeroes = dict_get_str_boolean(
+ xdata, "buf-has-zeroes", _gf_false);
+ replies[i].fips_mode_rchecksum = dict_get_str_boolean(
+ xdata, "fips-mode-rchecksum", _gf_false);
+ }
+ if (strong) {
+ if (replies[i].fips_mode_rchecksum) {
+ memcpy(local->replies[i].checksum, strong, SHA256_DIGEST_LENGTH);
+ } else {
+ memcpy(local->replies[i].checksum, strong, MD5_DIGEST_LENGTH);
+ }
+ }
+
+ syncbarrier_wake(&local->barrier);
+ return 0;
+}
+
+static gf_boolean_t
+__afr_can_skip_data_block_heal(call_frame_t *frame, xlator_t *this, fd_t *fd,
+ int source, unsigned char *healed_sinks,
+ off_t offset, size_t size, struct iatt *poststat)
+{
+ afr_private_t *priv = NULL;
+ afr_local_t *local = NULL;
+ unsigned char *wind_subvols = NULL;
+ gf_boolean_t checksum_match = _gf_true;
+ struct afr_reply *replies = NULL;
+ dict_t *xdata = NULL;
+ int i = 0;
+
+ priv = this->private;
+ local = frame->local;
+ replies = local->replies;
+
+ xdata = dict_new();
+ if (!xdata)
+ goto out;
+ if (dict_set_int32_sizen(xdata, "check-zero-filled", 1)) {
+ dict_unref(xdata);
+ goto out;
+ }
+
+ wind_subvols = alloca0(priv->child_count);
+ for (i = 0; i < priv->child_count; i++) {
+ if (i == source || healed_sinks[i])
+ wind_subvols[i] = 1;
+ }
+
+ AFR_ONLIST(wind_subvols, frame, __checksum_cbk, rchecksum, fd, offset, size,
+ xdata);
+ if (xdata)
+ dict_unref(xdata);
+
+ if (!replies[source].valid || replies[source].op_ret != 0)
+ return _gf_false;
-static inline gf_boolean_t
-afr_sh_data_proceed (unsigned int success_count)
-{
- return (success_count >= AFR_SH_MIN_PARTICIPANTS);
+ for (i = 0; i < priv->child_count; i++) {
+ if (i == source)
+ continue;
+ if (replies[i].valid) {
+ if (memcmp(replies[source].checksum, replies[i].checksum,
+ replies[source].fips_mode_rchecksum
+ ? SHA256_DIGEST_LENGTH
+ : MD5_DIGEST_LENGTH)) {
+ checksum_match = _gf_false;
+ break;
+ }
+ }
+ }
+
+ if (checksum_match) {
+ if (HAS_HOLES(poststat))
+ return _gf_true;
+
+ /* For non-sparse files, we might be better off writing the
+ * zeroes to sinks to avoid mismatch of disk-usage in bricks. */
+ if (local->replies[source].buf_has_zeroes)
+ return _gf_false;
+ else
+ return _gf_true;
+ }
+out:
+ return _gf_false;
+}
+
+static gf_boolean_t
+__afr_is_sink_zero_filled(xlator_t *this, fd_t *fd, size_t size, off_t offset,
+ int sink)
+{
+ afr_private_t *priv = NULL;
+ struct iobref *iobref = NULL;
+ struct iovec *iovec = NULL;
+ int count = 0;
+ int ret = 0;
+ gf_boolean_t zero_filled = _gf_false;
+
+ priv = this->private;
+ ret = syncop_readv(priv->children[sink], fd, size, offset, 0, &iovec,
+ &count, &iobref, NULL, NULL, NULL);
+ if (ret < 0)
+ goto out;
+ ret = iov_0filled(iovec, count);
+ if (!ret)
+ zero_filled = _gf_true;
+out:
+ if (iovec)
+ GF_FREE(iovec);
+ if (iobref)
+ iobref_unref(iobref);
+ return zero_filled;
}
-extern int
-sh_loop_finish (call_frame_t *loop_frame, xlator_t *this);
-
-int
-afr_post_sh_big_lock_success (call_frame_t *frame, xlator_t *this);
-
-int
-afr_post_sh_big_lock_failure (call_frame_t *frame, xlator_t *this);
-
-int
-afr_sh_data_finish (call_frame_t *frame, xlator_t *this);
-
-int
-afr_sh_data_done (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
+static int
+__afr_selfheal_data_read_write(call_frame_t *frame, xlator_t *this, fd_t *fd,
+ int source, unsigned char *healed_sinks,
+ off_t offset, size_t size,
+ struct afr_reply *replies, int type)
+{
+ struct iovec *iovec = NULL;
+ int count = 0;
+ struct iobref *iobref = NULL;
+ int ret = 0;
+ int i = 0;
+ afr_private_t *priv = NULL;
+
+ priv = this->private;
+
+ ret = syncop_readv(priv->children[source], fd, size, offset, 0, &iovec,
+ &count, &iobref, NULL, NULL, NULL);
+ if (ret <= 0)
+ return ret;
- local = frame->local;
- sh = &local->self_heal;
+ for (i = 0; i < priv->child_count; i++) {
+ if (!healed_sinks[i])
+ continue;
+
+ /*
+ * TODO: Use fiemap() and discard() to heal holes
+ * in the future.
+ *
+ * For now,
+ *
+ * - if the source had any holes at all,
+ * AND
+ * - if we are writing past the original file size
+ * of the sink
+ * AND
+ * - is NOT the last block of the source file. if
+ * the block contains EOF, it has to be written
+ * in order to set the file size even if the
+ * last block is 0-filled.
+ * AND
+ * - if the read buffer is filled with only 0's
+ *
+ * then, skip writing to this source. We don't depend
+ * on the write to happen to update the size as we
+ * have performed an ftruncate() upfront anyways.
+ */
+#define is_last_block(o, b, s) ((s >= o) && (s <= (o + b)))
+ if (HAS_HOLES((&replies[source].poststat)) &&
+ offset >= replies[i].poststat.ia_size &&
+ !is_last_block(offset, size, replies[source].poststat.ia_size) &&
+ (iov_0filled(iovec, count) == 0))
+ continue;
+
+ /* Avoid filling up sparse regions of the sink with 0-filled
+ * writes.*/
+ if (type == AFR_SELFHEAL_DATA_FULL &&
+ HAS_HOLES((&replies[source].poststat)) &&
+ ((offset + size) <= replies[i].poststat.ia_size) &&
+ (iov_0filled(iovec, count) == 0) &&
+ __afr_is_sink_zero_filled(this, fd, size, offset, i)) {
+ continue;
+ }
+
+ ret = syncop_writev(priv->children[i], fd, iovec, count, offset, iobref,
+ 0, NULL, NULL, NULL, NULL);
+ if (ret != iov_length(iovec, count)) {
+ /* write() failed on this sink. unset the corresponding
+ member in sinks[] (which is healed_sinks[] in the
+ caller) so that this server does NOT get considered
+ as successfully healed.
