diff mbox

binary execution from DAX mount hang since next-20160407

Message ID alpine.LSU.2.11.1604150104350.5801@eggly.anvils (mailing list archive)
State New, archived
Headers show

Commit Message

Hugh Dickins April 15, 2016, 8:38 a.m. UTC
On Fri, 15 Apr 2016, Xiong Zhou wrote:

> Hi, all
> 
> Since tag next-20160407 in linux-next repo, executing binary
> from/in DAX mount hangs.
> 
> It does not hang if mount without dax option.
> It hangs with both xfs and ext4.
> It does not hang if execute from a -t tmpfs mount.
> It does not hang on next-20160406 and still hangs on 0414 tree.
> 
> # ps -axjf
> ...
> S+       0   0:00  |       \_ sh -x thl.sh
> R+       0  42:33  |           \_ [hl]
> ..
> # cat thl.sh
> mkfs.ext4 /dev/pmem0
> mount -o dax /dev/pmem0 /daxmnt
> cp hl /daxmnt
> /daxmnt/hl
> # cat hl.c
> void main()
> {
>         printf("ok\n");
> }
> # cc hl.c -o hl
> 
> Bisecting commits between 0406 and 0407 tag, points to this:
> 
> d7c7d56ca61aec18e5e0cb3a64e50073c42195f7 is the first bad commit
> commit d7c7d56ca61aec18e5e0cb3a64e50073c42195f7
> Author: Hugh Dickins <hughd@google.com>
> Date:   Thu Apr 7 14:00:12 2016 +1000
> 
>     huge tmpfs: avoid premature exposure of new pagetable
> 
> Bisect log and config are attatched.

Excellent and very helpful bug report: thank you very much for taking
the trouble to make such a good report.

I see why this happens now: I've not been paying enough attention
to the DAX changes.

The fix requires a repositioning of where I allocate the new page
table: which is a change we knew we had to make for other reasons,
but it did not appear to be a high priority compared with other things
- until your bug report showing that I have broken DAX rather badly.

In return for your excellent bug report, I can immediately offer
the most shameful patch I have ever posted: which has the virtue of
simplicity, and will work so long as you have plenty of free memory;
but might deadlock if it has to go into page reclaim (or maybe not:
perhaps the DAXness would leave it as merely a lockdep violation).

Maybe not so much worse than the current hang, but still shameful:
I'm appending it here just in case you're in a hurry to see your "ok"
program working on DAX; but I think I'd better rearrange priorities
and try to provide a proper fix as soon as possible.

Never-to-be-Signed-off-by: an anonymous hacker

--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Murphy Zhou April 18, 2016, 2:50 a.m. UTC | #1
On Fri, Apr 15, 2016 at 4:38 PM, Hugh Dickins <hughd@google.com> wrote:
> On Fri, 15 Apr 2016, Xiong Zhou wrote:
>
>> Hi, all
>>
>> Since tag next-20160407 in linux-next repo, executing binary
>> from/in DAX mount hangs.
>>
>> It does not hang if mount without dax option.
>> It hangs with both xfs and ext4.
>> It does not hang if execute from a -t tmpfs mount.
>> It does not hang on next-20160406 and still hangs on 0414 tree.
>>
>> # ps -axjf
>> ...
>> S+       0   0:00  |       \_ sh -x thl.sh
>> R+       0  42:33  |           \_ [hl]
>> ..
>> # cat thl.sh
>> mkfs.ext4 /dev/pmem0
>> mount -o dax /dev/pmem0 /daxmnt
>> cp hl /daxmnt
>> /daxmnt/hl
>> # cat hl.c
>> void main()
>> {
>>         printf("ok\n");
>> }
>> # cc hl.c -o hl
>>
>> Bisecting commits between 0406 and 0407 tag, points to this:
>>
>> d7c7d56ca61aec18e5e0cb3a64e50073c42195f7 is the first bad commit
>> commit d7c7d56ca61aec18e5e0cb3a64e50073c42195f7
>> Author: Hugh Dickins <hughd@google.com>
>> Date:   Thu Apr 7 14:00:12 2016 +1000
>>
>>     huge tmpfs: avoid premature exposure of new pagetable
>>
>> Bisect log and config are attatched.
>
> Excellent and very helpful bug report: thank you very much for taking
> the trouble to make such a good report.
>
> I see why this happens now: I've not been paying enough attention
> to the DAX changes.
>
> The fix requires a repositioning of where I allocate the new page
> table: which is a change we knew we had to make for other reasons,
> but it did not appear to be a high priority compared with other things
> - until your bug report showing that I have broken DAX rather badly.
>
> In return for your excellent bug report, I can immediately offer
> the most shameful patch I have ever posted: which has the virtue of
> simplicity, and will work so long as you have plenty of free memory;
> but might deadlock if it has to go into page reclaim (or maybe not:
> perhaps the DAXness would leave it as merely a lockdep violation).
>
> Maybe not so much worse than the current hang, but still shameful:
> I'm appending it here just in case you're in a hurry to see your "ok"
> program working on DAX; but I think I'd better rearrange priorities
> and try to provide a proper fix as soon as possible.

No hurry, :)  Take your time.

>
> Never-to-be-Signed-off-by: an anonymous hacker
>
> --- 4.6-rc2-mm1/mm/memory.c     2016-04-10 10:12:06.167769232 -0700
> +++ linux/mm/memory.c   2016-04-15 00:54:06.427085026 -0700
> @@ -2874,7 +2874,7 @@ static int __do_fault(struct vm_area_str
>                 ret = VM_FAULT_HWPOISON;
>                 goto err;
>         }
> -
> + out:
>         /*
>          * Use pte_alloc instead of pte_alloc_map, because we can't
>          * run pte_offset_map on the pmd, if an huge pmd could
> @@ -2892,7 +2892,7 @@ static int __do_fault(struct vm_area_str
>                 ret = VM_FAULT_NOPAGE;
>                 goto err;
>         }
> - out:
> +
>         *page = vmf.page;
>         return ret;
>  err:

Yes, "ok" is printed ok!
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

--- 4.6-rc2-mm1/mm/memory.c	2016-04-10 10:12:06.167769232 -0700
+++ linux/mm/memory.c	2016-04-15 00:54:06.427085026 -0700
@@ -2874,7 +2874,7 @@  static int __do_fault(struct vm_area_str
 		ret = VM_FAULT_HWPOISON;
 		goto err;
 	}
-
+ out:
 	/*
 	 * Use pte_alloc instead of pte_alloc_map, because we can't
 	 * run pte_offset_map on the pmd, if an huge pmd could
@@ -2892,7 +2892,7 @@  static int __do_fault(struct vm_area_str
 		ret = VM_FAULT_NOPAGE;
 		goto err;
 	}
- out:
+
 	*page = vmf.page;
 	return ret;
 err: