ARC: mm: do_page_fault fixes #1: relinquish mmap_sem if signal arrives while handle_m...
authorVineet Gupta <vgupta@synopsys.com>
Tue, 11 Dec 2018 00:56:45 +0000 (16:56 -0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 16 Sep 2019 06:22:02 +0000 (08:22 +0200)
commit8c6fb55abf2c8272cae2464b0d3f0537792e79c3
tree898ed687b766028d8bc81efdddb92816ba37d622
parent96af7d92d8d27d97edc65ce53d8dd8c4d0f4fa58
ARC: mm: do_page_fault fixes #1: relinquish mmap_sem if signal arrives while handle_mm_fault

[ Upstream commit 4d447455e73b47c43dd35fcc38ed823d3182a474 ]

do_page_fault() forgot to relinquish mmap_sem if a signal came while
handling handle_mm_fault() - due to say a ctl+c or oom etc.
This would later cause a deadlock by acquiring it twice.

This came to light when running libc testsuite tst-tls3-malloc test but
is likely also the cause for prior seen LTP failures. Using lockdep
clearly showed what the issue was.

| # while true; do ./tst-tls3-malloc ; done
| Didn't expect signal from child: got `Segmentation fault'
| ^C
| ============================================
| WARNING: possible recursive locking detected
| 4.17.0+ #25 Not tainted
| --------------------------------------------
| tst-tls3-malloc/510 is trying to acquire lock:
606c7728 (&mm->mmap_sem){++++}, at: __might_fault+0x28/0x5c
|
|but task is already holding lock:
|606c7728 (&mm->mmap_sem){++++}, at: do_page_fault+0x9c/0x2a0
|
| other info that might help us debug this:
|  Possible unsafe locking scenario:
|
|       CPU0
|       ----
|  lock(&mm->mmap_sem);
|  lock(&mm->mmap_sem);
|
| *** DEADLOCK ***
|

------------------------------------------------------------
What the change does is not obvious (note to myself)

prior code was

| do_page_fault
|
|   down_read() <-- lock taken
|   handle_mm_fault <-- signal pending as this runs
|   if fatal_signal_pending
|       if VM_FAULT_ERROR
|           up_read
|       if user_mode
|          return <-- lock still held, this was the BUG

New code

| do_page_fault
|
|   down_read() <-- lock taken
|   handle_mm_fault <-- signal pending as this runs
|   if fatal_signal_pending
|       if VM_FAULT_RETRY
|          return       <-- not same case as above, but still OK since
|                           core mm already relinq lock for FAULT_RETRY
|    ...
|
|   < Now falls through for bug case above >
|
|   up_read() <-- lock relinquished

Cc: stable@vger.kernel.org
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
arch/arc/mm/fault.c