+ */
+ healed_sinks[i] = 0;
+ }
+ }
+ if (iovec)
+ GF_FREE(iovec);
+ if (iobref)
+ iobref_unref(iobref);
+
+ return ret;
+}
+
+static gf_boolean_t
+afr_source_sinks_locked(xlator_t *this, unsigned char *locked_on, int source,
+ unsigned char *healed_sinks)
+{
+ afr_private_t *priv = this->private;
+ int i = 0;
+
+ if (!locked_on[source])
+ return _gf_false;
- sh->completion_cbk (frame, this);
+ for (i = 0; i < priv->child_count; i++) {
+ if (healed_sinks[i] && locked_on[i])
+ return _gf_true;
+ }
- return 0;
+ return _gf_false;
}
-
-int
-afr_sh_data_flush_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
- int32_t op_ret, int32_t op_errno, dict_t *xdata)
+static int
+afr_selfheal_data_block(call_frame_t *frame, xlator_t *this, fd_t *fd,
+ int source, unsigned char *healed_sinks, off_t offset,
+ size_t size, int type, struct afr_reply *replies)
{
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- int call_count = 0;
- int child_index = (long) cookie;
-
- local = frame->local;
- priv = this->private;
-
- LOCK (&frame->lock);
- {
- if (op_ret == -1) {
- gf_log (this->name, GF_LOG_ERROR,
- "flush failed on %s on subvolume %s: %s",
- local->loc.path, priv->children[child_index]->name,
- strerror (op_errno));
- }
- }
- UNLOCK (&frame->lock);
-
- call_count = afr_frame_return (frame);
-
- if (call_count == 0) {
- afr_sh_data_done (frame, this);
- }
+ int ret = -1;
+ afr_private_t *priv = NULL;
+ unsigned char *data_lock = NULL;
- return 0;
-}
-
-int
-afr_sh_data_close (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- afr_self_heal_t *sh = NULL;
- int i = 0;
- int call_count = 0;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- if (!sh->healing_fd) {
- //This happens when file is non-reg
- afr_sh_data_done (frame, this);
- return 0;
- }
- call_count = afr_set_elem_count_get (sh->success,
- priv->child_count);
- local->call_count = call_count;
+ priv = this->private;
+ data_lock = alloca0(priv->child_count);
- if (call_count == 0) {
- afr_sh_data_done (frame, this);
- return 0;
+ ret = afr_selfheal_inodelk(frame, this, fd->inode, this->name, offset, size,
+ data_lock);
+ {
+ if (!afr_source_sinks_locked(this, data_lock, source, healed_sinks)) {
+ ret = -ENOTCONN;
+ goto unlock;
}
- for (i = 0; i < priv->child_count; i++) {
- if (!sh->success[i])
- continue;
- gf_log (this->name, GF_LOG_DEBUG,
- "closing fd of %s on %s",
- local->loc.path, priv->children[i]->name);
-
- STACK_WIND_COOKIE (frame, afr_sh_data_flush_cbk,
- (void *) (long) i,
- priv->children[i],
- priv->children[i]->fops->flush,
- sh->healing_fd, NULL);
-
- if (!--call_count)
- break;
+ if (type == AFR_SELFHEAL_DATA_DIFF &&
+ __afr_can_skip_data_block_heal(frame, this, fd, source,
+ healed_sinks, offset, size,
+ &replies[source].poststat)) {
+ ret = 0;
+ goto unlock;
}
- return 0;
+ ret = __afr_selfheal_data_read_write(
+ frame, this, fd, source, healed_sinks, offset, size, replies, type);
+ }
+unlock:
+ afr_selfheal_uninodelk(frame, this, fd->inode, this->name, offset, size,
+ data_lock);
+ return ret;
}
-int
-afr_sh_data_setattr_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
- int32_t op_ret, int32_t op_errno, struct iatt *statpre,
- struct iatt *statpost, dict_t *xdata)
+static int
+afr_selfheal_data_fsync(call_frame_t *frame, xlator_t *this, fd_t *fd,
+ unsigned char *healed_sinks)
{
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int i = 0;
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- int call_count = 0;
- int child_index = (long) cookie;
-
- local = frame->local;
- priv = this->private;
-
- LOCK (&frame->lock);
- {
- if (op_ret == -1) {
- gf_log (this->name, GF_LOG_INFO,
- "setattr failed on %s on subvolume %s: %s",
- local->loc.path, priv->children[child_index]->name,
- strerror (op_errno));
- }
- }
- UNLOCK (&frame->lock);
+ local = frame->local;
+ priv = this->private;
- call_count = afr_frame_return (frame);
+ if (!priv->ensure_durability)
+ return 0;
- if (call_count == 0) {
- afr_sh_data_finish (frame, this);
- }
+ AFR_ONLIST(healed_sinks, frame, afr_sh_generic_fop_cbk, fsync, fd, 0, NULL);
- return 0;
+ for (i = 0; i < priv->child_count; i++)
+ if (healed_sinks[i] && local->replies[i].op_ret != 0)
+ /* fsync() failed. Do NOT consider this server
+ as successfully healed. Mark it so.
+ */
+ healed_sinks[i] = 0;
+ return 0;
}
-int
-afr_sh_data_setattr (call_frame_t *frame, xlator_t *this, struct iatt* stbuf)
+static int
+afr_data_self_heal_type_get(afr_private_t *priv, unsigned char *healed_sinks,
+ int source, struct afr_reply *replies)
{
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- afr_self_heal_t *sh = NULL;
- int i = 0;
- int call_count = 0;
- int32_t valid = 0;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- valid = (GF_SET_ATTR_ATIME | GF_SET_ATTR_MTIME);
-
- call_count = afr_set_elem_count_get (sh->success,
- priv->child_count);
- local->call_count = call_count;
-
- if (call_count == 0) {
- GF_ASSERT (0);
- afr_sh_data_finish (frame, this);
- return 0;
- }
+ int type = AFR_SELFHEAL_DATA_FULL;
+ int i = 0;
+ if (priv->data_self_heal_algorithm == AFR_SELFHEAL_DATA_DYNAMIC) {
+ type = AFR_SELFHEAL_DATA_FULL;
for (i = 0; i < priv->child_count; i++) {
- if (!sh->success[i])
- continue;
-
- STACK_WIND_COOKIE (frame, afr_sh_data_setattr_cbk,
- (void *) (long) i,
- priv->children[i],
- priv->children[i]->fops->setattr,
- &local->loc, stbuf, valid, NULL);
-
- if (!--call_count)
- break;
+ if (!healed_sinks[i] && i != source)
+ continue;
+ if (replies[i].poststat.ia_size) {
+ type = AFR_SELFHEAL_DATA_DIFF;
+ break;
+ }
}
-
- return 0;
+ } else {
+ type = priv->data_self_heal_algorithm;
+ }
+ return type;
}
-int
-afr_sh_data_setattr_fstat_cbk (call_frame_t *frame, void *cookie,
- xlator_t *this, int32_t op_ret, int32_t op_errno,
- struct iatt *buf, dict_t *xdata)
+static int
+afr_selfheal_data_do(call_frame_t *frame, xlator_t *this, fd_t *fd, int source,
+ unsigned char *healed_sinks, struct afr_reply *replies)
{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- int child_index = (long) cookie;
-
- local = frame->local;
- sh = &local->self_heal;
+ afr_private_t *priv = NULL;
+ off_t off = 0;
+ size_t block = 0;
+ int type = AFR_SELFHEAL_DATA_FULL;
+ int ret = -1;
+ call_frame_t *iter_frame = NULL;
+ unsigned char arbiter_sink_status = 0;
- GF_ASSERT (sh->source == child_index);
- if (op_ret != -1) {
- sh->buf[child_index] = *buf;
- afr_sh_data_setattr (frame, this, buf);
- } else {
- gf_log (this->name, GF_LOG_ERROR, "%s: Failed to set "
- "time-stamps after self-heal", local->loc.path);
- afr_sh_data_fail (frame, this);
- }
+ gf_msg(this->name, GF_LOG_INFO, 0, AFR_MSG_SELF_HEAL_INFO,
+ "performing data selfheal on %s", uuid_utoa(fd->inode->gfid));
- return 0;
-}
-
-/*
- * If there are any writes after the self-heal is triggered then the
- * stbuf stored in local->self_heal.buf[] will be invalid so we do one more
- * stat on the source and then set the [am]times
- */
-int
-afr_sh_set_timestamps (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- afr_self_heal_t *sh = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- STACK_WIND_COOKIE (frame, afr_sh_data_setattr_fstat_cbk,
