diff options
Diffstat (limited to 'xlators/storage/posix/src/posix.c')
| -rw-r--r-- | xlators/storage/posix/src/posix.c | 431 |
1 files changed, 402 insertions, 29 deletions
diff --git a/xlators/storage/posix/src/posix.c b/xlators/storage/posix/src/posix.c index 05101fd31..fb45c7a67 100644 --- a/xlators/storage/posix/src/posix.c +++ b/xlators/storage/posix/src/posix.c @@ -23,6 +23,8 @@ #include <pthread.h> #include <ftw.h> #include <sys/stat.h> +#include <signal.h> +#include <sys/uio.h> #ifndef GF_BSD_HOST_OS #include <alloca.h> @@ -50,6 +52,7 @@ #include "glusterfs3-xdr.h" #include "hashfn.h" #include "posix-aio.h" +#include "glusterfs-acl.h" extern char *marker_xattrs[]; #define ALIGN_SIZE 4096 @@ -128,7 +131,7 @@ posix_lookup (call_frame_t *frame, xlator_t *this, MAKE_ENTRY_HANDLE (real_path, par_path, this, loc, &buf); if (uuid_is_null (loc->inode->gfid)) { - posix_gfid_set (this, real_path, loc, xdata); + posix_gfid_heal (this, real_path, loc, xdata); MAKE_ENTRY_HANDLE (real_path, par_path, this, loc, &buf); } @@ -614,6 +617,166 @@ out: return ret; } +char* +_page_aligned_alloc (size_t size, char **aligned_buf) +{ + char *alloc_buf = NULL; + char *buf = NULL; + + alloc_buf = GF_CALLOC (1, (size + ALIGN_SIZE), gf_posix_mt_char); + if (!alloc_buf) + goto out; + /* page aligned buffer */ + buf = GF_ALIGN_BUF (alloc_buf, ALIGN_SIZE); + *aligned_buf = buf; +out: + return alloc_buf; +} + +static int32_t +_posix_do_zerofill(int fd, off_t offset, size_t len, int o_direct) +{ + size_t num_vect = 0; + int32_t num_loop = 1; + int32_t idx = 0; + int32_t op_ret = -1; + int32_t vect_size = VECTOR_SIZE; + size_t remain = 0; + size_t extra = 0; + struct iovec *vector = NULL; + char *iov_base = NULL; + char *alloc_buf = NULL; + + if (len == 0) + return 0; + if (len < VECTOR_SIZE) + vect_size = len; + + num_vect = len / (vect_size); + remain = len % vect_size ; + if (num_vect > MAX_NO_VECT) { + extra = num_vect % MAX_NO_VECT; + num_loop = num_vect / MAX_NO_VECT; + num_vect = MAX_NO_VECT; + } + + vector = GF_CALLOC (num_vect, sizeof(struct iovec), + gf_common_mt_iovec); + if (!vector) + return -1; + if (o_direct) { + alloc_buf = _page_aligned_alloc(vect_size, &iov_base); + if (!alloc_buf) { + gf_log ("_posix_do_zerofill", GF_LOG_DEBUG, + "memory alloc failed, vect_size %d: %s", + vect_size, strerror(errno)); + GF_FREE(vector); + return -1; + } + } else { + iov_base = GF_CALLOC (vect_size, sizeof(char), + gf_common_mt_char); + if (!iov_base) { + GF_FREE(vector); + return -1; + } + } + + for (idx = 0; idx < num_vect; idx++) { + vector[idx].iov_base = iov_base; + vector[idx].iov_len = vect_size; + } + lseek(fd, offset, SEEK_SET); + for (idx = 0; idx < num_loop; idx++) { + op_ret = writev(fd, vector, num_vect); + if (op_ret < 0) + goto err; + } + if (extra) { + op_ret = writev(fd, vector, extra); + if (op_ret < 0) + goto err; + } + if (remain) { + vector[0].iov_len = remain; + op_ret = writev(fd, vector , 1); + if (op_ret < 0) + goto err; + } +err: + if (o_direct) + GF_FREE(alloc_buf); + else + GF_FREE(iov_base); + GF_FREE(vector); + return op_ret; +} + +static int32_t +posix_do_zerofill(call_frame_t *frame, xlator_t *this, fd_t *fd, + off_t offset, size_t len, struct iatt *statpre, + struct iatt *statpost) +{ + struct posix_fd *pfd = NULL; + int32_t ret = -1; + + DECLARE_OLD_FS_ID_VAR; + + SET_FS_ID (frame->root->uid, frame->root->gid); + + VALIDATE_OR_GOTO (frame, out); + VALIDATE_OR_GOTO (this, out); + VALIDATE_OR_GOTO (fd, out); + + ret = posix_fd_ctx_get (fd, this, &pfd); + if (ret < 0) { + gf_log (this->name, GF_LOG_DEBUG, + "pfd is NULL from fd=%p", fd); + goto out; + } + + ret = posix_fdstat (this, pfd->fd, statpre); + if (ret == -1) { + ret = -errno; + gf_log (this->name, GF_LOG_ERROR, + "pre-operation fstat failed on fd = %p: %s", fd, + strerror (errno)); + goto out; + } + ret = _posix_do_zerofill(pfd->fd, offset, len, pfd->flags & O_DIRECT); + if (ret < 0) { + ret = -errno; + gf_log(this->name, GF_LOG_ERROR, + "zerofill failed on fd %d length %ld %s", + pfd->fd, len, strerror(errno)); + goto out; + } + if (pfd->flags & (O_SYNC|O_DSYNC)) { + ret = fsync (pfd->fd); + if (ret) { + gf_log (this->name, GF_LOG_ERROR, + "fsync() in writev on fd %d failed: %s", + pfd->fd, strerror (errno)); + ret = -errno; + goto out; + } + } + + ret = posix_fdstat (this, pfd->fd, statpost); + if (ret == -1) { + ret = -errno; + gf_log (this->name, GF_LOG_ERROR, + "post operation fstat failed on fd=%p: %s", fd, + strerror (errno)); + goto out; + } + +out: + SET_TO_OLD_FS_ID (); + + return ret; +} + static int32_t _posix_fallocate(call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t keep_size, off_t offset, size_t len, dict_t *xdata) @@ -628,7 +791,7 @@ _posix_fallocate(call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t keep_siz ret = posix_do_fallocate(frame, this, fd, flags, offset, len, &statpre, &statpost); - if (ret < 0) + if (ret < 0) goto err; STACK_UNWIND_STRICT(fallocate, frame, 0, 0, &statpre, &statpost, NULL); @@ -650,7 +813,7 @@ posix_discard(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, ret = posix_do_fallocate(frame, this, fd, flags, offset, len, &statpre, &statpost); - if (ret < 0) + if (ret < 0) goto err; STACK_UNWIND_STRICT(discard, frame, 0, 0, &statpre, &statpost, NULL); @@ -662,6 +825,28 @@ err: } +static int32_t +posix_zerofill(call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, + size_t len, dict_t *xdata) +{ + int32_t ret = 0; + struct iatt statpre = {0,}; + struct iatt statpost = {0,}; + + ret = posix_do_zerofill(frame, this, fd, offset, len, + &statpre, &statpost); + if (ret < 0) + goto err; + + STACK_UNWIND_STRICT(zerofill, frame, 0, 0, &statpre, &statpost, NULL); + return 0; + +err: + STACK_UNWIND_STRICT(zerofill, frame, -1, -ret, NULL, NULL, NULL); + return 0; + +} + int32_t posix_opendir (call_frame_t *frame, xlator_t *this, loc_t *loc, fd_t *fd, dict_t *xdata) @@ -2070,11 +2255,7 @@ posix_readv (call_frame_t *frame, xlator_t *this, } /* Hack to notify higher layers of EOF. */ - if (stbuf.ia_size == 0) - op_errno = ENOENT; - else if ((offset + vec.iov_len) == stbuf.ia_size) - op_errno = ENOENT; - else if (offset > stbuf.ia_size) + if (!stbuf.ia_size || (offset + vec.iov_len) >= stbuf.ia_size) op_errno = ENOENT; op_ret = vec.iov_len; @@ -2119,22 +2300,6 @@ err: return op_ret; } -char* -_page_aligned_alloc (size_t size, char **aligned_buf) -{ - char *alloc_buf = NULL; - char *buf = NULL; - - alloc_buf = GF_CALLOC (1, (size + ALIGN_SIZE), gf_posix_mt_char); - if (!alloc_buf) - goto out; - /* page aligned buffer */ - buf = GF_ALIGN_BUF (alloc_buf, ALIGN_SIZE); - *aligned_buf = buf; -out: - return alloc_buf; -} - int32_t __posix_writev (int fd, struct iovec *vector, int count, off_t startoff, int odirect) @@ -2183,6 +2348,48 @@ err: return op_ret; } +dict_t* +_fill_writev_xdata (fd_t *fd, dict_t *xdata, xlator_t *this, int is_append) +{ + dict_t *rsp_xdata = NULL; + int32_t ret = 0; + inode_t *inode = NULL; + + if (fd) + inode = fd->inode; + + if (!fd || !fd->inode || uuid_is_null (fd->inode->gfid)) { + gf_log_callingfn (this->name, GF_LOG_ERROR, "Invalid Args: " + "fd: %p inode: %p gfid:%s", fd, inode?inode:0, + inode?uuid_utoa(inode->gfid):"N/A"); + goto out; + } + + if (!xdata || !dict_get (xdata, GLUSTERFS_OPEN_FD_COUNT)) + goto out; + + rsp_xdata = dict_new(); + if (!rsp_xdata) + goto out; + + ret = dict_set_uint32 (rsp_xdata, GLUSTERFS_OPEN_FD_COUNT, + fd->inode->fd_count); + if (ret < 0) { + gf_log (this->name, GF_LOG_WARNING, "%s: Failed to set " + "dictionary value for %s", uuid_utoa (fd->inode->gfid), + GLUSTERFS_OPEN_FD_COUNT); + } + + ret = dict_set_uint32 (rsp_xdata, GLUSTERFS_WRITE_IS_APPEND, + is_append); + if (ret < 0) { + gf_log (this->name, GF_LOG_WARNING, "%s: Failed to set " + "dictionary value for %s", uuid_utoa (fd->inode->gfid), + GLUSTERFS_WRITE_IS_APPEND); + } +out: + return rsp_xdata; +} int32_t posix_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, @@ -2197,6 +2404,9 @@ posix_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, struct iatt preop = {0,}; struct iatt postop = {0,}; int ret = -1; + dict_t *rsp_xdata = NULL; + int is_append = 0; + gf_boolean_t locked = _gf_false; VALIDATE_OR_GOTO (frame, out); VALIDATE_OR_GOTO (this, out); @@ -2218,6 +2428,17 @@ posix_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, _fd = pfd->fd; + if (xdata && dict_get (xdata, GLUSTERFS_WRITE_IS_APPEND)) { + /* The write_is_append check and write must happen + atomically. Else another write can overtake this + write after the check and get written earlier. + + So lock before preop-stat and unlock after write. + */ + locked = _gf_true; + LOCK(&fd->inode->lock); + } + op_ret = posix_fdstat (this, _fd, &preop); if (op_ret == -1) { op_errno = errno; @@ -2227,8 +2448,19 @@ posix_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, goto out; } + if (locked) { + if (preop.ia_size == offset || (fd->flags & O_APPEND)) + is_append = 1; + } + op_ret = __posix_writev (_fd, vector, count, offset, (pfd->flags & O_DIRECT)); + + if (locked) { + UNLOCK (&fd->inode->lock); + locked = _gf_false; + } + if (op_ret < 0) { op_errno = -op_ret; op_ret = -1; @@ -2244,6 +2476,7 @@ posix_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, UNLOCK (&priv->lock); if (op_ret >= 0) { + rsp_xdata = _fill_writev_xdata (fd, xdata, this, is_append); /* wiretv successful, we also need to get the stat of * the file we wrote to */ @@ -2273,9 +2506,16 @@ posix_writev (call_frame_t *frame, xlator_t *this, fd_t *fd, out: + if (locked) { + UNLOCK (&fd->inode->lock); + locked = _gf_false; + } + STACK_UNWIND_STRICT (writev, frame, op_ret, op_errno, &preop, &postop, - NULL); + rsp_xdata); + if (rsp_xdata) + dict_unref (rsp_xdata); return 0; } @@ -2402,6 +2642,33 @@ out: } +int +posix_batch_fsync (call_frame_t *frame, xlator_t *this, + fd_t *fd, int datasync, dict_t *xdata) +{ + call_stub_t *stub = NULL; + struct