| Message ID | 20170512070838.5037-1-avagin@openvz.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, 2017-05-12 at 00:08 -0700, Andrei Vagin wrote: > With this patch, we don't try to umount all mounts of a tree together. > Instead of this we umount them one by one. In this case, we see a significant > improvement in performance for the worsе case. Indeed, umount has been very slow for a while now. Even a moderately large number of mounts (~10000) become painfully slow. Re you still perusing this? Anything I can do to help? Eric, what are your thoughts on this latest attempt? > > The reason of this optimization is that umount() can hold namespace_sem > for a long time, this semaphore is global, so it affects all users. > Recently Eric W. Biederman added a per mount namespace limit on the > number of mounts. The default number of mounts allowed per mount > namespace at 100,000. Currently this value is allowed to construct a tree > which requires hours to be umounted. > > In a worse case the current complexity of umount_tree() is O(n^3). > * Enumirate all mounts in a target tree (propagate_umount) > * Enumirate mounts to find where these changes have to > be propagated (mark_umount_candidates) > * Enumirate mounts to find a requered mount by parent and dentry > (__lookup_mnt). __lookup_mnt() searches a mount in m_hash, but > the number of mounts is much bigger than a size of the hash. > > The worse case is when all mounts from the tree live in the same shared > group. In this case we have to enumirate all mounts on each step. > > There is CVE-2016-6213 about this case. > > Here are results for the kernel with this patch > $ for i in `seq 10 15`; do unshare -m sh ./run.sh $i; done > umount -l /mnt/1 -> 0:00.00 > umount -l /mnt/1 -> 0:00.01 > umount -l /mnt/1 -> 0:00.01 > umount -l /mnt/1 -> 0:00.03 > umount -l /mnt/1 -> 0:00.07 > umount -l /mnt/1 -> 0:00.14 > > Here are results for the kernel without this patch > $ for i in `seq 10 15`; do unshare -m sh ./run.sh $i; done > umount -l /mnt/1 -> 0:00.04 > umount -l /mnt/1 -> 0:00.17 > umount -l /mnt/1 -> 0:00.75 > umount -l /mnt/1 -> 0:05.96 > umount -l /mnt/1 -> 0:34.40 > umount -l /mnt/1 -> 3:46.27 > > And here is a test script: > $ cat run.sh > set -e -m > > mount -t tmpfs zdtm /mnt > mkdir -p /mnt/1 /mnt/2 > mount -t tmpfs zdtm /mnt/1 > mount --make-shared /mnt/1 > mkdir /mnt/1/1 > > for i in `seq $1`; do > ./mount --bind /mnt/1/1 /mnt/1/1 > done > > echo -n "umount -l /mnt/1 -> " > /usr/bin/time -f '%E' ./umount -l /mnt/1 > > And we need these simple mount and umount tools, because the standard > ones read /proc/self/mountinfo, but this is extremely slow when we have > thousands of mounts. > $ cat mount.c > #include <sys/mount.h> > #include <stdlib.h> > > int main(int argc, char **argv) > { > return mount(argv[2], argv[3], NULL, MS_BIND, NULL); > } > > $ cat umount.c > #include <sys/mount.h> > > int main(int argc, char **argv) > { > return umount2(argv[2], MNT_DETACH); > } > > Here is a previous attempt to optimize this code: > https://lkml.org/lkml/2016/10/10/495 > > Signed-off-by: Andrei Vagin <avagin@openvz.org> > --- > fs/namespace.c | 81 +++++++++++++++++++++++++++++++------------------------ > --- > 1 file changed, 43 insertions(+), 38 deletions(-) > > diff --git a/fs/namespace.c b/fs/namespace.c > index 3bf0cd2..4e6f258 100644 > --- a/fs/namespace.c > +++ b/fs/namespace.c > @@ -1474,56 +1474,61 @@ static bool disconnect_mount(struct mount *mnt, enum > umount_tree_flags how) > */ > static void umount_tree(struct mount *mnt, enum umount_tree_flags how) > { > - LIST_HEAD(tmp_list); > struct mount *p; > + int done = 0; > > if (how & UMOUNT_PROPAGATE) > propagate_mount_unlock(mnt); > > /* Gather the mounts to umount */ > - for (p = mnt; p; p = next_mnt(p, mnt)) { > + while (!done) { > + LIST_HEAD(tmp_list); > + > + p = mnt; > + while (!list_empty(&p->mnt_mounts)) > + p = list_entry(p->mnt_mounts.next, struct mount, > mnt_child); > + > p->mnt.mnt_flags |= MNT_UMOUNT; > list_move(&p->mnt_list, &tmp_list); > - } > - > - /* Hide the mounts