- (void *) (long) sh->source,
- priv->children[sh->source],
- priv->children[sh->source]->fops->fstat,
- sh->healing_fd, NULL);
- return 0;
-}
-
-//Fun fact, lock_cbk is being used for both lock & unlock
-int
-afr_sh_data_unlock (call_frame_t *frame, xlator_t *this,
- afr_lock_cbk_t lock_cbk)
-{
- afr_local_t *local = NULL;
- afr_internal_lock_t *int_lock = NULL;
- afr_self_heal_t *sh = NULL;
+ priv = this->private;
+ if (priv->arbiter_count) {
+ arbiter_sink_status = healed_sinks[ARBITER_BRICK_INDEX];
+ healed_sinks[ARBITER_BRICK_INDEX] = 0;
+ }
- local = frame->local;
- int_lock = &local->internal_lock;
- sh = &local->self_heal;
+ block = 128 * 1024 * priv->data_self_heal_window_size;
- GF_ASSERT (sh->data_lock_held);
+ type = afr_data_self_heal_type_get(priv, healed_sinks, source, replies);
- sh->data_lock_held = _gf_false;
- int_lock->lock_cbk = lock_cbk;
- afr_unlock (frame, this);
+ iter_frame = afr_copy_frame(frame);
+ if (!iter_frame) {
+ ret = -ENOMEM;
+ goto out;
+ }
- return 0;
-}
+ for (off = 0; off < replies[source].poststat.ia_size; off += block) {
+ if (AFR_COUNT(healed_sinks, priv->child_count) == 0) {
+ ret = -ENOTCONN;
+ goto out;
+ }
-int
-afr_sh_data_finish (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
+ ret = afr_selfheal_data_block(iter_frame, this, fd, source,
+ healed_sinks, off, block, type, replies);
+ if (ret < 0)
+ goto out;
- local = frame->local;
- sh = &local->self_heal;
+ AFR_STACK_RESET(iter_frame);
+ if (iter_frame->local == NULL) {
+ ret = -ENOTCONN;
+ goto out;
+ }
+ }
- gf_log (this->name, GF_LOG_DEBUG,
- "finishing data selfheal of %s", local->loc.path);
+ ret = afr_selfheal_data_fsync(frame, this, fd, healed_sinks);
- if (sh->data_lock_held)
- afr_sh_data_unlock (frame, this, afr_sh_data_close);
- else
- afr_sh_data_close (frame, this);
+out:
+ if (arbiter_sink_status)
+ healed_sinks[ARBITER_BRICK_INDEX] = arbiter_sink_status;
- return 0;
+ if (iter_frame)
+ AFR_STACK_DESTROY(iter_frame);
+ return ret;
}
-int
-afr_sh_data_fail (call_frame_t *frame, xlator_t *this)
+static int
+__afr_selfheal_truncate_sinks(call_frame_t *frame, xlator_t *this, fd_t *fd,
+ unsigned char *healed_sinks, uint64_t size)
{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
-
- gf_log (this->name, GF_LOG_DEBUG,
- "finishing failed data selfheal of %s", local->loc.path);
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int i = 0;
- sh->op_failed = 1;
- if (sh->data_lock_held)
- afr_sh_data_unlock (frame, this, afr_sh_data_close);
- else
- afr_sh_data_close (frame, this);
- return 0;
-}
+ local = frame->local;
+ priv = this->private;
-int
-afr_sh_data_erase_pending_cbk (call_frame_t *frame, void *cookie,
- xlator_t *this, int32_t op_ret,
- int32_t op_errno, dict_t *xattr, dict_t *xdata)
-{
- int call_count = 0;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- afr_private_t *priv = NULL;
- int32_t child_index = (long) cookie;
-
- priv = this->private;
- local = frame->local;
- sh = &local->self_heal;
- if (op_ret < 0) {
- gf_log (this->name, GF_LOG_ERROR, "Erasing of pending change "
- "log failed on %s for subvol %s, reason: %s",
- local->loc.path, priv->children[child_index]->name,
- strerror (op_errno));
- sh->op_failed = 1;
- }
+ /* This will send truncate on the arbiter brick as well if it is marked as
+ * sink. If changelog is enabled on the volume it captures truncate as a
+ * data transactions on the arbiter brick. This will help geo-rep to
+ * properly sync the data from master to slave if arbiter is the ACTIVE
+ * brick during syncing and which had got some entries healed for data as
+ * part of self heal.
+ */
+ AFR_ONLIST(healed_sinks, frame, afr_sh_generic_fop_cbk, ftruncate, fd, size,
+ NULL);
- call_count = afr_frame_return (frame);
-
- if (call_count == 0) {
- if (sh->op_failed) {
- if (sh->old_loop_frame)
- sh_loop_finish (sh->old_loop_frame, this);
- sh->old_loop_frame = NULL;
- afr_sh_data_fail (frame, this);
- goto out;
- }
- if (!IA_ISREG (sh->type)) {
- afr_sh_data_finish (frame, this);
- goto out;
- }
- GF_ASSERT (sh->old_loop_frame);
- afr_sh_data_lock (frame, this, 0, 0, _gf_true,
- afr_post_sh_big_lock_success,
- afr_post_sh_big_lock_failure);
- }
-out:
- return 0;
-}
+ for (i = 0; i < priv->child_count; i++)
+ if (healed_sinks[i] && local->replies[i].op_ret == -1)
+ /* truncate() failed. Do NOT consider this server
+ as successfully healed. Mark it so.
+ */
+ healed_sinks[i] = 0;
-int
-afr_sh_data_erase_pending (call_frame_t *frame, xlator_t *this)
-{
- afr_sh_erase_pending (frame, this, AFR_DATA_TRANSACTION,
- afr_sh_data_erase_pending_cbk,
- afr_sh_data_finish);
- return 0;
+ return 0;
}
-
-static struct afr_sh_algorithm *
-sh_algo_from_name (xlator_t *this, char *name)
+gf_boolean_t
+afr_has_source_witnesses(xlator_t *this, unsigned char *sources,
+ uint64_t *witness)
{
- int i = 0;
-
- if (name == NULL)
- goto out;
-
- while (afr_self_heal_algorithms[i].name) {
- if (!strcmp (name, afr_self_heal_algorithms[i].name)) {
- return &afr_self_heal_algorithms[i];
- }
+ int i = 0;
+ afr_private_t *priv = NULL;
- i++;
- }
+ priv = this->private;
-out:
- return NULL;
+ for (i = 0; i < priv->child_count; i++) {
+ if (sources[i] && witness[i])
+ return _gf_true;
+ }
+ return _gf_false;
}
-
-static int
-sh_zero_byte_files_exist (afr_local_t *local, int child_count)
+static gf_boolean_t
+afr_does_size_mismatch(xlator_t *this, unsigned char *sources,
+ struct afr_reply *replies)
{
- int i = 0;
- int ret = 0;
- afr_self_heal_t *sh = NULL;
-
- sh = &local->self_heal;
- for (i = 0; i < child_count; i++) {
- if (!local->child_up[i] || sh->child_errno[i])
- continue;
- if (sh->buf[i].ia_size == 0) {
- ret = 1;
- break;
- }
- }
+ int i = 0;
+ afr_private_t *priv = NULL;
+ struct iatt *min = NULL;
+ struct iatt *max = NULL;
- return ret;
-}
+ priv = this->private;
+ for (i = 0; i < priv->child_count; i++) {
+ if (!replies[i].valid)
+ continue;
-struct afr_sh_algorithm *
-afr_sh_data_pick_algo (call_frame_t *frame, xlator_t *this)
-{
- afr_private_t * priv = NULL;
- struct afr_sh_algorithm * algo = NULL;
- afr_local_t * local = NULL;
- afr_self_heal_t * sh = NULL;
-
- priv = this->private;
- local = frame->local;
- sh = &local->self_heal;
- algo = sh_algo_from_name (this, priv->data_self_heal_algorithm);
-
- if (algo == NULL) {
- /* option not set, so fall back on heuristics */
-
- if (sh_zero_byte_files_exist (local, priv->child_count)
- || (sh->file_size <= (priv->data_self_heal_window_size *
- this->ctx->page_size))) {
-
- /*
- * If the file does not exist on one of the subvolumes,
- * or a zero-byte file exists (created by entry self-heal)
- * the entire content has to be copied anyway, so there
- * is no benefit from using the "diff" algorithm.
- *
- * If the file size is about the same as page size,
- * the entire file can be read and written with a few
- * (pipelined) STACK_WINDs, which will be faster
- * than "diff" which has to read checksums and then
- * read and write.
