Skip to contentSkip to conversation

zram: zsmalloc calls sleeping function from atomic context

DISCUSSION Sergey Senozhatsky 4 participants 4 messages Mar 17, 2014 – Mar 18, 2014

lore
Share & tools
lore
#0 lore raw

Hello gents,

I just noticed that starting from commit

commit 3d693a5127e79e79da7c34dc0c776bc620697ce5

Author: Andrew Morton <[email protected]>
Date:   Mon Mar 17 11:23:56 2014 +1100

mm-vmalloc-avoid-soft-lockup-warnings-when-vunmaping-large-ranges-fix

add a might_sleep() to catch atomic callers more promptly

and

commit 032dda8b6c4021d4be63bcc483b47fd26c6f48a2

Author: David Vrabel <[email protected]>
Date:   Mon Mar 17 11:23:56 2014 +1100

mm/vmalloc: avoid soft lockup warnings when vunmap()'ing large ranges

If vunmap() is used to unmap a large (e.g., 50 GB) region, it may take sufficiently long that it triggers soft lockup warnings.

Add a cond_resched() into vunmap_pmd_range() so the calling task may be resheduled after unmapping each PMD entry. This is how zap_pmd_range() fixes the same problem for userspace mappings.

All callers may sleep except for the APEI GHES driver (apei/ghes.c) which calls unmap_kernel_range_no_flush() from NMI and IRQ contexts. This driver only unmaps a single pages so don't call cond_resched() if the unmap doesn't cross a PMD boundary.

w/ CONFIG_PGTABLE_MAPPING=y zs_unmap_object() calls unmap_kernel_range() under rwlock, producing the following warning (basically we perform every read()/write() under rwlock, so I can see lots of these warnings):

