has been pointing out constant problems. The changes have been going into
the stack tracer, but it has been discovered that the problem isn't
with the stack tracer itself, but it is with calling save_stack_trace()
from within the internals of RCU. The stack tracer is the one that
can trigger the issue the easiest, but examining the problem further,
it could also happen from a WARN() in the wrong place, or even if
an NMI happened in this area and it did an rcu_read_lock().
The critical area is where RCU is not watching. Which can happen while
going to and from idle, or bringing up or taking down a CPU.
The final fix was to put the protection in kernel_text_address() as it
is the one that requires RCU to be watching while doing the stack trace.
To make this work properly, Paul had to allow rcu_irq_enter() happen after
rcu_nmi_enter(). This should have been done anyway, since an NMI can
page fault (reading vmalloc area), and a page fault triggers rcu_irq_enter().
One patch is just a consolidation of code so that the fix only needed
to be done in one location.
-----BEGIN PGP SIGNATURE-----
iQFIBAABCAAyFiEEQEw9Eu0DdyUUkuUUybkF8mrZjcsFAlnGyXoUHHJvc3RlZHRA
Z29vZG1pcy5vcmcACgkQybkF8mrZjctKtwf8CeKGqOdlqkZEafIpWaIASXmAVMO/
WE+hQK+rCydWFvzADgb/rOmsR0ou8WGEXcuUPxVxmvMyqhKhZ6AU1hE/7Y8P0pMq
F4bev+j3lAJC65ezFAh+ZQcIjaRIH4MFVPsUTaibSPSN7xziMNIpbf9VOVfpUm8A
jf9p6YAmyhFVi6DstCc29SWnywEVwC2ZWRVKRPXKry8/dPxjfVcLclGX680Eqi9I
EnYaOdC/mGbtvHPOUSs/P0cfxExHmyEErQHeOV8FPymj6KJ6+KoYIiELNlTHUBj/
eeKzrHc/b3j+lz0RPlA8WxYmpmEm4SE5cV3vRebdBNUBrABSN1RxeOozyQ==
=1KkS
-----END PGP SIGNATURE-----
Merge tag 'trace-v4.14-rc1-2' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull tracing fixes from Steven Rostedt:
"Stack tracing and RCU has been having issues with each other and
lockdep has been pointing out constant problems.
The changes have been going into the stack tracer, but it has been
discovered that the problem isn't with the stack tracer itself, but it
is with calling save_stack_trace() from within the internals of RCU.
The stack tracer is the one that can trigger the issue the easiest,
but examining the problem further, it could also happen from a WARN()
in the wrong place, or even if an NMI happened in this area and it did
an rcu_read_lock().
The critical area is where RCU is not watching. Which can happen while
going to and from idle, or bringing up or taking down a CPU.
The final fix was to put the protection in kernel_text_address() as it
is the one that requires RCU to be watching while doing the stack
trace.
To make this work properly, Paul had to allow rcu_irq_enter() happen
after rcu_nmi_enter(). This should have been done anyway, since an NMI
can page fault (reading vmalloc area), and a page fault triggers
rcu_irq_enter().
One patch is just a consolidation of code so that the fix only needed
to be done in one location"
* tag 'trace-v4.14-rc1-2' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
tracing: Remove RCU work arounds from stack tracer
extable: Enable RCU if it is not watching in kernel_text_address()
extable: Consolidate *kernel_text_address() functions
rcu: Allow for page faults in NMI handlers
Now that there are potentially long delays between when a remoteport or
targetport delete calls is made and when the callback occurs (dev_loss_tmo
timeout), no longer block in the delete routines and move the final nport
puts to the callbacks.
Moved the fcloop_nport_get/put/free routines to avoid forward declarations.
Ensure port_info structs used in registrations are nulled in case fields
are not set (ex: devloss_tmo values).
