Date:   Thu, 23 Sep 2021 20:09:21 +1000
From:   "NeilBrown" <neilb@...e.de>
To:     "Coly Li" <colyli@...e.de>
Cc:     linux-kernel@...r.kernel.org, linux-block@...r.kernel.org,
        linux-raid@...r.kernel.org, nvdimm@...ts.linux.dev,
        antlists@...ngman.org.uk,
        "Dan Williams" <dan.j.williams@...el.com>,
        "Hannes Reinecke" <hare@...e.de>, "Jens Axboe" <axboe@...nel.dk>,
        "Richard Fan" <richard.fan@...e.com>,
        "Vishal L Verma" <vishal.l.verma@...el.com>,
        "Greg Kroah-Hartman" <gregkh@...uxfoundation.org>,
        rafael@...nel.org
Subject: Re: Too large badblocks sysfs file (was: [PATCH v3 0/7] badblocks
 improvement for multiple bad block ranges)

On Thu, 23 Sep 2021, Coly Li wrote:
> Hi all the kernel gurus, and folks in mailing lists,
> 
> This is a question about exporting 4KB+ text information via sysfs 
> interface. I need advice on how to handle the problem.

Why do you think there is a problem?
As documented in Documentation/admin-guide/md.rst, the truncation at 1
page is expected and by design.

The "unacknowledge-bad-blocks" file is the important one that is needed
for correct behaviour.  Being able to read a single block is sufficient,
though being able to read more than one could provide better performance
in some cases.

The "bad-blocks" file primarily exist to provide visibility into the
state of the system - useful during development.  It can be written to
to add bad blocks.  I never *needs* to be read from.

The authoritative source of information about the set of bad blocks is
the on-disk data the can be and should be read directly...

Except that mdadm does.  That was a mistake.  check_for_cleared_bb() is
wrong.  I wonder why it was added.  The commit message doesn't give any
justification.

NeilBrown


> 
> Recently I work on the bad blocks API (block/badblocks.c) improvement, 
> there is a sysfs file to export the bad block ranges for me raid. E.g 
> for a md raid1 device, file
>      /sys/block/md0/md/rd0/bad_blocks
> may contain the following text content,
>      64 32
>     128 8
> The above lines mean there are two bad block ranges, one starts at LBA 
> 64, length 32 sectors, another one starts at LBA 128 and length 8 
> sectors. All the content is generated from the internal bad block 
> records with 512 elements. In my testing the worst case only 185 from 
> 512 records can be displayed via the sysfs file if the LBA string is 
> very long, e.g.the following content,
>    17668164135030776 512
>    17668164135029776 512
>    17668164135028776 512
>    17668164135027776 512
>    ... ...
> The bad block ranges stored in internal bad blocks array are correct, 
> but the output message is truncated. This is the problem I encountered.
> 
> I don't see sysfs has seq_file support (correct me if I am wrong), and I 
> know it is improper to transfer 4KB+ text via sysfs interface, but the 
> code is here already for long time.
> 
> There are 2 ideas to fix showing up in my brain,
> 1) Do not fix the problem
>      Normally it is rare that a storage media has 100+ bad block ranges, 
> maybe in real world all the existing bad blocks information won't exceed 
> the page size limitation of sysfs file.
> 2) Add seq_file support to sysfs interface if there is no
> 
> It is probably there is other better solution to fix. So I do want to 
> get hint/advice from you.
> 
> Thanks in advance for any comment :-)
> 
> Coly Li
> 
> On 9/14/21 12:36 AM, Coly Li wrote:
> > This is the second effort to improve badblocks code APIs to handle
> > multiple ranges in bad block table.
> >
> > There are 2 changes from previous version,
> > - Fixes 2 bugs in front_overwrite() which are detected by the user
> >    space testing code.
> > - Provide the user space testing code in last patch.
> >
> > There is NO in-memory or on-disk format change in the whole series, all
> > existing API and data structures are consistent. This series just only
> > improve the code algorithm to handle more corner cases, the interfaces
> > are same and consistency to all existing callers (md raid and nvdimm
> > drivers).
> >
> > The original motivation of the change is from the requirement from our
> > customer, that current badblocks routines don't handle multiple ranges.
> > For example if the bad block setting range covers multiple ranges from
> > bad block table, only the first two bad block ranges merged and rested
> > ranges are intact. The expected behavior should be all the covered
> > ranges to be handled.
> >
> > All the patches are tested by modified user space code and the code
> > logic works as expected. The modified user space testing code is
> > provided in last patch. The testing code detects 2 defects in helper
> > front_overwrite() and fixed in this version.
> >
> > The whole change is divided into 6 patches to make the code review more
> > clear and easier. If people prefer, I'd like to post a single large
> > patch finally after the code review accomplished.
> >
> > This version is seriously tested, and so far no more defect observed.
> >
> >
> > Coly Li
> >
> > Cc: Dan Williams <dan.j.williams@...el.com>
> > Cc: Hannes Reinecke <hare@...e.de>
> > Cc: Jens Axboe <axboe@...nel.dk>
> > Cc: NeilBrown <neilb@...e.de>
> > Cc: Richard Fan <richard.fan@...e.com>
> > Cc: Vishal L Verma <vishal.l.verma@...el.com>
> > ---
> > Changelog:
> > v3: add tester Richard Fan <richard.fan@...e.com>
> > v2: the improved version, and with testing code.
> > v1: the first completed version.
> >
> >
> > Coly Li (6):
> >    badblocks: add more helper structure and routines in badblocks.h
> >    badblocks: add helper routines for badblock ranges handling
> >    badblocks: improvement badblocks_set() for multiple ranges handling
> >    badblocks: improve badblocks_clear() for multiple ranges handling
> >    badblocks: improve badblocks_check() for multiple ranges handling
> >    badblocks: switch to the improved badblock handling code
> > Coly Li (1):
> >    test: user space code to test badblocks APIs
> >
> >   block/badblocks.c         | 1599 ++++++++++++++++++++++++++++++-------
> >   include/linux/badblocks.h |   32 +
> >   2 files changed, 1340 insertions(+), 291 deletions(-)
> >
> 
> 

Powered by blists - more mailing lists