[  631.541177] BUG: sleeping function called from invalid context at mm/vmalloc.c:74
[  631.541181] in_atomic(): 1, irqs_disabled(): 0, pid: 94, name: kworker/u8:2
[  631.541183] Preemption disabled at:[<ffffffffa00ca0ad>] zram_bvec_rw.isra.14+0x2be/0x4fc [zram]
[  631.541193] CPU: 2 PID: 94 Comm: kworker/u8:2 Tainted: G           O  3.14.0-rc6-next-20140317-dbg-dirty #182
[  631.541195] Hardware name: Acer             Aspire 5741G    /Aspire 5741G    , BIOS V1.20 02/08/2011
[  631.541202] Workqueue: writeback bdi_writeback_workfn (flush-254:0)
[  631.541205]  0000000000000000 ffff88015211b748 ffffffff813ba01d 0000000000000000
[  631.541208]  ffff88015211b768 ffffffff81057ecb ffffc9000003e000 ffffc9000003e000
[  631.541212]  ffff88015211b7d8 ffffffff810cc491 ffffc9000003dfff ffff88015211b800
[  631.541216] Call Trace:
[  631.541223]  [<ffffffff813ba01d>] dump_stack+0x4e/0x7a
[  631.541229]  [<ffffffff81057ecb>] __might_sleep+0x14e/0x153
[  631.541234]  [<ffffffff810cc491>] vunmap_page_range+0x133/0x25d
[  631.541237]  [<ffffffff810cd81b>] unmap_kernel_range+0x16/0x26
[  631.541241]  [<ffffffff810de6f6>] zs_unmap_object+0xd8/0xff
[  631.541245]  [<ffffffffa00ca120>] zram_bvec_rw.isra.14+0x331/0x4fc [zram]
[  631.541248]  [<ffffffffa00ca439>] zram_make_request+0x14e/0x228 [zram]
[  631.541252]  [<ffffffff810a8088>] ? mempool_alloc+0x6d/0x130
[  631.541257]  [<ffffffff811e9395>] generic_make_request+0x97/0xd6
[  631.541259]  [<ffffffff811e94c6>] submit_bio+0xf2/0x131
[  631.541263]  [<ffffffff81106306>] _submit_bh+0x1c1/0x1eb
[  631.541266]  [<ffffffff8110633b>] submit_bh+0xb/0xd
[  631.541269]  [<ffffffff811078d9>] __block_write_full_page+0x1ad/0x2c8
[  631.541273]  [<ffffffff8110a118>] ? I_BDEV+0xd/0xd
[  631.541276]  [<ffffffff81105041>] ? end_buffer_write_sync+0x61/0x61
[  631.541278]  [<ffffffff8110a118>] ? I_BDEV+0xd/0xd
[  631.541282]  [<ffffffff81107bb4>] block_write_full_page_endio+0xdc/0xe8
[  631.541284]  [<ffffffff81107bd0>] block_write_full_page+0x10/0x12
[  631.541287]  [<ffffffff8110a6e5>] blkdev_writepage+0x13/0x15
[  631.541292]  [<ffffffff810acfb8>] __writepage+0xe/0x2c
[  631.541295]  [<ffffffff810ad35f>] write_cache_pages+0x25c/0x367
[  631.541297]  [<ffffffff810acfaa>] ? mapping_tagged+0xf/0xf
[  631.541301]  [<ffffffff810ad4a3>] generic_writepages+0x39/0x51
[  631.541304]  [<ffffffff810ae6d4>] do_writepages+0x19/0x27
[  631.541307]  [<ffffffff810ff6d4>] __writeback_single_inode+0x3c/0xee
[  631.541310]  [<ffffffff811000d7>] writeback_sb_inodes+0x1bf/0x2f9
[  631.541313]  [<ffffffff8110028b>] __writeback_inodes_wb+0x7a/0xb0
[  631.541316]  [<ffffffff811003c0>] wb_writeback+0xff/0x190
[  631.541319]  [<ffffffff810595f3>] ? get_parent_ip+0xd/0x3c
[  631.541322]  [<ffffffff811008f5>] bdi_writeback_workfn+0xcd/0x28d
[  631.541325]  [<ffffffff8105b32b>] ? try_to_wake_up+0x1f4/0x203
[  631.541330]  [<ffffffff8104d213>] process_one_work+0x1c9/0x2e9
[  631.541332]  [<ffffffff8104d7ad>] worker_thread+0x1d3/0x2bd
[  631.541335]  [<ffffffff8104d5da>] ? rescuer_thread+0x27d/0x27d
[  631.541338]  [<ffffffff81051e75>] kthread+0xd6/0xde
[  631.541341]  [<ffffffff81051d9f>] ? kthread_create_on_node+0x162/0x162
[  631.541345]  [<ffffffff813bf8bc>] ret_from_fork+0x7c/0xb0
[  631.541348]  [<ffffffff81051d9f>] ? kthread_create_on_node+0x162/0x162
	-ss
#1 lore raw
Quoted 48 lines · On Mon, 17 Mar 2014 17:43:58 +0300 Sergey Senozhatsky…

Hello gents,

I just noticed that starting from commit

commit 3d693a5127e79e79da7c34dc0c776bc620697ce5

Author: Andrew Morton <[email protected]>
Date:   Mon Mar 17 11:23:56 2014 +1100

mm-vmalloc-avoid-soft-lockup-warnings-when-vunmaping-large-ranges-fix

add a might_sleep() to catch atomic callers more promptly

and

commit 032dda8b6c4021d4be63bcc483b47fd26c6f48a2

Author: David Vrabel <[email protected]>
Date:   Mon Mar 17 11:23:56 2014 +1100

...

w/ CONFIG_PGTABLE_MAPPING=y zs_unmap_object() calls unmap_kernel_range() under rwlock, producing the following warning (basically we perform every read()/write() under rwlock, so I can see lots of these warnings):

