BugLink: https://bugs.launchpad.net/bugs/1775165 [Note to upstream] I understand that this patch is a little long for -stable, but this patch series fixes a real issue, seen by real users, is testable, and is made up from upstream commits. Please consider it. [Impact] When userspace tasks which are processing fanotify permission events act incorrectly, the fsnotify_mark_srcu SRCU is held indefinitely which causes the whole notification subsystem to hang. This has been seen in production, and it can also be seen when running the Linux Test Project testsuite, specifically fanotify07. [Fix] Instead of holding the SRCU lock while waiting for userspace to respond, which may never happen, or not in the order we are expecting, we drop the fsnotify_mark_srcu SRCU lock before waiting for userspace response, and then reacquire the lock again when userspace responds. The fixes are from a series of upstream commits: 05f0e38724e8449184acd8fbf0473ee5a07adc6c (cherry-pick) 9385a84d7e1f658bb2d96ab798393e4b16268aaa (backport) abc77577a669f424c5d0c185b9994f2621c52aa4 (backport) The following are upstream commits necessary for the fixes to function: 35e481761cdc688dbee0ef552a13f49af8eba6cc (backport) 0918f1c309b86301605650c836ddd2021d311ae2 (cherry-pick) [Testcase] You can reproduce the problem pretty quickly with the Linux Test Project: Steps (with root): 1. sudo apt-get install git xfsprogs -y 2. git clone --depth=1 https://github.com/linux-test-project/ltp.git 3. cd ltp 4. make autotools 5. ./configure 6. make; make install 7. cd /opt/ltp 8. echo -e "fanotify07 fanotify07 \nfanotify08 fanotify08" > /tmp/jobs 9. ./runltp -f /tmp/jobs On a stock Xenial kernel, the system will hang, and the testcase will look like: <<<test_start>>> tag=fanotify07 stime=1554326200 cmdline="fanotify07 " contacts="" analysis=exit <<<test_output>>> tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Test timeouted, sending SIGKILL! Cannot kill test processes! Congratulation, likely test hit a kernel bug. Exitting uncleanly... <<<execution_status>>> initiation_status="ok" duration=350 termination_type=exited termination_id=1 corefile=no cutime=0 cstime=0 <<<test_end>>> Looking at dmesg, we see the following call stack [ 790.772792] LTP: starting fanotify07 (fanotify07 ) [ 960.140455] INFO: task fsnotify_mark:36 blocked for more than 120 seconds. [ 960.140867] Not tainted 4.4.0-142-generic #168-Ubuntu [ 960.141185] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 960.141498] fsnotify_mark D ffff8800b6703c98 0 36 2 0x00000000 [ 960.141516] ffff8800b6703c98 ffff88013a558a00 ffff8800b7797000 ffff8800b66f8000 [ 960.141524] ffff8800b6704000 7fffffffffffffff ffff8800b6703de0 ffff8800b66f8000 [ 960.141528] 0000000000000000 ffff8800b6703cb0 ffffffff8185cb45 ffff8800b6703de8 [ 960.141532] Call Trace: [ 960.141580] [<ffffffff8185cb45>] schedule+0x35/0x80 [ 960.141588] [<ffffffff818600f4>] schedule_timeout+0x1b4/0x270 [ 960.141617] [<ffffffff810f57ac>] ? mod_timer+0x10c/0x240 [ 960.141621] [<ffffffff8185c60d>] ? __schedule+0x30d/0x810 [ 960.141625] [<ffffffff8185d652>] wait_for_completion+0xb2/0x190 [ 960.141636] [<ffffffff810b1f10>] ? wake_up_q+0x70/0x70 [ 960.141641] [<ffffffff810eb140>] __synchronize_srcu+0x100/0x1a0 [ 960.141645] [<ffffffff810ea400>] ? trace_raw_output_rcu_utilization+0x60/0x60 [ 960.141664] [<ffffffff81260870>] ? fsnotify_put_mark+0x40/0x40 [ 960.141669] [<ffffffff810eb204>] synchronize_srcu+0x24/0x30 [ 960.141672] [<ffffffff812608f4>] fsnotify_mark_destroy+0x84/0x130 [ 960.141680] [<ffffffff810ca000>] ? wake_atomic_t_function+0x60/0x60 [ 960.141691] [<ffffffff810a6227>] kthread+0xe7/0x100 [ 960.141694] [<ffffffff8185c601>] ? __schedule+0x301/0x810 [ 960.141699] [<ffffffff810a6140>] ? kthread_create_on_node+0x1e0/0x1e0 [ 960.141703] [<ffffffff818618e5>] ret_from_fork+0x55/0x80 [ 960.141706] [<ffffffff810a6140>] ? kthread_create_on_node+0x1e0/0x1e0 The vanilla 4.4 kernel also shows the same call stack. On a patched kernel, the test will pass successfully, and there will be no messages in dmesg. [Regression Potential] This makes modifications to how locking is performed in fsnotify / fanotify and there may be some cause for regression. Running all fanotify Linux Test Project tests shows that there are no extra failures caused by the patches, and instead fewer failures are seen due to the bugfix. Running the entire Linux Test Project testsuite actually works and runs to completion, something which doesn't happen in a unpatched kernel since it will hang on the fanotify07 test. The patches are taken from upstream, and all necessary commits have been taken into account, so I am happy with the potential risks and that testing has been completed. Jan Kara (4): fsnotify: avoid spurious EMFILE errors from inotify_init() fsnotify: Provide framework for dropping SRCU lock in ->handle_event fsnotify: Pass fsnotify_iter_info into handle_event handler fanotify: Release SRCU lock when waiting for userspace response Jeff Layton (1): fsnotify: turn fsnotify reaper thread into a workqueue job fs/notify/dnotify/dnotify.c | 3 +- fs/notify/fanotify/fanotify.c | 20 ++- fs/notify/fsnotify.c | 19 ++- fs/notify/fsnotify.h | 13 ++ fs/notify/group.c | 18 ++- fs/notify/inotify/inotify.h | 3 +- fs/notify/inotify/inotify_fsnotify.c | 3 +- fs/notify/inotify/inotify_user.c | 2 +- fs/notify/mark.c | 194 +++++++++++++++++++++------ include/linux/fsnotify_backend.h | 10 +- kernel/audit_fsnotify.c | 3 +- kernel/audit_tree.c | 3 +- kernel/audit_watch.c | 3 +- 13 files changed, 230 insertions(+), 64 deletions(-) -- 2.19.1
