lkml.org 
[lkml]   [2019]   [Apr]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: Possible bio merging breakage in mp bio rework
From
Date


On 2019/4/6 下午2:09, Nikolay Borisov wrote:
>
>
> On 6.04.19 г. 3:16 ч., Ming Lei wrote:
>> Hi Nikolay,
>>
>> On Fri, Apr 05, 2019 at 07:04:18PM +0300, Nikolay Borisov wrote:
>>> Hello Ming,
>>>
>>> Following the mp biovec rework what is the maximum
>>> data that a bio could contain? Should it be PAGE_SIZE * bio_vec
>>
>> There isn't any maximum data limit on the bio submitted from fs,
>> and block layer will make the final bio sent to driver correct
>> by applying all kinds of queue limit, such as max segment size,
>> max segment number, max sectors, ...
>>
>>> or something else? Currently I can see bios as large as 127 megs
>>> on sequential workloads, I got prompted to this since btrfs has a
>>> memory allocation that is dependent on the data in the bio and this
>>> particular memory allocation started failing with order 6 allocs.
>>
>> Could you share us the code? I don't see why order 6 allocs is a must.
>
> When a bio is submitted btrfs has to calculate the checksum for it, this
> happens in btrfs_csum_one_bio. Said checksums are stored in an
> kmalloc'ed array, whose size is calculated as:
>
> 32 + bio_size / btrfs' block size (usually 4k). So for a 127mb bio that
> would be: 32 * ((134184960÷4096) * 4) = 127k. We'd make an order 3
> allocation. Admittedly the code in btrfs should know better rather than
> make unbounded allocations without a fallback, but bio suddenly becoming
> rather unbounded in their size caught us offhand.

Can we switch between kmalloc() for small csum while using pages for
larger csum?

Thanks,
Qu

>
>
>>
>>> Further debugging showed that with the following xfs_io command line:
>>>
>>>
>>> xfs_io -f -c "pwrite -S 0x61 -b 4m 0 10g" /media/scratch/file1
>>>
>>> I can easily see very large bios:
>>>
>>> [ 188.366540] kworker/-7 3.... 34847519us : btrfs_submit_bio_hook: bio: ffff8dffe9940bb0 bi_iter.bi_size = 134184960 bi_vcn: 28 bi_vcnt_max: 256
>>> [ 188.367129] kworker/-658 2.... 34946536us : btrfs_submit_bio_hook: bio: ffff8dffe9940370 bi_iter.bi_size = 134246400 bi_vcn: 28 bi_vcnt_max: 256
>>> [ 188.367714] kworker/-7 3.... 35107967us : btrfs_submit_bio_hook: bio: ffff8dffe9940bb0 bi_iter.bi_size = 134184960 bi_vcn: 30 bi_vcnt_max: 256
>>> [ 188.368319] kworker/-658 2.... 35229894us : btrfs_submit_bio_hook: bio: ffff8dffe9940370 bi_iter.bi_size = 134246400 bi_vcn: 32 bi_vcnt_max: 256
>>> [ 188.368909] kworker/-7 3.... 35374809us : btrfs_submit_bio_hook: bio: ffff8dffe9940bb0 bi_iter.bi_size = 134184960 bi_vcn: 25 bi_vcnt_max: 256
>>> [ 188.369498] kworker/-658 2.... 35516194us : btrfs_submit_bio_hook: bio: ffff8dffe9940370 bi_iter.bi_size = 134246400 bi_vcn: 31 bi_vcnt_max: 256
>>> [ 188.370086] kworker/-7 3.... 35663669us : btrfs_submit_bio_hook: bio: ffff8dffe9940bb0 bi_iter.bi_size = 134184960 bi_vcn: 32 bi_vcnt_max: 256
>>> [ 188.370696] kworker/-658 2.... 35791006us : btrfs_submit_bio_hook: bio: ffff8dffe9940370 bi_iter.bi_size = 100655104 bi_vcn: 24 bi_vcnt_max: 256
>>> [ 188.371335] kworker/-658 2.... 35816114us : btrfs_submit_bio_hook: bio: ffff8dffe99434f0 bi_iter.bi_size = 33591296 bi_vcn: 5 bi_vcnt_max: 256
>>>
>>>
>>> So that's 127 megs in a single bio? This stems from the new merging logic.
>>> 07173c3ec276 ("block: enable multipage bvecs") made it so that physically
>>> contiguous pages added to the bio would just modify bi_iter.bi_size and the
>>> initial page's bio_vec's bv_len. There's no longer the
>>> page == bv->bv_page portion of the check.
>>
>> bio_add_page() tries best to put physically contiguous pages into one bvec, and
>> I don't see anything is wrong in the log.
>>
>> Could you show us what the real problem is?
>>
>> Thanks,
>> Ming
>>

\
 
 \ /
  Last update: 2019-04-06 10:00    [W:0.063 / U:3.612 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site