- */
-
- algo = sh_algo_from_name (this, "full");
-
- } else {
- algo = sh_algo_from_name (this, "diff");
- }
- }
+ if (replies[i].op_ret < 0)
+ continue;
- return algo;
-}
+ if (!sources[i])
+ continue;
+ if (AFR_IS_ARBITER_BRICK(priv, i) && (replies[i].poststat.ia_size == 0))
+ continue;
-int
-afr_sh_data_sync_prepare (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- struct afr_sh_algorithm *sh_algo = NULL;
+ if (!min)
+ min = &replies[i].poststat;
- local = frame->local;
- sh = &local->self_heal;
+ if (!max)
+ max = &replies[i].poststat;
- sh->algo_completion_cbk = afr_sh_data_erase_pending;
- sh->algo_abort_cbk = afr_sh_data_fail;
+ if (min->ia_size > replies[i].poststat.ia_size)
+ min = &replies[i].poststat;
- sh_algo = afr_sh_data_pick_algo (frame, this);
+ if (max->ia_size < replies[i].poststat.ia_size)
+ max = &replies[i].poststat;
+ }
- sh->algo = sh_algo;
- sh_algo->fn (frame, this);
+ if (min && max) {
+ if (min->ia_size != max->ia_size)
+ return _gf_true;
+ }
- return 0;
+ return _gf_false;
}
-int
-afr_sh_data_trim_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)
+static void
+afr_mark_biggest_witness_as_source(xlator_t *this, unsigned char *sources,
+ uint64_t *witness)
{
- int call_count = 0;
- int child_index = 0;
- afr_private_t *priv = NULL;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
-
- priv = this->private;
- local = frame->local;
- sh = &local->self_heal;
-
- child_index = (long) cookie;
-
- LOCK (&frame->lock);
- {
- if (op_ret == -1) {
- gf_log (this->name, GF_LOG_ERROR,
- "ftruncate of %s on subvolume %s failed (%s)",
- local->loc.path,
- priv->children[child_index]->name,
- strerror (op_errno));
- sh->op_failed = 1;
- } else {
- gf_log (this->name, GF_LOG_DEBUG,
- "ftruncate of %s on subvolume %s completed",
- local->loc.path,
- priv->children[child_index]->name);
- }
- }
- UNLOCK (&frame->lock);
+ int i = 0;
+ afr_private_t *priv = NULL;
+ uint64_t biggest_witness = 0;
- call_count = afr_frame_return (frame);
+ priv = this->private;
+ /* Find source with biggest witness count */
+ for (i = 0; i < priv->child_count; i++) {
+ if (!sources[i])
+ continue;
+ if (biggest_witness < witness[i])
+ biggest_witness = witness[i];
+ }
- if (call_count == 0) {
- if (sh->op_failed)
- afr_sh_data_fail (frame, this);
- else
- afr_sh_data_sync_prepare (frame, this);
- }
+ /* Mark files with less witness count as not source */
+ for (i = 0; i < priv->child_count; i++) {
+ if (!sources[i])
+ continue;
+ if (witness[i] < biggest_witness)
+ sources[i] = 0;
+ }
- return 0;
+ return;
}
-
-int
-afr_sh_data_trim_sinks (call_frame_t *frame, xlator_t *this)
+/* This is a tie breaker function. Only one source be assigned here */
+static void
+afr_mark_newest_file_as_source(xlator_t *this, unsigned char *sources,
+ struct afr_reply *replies)
{
- afr_private_t * priv = NULL;
- afr_local_t * local = NULL;
- afr_self_heal_t *sh = NULL;
- int *sources = NULL;
- int call_count = 0;
- int i = 0;
-
-
- priv = this->private;
- local = frame->local;
- sh = &local->self_heal;
+ int i = 0;
+ afr_private_t *priv = NULL;
+ int source = -1;
+ uint32_t max_ctime = 0;
- sources = sh->sources;
- call_count = sh->active_sinks;
+ priv = this->private;
+ /* Find source with latest ctime */
+ for (i = 0; i < priv->child_count; i++) {
+ if (!sources[i])
+ continue;
- local->call_count = call_count;
-
- for (i = 0; i < priv->child_count; i++) {
- if (sources[i] || !local->child_up[i])
- continue;
-
- STACK_WIND_COOKIE (frame, afr_sh_data_trim_cbk,
- (void *) (long) i,
- priv->children[i],
- priv->children[i]->fops->ftruncate,
- sh->healing_fd, sh->file_size,
- NULL);
-
- if (!--call_count)
- break;
+ if (max_ctime <= replies[i].poststat.ia_ctime) {
+ source = i;
+ max_ctime = replies[i].poststat.ia_ctime;
}
+ }
- return 0;
+ /* Only mark one of the files as source to break ties */
+ memset(sources, 0, sizeof(*sources) * priv->child_count);
+ sources[source] = 1;
}
-int
-afr_sh_inode_set_read_ctx (afr_self_heal_t *sh, xlator_t *this)
-{
- afr_private_t *priv = NULL;
- int ret = 0;
- int i = 0;
-
- priv = this->private;
- sh->source = afr_sh_select_source (sh->sources, priv->child_count);
- if (sh->source < 0) {
- ret = -1;
- goto out;
- }
-
- /* detect changes not visible through pending flags -- JIC */
- for (i = 0; i < priv->child_count; i++) {
- if (i == sh->source || sh->child_errno[i])
- continue;
-
- if (SIZE_DIFFERS (&sh->buf[i], &sh->buf[sh->source]))
- sh->sources[i] = 0;
- }
+static int
+__afr_selfheal_data_finalize_source(
+ call_frame_t *frame, xlator_t *this, inode_t *inode, unsigned char *sources,
+ unsigned char *sinks, unsigned char *healed_sinks, unsigned char *locked_on,
+ unsigned char *undid_pending, struct afr_reply *replies, uint64_t *witness)
+{
+ afr_private_t *priv = NULL;
+ int source = -1;
+ int sources_count = 0;
+ priv = this->private;
+
+ sources_count = AFR_COUNT(sources, priv->child_count);
+
+ if ((AFR_CMP(locked_on, healed_sinks, priv->child_count) == 0) ||
+ !sources_count) {
+ /* split brain */
+ source = afr_mark_split_brain_source_sinks(
+ frame, this, inode, sources, sinks, healed_sinks, locked_on,
+ replies, AFR_DATA_TRANSACTION);
+ if (source < 0) {
+ gf_event(EVENT_AFR_SPLIT_BRAIN,
+ "client-pid=%d;"
+ "subvol=%s;type=data;"
+ "file=%s",
+ this->ctx->cmd_args.client_pid, this->name,
+ uuid_utoa(inode->gfid));
+ return -EIO;
+ }
+
+ _afr_fav_child_reset_sink_xattrs(
+ frame, this, inode, source, healed_sinks, undid_pending,
+ AFR_DATA_TRANSACTION, locked_on, replies);
+ goto out;
+ }
+
+ /* No split brain at this point. If we were called from
+ * afr_heal_splitbrain_file(), abort.*/
+ if (afr_dict_contains_heal_op(frame))
+ return -EIO;
+
+ /* If there are no witnesses/size-mismatches on sources we are done*/
+ if (!afr_does_size_mismatch(this, sources, replies) &&
+ !afr_has_source_witnesses(this, sources, witness))
+ goto out;
+
+ afr_mark_largest_file_as_source(this, sources, replies);
+ afr_mark_biggest_witness_as_source(this, sources, witness);
+ afr_mark_newest_file_as_source(this, sources, replies);
+ if (priv->arbiter_count)
+ /* Choose non-arbiter brick as source for empty files. */
+ afr_mark_source_sinks_if_file_empty(this, sources, sinks, healed_sinks,
+ locked_on, replies,
+ AFR_DATA_TRANSACTION);
- afr_reset_children (sh->fresh_children, priv->child_count);
- afr_get_fresh_children (sh->success_children, sh->sources,
- sh->fresh_children, priv->child_count);
- afr_inode_set_read_ctx (this, sh->inode, sh->source,
- sh->fresh_children);
out:
- return ret;
-}
-
-void
-afr_sh_data_fix (call_frame_t *frame, xlator_t *this)
-{
- int source = 0;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- afr_private_t *priv = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- source = sh->source;
- sh->block_size = this->ctx->page_size;
- sh->file_size = sh->buf[source].ia_size;
-
- if (FILE_HAS_HOLES (&sh->buf[source]))
- sh->file_has_holes = 1;
-
- if (sh->background && sh->unwind && !sh->unwound) {
- sh->unwind (sh->orig_frame, this, sh->op_ret, sh->op_errno,
- sh->op_failed);
- sh->unwound = _gf_true;
- }
+ afr_mark_active_sinks(this, sources, locked_on, healed_sinks);
+ source = afr_choose_source_by_policy(priv, sources, AFR_DATA_TRANSACTION);
- afr_sh_mark_source_sinks (frame, this);
- if (sh->active_sinks == 0) {
- gf_log (this->name, GF_LOG_INFO,
- "no active sinks for performing self-heal on file %s",
- local->loc.path);
- afr_sh_data_finish (frame, this);
- return;
- }
-
- gf_log (this->name, GF_LOG_DEBUG,
- "self-healing file %s from subvolume %s to %d other",
- local->loc.path, priv->children[sh->source]->name,
- sh->active_sinks);
-
- sh->actual_sh_started = _gf_true;
- afr_sh_data_trim_sinks (frame, this);
+ return source;
}
+/*
+ * __afr_selfheal_data_prepare:
+ *
+ * This function inspects the on-disk xattrs and determines which subvols
+ * are sources and sinks.
