Skip to contentSkip to conversation

tasks getting stuck on mmap_sem?

DISCUSSION Dimitris Michailidis 2 participants 3 messages Aug 17, 2010

lore
Share & tools
lore
#0 lore raw
Quoted 21 lines · David Miller

I just started seeing this with current GIT, is it related to the page table locking shuffle of a few days ago?

[21029.265646] INFO: task git:16643 blocked for more than 120 seconds.
[21029.267054] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[21029.268412] git           D 000000000072f804     0 16643  16273 0x210001101000080
[21029.269809] Call Trace:
[21029.271062]  [000000000072f778] rwsem_down_failed_common+0x90/0xb8
[21029.272390]  [000000000072f804] __down_read+0x30/0x40
[21029.273689]  [0000000000731148] do_sparc64_fault+0x238/0x788
[21029.274996]  [00000000004079e8] sparc64_realfault_common+0x10/0x20
[21029.276347] INFO: task git:16644 blocked for more than 120 seconds.
[21029.277655] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[21029.278993] git           D 000000000072f874     0 16644  16273 0x210041101000080
[21029.280348] Call Trace:
[21029.281542]  [000000000072f778] rwsem_down_failed_common+0x90/0xb8
[21029.282818]  [000000000072f874] __down_write+0x34/0x44
[21029.284057]  [00000000004d9c44] SyS_mmap_pgoff+0xc8/0x118
[21029.285293]  [0000000000406154] linux_sparc_syscall32+0x34/0x40
 ...
Quoted 2 lines

etc. it's a set of git threads computing deltas for a remote pull request to the machine.

This may be similar to the problems IA-64 had last week as a result of the rwsem changes, eg see commit 01d69a82e1d3c938da16bf55aab000672243aa24. I see arch/sparc/include/asm/atomic_64.h uses ints with atomic64_t, which was a problem for IA-64. Just guessing.

#1 lore raw

From: Dimitris Michailidis <[email protected]>
Date: Mon, 16 Aug 2010 22:22:04 -0700

Quoted 5 lines

This may be similar to the problems IA-64 had last week as a result of the rwsem changes, eg see commit 01d69a82e1d3c938da16bf55aab000672243aa24. I see arch/sparc/include/asm/atomic_64.h uses ints with atomic64_t, which was a problem for IA-64. Just guessing.

Thanks Dimitris I'll take a look.

#2 lore raw
In reply to Dimitris Michailidis

From: Dimitris Michailidis <[email protected]>
Date: Mon, 16 Aug 2010 22:22:04 -0700

Quoted 5 lines

This may be similar to the problems IA-64 had last week as a result of the rwsem changes, eg see commit 01d69a82e1d3c938da16bf55aab000672243aa24. I see arch/sparc/include/asm/atomic_64.h uses ints with atomic64_t, which was a problem for IA-64. Just guessing.

Well, the return types for the atomic64 routines were wrong and I fixed that.

But our rwsem implementation on sparc64 uses a 32-bit counter so that couldn't be the issue. And indeed I still get the problem of tasks being stuck on rwsems even after the atomic64 return types are fixed.

But I bet that rwsem commit 424acaaeb3a3932d64a9b4bd59df6cf72c22d8f3 introduced some kind of problem in another manner. I'll try to dig some more.