| Message ID | 20200806215837.3968445-1-kbusch@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | block: fix get_max_io_size() | expand |
On 8/6/20 3:58 PM, Keith Busch wrote: > A previous commit aligning splits to physical block sizes inadvertently > modified one return case such that that it now returns 0 length splits > when the number of sectors doesn't exceed the physical offset. This > later hits a BUG in bio_split(). Restore the previous working behavior. Yikes! I wonder how that lived so long... Applied.
On 2020-08-06 14:58, Keith Busch wrote: > A previous commit aligning splits to physical block sizes inadvertently > modified one return case such that that it now returns 0 length splits > when the number of sectors doesn't exceed the physical offset. This > later hits a BUG in bio_split(). Restore the previous working behavior. > > Reported-by: Eric Deal <eric.deal@wdc.com> > Cc: Bart Van Assche <bvanassche@acm.org> > Cc: stable@vger.kernel.org > Fixes: 9cc5169cd478b ("block: Improve physical block alignment of split bios") > Signed-off-by: Keith Busch <kbusch@kernel.org> > --- > block/blk-merge.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/block/blk-merge.c b/block/blk-merge.c > index 5196dc145270..d7fef954d42f 100644 > --- a/block/blk-merge.c > +++ b/block/blk-merge.c > @@ -154,7 +154,7 @@ static inline unsigned get_max_io_size(struct request_queue *q, > if (max_sectors > start_offset) > return max_sectors - start_offset; > > - return sectors & (lbs - 1); > + return sectors & ~(lbs - 1); > } I think we agree that get_max_io_size() should never return zero. However, the above change seems wrong to me because it will cause get_max_io_size() to return zero if the logical block size is larger than 512 bytes and if sectors < lbs. How about changing the return statement as follows (untested): return max(sectors & (lbs - 1), sectors); Thanks, Bart.
On 2020-08-06 17:28, Bart Van Assche wrote: > On 2020-08-06 14:58, Keith Busch wrote: >> A previous commit aligning splits to physical block sizes inadvertently >> modified one return case such that that it now returns 0 length splits >> when the number of sectors doesn't exceed the physical offset. This >> later hits a BUG in bio_split(). Restore the previous working behavior. >> >> Reported-by: Eric Deal <eric.deal@wdc.com> >> Cc: Bart Van Assche <bvanassche@acm.org> >> Cc: stable@vger.kernel.org >> Fixes: 9cc5169cd478b ("block: Improve physical block alignment of split bios") >> Signed-off-by: Keith Busch <kbusch@kernel.org> >> --- >> block/blk-merge.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/block/blk-merge.c b/block/blk-merge.c >> index 5196dc145270..d7fef954d42f 100644 >> --- a/block/blk-merge.c >> +++ b/block/blk-merge.c >> @@ -154,7 +154,7 @@ static inline unsigned get_max_io_size(struct request_queue *q, >> if (max_sectors > start_offset) >> return max_sectors - start_offset; >> >> - return sectors & (lbs - 1); >> + return sectors & ~(lbs - 1); >> } > > I think we agree that get_max_io_size() should never return zero. However, the above > change seems wrong to me because it will cause get_max_io_size() to return zero if > the logical block size is larger than 512 bytes and if sectors < lbs. How about > changing the return statement as follows (untested): This should work better than what was mentioned in my previous email: - return sectors & (lbs - 1); + return sectors; Thanks, Bart.
On Thu, Aug 06, 2020 at 05:28:17PM -0700, Bart Van Assche wrote: > I think we agree that get_max_io_size() should never return zero. However, the above > change seems wrong to me because it will cause get_max_io_size() to return zero if > the logical block size is larger than 512 bytes and if sectors < lbs. I'm pretty sure we have more problems if 'sectors' isn't a multiple of the logical block size.
On Thu, Aug 06, 2020 at 06:25:50PM -0700, Bart Van Assche wrote: > This should work better than what was mentioned in my previous email: > > - return sectors & (lbs - 1); > + return sectors; It used to be something like that. There were some situations where it didn't work, which brought d0e5fbb01a67e, but I think the real problem was from mismatched queue_limits, which I think I addressed with 5f009d3f8e668, so maybe this is okay now.