+ *
+ * The return value is the index of the subvolume to be used as the source
+ * for self-healing, or -1 if no healing is necessary/split brain.
+ */
int
-afr_sh_data_fxattrop_fstat_done (call_frame_t *frame, xlator_t *this)
+__afr_selfheal_data_prepare(call_frame_t *frame, xlator_t *this, inode_t *inode,
+ unsigned char *locked_on, unsigned char *sources,
+ unsigned char *sinks, unsigned char *healed_sinks,
+ unsigned char *undid_pending,
+ struct afr_reply *replies, unsigned char *pflag)
{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- afr_private_t *priv = NULL;
- int nsources = 0;
- int ret = 0;
- int *old_sources = NULL;
- int tstamp_source = 0;
- int i = 0;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- gf_log (this->name, GF_LOG_DEBUG, "Pending matrix for: %s",
- lkowner_utoa (&frame->root->lk_owner));
- if (sh->sync_done) {
- //store sources before sync so that mtime can be set using the
- //iatt buf from one of them.
- old_sources = alloca (priv->child_count*sizeof (*old_sources));
- memcpy (old_sources, sh->sources,
- priv->child_count * sizeof (*old_sources));
- }
-
- nsources = afr_build_sources (this, sh->xattr, sh->buf, sh->pending_matrix,
- sh->sources, sh->success_children,
- AFR_DATA_TRANSACTION, NULL, _gf_true);
- if ((nsources == -1)
- && (priv->favorite_child != -1)
- && (sh->child_errno[priv->favorite_child] == 0)) {
+ int ret = -1;
+ int source = -1;
+ afr_private_t *priv = NULL;
+ uint64_t *witness = NULL;
- gf_log (this->name, GF_LOG_DEBUG,
- "Picking favorite child %s as authentic source to "
- "resolve conflicting data of %s",
- priv->children[priv->favorite_child]->name,
- local->loc.path);
+ priv = this->private;
- sh->sources[priv->favorite_child] = 1;
-
- nsources = afr_sh_source_count (sh->sources,
- priv->child_count);
- }
+ ret = afr_selfheal_unlocked_discover(frame, inode, inode->gfid, replies);
- if (nsources == -1) {
- afr_sh_print_split_brain_log (sh->pending_matrix, this,
- local->loc.path);
- afr_set_split_brain (this, sh->inode, DONT_KNOW, SPB);
-
- afr_sh_data_fail (frame, this);
- return 0;
- }
-
- afr_set_split_brain (this, sh->inode, DONT_KNOW, NO_SPB);
+ if (ret)
+ return ret;
- ret = afr_sh_inode_set_read_ctx (sh, this);
- if (ret) {
- gf_log (this->name, GF_LOG_DEBUG,
- "No active sources found.");
+ witness = alloca0(priv->child_count * sizeof(*witness));
+ ret = afr_selfheal_find_direction(frame, this, replies,
+ AFR_DATA_TRANSACTION, locked_on, sources,
+ sinks, witness, pflag);
+ if (ret)
+ return ret;
- afr_sh_data_fail (frame, this);
- return 0;
- }
+ /* Initialize the healed_sinks[] array optimistically to
+ the intersection of to-be-healed (i.e sinks[]) and
+ the list of servers which are up (i.e locked_on[]).
+ As we encounter failures in the healing process, we
+ will unmark the respective servers in the healed_sinks[]
+ array.
+ */
+ AFR_INTERSECT(healed_sinks, sinks, locked_on, priv->child_count);
- if (sh->sync_done) {
- /* Perform setattr from one of the old_sources if possible
- * Because only they have the correct mtime, the new sources
- * (i.e. old sinks) have mtime from last writev in sync.
- */
- tstamp_source = sh->source;
- for (i = 0; i < priv->child_count; i++) {
- if (old_sources[i] && sh->sources[i])
- tstamp_source = i;
- }
- afr_sh_data_setattr (frame, this, &sh->buf[tstamp_source]);
- } else {
- if (nsources == 0) {
- gf_log (this->name, GF_LOG_DEBUG,
- "No self-heal needed for %s",
- local->loc.path);
-
- afr_sh_data_finish (frame, this);
- return 0;
- }
-
- if (sh->do_data_self_heal &&
- afr_data_self_heal_enabled (priv->data_self_heal))
- afr_sh_data_fix (frame, this);
- else
- afr_sh_data_finish (frame, this);
- }
- return 0;
-}
+ source = __afr_selfheal_data_finalize_source(
+ frame, this, inode, sources, sinks, healed_sinks, locked_on,
+ undid_pending, replies, witness);
+ if (source < 0)
+ return -EIO;
-int
-afr_lookup_select_read_child_by_txn_type (xlator_t *this, afr_local_t *local,
- dict_t **xattr,
- afr_transaction_type txn_type,
- uuid_t gfid)
-{
- afr_private_t *priv = NULL;
- int read_child = -1;
- int32_t **pending_matrix = NULL;
- int32_t *sources = NULL;
- int32_t *success_children = NULL;
- struct iatt *bufs = NULL;
- int32_t nsources = 0;
- int32_t prev_read_child = -1;
- int32_t config_read_child = -1;
- int32_t subvol_status = 0;
-
- priv = this->private;
- bufs = local->cont.lookup.bufs;
- success_children = local->cont.lookup.success_children;
-
- pending_matrix = local->cont.lookup.pending_matrix;
- sources = local->cont.lookup.sources;
- memset (sources, 0, sizeof (*sources) * priv->child_count);
-
- nsources = afr_build_sources (this, xattr, bufs, pending_matrix,
- sources, success_children, txn_type,
- &subvol_status, _gf_false);
- if (subvol_status & SPLIT_BRAIN) {
- gf_log (this->name, GF_LOG_DEBUG, "%s: Possible split-brain",
- local->loc.path);
- switch (txn_type) {
- case AFR_DATA_TRANSACTION:
- local->cont.lookup.possible_spb = _gf_true;
- nsources = 1;
- sources[success_children[0]] = 1;
- break;
- case AFR_ENTRY_TRANSACTION:
- read_child = afr_get_no_xattr_dir_read_child (this,
- success_children,
- bufs);
- sources[read_child] = 1;
- nsources = 1;
- break;
- default:
- break;
- }
- }
- if (nsources < 0)
- goto out;
-
- prev_read_child = local->read_child_index;
- config_read_child = priv->read_child;
- read_child = afr_select_read_child_from_policy (success_children,
- priv->child_count,
- prev_read_child,
- config_read_child,
- sources,
- priv->hash_mode, gfid);
-out:
- gf_log (this->name, GF_LOG_DEBUG, "returning read_child: %d",
- read_child);
- return read_child;
+ return source;
}
-int
-afr_sh_data_fstat_cbk (call_frame_t *frame, void *cookie,
- xlator_t *this, int32_t op_ret, int32_t op_errno,
- struct iatt *buf, dict_t *xdata)
-{
- afr_private_t *priv = NULL;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- int call_count = -1;
- int child_index = (long) cookie;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- LOCK (&frame->lock);
- {
- if (op_ret != -1) {
- gf_log (this->name, GF_LOG_TRACE,
- "fstat of %s on %s succeeded",
- local->loc.path,
- priv->children[child_index]->name);
-
- sh->buf[child_index] = *buf;
- sh->success_children[sh->success_count] = child_index;
- sh->success_count++;
- } else {
- gf_log (this->name, GF_LOG_ERROR, "%s: fstat failed "
- "on %s, reason %s", local->loc.path,
- priv->children[child_index]->name,
- strerror (op_errno));
- sh->child_errno[child_index] = op_errno;
- }
- }
- UNLOCK (&frame->lock);
-
- call_count = afr_frame_return (frame);
-
- if (call_count == 0) {
- /* Previous versions of glusterfs might have set
- * the pending data xattrs which need to be erased
