linux-kernel - Re: Some new bio merging behaviors in __bio_try_merge

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

Message-ID: <cb392476-a00e-09ce-fa6b-9e088242ecc6@huawei.com>
Date:   Thu, 11 Apr 2019 15:43:02 +0800
From:   Gao Xiang <gaoxiang25@...wei.com>
To:     Ming Lei <ming.lei@...hat.com>
CC:     <linux-block@...r.kernel.org>, LKML <linux-kernel@...r.kernel.org>,
        "linux-erofs@...ts.ozlabs.org" <linux-erofs@...ts.ozlabs.org>,
        Jens Axboe <axboe@...nel.dk>, Chao Yu <yuchao0@...wei.com>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>
Subject: Re: Some new bio merging behaviors in __bio_try_merge_page



On 2019/4/11 15:08, Ming Lei wrote:
> Hi Gao Xiang,
> 
> On Thu, Apr 11, 2019 at 01:47:49PM +0800, Gao Xiang wrote:
>> Hi Ming,
>>
>> I found a erofs issue after commit 07173c3ec276
>> ("block: enable multipage bvecs") is merged. It seems that
>> it tries to merge more physical continuous pages in one iovec.
>>
>> However it breaks the current erofs_read_raw_page logic since it uses
>> nr_iovecs of bio_alloc to limit the maximum number of physical
> 
> I believe you can do the limit outside easily, such as by checking
> how many pages have been added to the bio.

Yes, I agree that. However, I noticed that bio_add_page is exported as
EXPORT_SYMBOL. I have no idea how many out-of-tree drivers break as well.

Is there no way to take old behavior into consideration in the block layer as well?

> 
>> continuous blocks as well. It was practicable since the old
>> __bio_try_merge_page only tries to merge in the same page.
>> it is a kAPI behavior change which also affects bio_alloc...
>>
>> ...
>> 231                 err = erofs_map_blocks(inode, &map, EROFS_GET_BLOCKS_RAW);
>> 232                 if (unlikely(err))
>> 233                         goto err_out;
>> ...
>> 284                 /* max # of continuous pages */
>> 285                 if (nblocks > DIV_ROUND_UP(map.m_plen, PAGE_SIZE))
>> 286                         nblocks = DIV_ROUND_UP(map.m_plen, PAGE_SIZE);
>> 287                 if (nblocks > BIO_MAX_PAGES)
>> 288                         nblocks = BIO_MAX_PAGES;
>> 289
>> 290                 bio = erofs_grab_bio(sb, blknr, nblocks, sb,
>> 291                                      read_endio, false);
> 
> Previously this bio is allowed to add at most 'nblocks' pages, however,
> now we are allowed to add at most 'nblocks' io vecs, instead of pages.
> 
>> 292                 if (IS_ERR(bio)) {
>> 293                         err = PTR_ERR(bio);
>> 294                         bio = NULL;
>> 295                         goto err_out;
>> 296                 }
>> 297         }
>> 298
>> 299         err = bio_add_page(bio, page, PAGE_SIZE, 0);
>> 300         /* out of the extent or bio is full */
>> 301         if (err < PAGE_SIZE)
>> 302                 goto submit_bio_retry;
>> ...
>>
>> After commit 07173c3ec276 ("block: enable multipage bvecs"), erofs could
>> read more beyond what erofs_map_blocks assigns, and out-of-bound data could
>> be read and it breaks tail-end inline determination.
> 
> I don't understand why, could you explain a bit why erofs reads more?

In current erofs, file could be break in two parts (non tail-end and tail-end blocks).
Considering this on-disk layout,
 ________________________________________________________________
|     non tail-end data              |   meta data               |
|____________________________________|_(inode) + non-inline data_|
                          ^          ^
                          |          |        
                          bio start  \-- what nr_iovecs indicates as well:
                                         the end of the non tail-end data.

Before this commit, it will stop just before the next meta data. However,
if the new bio merging behavior is introduced, it could add pages more than
nr_iovecs, thus meta data will be read as the normal data...


> 
> The amount depends on how many pages you added to the bio.
> 
>>
>> I can change the logic in erofs. However, out of curiosity, I have no idea
>> if some other places also are designed like this.
>>
>> IMO, it's better to provide a total count which indicates how many real
>> pages have been added in this bio. some thoughts?
> 
> As I mentioned, you may count how many pages added to bio, or you still
> can get the number via bio_segments(bio).

It is unsuitable to introduce bio_segments for each bio_add_page...

Thanks,
Gao Xiang

> 
> Thanks,
> Ming
>