Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Fri, Aug 7, 2015 at 5:30 PM, Kent Overstreet
<kent.overstreet at gmail.com> wrote:
> On Fri, Aug 07, 2015 at 04:40:06PM -0700, Ming Lin wrote:
>>
>> On Fri, 2015-08-07 at 09:30 +0200, Christoph Hellwig wrote:
>> > I'm for solution 3:
>> >
>> > - keep blk_bio_{discard,write_same}_split, but ensure we never built
>> > a > 4GB bio in blkdev_issue_{discard,write_same}.
>>
>> This has problem as I mentioned in solution 1.
>> We need to also make sure max discard size is of proper granularity.
>> See below example.
>>
>> 4G: 8388608 sectors
>> UINT_MAX: 8388607 sectors
>>
>> dm-thinp block size = default discard granularity = 128 sectors
>>
>> blkdev_issue_discard(sector=0, nr_sectors=8388608)
>>
>> 1. Only ensure bi_size not overflow
>>
>> It doesn't work.
>>
>> [start_sector, end_sector]
>> [0, 8388607]
>> [0, 8388606], then dm-thinp splits it to 2 bios
>> [0, 8388479]
>> [8388480, 8388606] ---> this has problem in process_discard_bio(),
>> because the discard size(7 sectors) covers less than a block(128 sectors)
>> [8388607, 8388607] ---> same problem
>>
>> 2. Ensure bi_size not overflow and max discard size is of proper granularity
>
> Ideally we'd get upper layers out of the business of knowing about the queue
> limits at all - that was the point of the patch series, after all.
>
> Instead of using UINT_MAX, would it work to just make the max 1 << 31 sectors?'
1 << 31 = 2G bytes = 0x400000 sectors.
Yes, that works as long as it's multiple of granularity.