posix_private *priv = NULL; + + priv = this->private; + + stub = fop_fsync_stub (frame, default_fsync, fd, datasync, xdata); + if (!stub) { + STACK_UNWIND_STRICT (fsync, frame, -1, ENOMEM, 0, 0, 0); + return 0; + } + + pthread_mutex_lock (&priv->fsync_mutex); + { + list_add_tail (&stub->list, &priv->fsyncs); + priv->fsync_queue_count++; + pthread_cond_signal (&priv->fsync_cond); + } + pthread_mutex_unlock (&priv->fsync_mutex); + + return 0; +} + + int32_t posix_fsync (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t datasync, dict_t *xdata) @@ -2413,6 +2680,7 @@ posix_fsync (call_frame_t *frame, xlator_t *this, int ret = -1; struct iatt preop = {0,}; struct iatt postop = {0,}; + struct posix_private *priv = NULL; DECLARE_OLD_FS_ID_VAR; @@ -2428,6 +2696,12 @@ posix_fsync (call_frame_t *frame, xlator_t *this, goto out; #endif + priv = this->private; + if (priv->batch_fsync_mode && xdata && dict_get (xdata, "batch-fsync")) { + posix_batch_fsync (frame, this, fd, datasync, xdata); + return 0; + } + ret = posix_fd_ctx_get (fd, this, &pfd); if (ret < 0) { op_errno = -ret; @@ -2653,8 +2927,9 @@ posix_getxattr (call_frame_t *frame, xlator_t *this, if (ret < 0) { op_ret = -1; op_errno = -ret; - gf_log (this->name, GF_LOG_WARNING, - "Failed to get rea filename (%s, %s): %s", + gf_log (this->name, (op_errno == ENOENT) ? + GF_LOG_DEBUG : GF_LOG_WARNING, + "Failed to get real filename (%s, %s): %s", loc->path, name, strerror (op_errno)); goto out; } @@ -2768,7 +3043,8 @@ posix_getxattr (call_frame_t *frame, xlator_t *this, "supported (try remounting" " brick with 'user_xattr' " "flag)"); - } else if (op_errno == ENOATTR) { + } else if (op_errno == ENOATTR || + op_errno == ENODATA) { gf_log (this->name, GF_LOG_DEBUG, "No such attribute:%s for file %s", key, real_path); @@ -4288,6 +4564,27 @@ posix_set_owner (xlator_t *this, uid_t uid, gid_t gid) return ret; } + +static int +set_batch_fsync_mode (struct posix_private *priv, const char *str) +{ + if (strcmp (str, "none") == 0) + priv->batch_fsync_mode = BATCH_NONE; + else if (strcmp (str, "syncfs") == 0) + priv->batch_fsync_mode = BATCH_SYNCFS; + else if (strcmp (str, "syncfs-single-fsync") == 0) + priv->batch_fsync_mode = BATCH_SYNCFS_SINGLE_FSYNC; + else if (strcmp (str, "syncfs-reverse-fsync") == 0) + priv->batch_fsync_mode = BATCH_SYNCFS_REVERSE_FSYNC; + else if (strcmp (str, "reverse-fsync") == 0) + priv->batch_fsync_mode = BATCH_REVERSE_FSYNC; + else + return -1; + + return 0; +} + + int reconfigure (xlator_t *this, dict_t *options) { @@ -4295,6 +4592,7 @@ reconfigure (xlator_t *this, dict_t *options) struct posix_private *priv = NULL; uid_t uid = -1; gid_t gid = -1; + char *batch_fsync_mode_str = NULL; priv = this->private; @@ -4302,6 +4600,18 @@ reconfigure (xlator_t *this, dict_t *options) GF_OPTION_RECONF ("brick-gid", gid, options, uint32, out); posix_set_owner (this, uid, gid); + GF_OPTION_RECONF ("batch-fsync-delay-usec", priv->batch_fsync_delay_usec, + options, uint32, out); + + GF_OPTION_RECONF ("batch-fsync-mode", batch_fsync_mode_str, + options, str, out); + + if (set_batch_fsync_mode (priv, batch_fsync_mode_str) != 0) { + gf_log (this->name, GF_LOG_ERROR, "Unknown mode string: %s", + batch_fsync_mode_str); + goto out; + } + GF_OPTION_RECONF ("linux-aio", priv->aio_configured, options, bool, out); @@ -4320,6 +4630,10 @@ reconfigure (xlator_t *this, dict_t *options) " fallback to <hostname>:<export>"); } + GF_OPTION_RECONF ("health-check-interval", priv->health_check_interval, + options, uint32, out); + posix_spawn_health_check_thread (this); + ret = 0; out: return ret; @@ -4349,6 +4663,7 @@ init (xlator_t *this) char *guuid = NULL; uid_t uid = -1; gid_t gid = -1; + char *batch_fsync_mode_str; dir_data = dict_get (this->options, "directory"); @@ -4501,7 +4816,7 @@ init (xlator_t *this) } } - size = sys_lgetxattr (dir_data->data, "system.posix_acl_access", + size = sys_lgetxattr (dir_data->data, POSIX_ACL_ACCESS_XATTR, NULL, 0); if ((size < 0) && (errno == ENOTSUP)) gf_log (this->name, GF_LOG_WARNING, @@ -4690,11 +5005,39 @@ init (xlator_t *this) " fallback to <hostname>:<export>"); } + _private->health_check_active = _gf_false; + GF_OPTION_INIT ("health-check-interval", + _private->health_check_interval, uint32, out); + if (_private->health_check_interval) + posix_spawn_health_check_thread (this); + pthread_mutex_init (&_private->janitor_lock, NULL); pthread_cond_init (&_private->janitor_cond, NULL); INIT_LIST_HEAD (&_private->janitor_fds); posix_spawn_janitor_thread (this); + + pthread_mutex_init (&_private->fsync_mutex, NULL); + pthread_cond_init (&_private->fsync_cond, NULL); + INIT_LIST_HEAD (&_private->fsyncs); + + ret = gf_thread_create (&_private->fsyncer, NULL, posix_fsyncer, this); + if (ret) { + gf_log (this->name, GF_LOG_ERROR, "fsyncer thread" + " creation failed (%s)", strerror (errno)); + goto out; + } + + GF_OPTION_INIT ("batch-fsync-mode", batch_fsync_mode_str, str, out); + + if (set_batch_fsync_mode (_private, batch_fsync_mode_str) != 0) { + gf_log (this->name, GF_LOG_ERROR, "Unknown mode string: %s", + batch_fsync_mode_str); + goto out; + } + + GF_OPTION_INIT ("batch-fsync-delay-usec", _private->batch_fsync_delay_usec, + uint32, out); out: return ret; } @@ -4762,6 +5105,7 @@ struct xlator_fops fops = { .fsetattr = posix_fsetattr, .fallocate = _posix_fallocate, .discard = posix_discard, + .zerofill = posix_zerofill, }; struct xlator_cbks cbks = { @@ -4815,5 +5159,34 @@ struct volume_options options[] = { .description = "return glusterd's node-uuid in pathinfo xattr" " string instead of hostname" }, + { + .key = {"health-check-interval"}, + .type = GF_OPTION_TYPE_INT, + .min = 0, + .default_value = "30", + .validate = GF_OPT_VALIDATE_MIN, + .description = "Interval in seconds for a filesystem health check, " + "set to 0 to disable" + }, + { .key = {"batch-fsync-mode"}, + .type = GF_OPTION_TYPE_STR, + .default_value = "reverse-fsync", + .description = "Possible values:\n" + "\t- syncfs: Perform one syncfs() on behalf oa batch" + "of fsyncs.\n" + "\t- syncfs-single-fsync: Perform one syncfs() on behalf of a batch" + " of fsyncs and one fsync() per batch.\n" + "\t- syncfs-reverse-fsync: Preform one syncfs() on behalf of a batch" + " of fsyncs and fsync() each file in the batch in reverse order.\n" + " in reverse order.\n" + "\t- reverse-fsync: Perform fsync() of each file in the batch in" + " reverse order." + }, + { .key = {"batch-fsync-delay-usec"}, + .type = GF_OPTION_TYPE_INT, + .default_value = "0", + .description = "Num of usecs to wait for aggregating fsync" + " requests", + }, { .key = {NULL} } }; |