- */
- if (!afr_sh_data_proceed (sh->success_count)) {
- gf_log (this->name, GF_LOG_ERROR, "inspecting metadata "
- "succeeded on < %d children, aborting "
- "self-heal for %s", AFR_SH_MIN_PARTICIPANTS,
- local->loc.path);
- afr_sh_data_fail (frame, this);
- goto out;
- }
- afr_sh_data_fxattrop_fstat_done (frame, this);
- }
+static int
+__afr_selfheal_data(call_frame_t *frame, xlator_t *this, fd_t *fd,
+ unsigned char *locked_on)
+{
+ afr_private_t *priv = NULL;
+ int ret = -1;
+ unsigned char *sources = NULL;
+ unsigned char *sinks = NULL;
+ unsigned char *data_lock = NULL;
+ unsigned char *healed_sinks = NULL;
+ unsigned char *undid_pending = NULL;
+ struct afr_reply *locked_replies = NULL;
+ int source = -1;
+ gf_boolean_t did_sh = _gf_true;
+ gf_boolean_t is_arbiter_the_only_sink = _gf_false;
+ gf_boolean_t empty_file = _gf_false;
+
+ priv = this->private;
+
+ sources = alloca0(priv->child_count);
+ sinks = alloca0(priv->child_count);
+ healed_sinks = alloca0(priv->child_count);
+ data_lock = alloca0(priv->child_count);
+ undid_pending = alloca0(priv->child_count);
+
+ locked_replies = alloca0(sizeof(*locked_replies) * priv->child_count);
+
+ ret = afr_selfheal_inodelk(frame, this, fd->inode, this->name, 0, 0,
+ data_lock);
+ {
+ if (ret < priv->child_count) {
+ gf_msg_debug(this->name, 0,
+ "%s: Skipping "
+ "self-heal as only %d number "
+ "of subvolumes "
+ "could be locked",
+ uuid_utoa(fd->inode->gfid), ret);
+ ret = -ENOTCONN;
+ goto unlock;
+ }
+
+ ret = __afr_selfheal_data_prepare(frame, this, fd->inode, data_lock,
+ sources, sinks, healed_sinks,
+ undid_pending, locked_replies, NULL);
+ if (ret < 0)
+ goto unlock;
+
+ if (AFR_COUNT(healed_sinks, priv->child_count) == 0) {
+ did_sh = _gf_false;
+ goto unlock;
+ }
+
+ source = ret;
+
+ if (AFR_IS_ARBITER_BRICK(priv, source)) {
+ empty_file = afr_is_file_empty_on_all_children(priv,
+ locked_replies);
+ if (empty_file)
+ goto restore_time;
+
+ did_sh = _gf_false;
+ goto unlock;
+ }
+
+ ret = __afr_selfheal_truncate_sinks(
+ frame, this, fd, healed_sinks,
+ locked_replies[source].poststat.ia_size);
+ if (ret < 0)
+ goto unlock;
+
+ if (priv->arbiter_count &&
+ AFR_COUNT(healed_sinks, priv->child_count) == 1 &&
+ healed_sinks[ARBITER_BRICK_INDEX]) {
+ is_arbiter_the_only_sink = _gf_true;
+ goto restore_time;
+ }
+ ret = 0;
+ }
+unlock:
+ afr_selfheal_uninodelk(frame, this, fd->inode, this->name, 0, 0, data_lock);
+ if (ret < 0)
+ goto out;
+
+ if (!did_sh)
+ goto out;
+
+ ret = afr_selfheal_data_do(frame, this, fd, source, healed_sinks,
+ locked_replies);
+ if (ret)
+ goto out;
+restore_time:
+ afr_selfheal_restore_time(frame, this, fd->inode, source, healed_sinks,
+ locked_replies);
+
+ if (!is_arbiter_the_only_sink && !empty_file) {
+ ret = afr_selfheal_inodelk(frame, this, fd->inode, this->name, 0, 0,
+ data_lock);
+ if (ret < priv->child_count) {
+ ret = -ENOTCONN;
+ did_sh = _gf_false;
+ goto skip_undo_pending;
+ }
+ }
+ ret = afr_selfheal_undo_pending(
+ frame, this, fd->inode, sources, sinks, healed_sinks, undid_pending,
+ AFR_DATA_TRANSACTION, locked_replies, data_lock);
+skip_undo_pending:
+ afr_selfheal_uninodelk(frame, this, fd->inode, this->name, 0, 0, data_lock);
out:
- return 0;
-}
+ if (did_sh)
+ afr_log_selfheal(fd->inode->gfid, this, ret, "data", source, sources,
+ healed_sinks);
+ else
+ ret = 1;
-int
-afr_sh_data_fstat (call_frame_t *frame, xlator_t *this)
-{
- afr_self_heal_t *sh = NULL;
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- int call_count = 0;
- int i = 0;
- int child = 0;
- int32_t *fstat_children = NULL;
-
- priv = this->private;
- local = frame->local;
- sh = &local->self_heal;
-
- fstat_children = memdup (sh->success_children,
- sizeof (*fstat_children) * priv->child_count);
- if (!fstat_children) {
- afr_sh_data_fail (frame, this);
- goto out;
- }
- call_count = sh->success_count;
- local->call_count = call_count;
+ if (locked_replies)
+ afr_replies_wipe(locked_replies, priv->child_count);
- memset (sh->buf, 0, sizeof (*sh->buf) * priv->child_count);
- afr_reset_children (sh->success_children, priv->child_count);
- sh->success_count = 0;
- for (i = 0; i < priv->child_count; i++) {
- child = fstat_children[i];
- if (child == -1)
- break;
- STACK_WIND_COOKIE (frame, afr_sh_data_fstat_cbk,
- (void *) (long) child,
- priv->children[child],
- priv->children[child]->fops->fstat,
- sh->healing_fd, NULL);
- --call_count;
- }
- GF_ASSERT (!call_count);
-out:
- GF_FREE (fstat_children);
- return 0;
-}
-
-void
-afr_sh_common_fxattrop_resp_handler (call_frame_t *frame, void *cookie,
- xlator_t *this, int32_t op_ret,
- int32_t op_errno, dict_t *xattr)
-{
- afr_private_t *priv = NULL;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- int child_index = (long) cookie;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- LOCK (&frame->lock);
- {
- if (op_ret != -1) {
- gf_log (this->name, GF_LOG_TRACE,
- "fxattrop of %s on %s succeeded",
- local->loc.path,
- priv->children[child_index]->name);
-
- sh->xattr[child_index] = dict_ref (xattr);
- sh->success_children[sh->success_count] = child_index;
- sh->success_count++;
- } else {
- gf_log (this->name, GF_LOG_ERROR, "fxattrop of %s "
- "failed on %s, reason %s", local->loc.path,
- priv->children[child_index]->name,
- strerror (op_errno));
- sh->child_errno[child_index] = op_errno;
- }
- }
- UNLOCK (&frame->lock);
+ return ret;
}
int
-afr_sh_data_fxattrop_cbk (call_frame_t *frame, void *cookie,
- xlator_t *this, int32_t op_ret, int32_t op_errno,
- dict_t *xattr, dict_t *xdata)
+afr_selfheal_data_open_cbk(call_frame_t *frame, void *cookie, xlator_t *this,
+ int32_t op_ret, int32_t op_errno, fd_t *fd,
+ dict_t *xdata)
{
- int call_count = -1;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
-
- afr_sh_common_fxattrop_resp_handler (frame, cookie, this, op_ret,
- op_errno, xattr);
-
- call_count = afr_frame_return (frame);
- if (call_count == 0) {
- if (!afr_sh_data_proceed (sh->success_count)) {
- gf_log (this->name, GF_LOG_ERROR, "%s, inspecting "
- "change log succeeded on < %d children",
- local->loc.path, AFR_SH_MIN_PARTICIPANTS);
- afr_sh_data_fail (frame, this);
- goto out;
- }
- afr_sh_data_fstat (frame, this);
- }
-out:
- return 0;
-}
+ afr_local_t *local = NULL;
+ int i = (long)cookie;
+ local = frame->local;
-int
-afr_sh_data_fxattrop (call_frame_t *frame, xlator_t *this)
-{
- afr_self_heal_t *sh = NULL;
- afr_local_t *local = NULL;
- afr_private_t *priv = NULL;
- dict_t **xattr_req;
- int32_t *zero_pending = NULL;
- int call_count = 0;
- int i = 0;
- int ret = 0;
- int j;
-
- priv = this->private;
- local = frame->local;
- sh = &local->self_heal;
-
