Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1613527 > unrolled thread

Re: [PATCH 1/4] mm/vmalloc: allow to call vfree() in atomic context

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [PATCH 1/4] mm/vmalloc: allow to call vfree() in atomic context Andrew Morton <akpm@linux-foundation.org> - 2017-03-31 00:30 +0200
    Re: [PATCH 1/4] mm/vmalloc: allow to call vfree() in atomic context Joel Fernandes <joelaf@google.com> - 2017-03-31 09:20 +0200
    Re: [PATCH 1/4] mm/vmalloc: allow to call vfree() in atomic context Michal Hocko <mhocko@kernel.org> - 2017-04-04 11:40 +0200

#1613527 — Re: [PATCH 1/4] mm/vmalloc: allow to call vfree() in atomic context

On Thu, 30 Mar 2017 13:27:16 +0300 Andrey Ryabinin <aryabinin@virtuozzo.com> wrote:

> Commit 5803ed292e63 ("mm: mark all calls into the vmalloc subsystem
> as potentially sleeping") added might_sleep() to remove_vm_area() from
> vfree(), and commit 763b218ddfaf ("mm: add preempt points into
> __purge_vmap_area_lazy()") actually made vfree() potentially sleeping.
> 
> This broke vmwgfx driver which calls vfree() under spin_lock().
> 
>     BUG: sleeping function called from invalid context at mm/vmalloc.c:1480
>     in_atomic(): 1, irqs_disabled(): 0, pid: 341, name: plymouthd
>     2 locks held by plymouthd/341:
>      #0:  (drm_global_mutex){+.+.+.}, at: [<ffffffffc01c274b>] drm_release+0x3b/0x3b0 [drm]
>      #1:  (&(&tfile->lock)->rlock){+.+...}, at: [<ffffffffc0173038>] ttm_object_file_release+0x28/0x90 [ttm]
> 
>     Call Trace:
>      dump_stack+0x86/0xc3
>      ___might_sleep+0x17d/0x250
>      __might_sleep+0x4a/0x80
>      remove_vm_area+0x22/0x90
>      __vunmap+0x2e/0x110
>      vfree+0x42/0x90
>      kvfree+0x2c/0x40
>      drm_ht_remove+0x1a/0x30 [drm]
>      ttm_object_file_release+0x50/0x90 [ttm]
>      vmw_postclose+0x47/0x60 [vmwgfx]
>      drm_release+0x290/0x3b0 [drm]
>      __fput+0xf8/0x210
>      ____fput+0xe/0x10
>      task_work_run+0x85/0xc0
>      exit_to_usermode_loop+0xb4/0xc0
>      do_syscall_64+0x185/0x1f0
>      entry_SYSCALL64_slow_path+0x25/0x25
>
> This can be fixed in vmgfx, but it would be better to make vfree()
> non-sleeping again because we may have other bugs like this one.

I tend to disagree: adding yet another schedule_work() introduces
additional overhead and adds some risk of ENOMEM errors which wouldn't
occur with a synchronous free.

> __purge_vmap_area_lazy() is the only function in the vfree() path that
> wants to be able to sleep. So it make sense to schedule
> __purge_vmap_area_lazy() via schedule_work() so it runs only in sleepable
> context.

vfree() already does

	if (unlikely(in_interrupt()))
		__vfree_deferred(addr);

so it seems silly to introduce another defer-to-kernel-thread thing
when we already have one.

> This will have a minimal effect on the regular vfree() path.
> since __purge_vmap_area_lazy() is rarely called.

hum, OK, so perhaps the overhead isn't too bad.

Remind me: where does __purge_vmap_area_lazy() sleep?


Seems to me that a better fix would be to make vfree() atomic, if poss.

Otherwise, to fix callers so they call vfree from sleepable context. 
That will reduce kernel latencies as well.