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Comments were incorrect:
- defer_rcv was in host port template. moved to target port template
- Added Mandatory statements for target port template items
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When searching for queue id's ensure they are within the expected range.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Avoid calling the put routine, as it may traverse to free routines while
holding the target lock.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
By calling nvme_stop_ctrl on a already failed controller will wait for the
scan work to complete (only by identify timeout expiration which is 60
seconds). This is unnecessary when we already know that the controller has
failed.
Reported-by: Yi Zhang <yizhan@redhat.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we failed to transition to state LIVE after a successful reconnect,
then controller deletion already started. In this case there is no
point moving forward with reconnect.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
async_event_work might race as it is executed from two different
workqueues at the moment.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Fix bug in sqhd patch.
It wasn't the sq that was at risk. In the case where the admin queue
connect command fails, the sq->size field is not set. Therefore, this
becomes a divide by zero error.
Add a quick check to bypass under this failure condition.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The switch to rhashtables (commit 88ffbf3e03) broke the debugfs glock
dump (/sys/kernel/debug/gfs2/<device>/glocks) for dumps bigger than a
single buffer: the right function for restarting an rhashtable iteration
from the beginning of the hash table is rhashtable_walk_enter;
rhashtable_walk_stop + rhashtable_walk_start will just resume from the
current position.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Cc: stable@vger.kernel.org # v4.3+
When timer_create() fails on a bootime or realtime clock, setup_timer()
returns 0 as if timer has been set. Callers wait forever for the timer
to expire.
This hang is seen on a system that doesn't have support for:
CLOCK_REALTIME_ALARM ABSTIME missing CAP_WAKE_ALARM? : [UNSUPPORTED]
Test hangs waiting for a timer that hasn't been set to expire. Fix
setup_timer() to return 1, add handling in callers to detect the
unsupported case and return 0 without waiting to not fail the test.
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
do_timer_oneshot() uses select() as a timer with FD_SETSIZE and readfs
is cleared with FD_ZERO without FD_SET.
When stdout and stderr are redirected, the test hangs in select forever.
Fix the problem calling select() with readfds empty and nfds zero. This
is sufficient for using select() for timer.
With this fix "./set-timer-lat > /dev/null 2>&1" no longer hangs.
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
Acked-by: Greg Hackmann <ghackmann@google.com>
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
to fix the following issue:
------------------
TAP version 13
selftests: run_tests.sh
========================================
selftests: Warning: file run_tests.sh is not executable, correct this.
not ok 1..1 selftests: run_tests.sh [FAIL]
------------------
Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
The 2.26 release of glibc changed how siginfo_t is defined, and the earlier
work-around to using the kernel definition are no longer needed. The old
way needs to stay around for a while, though.
Reported-by: Seth Forshee <seth.forshee@canonical.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Will Drewry <wad@chromium.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: linux-kselftest@vger.kernel.org
Cc: stable@vger.kernel.org
Signed-off-by: Kees Cook <keescook@chromium.org>
Tested-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
Fix for loops in targets to run silently to avoid cluttering the test
results.
Suppresses the following from targets:
for DIR in functional; do \
BUILD_TARGET=./tools/testing/selftests/futex/$DIR; \
mkdir $BUILD_TARGET -p; \
make OUTPUT=$BUILD_TARGET -C $DIR all;\
done
./tools/testing/selftests/futex/run.sh
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
Reviewed-by: Darren Hart (VMware) <dvhart@infradead.org>
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
Fix for loops in targets to run silently to avoid cluttering the test
results.
Suppresses the following from targets: e.g run from breakpoints
for TARGET in breakpoints; do \
BUILD_TARGET=$BUILD/$TARGET; \
mkdir $BUILD_TARGET -p; \
make OUTPUT=$BUILD_TARGET -C $TARGET;\
done;
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
Use full path including $(OUTPUT) to run tests from Makefile for
normal case when objects reside in the source tree as well as when
objects are relocated with make O=dir. In both cases $(OUTPUT) will
be set correctly by lib.mk.
