lkml.org 
[lkml]   [2020]   [Jan]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v4 6/7] dm: Directly disable max_allocate_sectors for now
From
Date
On 21.01.2020 17:43, Mike Snitzer wrote:
> On Tue, Jan 21 2020 at 9:20am -0500,
> Kirill Tkhai <ktkhai@virtuozzo.com> wrote:
>
>> On 21.01.2020 16:48, Mike Snitzer wrote:
>>> On Tue, Jan 21 2020 at 8:33am -0500,
>>> Kirill Tkhai <ktkhai@virtuozzo.com> wrote:
>>>
>>>> On 21.01.2020 15:36, Kirill Tkhai wrote:
>>>>> On 21.01.2020 15:24, Mike Snitzer wrote:
>>>>>> On Tue, Jan 21 2020 at 5:42am -0500,
>>>>>> Kirill Tkhai <ktkhai@virtuozzo.com> wrote:
>>>>>>
>>>>>>> Since dm inherits limits from underlining block devices,
>>>>>>> this patch directly disables max_allocate_sectors for dm
>>>>>>> till full allocation support is implemented.
>>>>>>>
>>>>>>> This prevents high-level primitives (generic_make_request_checks(),
>>>>>>> __blkdev_issue_write_zeroes(), ...) from sending REQ_ALLOCATE
>>>>>>> requests.
>>>>>>>
>>>>>>> Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
>>>>>>> ---
>>>>>>> drivers/md/dm-table.c | 2 ++
>>>>>>> drivers/md/md.h | 1 +
>>>>>>> 2 files changed, 3 insertions(+)
>>>>>>
>>>>>> You're mixing DM and MD changes in the same patch.
>>>>>>
>>>>>> But I'm wondering if it might be best to set this default for stacking
>>>>>> devices in blk_set_stacking_limits()?
>>>>>>
>>>>>> And then it is up to each stacking driver to override as needed.
>>>>>
>>>>> Hm. Sound like a good idea. This "lim->max_allocate_sectors = 0" in blk_set_stacking_limits()
>>>>> should work for dm's dm_calculate_queue_limits(), since it calls blk_stack_limits(), which is:
>>>>>
>>>>> t->max_allocate_sectors = min(t->max_allocate_sectors,
>>>>> b->max_allocate_sectors);
>>>>>
>>>>> Could you please tell is this fix is also enough for md?
>>>>
>>>> It looks like it's enough since queue defaults are set in md_alloc()->blk_set_stacking_limits().
>>>> In case of we set "max_allocate_sectors = 0", in further it can be changed only manually,
>>>> but nobody does this.
>>>
>>> Yes, it will work to disable this capability for MD and DM.
>>>
>>> But if/when a stacked device _dooes_ want to support this then it'll be
>>> awkward to override this stacking default to allow blk_stack_limits()
>>> to properly stack up this limit. blk_limits are extremely fiddley so
>>> this isn't necessarily new. But by explicitly defaulting to 0 and then
>>> having blk_stack_limits use min() for this limit: it results in stacking
>>> drivers needing to clumsily unwind the default. E.g. DM will need to
>>> tweak its blk_stack_limits() related code to allow override that
>>> actually _does_ stack up the underlying devices' capability (and not
>>> just impose its own limit that ignores the underlying devices).
>>>
>>> So I'm not convinced this is the right way to go (be it the v4 approach
>>> you took or the cleaner use of blk_set_stacking_limits I suggested).
>>
>> Is there a strong vision about the way we should go? Or you leave this choose
>> up to me?
>
> I don't have time to work through it at the moment (e.g. implementing
> dm-thinp support to know what the block core code should be) so I'll
> just defer to you on a disabling it for now.
>
>>> And to be clear, I'm interested in having DM thinp support this
>>> capability to preallocate blocks.
>>
>> My opinion is it would be better to not mix several subsystem related
>> support in a single patch set. Both of the approaches (v4 or that you
>> suggested) do not prevents us to implement allocation support in next
>> patch series. After we have the base functionality enabled, we may add
>> support in other subsystems and drivers one by one with more focus
>> on the subsystem specificities and with the best possible attention.
>
> Yeah, I'm aware nothing is ever set in stone.
>
> Setting to 0 in blk_set_stacking_limits() is OK for now.

I get your point. Thanks for the suggestion and comments, Mike.

Kirill

\
 
 \ /
  Last update: 2020-01-21 16:14    [W:0.063 / U:0.316 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site