[toc] | [next] | [standalone]

#1613681

Hi Andrew,

On Thu, Mar 30, 2017 at 3:22 PM, Andrew Morton
<akpm@linux-foundation.org> wrote:
> On Thu, 30 Mar 2017 13:27:16 +0300 Andrey Ryabinin <aryabinin@virtuozzo.com> wrote:
>
>> Commit 5803ed292e63 ("mm: mark all calls into the vmalloc subsystem
>> as potentially sleeping") added might_sleep() to remove_vm_area() from
>> vfree(), and commit 763b218ddfaf ("mm: add preempt points into
>> __purge_vmap_area_lazy()") actually made vfree() potentially sleeping.
>>
>> This broke vmwgfx driver which calls vfree() under spin_lock().
>>
>>     BUG: sleeping function called from invalid context at mm/vmalloc.c:1480
>>     in_atomic(): 1, irqs_disabled(): 0, pid: 341, name: plymouthd
>>     2 locks held by plymouthd/341:
>>      #0:  (drm_global_mutex){+.+.+.}, at: [<ffffffffc01c274b>] drm_release+0x3b/0x3b0 [drm]
>>      #1:  (&(&tfile->lock)->rlock){+.+...}, at: [<ffffffffc0173038>] ttm_object_file_release+0x28/0x90 [ttm]
>>
>>     Call Trace:
>>      dump_stack+0x86/0xc3
>>      ___might_sleep+0x17d/0x250
>>      __might_sleep+0x4a/0x80
>>      remove_vm_area+0x22/0x90
>>      __vunmap+0x2e/0x110
>>      vfree+0x42/0x90
>>      kvfree+0x2c/0x40
>>      drm_ht_remove+0x1a/0x30 [drm]
>>      ttm_object_file_release+0x50/0x90 [ttm]
>>      vmw_postclose+0x47/0x60 [vmwgfx]
>>      drm_release+0x290/0x3b0 [drm]
>>      __fput+0xf8/0x210
>>      ____fput+0xe/0x10
>>      task_work_run+0x85/0xc0
>>      exit_to_usermode_loop+0xb4/0xc0
>>      do_syscall_64+0x185/0x1f0
>>      entry_SYSCALL64_slow_path+0x25/0x25
>>
>> This can be fixed in vmgfx, but it would be better to make vfree()
>> non-sleeping again because we may have other bugs like this one.
>
> I tend to disagree: adding yet another schedule_work() introduces
> additional overhead and adds some risk of ENOMEM errors which wouldn't
> occur with a synchronous free.
>
>> __purge_vmap_area_lazy() is the only function in the vfree() path that
>> wants to be able to sleep. So it make sense to schedule
>> __purge_vmap_area_lazy() via schedule_work() so it runs only in sleepable
>> context.
>
> vfree() already does
>
>         if (unlikely(in_interrupt()))
>                 __vfree_deferred(addr);
>
> so it seems silly to introduce another defer-to-kernel-thread thing
> when we already have one.
>
>> This will have a minimal effect on the regular vfree() path.
>> since __purge_vmap_area_lazy() is rarely called.
>
> hum, OK, so perhaps the overhead isn't too bad.
>
> Remind me: where does __purge_vmap_area_lazy() sleep?

Because it will make for a possibly time consuming critical section.
This was hurting real-time workloads which are sensitive to latencies
(commit f9e09977671b618a "mm: turn vmap_purge_lock into a mutex" fixed
it).

Thanks,
Joel

[toc] | [prev] | [next] | [standalone]

#1615824

On Thu 30-03-17 15:22:29, Andrew Morton wrote:
> On Thu, 30 Mar 2017 13:27:16 +0300 Andrey Ryabinin <aryabinin@virtuozzo.com> wrote:
[...]
> > This can be fixed in vmgfx, but it would be better to make vfree()
> > non-sleeping again because we may have other bugs like this one.
> 
> I tend to disagree: adding yet another schedule_work() introduces
> additional overhead and adds some risk of ENOMEM errors which wouldn't
> occur with a synchronous free.

I do not think ENOMEM would be a problem. We are talking about lazy
handling already. Besides that the allocation path also does this lazy
free AFAICS.

> > __purge_vmap_area_lazy() is the only function in the vfree() path that
> > wants to be able to sleep. So it make sense to schedule
> > __purge_vmap_area_lazy() via schedule_work() so it runs only in sleepable
> > context.
> 
> vfree() already does
> 
> 	if (unlikely(in_interrupt()))
> 		__vfree_deferred(addr);
> 
> so it seems silly to introduce another defer-to-kernel-thread thing
> when we already have one.

But this only cares about the IRQ context and this patch aims at atomic
context in general. I agree it would have been better to reduce this
deferred behavior to only _atomic_ context but we not have a reliable
way to detect that on non-preemptive kernels AFAIR.
-- 
Michal Hocko
SUSE Labs

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web