from mnt_mounts */ > - list_for_each_entry(p, &tmp_list, mnt_list) { > list_del_init(&p->mnt_child); > - } > > - /* Add propogated mounts to the tmp_list */ > - if (how & UMOUNT_PROPAGATE) > - propagate_umount(&tmp_list); > - > - while (!list_empty(&tmp_list)) { > - struct mnt_namespace *ns; > - bool disconnect; > - p = list_first_entry(&tmp_list, struct mount, mnt_list); > - list_del_init(&p->mnt_expire); > - list_del_init(&p->mnt_list); > - ns = p->mnt_ns; > - if (ns) { > - ns->mounts--; > - __touch_mnt_namespace(ns); > - } > - p->mnt_ns = NULL; > - if (how & UMOUNT_SYNC) > - p->mnt.mnt_flags |= MNT_SYNC_UMOUNT; > - > - disconnect = disconnect_mount(p, how); > - > - pin_insert_group(&p->mnt_umount, &p->mnt_parent->mnt, > - disconnect ? &unmounted : NULL); > - if (mnt_has_parent(p)) { > - mnt_add_count(p->mnt_parent, -1); > - if (!disconnect) { > - /* Don't forget about p */ > - list_add_tail(&p->mnt_child, &p->mnt_parent- > >mnt_mounts); > - } else { > - umount_mnt(p); > + /* Add propogated mounts to the tmp_list */ > + if (how & UMOUNT_PROPAGATE) > + propagate_umount(&tmp_list); > + > + if (p == mnt) > + done = 1; > + > + while (!list_empty(&tmp_list)) { > + struct mnt_namespace *ns; > + bool disconnect; > + p = list_first_entry(&tmp_list, struct mount, > mnt_list); > + list_del_init(&p->mnt_expire); > + list_del_init(&p->mnt_list); > + ns = p->mnt_ns; > + if (ns) { > + ns->mounts--; > + __touch_mnt_namespace(ns); > + } > + p->mnt_ns = NULL; > + if (how & UMOUNT_SYNC) > + p->mnt.mnt_flags |= MNT_SYNC_UMOUNT; > + > + disconnect = disconnect_mount(p, how); > + > + pin_insert_group(&p->mnt_umount, &p->mnt_parent->mnt, > + disconnect ? &unmounted : NULL); > + if (mnt_has_parent(p)) { > + mnt_add_count(p->mnt_parent, -1); > + if (!disconnect) { > + /* Don't forget about p */ > + list_add_tail(&p->mnt_child, &p- > >mnt_parent->mnt_mounts); > + } else { > + umount_mnt(p); > + } > } > + change_mnt_propagation(p, MS_PRIVATE); > } > - change_mnt_propagation(p, MS_PRIVATE); > } > } >
Ian Kent <raven@themaw.net> writes: > On Fri, 2017-05-12 at 00:08 -0700, Andrei Vagin wrote: >> With this patch, we don't try to umount all mounts of a tree together. >> Instead of this we umount them one by one. In this case, we see a significant >> improvement in performance for the worsе case. > > Indeed, umount has been very slow for a while now. > Even a moderately large number of mounts (~10000) become painfully slow. > > Re you still perusing this? > Anything I can do to help? > > Eric, what are your thoughts on this latest attempt? I have something slightly more recent. Please checkout my for-next branch of my userns tree: git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace.git for-next There is one open area of semantics that I looking at with Ram Pai in the hopes we can drive consensus before we take any patches for better checkpoint-restart support. Eric
diff --git a/fs/namespace.c b/fs/namespace.c index 3bf0cd2..4e6f258 100644 --- a/fs/namespace.c +++ b/fs/namespace.c @@ -1474,56 +1474,61 @@ static bool disconnect_mount(struct mount *mnt, enum umount_tree_flags how) */ static void umount_tree(struct mount *mnt, enum umount_tree_flags how) { - LIST_HEAD(tmp_list); struct mount *p; + int done = 0; if (how & UMOUNT_PROPAGATE) propagate_mount_unlock(mnt); /* Gather the mounts to umount */ - for (p = mnt; p; p = next_mnt(p, mnt)) { + while (!done) { + LIST_HEAD(tmp_list); + + p = mnt; + while (!list_empty(&p->mnt_mounts)) + p = list_entry(p->mnt_mounts.next, struct mount, mnt_child); + p->mnt.mnt_flags |= MNT_UMOUNT; list_move(&p->mnt_list, &tmp_list); - } - - /* Hide the mounts from mnt_mounts */ - list_for_each_entry(p, &tmp_list, mnt_list) { list_del_init(&p->mnt_child); - } - /* Add propogated mounts to the tmp_list */ - if (how & UMOUNT_PROPAGATE) - propagate_umount(&tmp_list); - - while (!list_empty(&tmp_list)) { - struct mnt_namespace *ns; - bool disconnect; - p = list_first_entry(&tmp_list, struct mount, mnt_list); - list_del_init(&p->mnt_expire); - list_del_init(&p->mnt_list); - ns = p->mnt_ns; - if (ns) { - ns->mounts--; - __touch_mnt_namespace(ns); - } - p->mnt_ns = NULL; - if (how & UMOUNT_SYNC) - p->mnt.mnt_flags |= MNT_SYNC_UMOUNT; - - disconnect = disconnect_mount(p, how); - - pin_insert_group(&p->mnt_umount, &p->mnt_parent->mnt, - disconnect ? &unmounted : NULL); - if (mnt_has_parent(p)) { - mnt_add_count(p->mnt_parent, -1); - if (!disconnect) { - /* Don't forget about p */ - list_add_tail(&p->mnt_child, &p->mnt_parent->mnt_mounts); - } else { - umount_mnt(p); + /* Add propogated mounts to the tmp_list */ + if (how & UMOUNT_PROPAGATE) + propagate_umount(&tmp_list); + + if (p == mnt) + done = 1; + + while (!list_empty(&tmp_list)) { + struct mnt_namespace *ns; + bool disconnect; + p = list_first_entry(&tmp_list, struct mount, mnt_list); + list_del_init(&p->mnt_expire); + list_del_init(&p->mnt_list); + ns = p->mnt_ns; + if (ns) { + ns->mounts--; + __touch_mnt_namespace(ns); + } + p->mnt_ns = NULL; + if (how & UMOUNT_SYNC) + p->mnt.mnt_flags |= MNT_SYNC_UMOUNT; + + disconnect = disconnect_mount(p, how); + + pin_insert_group(&p->mnt_umount, &p->mnt_parent->mnt, + disconnect ? &unmounted : NULL); + if (mnt_has_parent(p)) { + mnt_add_count(p->mnt_parent, -1); + if (!disconnect) { + /* Don't forget about p */ + list_add_tail(&p->mnt_child, &p->mnt_parent->mnt_mounts); + } else { + umount_mnt(p); + } } + change_mnt_propagation(p, MS_PRIVATE); } - change_mnt_propagation(p, MS_PRIVATE); } }
With this patch, we don't try to umount all mounts of a tree together. Instead of this we umount them one by one. In this case, we see a significant improvement in performance for the worsе case. The reason of this optimization is that umount() can hold namespace_sem for a long time, this semaphore is global, so it affects all users. Recently Eric W. Biederman added a per mount namespace limit on the number of mounts. The default number of mounts allowed per mount namespace at 100,000. Currently this value is allowed to construct a tree which requires hours to be umounted. In a worse case the current complexity of umount_tree() is O(n^3). * Enumirate all mounts in a target tree (propagate_umount) * Enumirate mounts to find where these changes have to be propagated (mark_umount_candidates) * Enumirate mounts to find a requered mount by parent and dentry (__lookup_mnt). __lookup_mnt() searches a mount in m_hash, but the number of mounts is much bigger than a size of the hash. The worse case is when all mounts from the tree live in the same shared group. In this case we have to enumirate all mounts on each step. There is CVE-2016-6213 about this case. Here are results for the kernel with this patch $ for i in `seq 10 15`; do unshare -m sh ./run.sh $i; done umount -l /mnt/1 -> 0:00.00 umount -l /mnt/1 -> 0:00.01 umount -l /mnt/1 -> 0:00.01 umount -l /mnt/1 -> 0:00.03 umount -l /mnt/1 -> 0:00.07 umount -l /mnt/1 -> 0:00.14 Here are results for the kernel without this patch $ for i in `seq 10 15`; do unshare -m sh ./run.sh $i; done umount -l /mnt/1 -> 0:00.04 umount -l /mnt/1 -> 0:00.17 umount -l /mnt/1 -> 0:00.75 umount -l /mnt/1 -> 0:05.96 umount -l /mnt/1 -> 0:34.40 umount -l /mnt/1 -> 3:46.27 And here is a test script: $ cat run.sh set -e -m mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 for i in `seq $1`; do ./mount --bind /mnt/1/1 /mnt/1/1 done echo -n "umount -l /mnt/1 -> " /usr/bin/time -f '%E' ./umount -l /mnt/1 And we need these simple mount and umount tools, because the standard ones read /proc/self/mountinfo, but this is extremely slow when we have thousands of mounts. $ cat mount.c #include <sys/mount.h> #include <stdlib.h> int main(int argc, char **argv) { return mount(argv[2], argv[3], NULL, MS_BIND, NULL); } $ cat umount.c #include <sys/mount.h> int main(int argc, char **argv) { return umount2(argv[2], MNT_DETACH); } Here is a previous attempt to optimize this code: https://lkml.org/lkml/2016/10/10/495 Signed-off-by: Andrei Vagin <avagin@openvz.org> --- fs/namespace.c | 81 +++++++++++++++++++++++++++++++--------------------------- 1 file changed, 43 insertions(+), 38 deletions(-)