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
For make O=dir run_tests to work, test scripts from sub-directories
need to be copied over to the object directory. Running tests from the
object directory is necessary to avoid making the source tree dirty.
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
Reviewed-by: Darren Hart (VMware) <dvhart@infradead.org>
Signed-off-by: Shuah Khan <shuahkh@osg.samsung.com>
part_stat_show takes a part device not a disk, so we should use
part_to_disk.
Fixes: d62e26b3ffd2("block: pass in queue to inflight accounting")
Cc: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Omar Sandoval <osandov@fb.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently when mixing buffered reads and asynchronous direct writes it
is possible to end up with the situation where we have stale data in the
page cache while the new data is already written to disk. This is
permanent until the affected pages are flushed away. Despite the fact
that mixing buffered and direct IO is ill-advised it does pose a thread
for a data integrity, is unexpected and should be fixed.
Fix this by deferring completion of asynchronous direct writes to a
process context in the case that there are mapped pages to be found in
the inode. Later before the completion in dio_complete() invalidate
the pages in question. This ensures that after the completion the pages
in the written area are either unmapped, or populated with up-to-date
data. Also do the same for the iomap case which uses
iomap_dio_complete() instead.
This has a side effect of deferring the completion to a process context
for every AIO DIO that happens on inode that has pages mapped. However
since the consensus is that this is ill-advised practice the performance
implication should not be a problem.
This was based on proposal from Jeff Moyer, thanks!
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To support sqhd, for initiators that are following the spec and
paying attention to sqhd vs their sqtail values:
- add sqhd to struct nvmet_sq
- initialize sqhd to 0 in nvmet_sq_setup
- rather than propagate the 0's-based qsize value from the connect message
which requires a +1 in every sqhd update, and as nothing else references
it, convert to 1's-based value in nvmt_sq/cq_setup() calls.
- validate connect message sqsize being non-zero per spec.
- updated assign sqhd for every completion that goes back.
Also remove handling the NULL sq case in __nvmet_req_complete, as it can't
happen with the current code.
Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, driver code allows user to set 0 as KATO
(Keep Alive TimeOut), but this is not being respected.
This patch enforces the expected behavior.
Signed-off-by: Guilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently the nvme_req_needs_retry() applies several checks to see if
a retry is allowed. On of those is whether the current time has exceeded
the start time of the io plus the timeout length. This check, if an io
times out, means there is never a retry allowed for the io. Which means
applications see the io failure.
Remove this check and allow the io to timeout, like it does on other
protocols, and retries to be made.
On the FC transport, a frame can be lost for an individual io, and there
may be no other errors that escalate for the connection/association.
The io will timeout, which causes the transport to escalate into creating
a new association, but the io that timed out, due to this retry logic, has
already failed back to the application and things are hosed.
Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If an nvme async_event command completes, in most cases, a new
async event is posted. However, if the controller enters a
resetting or reconnecting state, there is nothing to block the
scheduled work element from posting the async event again. Nor are
there calls from the transport to stop async events when an
association dies.
In the case of FC, where the association is torn down, the aer must
be aborted on the FC link and completes through the normal job
completion path. Thus the terminated async event ends up being
rescheduled even though the controller isn't in a valid state for
the aer, and the reposting gets the transport into a partially torn
down data structure.
It's possible to hit the scenario on rdma, although much less likely
due to an aer completing right as the association is terminated and
as the association teardown reclaims the blk requests via
nvme_cancel_request() so its immediate, not a link-related action
like on FC.
Fix by putting controller state checks in both the async event
completion routine where it schedules the async event and in the
async event work routine before it calls into the transport. It's
effectively a "stop_async_events()" behavior. The transport, when
it creates a new association with the subsystem will transition
the state back to live and is already restarting the async event
posting.