[  631.541177] BUG: sleeping function called from invalid context at mm/vmalloc.c:74
[  631.541181] in_atomic(): 1, irqs_disabled(): 0, pid: 94, name: kworker/u8:2
[  631.541183] Preemption disabled at:[<ffffffffa00ca0ad>] zram_bvec_rw.isra.14+0x2be/0x4fc [zram]
[  631.541193] CPU: 2 PID: 94 Comm: kworker/u8:2 Tainted: G           O  3.14.0-rc6-next-20140317-dbg-dirty #182
[  631.541195] Hardware name: Acer             Aspire 5741G    /Aspire 5741G    , BIOS V1.20 02/08/2011
[  631.541202] Workqueue: writeback bdi_writeback_workfn (flush-254:0)
[  631.541205]  0000000000000000 ffff88015211b748 ffffffff813ba01d 0000000000000000
[  631.541208]  ffff88015211b768 ffffffff81057ecb ffffc9000003e000 ffffc9000003e000
[  631.541212]  ffff88015211b7d8 ffffffff810cc491 ffffc9000003dfff ffff88015211b800
[  631.541216] Call Trace:
[  631.541223]  [<ffffffff813ba01d>] dump_stack+0x4e/0x7a
[  631.541229]  [<ffffffff81057ecb>] __might_sleep+0x14e/0x153
[  631.541234]  [<ffffffff810cc491>] vunmap_page_range+0x133/0x25d
[  631.541237]  [<ffffffff810cd81b>] unmap_kernel_range+0x16/0x26
[  631.541241]  [<ffffffff810de6f6>] zs_unmap_object+0xd8/0xff
[  631.541245]  [<ffffffffa00ca120>] zram_bvec_rw.isra.14+0x331/0x4fc [zram]
[  631.541248]  [<ffffffffa00ca439>] zram_make_request+0x14e/0x228 [zram]
[  631.541252]  [<ffffffff810a8088>] ? mempool_alloc+0x6d/0x130
[  631.541257]  [<ffffffff811e9395>] generic_make_request+0x97/0xd6
[  631.541259]  [<ffffffff811e94c6>] submit_bio+0xf2/0x131
Quoted 2 lines

...

OK, thanks. David, there's our atomic unmap and there are probably others. Converting a previously-atomic utility function into one which can sleep is going to be difficult.

One "fix" would be to make unmaps of (say) less than 16MB atomic, but unmaps of larger regions can do cond_resched(). So vunmap_pmd_range() will do

	if (end - addr < 16MB)
		might_sleep();

but I can't believe I even mentioned that.

So what to do? Add a new interface, perhaps: "vunmap_large()", perhaps. Change that to pass a boolean "may_reschedule" down the various levels.

Or can this code which vmaps 50GB be changed to unmap it in 16MB chunks via unmap_kernel_range(), with a cond_resched() in the loop?

I'll drop the patches while we sort this out.

btw, I note that vunmap() itself already has a might_sleep() in it, and I can't work out why - I don't think it _does_ sleep. The changelog to 34754b69a6f87aa6aa is, in toto:

"x86: make vmap yell louder when it is used under irqs_disabled()"

No explanation *why*. And why didn't it use WARN_ON(irqs_disabled())?

#2 lore raw
Quoted 4 lines · Andrew Morton

OK, thanks. David, there's our atomic unmap and there are probably others. Converting a previously-atomic utility function into one which can sleep is going to be difficult.

I think we should drop these patches. I think Fujitsu were doing something particularly odd with an out-of-tree driver.

Quoted 2 lines

Or can this code which vmaps 50GB be changed to unmap it in 16MB chunks via unmap_kernel_range(), with a cond_resched() in the loop?

This sounds like something the people from Fujitsu can explore.

David

#3 lore raw
In reply to Andrew Morton
Quoted 7 lines · Andrew Morton

btw, I note that vunmap() itself already has a might_sleep() in it, and I can't work out why - I don't think it _does_ sleep. The changelog to 34754b69a6f87aa6aa is, in toto:

"x86: make vmap yell louder when it is used under irqs_disabled()"

No explanation *why*. And why didn't it use WARN_ON(irqs_disabled())?

Man I suck.. and 5 years ago too. I can barely remember last week.

Lets see if the email archive has clues.

vmap()
get_vm_area_caller()

    __get_vm_area_node(.gfp = GFP_KERNEL)
      kmalloc_node(gfp);

It does sleep.

https://lkml.org/lkml/2009/2/13/115