- call_count = afr_up_children_count (local->child_up,
- priv->child_count);
-
- local->call_count = call_count;
-
- xattr_req = GF_CALLOC(priv->child_count, sizeof(struct dict_t *),
- gf_afr_mt_dict_t);
- if (!xattr_req)
- goto out;
-
- for (i = 0; i < priv->child_count; i++) {
- xattr_req[i] = dict_new();
- if (!xattr_req[i]) {
- ret = -1;
- goto out;
- }
- }
-
- for (i = 0; i < priv->child_count; i++) {
- for (j = 0; j < priv->child_count; j++) {
- zero_pending = GF_CALLOC (3, sizeof (*zero_pending),
- gf_afr_mt_int32_t);
- if (!zero_pending) {
- ret = -1;
- goto out;
- }
- ret = dict_set_dynptr (xattr_req[i], priv->pending_key[j],
- zero_pending,
- 3 * sizeof (*zero_pending));
- if (ret < 0) {
- gf_log (this->name, GF_LOG_WARNING,
- "Unable to set dict value");
- goto out;
- } else {
- zero_pending = NULL;
- }
- }
- }
-
- afr_reset_xattr (sh->xattr, priv->child_count);
- afr_reset_children (sh->success_children, priv->child_count);
- memset (sh->child_errno, 0,
- sizeof (*sh->child_errno) * priv->child_count);
- sh->success_count = 0;
- for (i = 0; i < priv->child_count; i++) {
- if (local->child_up[i]) {
- STACK_WIND_COOKIE (frame, afr_sh_data_fxattrop_cbk,
- (void *) (long) i,
- priv->children[i],
- priv->children[i]->fops->fxattrop,
- sh->healing_fd, GF_XATTROP_ADD_ARRAY,
- xattr_req[i], NULL);
-
- if (!--call_count)
- break;
- }
- }
+ local->replies[i].valid = 1;
+ local->replies[i].op_ret = op_ret;
+ local->replies[i].op_errno = op_errno;
-out:
- if (xattr_req) {
- for (i = 0; i < priv->child_count; i++)
- if (xattr_req[i])
- dict_unref(xattr_req[i]);
- GF_FREE(xattr_req);
- }
-
- if (ret) {
- GF_FREE (zero_pending);
- afr_sh_data_fail (frame, this);
- }
+ syncbarrier_wake(&local->barrier);
- return 0;
-}
-
-int
-afr_sh_data_big_lock_success (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
-
- sh->data_lock_held = _gf_true;
- afr_sh_data_fxattrop (frame, this);
- return 0;
+ return 0;
}
int
-afr_sh_data_post_blocking_inodelk_cbk (call_frame_t *frame, xlator_t *this)
+afr_selfheal_data_open(xlator_t *this, inode_t *inode, fd_t **fd)
{
- afr_internal_lock_t *int_lock = NULL;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
+ int ret = 0;
+ fd_t *fd_tmp = NULL;
+ loc_t loc = {
+ 0,
+ };
+ call_frame_t *frame = NULL;
+ afr_local_t *local = NULL;
+ afr_private_t *priv = NULL;
+ int i = 0;
- local = frame->local;
- int_lock = &local->internal_lock;
- sh = &local->self_heal;
+ priv = this->private;
- if (int_lock->lock_op_ret < 0) {
- gf_log (this->name, GF_LOG_ERROR, "Blocking data inodelks "
- "failed for %s. by %s",
- local->loc.path, lkowner_utoa (&frame->root->lk_owner));
+ fd_tmp = fd_create(inode, 0);
+ if (!fd_tmp)
+ return -ENOMEM;
- sh->data_lock_failure_handler (frame, this);
- } else {
-
- gf_log (this->name, GF_LOG_DEBUG, "Blocking data inodelks "
- "done for %s by %s. Proceding to self-heal",
- local->loc.path, lkowner_utoa (&frame->root->lk_owner));
+ loc.inode = inode_ref(inode);
+ gf_uuid_copy(loc.gfid, inode->gfid);
- sh->data_lock_success_handler (frame, this);
- }
+ frame = afr_frame_create(this, &ret);
+ if (!frame) {
+ ret = -ret;
+ fd_unref(fd_tmp);
+ goto out;
+ }
+ local = frame->local;
- return 0;
-}
+ AFR_ONLIST(local->child_up, frame, afr_selfheal_data_open_cbk, open, &loc,
+ O_RDWR | O_LARGEFILE, fd_tmp, NULL);
-int
-afr_sh_data_post_nonblocking_inodelk_cbk (call_frame_t *frame, xlator_t *this)
-{
- afr_internal_lock_t *int_lock = NULL;
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
-
- local = frame->local;
- int_lock = &local->internal_lock;
- sh = &local->self_heal;
-
- if (int_lock->lock_op_ret < 0) {
- gf_log (this->name, GF_LOG_DEBUG, "Non Blocking data inodelks "
- "failed for %s. by %s",
- local->loc.path, lkowner_utoa (&frame->root->lk_owner));
-
- if (!sh->data_lock_block) {
- sh->data_lock_failure_handler(frame, this);
- } else {
- int_lock->lock_cbk =
- afr_sh_data_post_blocking_inodelk_cbk;
- afr_blocking_lock (frame, this);
- }
- } else {
+ ret = -ENOTCONN;
+ for (i = 0; i < priv->child_count; i++) {
+ if (!local->replies[i].valid)
+ continue;
- gf_log (this->name, GF_LOG_DEBUG, "Non Blocking data inodelks "
- "done for %s by %s. Proceeding to self-heal",
- local->loc.path, lkowner_utoa (&frame->root->lk_owner));
- sh->data_lock_success_handler (frame, this);
+ if (local->replies[i].op_ret < 0) {
+ ret = -local->replies[i].op_errno;
+ continue;
}
- return 0;
-}
-
-int
-afr_sh_data_lock_rec (call_frame_t *frame, xlator_t *this, off_t start, off_t len)
-{
- afr_internal_lock_t *int_lock = NULL;
- afr_local_t *local = NULL;
-
- local = frame->local;
- int_lock = &local->internal_lock;
-
- int_lock->transaction_lk_type = AFR_SELFHEAL_LK;
- int_lock->selfheal_lk_type = AFR_DATA_SELF_HEAL_LK;
-
- afr_set_lock_number (frame, this);
-
- int_lock->lk_flock.l_start = start;
- int_lock->lk_flock.l_len = len;
- int_lock->lk_flock.l_type = F_WRLCK;
- int_lock->lock_cbk = afr_sh_data_post_nonblocking_inodelk_cbk;
-
- afr_nonblocking_inodelk (frame, this);
-
- return 0;
-}
-
-int
-afr_post_sh_big_lock_success (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
-
- GF_ASSERT (sh->old_loop_frame);
- sh_loop_finish (sh->old_loop_frame, this);
- sh->old_loop_frame = NULL;
- sh->data_lock_held = _gf_true;
- sh->sync_done = _gf_true;
- afr_sh_data_fxattrop (frame, this);
- return 0;
-}
+ ret = 0;
+ break;
+ }
-int
-afr_post_sh_big_lock_failure (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
+ if (ret < 0) {
+ fd_unref(fd_tmp);
+ goto out;
+ } else {
+ fd_bind(fd_tmp);
+ }
- local = frame->local;
- sh = &local->self_heal;
-
- GF_ASSERT (sh->old_loop_frame);
- sh_loop_finish (sh->old_loop_frame, this);
- sh->old_loop_frame = NULL;
- afr_sh_set_timestamps (frame, this);
- return 0;
-}
-
-
-int
-afr_sh_data_lock (call_frame_t *frame, xlator_t *this,
- off_t start, off_t len, gf_boolean_t block,
- afr_lock_cbk_t success_handler,
- afr_lock_cbk_t failure_handler)
-{
- afr_local_t * local = NULL;
- afr_self_heal_t * sh = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
-
- sh->data_lock_success_handler = success_handler;
- sh->data_lock_failure_handler = failure_handler;
- sh->data_lock_block = block;
- return afr_sh_data_lock_rec (frame, this, start, len);
-}
-
-int
-afr_sh_data_open_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
- int32_t op_ret, int32_t op_errno, fd_t *fd, dict_t *xdata)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- afr_private_t *priv = NULL;
- int call_count = 0;
- int child_index = 0;
- gf_boolean_t block = _gf_true;
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- child_index = (long) cookie;
-
- /* TODO: some of the open's might fail.