Signed-off-by: James Smart <james.smart@broadcom.com>
[hch: remove taking a lock over reading the controller state]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The WARN_ONCE macro returns true if the condition is true, not if the
warn was raised, so we're printing the scatter list every time it's
invalid. This is excessive and makes debugging harder, so this patch
prints it just once.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A spurious interrupt before the nvme driver has initialized the completion
queue may inadvertently cause the driver to believe it has a completion
to process. This may result in a NULL dereference since the nvmeq's tags
are not set at this point.
The patch initializes the host's CQ memory so that a spurious interrupt
isn't mistaken for a real completion.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
fc transport is treating NVMET_NR_QUEUES as maximum queue count, e.g.
admin queue plus NVMET_NR_QUEUES-1 io queues. But NVMET_NR_QUEUES is
the number of io queues, so maximum queue count is really
NVMET_NR_QUEUES+1.
Fix the handling in the target fc transport
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Sync with NVM Express spec change and FC-NVME 1.18.
FC transport sets SGL type to Transport SGL Data Block Descriptor and
subtype to transport-specific value 0x0A.
Removed the warn-on's on the PRP fields. They are unneeded. They were
to check for values from the upper layer that weren't set right, and
for the most part were fine. But, with Async events, which reuse the
same structure and 2nd time issued the SGL overlay converted them to
the Transport SGL values - the warn-on's were errantly firing.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add transport SGL defintions from NVMe TP 4008, required for
the final NVMe-FC standard.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The NVM express group recinded the reserved range for the transport.
Remove the FC-centric values that had been defined.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The qla2xxx driver uses the FC-specific error when it needed to return an
error to the FC-NVME transport. Convert to use a generic value instead.
Signed-off-by: James Smart <james.smart@broadcom.com>
Acked-by: Himanshu Madhani <himanshu.madhani@cavium.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The lpfc driver uses the FC-specific error when it needed to return an
error to the FC-NVME transport. Convert to use a generic value instead.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The FC-NVME transport loopback test module used the FC-specific error
codes in cases where it emulated a transport abort case. Instead of
using the FC-specific values, now use a generic value (NVME_SC_INTERNAL).
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The FC-NVME target transport used the FC-specific error codes in
return codes when the transport or lldd failed. Instead of using the
FC-specific values, now use a generic value (NVME_SC_INTERNAL).
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The FC-NVME transport used the FC-specific error codes in cases where
it had to fabricate an error to go back up stack. Instead of using the
FC-specific values, now use a generic value (NVME_SC_INTERNAL).
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When the request is completed, lo_complete_rq() checks cmd->use_aio.
However, if this is in fact an aio request, cmd->use_aio will have
already been reused as cmd->ref by lo_rw_aio*. Fix it by not using a
union. On x86_64, there's a hole after the union anyways, so this
doesn't make struct loop_cmd any bigger.
Fixes: 92d773324b ("block/loop: fix use after free")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The lockdep code had reported the following unsafe locking scenario:
CPU0 CPU1
---- ----
lock(s_active#228);
lock(&bdev->bd_mutex/1);
lock(s_active#228);
lock(&bdev->bd_mutex);
*** DEADLOCK ***
The deadlock may happen when one task (CPU1) is trying to delete a
partition in a block device and another task (CPU0) is accessing
tracing sysfs file (e.g. /sys/block/dm-1/trace/act_mask) in that
partition.
The s_active isn't an actual lock. It is a reference count (kn->count)
on the sysfs (kernfs) file. Removal of a sysfs file, however, require
a wait until all the references are gone. The reference count is
treated like a rwsem using lockdep instrumentation code.
The fact that a thread is in the sysfs callback method or in the
ioctl call means there is a reference to the opended sysfs or device
file. That should prevent the underlying block structure from being
removed.
Instead of using bd_mutex in the block_device structure, a new
blk_trace_mutex is now added to the request_queue structure to protect
access to the blk_trace structure.
