From 7241756f94c975c89f6dc1023d2403f531e6f5ac Mon Sep 17 00:00:00 2001 From: Jeff Darcy Date: Mon, 25 Apr 2016 12:17:12 -0400 Subject: dht: add "nuke" functionality for efficient server-side deletion This is a backport of the following two patches (of which the second is a trivial adjustment to a timeout for a test added by the first). http://review.gluster.org/13878 http://review.gluster.org/13935 This turns a special xattr into an rmdir with flags set. When that hits the posix translator on the server side, that causes the file/directory to be moved into the special "landfill" directory. From there, the posix janitor thread will take care of deleting it entirely on the server side - traversing it recursively if necessary. A couple of secondary issues were fixed to make this effective. * FUSE now ensures that setxattr values are NUL terminated. * The janitor thread now gets woken up immediately when something is placed in 'landfill' instead of only when file descriptors need to be closed. * The default landfill-emptying interval was reduced to 10s. To use the feature, issue a setxattr something like this: setfattr -n glusterfs.dht.nuke -v "" /mnt/glusterfs/vol/some_dir The value doesn't actually matter; the mere receipt of a request with this key is sufficient. Some day it might be useful to allow setting a required value as a sort of password, so that only those who know it can access the underlying special functionality. Change-Id: I4132a30d1faa53a6682399ad1d9041e2c4519951 BUG: 1330241 Signed-off-by: Jeff Darcy Reviewed-on: http://review.gluster.org/14065 Smoke: Gluster Build System NetBSD-regression: NetBSD Build System CentOS-regression: Gluster Build System Reviewed-by: N Balachandran Reviewed-by: Raghavendra G --- xlators/cluster/dht/src/dht-common.c | 45 ++++++++++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+) (limited to 'xlators/cluster') diff --git a/xlators/cluster/dht/src/dht-common.c b/xlators/cluster/dht/src/dht-common.c index 96f540d721c..3421c1de657 100644 --- a/xlators/cluster/dht/src/dht-common.c +++ b/xlators/cluster/dht/src/dht-common.c @@ -3736,6 +3736,42 @@ err: return 0; } +int +dht_nuke_dir (call_frame_t *frame, xlator_t *this, loc_t *loc, data_t *tmp) +{ + if (!IA_ISDIR(loc->inode->ia_type)) { + DHT_STACK_UNWIND (setxattr, frame, -1, ENOTSUP, NULL); + return 0; + } + + /* Setxattr didn't need the parent, but rmdir does. */ + loc->parent = inode_parent (loc->inode, NULL, NULL); + if (!loc->parent) { + DHT_STACK_UNWIND (setxattr, frame, -1, ENOENT, NULL); + return 0; + } + gf_uuid_copy (loc->pargfid, loc->parent->gfid); + + if (!loc->name && loc->path) { + loc->name = strrchr (loc->path, '/'); + if (loc->name) { + ++(loc->name); + } + } + + /* + * We do this instead of calling dht_rmdir_do directly for two reasons. + * The first is that we want to reuse all of the initialization that + * dht_rmdir does, so if it ever changes we'll just follow along. The + * second (i.e. why we don't use STACK_WIND_TAIL) is so that we don't + * obscure the fact that we came in via this path instead of a genuine + * rmdir. That makes debugging just a tiny bit easier. + */ + STACK_WIND (frame, default_rmdir_cbk, this, this->fops->rmdir, + loc, 1, NULL); + + return 0; +} int dht_setxattr (call_frame_t *frame, xlator_t *this, @@ -3960,6 +3996,11 @@ dht_setxattr (call_frame_t *frame, xlator_t *this, goto err; } + tmp = dict_get (xattr, "glusterfs.dht.nuke"); + if (tmp) { + return dht_nuke_dir (frame, this, loc, tmp); + } + if (IA_ISDIR (loc->inode->ia_type)) { for (i = 0; i < call_cnt; i++) { @@ -8231,6 +8272,10 @@ dht_rmdir (call_frame_t *frame, xlator_t *this, loc_t *loc, int flags, goto err; } + if (flags) { + return dht_rmdir_do (frame, this); + } + for (i = 0; i < conf->subvolume_cnt; i++) { STACK_WIND (frame, dht_rmdir_opendir_cbk, conf->subvolumes[i], -- cgit