- In that case, modify cleanup fn to send flush on those
- fd's which are already open */
-
- LOCK (&frame->lock);
- {
- if (op_ret == -1) {
- gf_log (this->name, GF_LOG_ERROR,
- "open of %s failed on child %s (%s)",
- local->loc.path,
- priv->children[child_index]->name,
- strerror (op_errno));
- sh->op_failed = 1;
- } else {
- gf_log (this->name, GF_LOG_TRACE,
- "open of %s succeeded on child %s",
- local->loc.path,
- priv->children[child_index]->name);
- }
- }
- UNLOCK (&frame->lock);
-
- call_count = afr_frame_return (frame);
-
- if (call_count == 0) {
- if (sh->op_failed) {
- afr_sh_data_fail (frame, this);
- return 0;
- }
-
- gf_log (this->name, GF_LOG_TRACE,
- "fd for %s opened, commencing sync",
- local->loc.path);
-
- /*
- * The read and write self-heal trigger codepaths do not provide
- * an unwind callback. We run a trylock in these codepaths
- * because we are sensitive to locking latency.
- */
- block = sh->unwind ? _gf_true : _gf_false;
- afr_sh_data_lock (frame, this, 0, 0, block,
- afr_sh_data_big_lock_success,
- afr_sh_data_fail);
- }
-
- return 0;
+ *fd = fd_tmp;
+out:
+ loc_wipe(&loc);
+ if (frame)
+ AFR_STACK_DESTROY(frame);
+ return ret;
}
-
int
-afr_sh_data_open (call_frame_t *frame, xlator_t *this)
+afr_selfheal_data(call_frame_t *frame, xlator_t *this, fd_t *fd)
{
- int i = 0;
- int call_count = 0;
- fd_t *fd = NULL;
- afr_local_t * local = NULL;
- afr_private_t * priv = NULL;
- afr_self_heal_t *sh = NULL;
+ afr_private_t *priv = NULL;
+ unsigned char *locked_on = NULL;
+ int ret = 0;
+ inode_t *inode = fd->inode;
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
+ priv = this->private;
- call_count = afr_up_children_count (local->child_up, priv->child_count);
- local->call_count = call_count;
+ locked_on = alloca0(priv->child_count);
- fd = fd_create (local->loc.inode, frame->root->pid);
- sh->healing_fd = fd;
-
- /* open sinks */
- for (i = 0; i < priv->child_count; i++) {
- if(!local->child_up[i])
- continue;
-
- STACK_WIND_COOKIE (frame, afr_sh_data_open_cbk,
- (void *) (long) i,
- priv->children[i],
- priv->children[i]->fops->open,
- &local->loc,
- O_RDWR|O_LARGEFILE, fd, NULL);
-
- if (!--call_count)
- break;
+ ret = afr_selfheal_tie_breaker_inodelk(frame, this, inode, priv->sh_domain,
+ 0, 0, locked_on);
+ {
+ if (ret < priv->child_count) {
+ gf_msg_debug(this->name, 0,
+ "%s: Skipping "
+ "self-heal as only %d number of "
+ "subvolumes could be locked",
+ uuid_utoa(fd->inode->gfid), ret);
+ /* Either less than two subvols available, or another
+ selfheal (from another server) is in progress. Skip
+ for now in any case there isn't anything to do.
+ */
+ ret = -ENOTCONN;
+ goto unlock;
}
- return 0;
-}
-
-void
-afr_sh_non_reg_fix (call_frame_t *frame, xlator_t *this,
- int32_t op_ret, int32_t op_errno)
-{
- afr_private_t *priv = NULL;
- afr_self_heal_t *sh = NULL;
- afr_local_t *local = NULL;
- int i = 0;
-
- if (op_ret < 0) {
- afr_sh_data_fail (frame, this);
- return;
- }
-
- local = frame->local;
- sh = &local->self_heal;
- priv = this->private;
-
- for (i = 0; i < priv->child_count ; i++) {
- if (1 == local->child_up[i])
- sh->success[i] = 1;
- }
-
- afr_sh_erase_pending (frame, this, AFR_DATA_TRANSACTION,
- afr_sh_data_erase_pending_cbk,
- afr_sh_data_finish);
-}
+ ret = __afr_selfheal_data(frame, this, fd, locked_on);
+ }
+unlock:
+ afr_selfheal_uninodelk(frame, this, inode, priv->sh_domain, 0, 0,
+ locked_on);
-int
-afr_sh_non_reg_lock_success (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
-
- local = frame->local;
- sh = &local->self_heal;
- sh->data_lock_held = _gf_true;
- afr_sh_common_lookup (frame, this, &local->loc,
- afr_sh_non_reg_fix, NULL,
- AFR_LOOKUP_FAIL_CONFLICTS |
- AFR_LOOKUP_FAIL_MISSING_GFIDS,
- NULL);
- return 0;
-}
-
-gf_boolean_t
-afr_can_start_data_self_heal (afr_self_heal_t *sh, afr_private_t *priv)
-{
- if (sh->force_confirm_spb)
- return _gf_true;
- if (sh->do_data_self_heal &&
- afr_data_self_heal_enabled (priv->data_self_heal))
- return _gf_true;
- return _gf_false;
-}
-
-int
-afr_self_heal_data (call_frame_t *frame, xlator_t *this)
-{
- afr_local_t *local = NULL;
- afr_self_heal_t *sh = NULL;
- afr_private_t *priv = this->private;
-
- local = frame->local;
- sh = &local->self_heal;
-
- if (afr_can_start_data_self_heal (sh, priv)) {
- if (IA_ISREG (sh->type)) {
- afr_sh_data_open (frame, this);
- } else {
- afr_sh_data_lock (frame, this, 0, 0, _gf_true,
- afr_sh_non_reg_lock_success,
- afr_sh_data_fail);
- }
- } else {
- gf_log (this->name, GF_LOG_TRACE,
- "not doing data self heal on %s",
- local->loc.path);
- afr_sh_data_done (frame, this);
- }
-
- return 0;
+ return ret;
}
generated by cgit (git 2.34.1) at 2026-01-04 10:17:07 +0000