Suggested-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Waiman Long <longman@redhat.com>
Acked-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Fix typo in patch subject line, and prune a comment detailing how
the code used to work.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In testing we noticed that nbd would spew if you ran a fio job against
the raw device itself. This is because fio calls a block device
specific ioctl, however the block layer will first pass this back to the
driver ioctl handler in case the driver wants to do something special.
Since the device was setup using netlink this caused us to spew every
time fio called this ioctl. Since we don't have special handling, just
error out for any non-nbd specific ioctl's that come in. This fixes the
spew.
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The job structure is allocated as part of the request, so we should not
free it in the error path of bsg_prepare_job.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Cc: stable@vger.kernel.org
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The code in __brd_direct_access multiplies the pgoff variable by page size
and divides it by 512. It can cause overflow on 32-bit architectures. The
overflow happens if we create ramdisk larger than 4G and use it as a
sparse device.
This patch replaces multiplication and division with multiplication by the
number of sectors per page.
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Fixes: 1647b9b959 ("brd: add dax_operations support")
Cc: stable@vger.kernel.org # 4.12+
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Fix the default for microblaze. Michal Simek mentioned default for
microblaze should be CPU_LITTLE_ENDIAN.
Fixes : commit 206d3642d8 ("arch/microblaze: add choice for endianness
and update Makefile")
Signed-off-by: Babu Moger <babu.moger@oracle.com>
Cc: Michal Simek <monstr@monstr.eu>
Signed-off-by: Michal Simek <michal.simek@xilinx.com>
Use vma_pages function on vma object instead of explicit computation.
Found by coccinelle spatch "api/vma_pages.cocci"
Signed-off-by: Thomas Meyer <thomas@m3y3r.de>
Signed-off-by: Michal Simek <michal.simek@xilinx.com>
Userspace can change the FPU state of a task using the ptrace() or
rt_sigreturn() system calls. Because reserved bits in the FPU state can
cause the XRSTOR instruction to fail, the kernel has to carefully
validate that no reserved bits or other invalid values are being set.
Unfortunately, there have been bugs in this validation code. For
example, we were not checking that the 'xcomp_bv' field in the
xstate_header was 0. As-is, such bugs are exploitable to read the FPU
registers of other processes on the system. To do so, an attacker can
create a task, assign to it an invalid FPU state, then spin in a loop
and monitor the values of the FPU registers. Because the task's FPU
registers are not being restored, sometimes the FPU registers will have
the values from another process.
This is likely to continue to be a problem in the future because the
validation done by the CPU instructions like XRSTOR is not immediately
visible to kernel developers. Nor will invalid FPU states ever be
encountered during ordinary use --- they will only be seen during
fuzzing or exploits. There can even be reserved bits outside the
xstate_header which are easy to forget about. For example, the MXCSR
register contains reserved bits, which were not validated by the
KVM_SET_XSAVE ioctl until commit a575813bfe ("KVM: x86: Fix load
damaged SSEx MXCSR register").
Therefore, mitigate this class of vulnerability by restoring the FPU
registers from init_fpstate if restoring from the task's state fails.
We actually used to do this, but it was (perhaps unwisely) removed by
commit 9ccc27a5d2 ("x86/fpu: Remove error return values from
copy_kernel_to_*regs() functions"). This new patch is also a bit
different. First, it only clears the registers, not also the bad
in-memory state; this is simpler and makes it easier to make the
mitigation cover all callers of __copy_kernel_to_fpregs(). Second, it
does the register clearing in an exception handler so that no extra
instructions are added to context switches. In fact, we *remove*
instructions, since previously we were always zeroing the register
containing 'err' even if CONFIG_X86_DEBUG_FPU was disabled.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Eric Biggers <ebiggers3@gmail.com>
Cc: Fenghua Yu <fenghua.yu@intel.com>
Cc: Kevin Hao <haokexin@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Halcrow <mhalcrow@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Wanpeng Li <wanpeng.li@hotmail.com>
Cc: Yu-cheng Yu <yu-cheng.yu@intel.com>
Cc: kernel-hardening@lists.openwall.com
Link: http://lkml.kernel.org/r/20170922174156.16780-4-ebiggers3@gmail.com
Link: http://lkml.kernel.org/r/20170923130016.21448-27-mingo@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
On x86, userspace can use the ptrace() or rt_sigreturn() system calls to
set a task's extended state (xstate) or "FPU" registers. ptrace() can
set them for another task using the PTRACE_SETREGSET request with
NT_X86_XSTATE, while rt_sigreturn() can set them for the current task.