On 2020-08-06 20:24, Keith Busch wrote: > On Thu, Aug 06, 2020 at 06:25:50PM -0700, Bart Van Assche wrote: >> This should work better than what was mentioned in my previous email: >> >> - return sectors & (lbs - 1); >> + return sectors; > > It used to be something like that. There were some situations where it > didn't work, which brought d0e5fbb01a67e, but I think the real problem > was from mismatched queue_limits, which I think I addressed with > 5f009d3f8e668, so maybe this is okay now. Hi Keith, How about replacing your patch with the (untested) patch below? Thanks, Bart. diff --git a/block/blk-merge.c b/block/blk-merge.c index 5196dc145270..2d10fa3768a3 100644 --- a/block/blk-merge.c +++ b/block/blk-merge.c @@ -135,10 +135,9 @@ static struct bio *blk_bio_write_same_split(struct request_queue *q, /* * Return the maximum number of sectors from the start of a bio that may be * submitted as a single request to a block device. If enough sectors remain, - * align the end to the physical block size. Otherwise align the end to the - * logical block size. This approach minimizes the number of non-aligned - * requests that are submitted to a block device if the start of a bio is not - * aligned to a physical block boundary. + * align the end to the physical block size. This approach minimizes the + * number of non-aligned requests that are submitted to a block device if the + * start of a bio is not aligned to a physical block boundary. */ static inline unsigned get_max_io_size(struct request_queue *q, struct bio *bio) @@ -146,7 +145,6 @@ static inline unsigned get_max_io_size(struct request_queue *q, unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector); unsigned max_sectors = sectors; unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT; - unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT; unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1); max_sectors += start_offset; @@ -154,7 +152,7 @@ static inline unsigned get_max_io_size(struct request_queue *q, if (max_sectors > start_offset) return max_sectors - start_offset; - return sectors & (lbs - 1); + return sectors; } static inline unsigned get_max_segment_size(const struct request_queue *q,
On Fri, Aug 07, 2020 at 07:18:49AM -0700, Bart Van Assche wrote: > Hi Keith, > > How about replacing your patch with the (untested) patch below? I believe that should be fine, but I broke the kernel last time I did something like that. I still think it was from incorrect queue_limits, but Linus disagreed. * http://lkml.iu.edu/hypermail/linux/kernel/1601.2/03994.html
On 2020-08-07 10:10, Keith Busch wrote: > On Fri, Aug 07, 2020 at 07:18:49AM -0700, Bart Van Assche wrote: >> Hi Keith, >> >> How about replacing your patch with the (untested) patch below? > > > I believe that should be fine, but I broke the kernel last time I did > something like that. I still think it was from incorrect queue_limits, > but Linus disagreed. > > * http://lkml.iu.edu/hypermail/linux/kernel/1601.2/03994.html Hi Keith, Thanks for the interesting link. Regarding Linus' comments about bio splitting: if the last return statement in get_max_io_size() is reached then it is guaranteed that sectors < pbs (physical block size). So I think that Linus' comment applies to the previous return statement instead of to the last ("return max_sectors - start_offset;"). However, I think it is already guaranteed that that value is a multiple of the logical block size because start_offset is a multiple of the logical block size and because of the following statement: "max_sectors &= ~(pbs - 1);". Bart.
On Fri, Aug 07, 2020 at 12:02:30PM -0700, Bart Van Assche wrote: > On 2020-08-07 10:10, Keith Busch wrote: > > On Fri, Aug 07, 2020 at 07:18:49AM -0700, Bart Van Assche wrote: > >> Hi Keith, > >> > >> How about replacing your patch with the (untested) patch below? > > > > > > I believe that should be fine, but I broke the kernel last time I did > > something like that. I still think it was from incorrect queue_limits, > > but Linus disagreed. > > > > * http://lkml.iu.edu/hypermail/linux/kernel/1601.2/03994.html > > Hi Keith, > > Thanks for the interesting link. Regarding Linus' comments about bio > splitting: if the last return statement in get_max_io_size() is reached > then it is guaranteed that sectors < pbs (physical block size). So I think > that Linus' comment applies to the previous return statement instead of to > the last ("return max_sectors - start_offset;"). However, I think it is > already guaranteed that that value is a multiple of the logical block size > because start_offset is a multiple of the logical block size and because > of the following statement: "max_sectors &= ~(pbs - 1);". This breaks if limits.max_sectors is not a multiple of the queue's logical block size and the physical block size is larger than max_sectors.
Hi Jens, The proposed alternatives continue to break with allowable (however unlikely) queue limits, where this should be safe for any possible settings. I think this should be okay to go as-is. On Thu, Aug 06, 2020 at 02:58:37PM -0700, Keith Busch wrote: > A previous commit aligning splits to physical block sizes inadvertently > modified one return case such that that it now returns 0 length splits > when the number of sectors doesn't exceed the physical offset. This > later hits a BUG in bio_split(). Restore the previous working behavior. > > Reported-by: Eric Deal <eric.deal@wdc.com> > Cc: Bart Van Assche <bvanassche@acm.org> > Cc: stable@vger.kernel.org > Fixes: 9cc5169cd478b ("block: Improve physical block alignment of split bios") > Signed-off-by: Keith Busch <kbusch@kernel.org> > --- > block/blk-merge.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/block/blk-merge.c b/block/blk-merge.c > index 5196dc145270..d7fef954d42f 100644 > --- a/block/blk-merge.c > +++ b/block/blk-merge.c > @@ -154,7 +154,7 @@ static inline unsigned get_max_io_size(struct request_queue *q, > if (max_sectors > start_offset) > return max_sectors - start_offset; > > - return sectors & (lbs - 1); > + return sectors & ~(lbs - 1); > } > > static inline unsigned get_max_segment_size(const struct request_queue *q, > --
On 8/18/20 9:39 AM, Keith Busch wrote: > Hi Jens, > > The proposed alternatives continue to break with allowable (however > unlikely) queue limits, where this should be safe for any possible > settings. I think this should be okay to go as-is. OK, let's try this again then. Queued up for 5.9.
diff --git a/block/blk-merge.c b/block/blk-merge.c index 5196dc145270..d7fef954d42f 100644 --- a/block/blk-merge.c +++ b/block/blk-merge.c @@ -154,7 +154,7 @@ static inline unsigned get_max_io_size(struct request_queue *q, if (max_sectors > start_offset) return max_sectors - start_offset; - return sectors & (lbs - 1); + return sectors & ~(lbs - 1); } static inline unsigned get_max_segment_size(const struct request_queue *q,
A previous commit aligning splits to physical block sizes inadvertently modified one return case such that that it now returns 0 length splits when the number of sectors doesn't exceed the physical offset. This later hits a BUG in bio_split(). Restore the previous working behavior. Reported-by: Eric Deal <eric.deal@wdc.com> Cc: Bart Van Assche <bvanassche@acm.org> Cc: stable@vger.kernel.org Fixes: 9cc5169cd478b ("block: Improve physical block alignment of split bios") Signed-off-by: Keith Busch <kbusch@kernel.org> --- block/blk-merge.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)