linux-kernel - Re: [RFC 1/1] pmem: Add cond_resched() in bio_for_each_segment loop in pmem_make

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Fri, 7 Aug 2020 14:53:10 +0530
From:   Ritesh Harjani <riteshh@...ux.ibm.com>
To:     Dave Chinner <david@...morbit.com>
Cc:     linux-nvdimm@...ts.01.org, linux-ext4@...r.kernel.org,
        linux-xfs@...r.kernel.org, linux-kernel@...r.kernel.org,
        "Aneesh Kumar K . V" <aneesh.kumar@...ux.ibm.com>
Subject: Re: [RFC 1/1] pmem: Add cond_resched() in bio_for_each_segment loop
 in pmem_make_request



On 8/4/20 3:21 AM, Dave Chinner wrote:
> On Mon, Aug 03, 2020 at 12:44:04PM +0530, Ritesh Harjani wrote:
>>
>>
>> On 8/3/20 4:31 AM, Dave Chinner wrote:
>>> On Wed, Jul 29, 2020 at 02:15:18PM +0530, Ritesh Harjani wrote:
>>>> For systems which do not have CONFIG_PREEMPT set and
>>>> if there is a heavy multi-threaded load/store operation happening
>>>> on pmem + sometimes along with device latencies, softlockup warnings like
>>>> this could trigger. This was seen on Power where pagesize is 64K.
>>>>
>>>> To avoid softlockup, this patch adds a cond_resched() in this path.
>>>>
>>>> <...>
>>>> watchdog: BUG: soft lockup - CPU#31 stuck for 22s!
>>>> <...>
>>>> CPU: 31 PID: 15627 <..> 5.3.18-20
>>>> <...>
>>>> NIP memcpy_power7+0x43c/0x7e0
>>>> LR memcpy_flushcache+0x28/0xa0
>>>>
>>>> Call Trace:
>>>> memcpy_power7+0x274/0x7e0 (unreliable)
>>>> memcpy_flushcache+0x28/0xa0
>>>> write_pmem+0xa0/0x100 [nd_pmem]
>>>> pmem_do_bvec+0x1f0/0x420 [nd_pmem]
>>>> pmem_make_request+0x14c/0x370 [nd_pmem]
>>>> generic_make_request+0x164/0x400
>>>> submit_bio+0x134/0x2e0
>>>> submit_bio_wait+0x70/0xc0
>>>> blkdev_issue_zeroout+0xf4/0x2a0
>>>> xfs_zero_extent+0x90/0xc0 [xfs]
>>>> xfs_bmapi_convert_unwritten+0x198/0x230 [xfs]
>>>> xfs_bmapi_write+0x284/0x630 [xfs]
>>>> xfs_iomap_write_direct+0x1f0/0x3e0 [xfs]
>>>> xfs_file_iomap_begin+0x344/0x690 [xfs]
>>>> dax_iomap_pmd_fault+0x488/0xc10
>>>> __xfs_filemap_fault+0x26c/0x2b0 [xfs]
>>>> __handle_mm_fault+0x794/0x1af0
>>>> handle_mm_fault+0x12c/0x220
>>>> __do_page_fault+0x290/0xe40
>>>> do_page_fault+0x38/0xc0
>>>> handle_page_fault+0x10/0x30
>>>>
>>>> Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@...ux.ibm.com>
>>>> Signed-off-by: Ritesh Harjani <riteshh@...ux.ibm.com>
>>>> ---
>>>>    drivers/nvdimm/pmem.c | 1 +
>>>>    1 file changed, 1 insertion(+)
>>>>
>>>> diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
>>>> index 2df6994acf83..fcf7af13897e 100644
>>>> --- a/drivers/nvdimm/pmem.c
>>>> +++ b/drivers/nvdimm/pmem.c
>>>> @@ -214,6 +214,7 @@ static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
>>>>    			bio->bi_status = rc;
>>>>    			break;
>>>>    		}
>>>> +		cond_resched();
>>>
>>> There are already cond_resched() calls between submitted bios in
>>> blkdev_issue_zeroout() via both __blkdev_issue_zero_pages() and
>>> __blkdev_issue_write_zeroes(), so I'm kinda wondering where the
>>> problem is coming from here.
>>
>> This problem is coming from that bio call- submit_bio()
>>
>>>
>>> Just how big is the bio being issued here that it spins for 22s
>>> trying to copy it?
>>
>> It's 256 (due to BIO_MAX_PAGES) * 64KB (pagesize) = 16MB.
>> So this is definitely not an easy trigger as per tester was mainly seen
>> on a VM.
> 
> Right, so a memcpy() of 16MB of data in a single bio is taking >22s?
> 
> If so, then you can't solve this problem with cond_resched() - if
> something that should only take half a millisecond to run is taking
> 22s of CPU time, there are bigger problems occurring that need
> fixing.
> 
> i.e. if someone is running a workload that has effectively
> livelocked the hardware cache coherency protocol across the entire
> machine, then changing kernel code that requires memory bandwidth to
> be available is not going to fix the problem. All is does is shoot
> the messenger that tells you there is something going wrong.

Thanks Dave. Appreciate your inputs in this area.
Yes, agreed on the fact that we should not shoot the messenger itself.
Will look more into this.

-ritesh