In either case, registers can be set to any value, but the kernel
assumes that the XSAVE area itself remains valid in the sense that the
CPU can restore it.
However, in the case where the kernel is using the uncompacted xstate
format (which it does whenever the XSAVES instruction is unavailable),
it was possible for userspace to set the xcomp_bv field in the
xstate_header to an arbitrary value. However, all bits in that field
are reserved in the uncompacted case, so when switching to a task with
nonzero xcomp_bv, the XRSTOR instruction failed with a #GP fault. This
caused the WARN_ON_FPU(err) in copy_kernel_to_xregs() to be hit. In
addition, since the error is otherwise ignored, the FPU registers from
the task previously executing on the CPU were leaked.
Fix the bug by checking that the user-supplied value of xcomp_bv is 0 in
the uncompacted case, and returning an error otherwise.
The reason for validating xcomp_bv rather than simply overwriting it
with 0 is that we want userspace to see an error if it (incorrectly)
provides an XSAVE area in compacted format rather than in uncompacted
format.
Note that as before, in case of error we clear the task's FPU state.
This is perhaps non-ideal, especially for PTRACE_SETREGSET; it might be
better to return an error before changing anything. But it seems the
"clear on error" behavior is fine for now, and it's a little tricky to
do otherwise because it would mean we couldn't simply copy the full
userspace state into kernel memory in one __copy_from_user().
This bug was found by syzkaller, which hit the above-mentioned
WARN_ON_FPU():
WARNING: CPU: 1 PID: 0 at ./arch/x86/include/asm/fpu/internal.h:373 __switch_to+0x5b5/0x5d0
CPU: 1 PID: 0 Comm: swapper/1 Not tainted 4.13.0 #453
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
task: ffff9ba2bc8e42c0 task.stack: ffffa78cc036c000
RIP: 0010:__switch_to+0x5b5/0x5d0
RSP: 0000:ffffa78cc08bbb88 EFLAGS: 00010082
RAX: 00000000fffffffe RBX: ffff9ba2b8bf2180 RCX: 00000000c0000100
RDX: 00000000ffffffff RSI: 000000005cb10700 RDI: ffff9ba2b8bf36c0
RBP: ffffa78cc08bbbd0 R08: 00000000929fdf46 R09: 0000000000000001
R10: 0000000000000000 R11: 0000000000000000 R12: ffff9ba2bc8e42c0
R13: 0000000000000000 R14: ffff9ba2b8bf3680 R15: ffff9ba2bf5d7b40
FS: 00007f7e5cb10700(0000) GS:ffff9ba2bf400000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000004005cc CR3: 0000000079fd5000 CR4: 00000000001406e0
Call Trace:
Code: 84 00 00 00 00 00 e9 11 fd ff ff 0f ff 66 0f 1f 84 00 00 00 00 00 e9 e7 fa ff ff 0f ff 66 0f 1f 84 00 00 00 00 00 e9 c2 fa ff ff <0f> ff 66 0f 1f 84 00 00 00 00 00 e9 d4 fc ff ff 66 66 2e 0f 1f
Here is a C reproducer. The expected behavior is that the program spin
forever with no output. However, on a buggy kernel running on a
processor with the "xsave" feature but without the "xsaves" feature
(e.g. Sandy Bridge through Broadwell for Intel), within a second or two
the program reports that the xmm registers were corrupted, i.e. were not
restored correctly. With CONFIG_X86_DEBUG_FPU=y it also hits the above
kernel warning.
#define _GNU_SOURCE
#include <stdbool.h>
#include <inttypes.h>
#include <linux/elf.h>
#include <stdio.h>
#include <sys/ptrace.h>
#include <sys/uio.h>
#include <sys/wait.h>
#include <unistd.h>
int main(void)
{
int pid = fork();
uint64_t xstate[512];
struct iovec iov = { .iov_base = xstate, .iov_len = sizeof(xstate) };
if (pid == 0) {
bool tracee = true;
for (int i = 0; i < sysconf(_SC_NPROCESSORS_ONLN) && tracee; i++)
tracee = (fork() != 0);
uint32_t xmm0[4] = { [0 ... 3] = tracee ? 0x00000000 : 0xDEADBEEF };
asm volatile(" movdqu %0, %%xmm0\n"
" mov %0, %%rbx\n"
"1: movdqu %%xmm0, %0\n"
" mov %0, %%rax\n"
" cmp %%rax, %%rbx\n"
" je 1b\n"
: "+m" (xmm0) : : "rax", "rbx", "xmm0");
printf("BUG: xmm registers corrupted! tracee=%d, xmm0=%08X%08X%08X%08X\n",
tracee, xmm0[0], xmm0[1], xmm0[2], xmm0[3]);
} else {
usleep(100000);
ptrace(PTRACE_ATTACH, pid, 0, 0);
wait(NULL);
ptrace(PTRACE_GETREGSET, pid, NT_X86_XSTATE, &iov);
xstate[65] = -1;
ptrace(PTRACE_SETREGSET, pid, NT_X86_XSTATE, &iov);
ptrace(PTRACE_CONT, pid, 0, 0);
wait(NULL);
}
return 1;
}
Note: the program only tests for the bug using the ptrace() system call.
The bug can also be reproduced using the rt_sigreturn() system call, but
only when called from a 32-bit program, since for 64-bit programs the
kernel restores the FPU state from the signal frame by doing XRSTOR
directly from userspace memory (with proper error checking).
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: <stable@vger.kernel.org> [v3.17+]
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Eric Biggers <ebiggers3@gmail.com>
Cc: Fenghua Yu <fenghua.yu@intel.com>
Cc: Kevin Hao <haokexin@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Halcrow <mhalcrow@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Wanpeng Li <wanpeng.li@hotmail.com>
Cc: Yu-cheng Yu <yu-cheng.yu@intel.com>
Cc: kernel-hardening@lists.openwall.com
Fixes: 0b29643a58 ("x86/xsaves: Change compacted format xsave area header")
Link: http://lkml.kernel.org/r/20170922174156.16780-2-ebiggers3@gmail.com
Link: http://lkml.kernel.org/r/20170923130016.21448-25-mingo@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull x86 fixes from Ingo Molnar:
"Another round of CR3/PCID related fixes (I think this addresses all
but one of the known problems with PCID support), an objtool fix plus
a Clang fix that (finally) solves all Clang quirks to build a bootable
x86 kernel as-is"
* 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/asm: Fix inline asm call constraints for Clang
objtool: Handle another GCC stack pointer adjustment bug
x86/mm/32: Load a sane CR3 before cpu_init() on secondary CPUs
x86/mm/32: Move setup_clear_cpu_cap(X86_FEATURE_PCID) earlier
x86/mm/64: Stop using CR3.PCID == 0 in ASID-aware code
x86/mm: Factor out CR3-building code
Pull irq fixes from Ingo Molnar:
"Three irqchip driver fixes, and an affinity mask helper function bug
fix affecting x86"
* 'irq-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
Revert "genirq: Restrict effective affinity to interrupts actually using it"
irqchip.mips-gic: Fix shared interrupt mask writes
irqchip/gic-v4: Fix building with ancient gcc
irqchip/gic-v3: Iterate over possible CPUs by for_each_possible_cpu()
