Message ID | alpine.LRH.2.11.1608102050200.10662@mail.ewheeler.net (mailing list archive) |
---|---|
State | Superseded, archived |
Delegated to: | Mike Snitzer |
Headers | show |
Yes, this should be backported. It was lost somehow. Mike, please put it to your git. Mikulas On Wed, 10 Aug 2016, Eric Wheeler wrote: > Hello Mikulas and dm-devel list, > > The simple patch below with is confirmed to fix James Johnston's issue and > doesn't appear to be in v4.8-rc1: > > This references the following patchwork entry: > https://patchwork.kernel.org/patch/9138595/ > > Can we get this pushed upstream for v4.8? > > -- > Eric Wheeler > > > On Fri, 27 May 2016, Mikulas Patocka wrote: > > dm-crypt: Fix error with too large bios > > > > When dm-crypt processes writes, it allocates a new bio in the function > > crypt_alloc_buffer. The bio is allocated from a bio set and it can have at > > most BIO_MAX_PAGES vector entries, however the incoming bio can be larger > > if it was allocated by other means. For example, bcache creates bios > > larger than BIO_MAX_PAGES. If the incoming bio is larger, bio_alloc_bioset > > fails and error is returned. > > > > To avoid the error, we test for too large bio in the function crypt_map > > and dm_accept_partial_bio to split the bio. dm_accept_partial_bio trims > > the current bio to the desired size and requests that the device mapper > > core sends another bio with the rest of the data. > > > > Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> > > Cc: stable@vger.kernel.org # v3.16+ > > Tested-by: James Johnston <johnstonj.public@codenest.com> > > I tested this patch by: > > 1. Building v4.7-rc1 from Torvalds git repo. Confirmed that original bug > still occurs on Ubuntu 15.10. > > 2. Applying your patch to v4.7-rc1. My kill sequence no longer works, > and the writeback cache is now successfully flushed to disk, and the > cache can be detached from the backing device. > > 3. To check data integrity, copied 250 MB of /dev/urandom to some file > on main volume. Then, dd copy this file to /dev/bcache0. Then, > detached the cache device from the backing device. Then, rebooted. > Then, dd copy /dev/bcache0 to another file on main volume. Then, > diff the files and confirm no changes. > > So it looks like it works, based on this admittedly brief testing. Thanks! > > Best regards, > > James Johnston > > > -- > dm-devel mailing list > dm-devel@redhat.com > https://www.redhat.com/mailman/listinfo/dm-devel > > Patch > > Index: linux-4.6/drivers/md/dm-crypt.c > =================================================================== > --- linux-4.6.orig/drivers/md/dm-crypt.c > +++ linux-4.6/drivers/md/dm-crypt.c > @@ -2137,6 +2137,10 @@ static int crypt_map(struct dm_target *t > struct dm_crypt_io *io; > struct crypt_config *cc = ti->private; > > + if (unlikely(bio->bi_iter.bi_size > BIO_MAX_SIZE) && > + (bio->bi_rw & (REQ_FLUSH | REQ_DISCARD | REQ_WRITE)) == REQ_WRITE) > + dm_accept_partial_bio(bio, BIO_MAX_SIZE >> SECTOR_SHIFT); > + > /* > * If bio is REQ_FLUSH or REQ_DISCARD, just bypass crypt queues. > * - for REQ_FLUSH device-mapper core ensures that no IO is in-flight > > > > -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
[top-posting just because others went wild with it] I don't have a strong opinion but I just assumed the local dm-crypt workaround wasn't the way forward. I didn't stage it because Christoph disagreed with it: https://lkml.org/lkml/2016/6/1/456 https://lkml.org/lkml/2016/6/1/477 Also, this would appear to be a more generic fix: "block: make sure big bio is splitted into at most 256 bvecs https://lkml.org/lkml/2016/8/12/154 (but Christoph disagrees there too, so the way forward isn't clear) Mike On Sat, Aug 13 2016 at 1:45pm -0400, Mikulas Patocka <mpatocka@redhat.com> wrote: > Yes, this should be backported. It was lost somehow. > > Mike, please put it to your git. > > Mikulas > > > On Wed, 10 Aug 2016, Eric Wheeler wrote: > > > Hello Mikulas and dm-devel list, > > > > The simple patch below with is confirmed to fix James Johnston's issue and > > doesn't appear to be in v4.8-rc1: > > > > This references the following patchwork entry: > > https://patchwork.kernel.org/patch/9138595/ > > > > Can we get this pushed upstream for v4.8? > > > > -- > > Eric Wheeler > > > > > > On Fri, 27 May 2016, Mikulas Patocka wrote: > > > dm-crypt: Fix error with too large bios > > > > > > When dm-crypt processes writes, it allocates a new bio in the function > > > crypt_alloc_buffer. The bio is allocated from a bio set and it can have at > > > most BIO_MAX_PAGES vector entries, however the incoming bio can be larger > > > if it was allocated by other means. For example, bcache creates bios > > > larger than BIO_MAX_PAGES. If the incoming bio is larger, bio_alloc_bioset > > > fails and error is returned. > > > > > > To avoid the error, we test for too large bio in the function crypt_map > > > and dm_accept_partial_bio to split the bio. dm_accept_partial_bio trims > > > the current bio to the desired size and requests that the device mapper > > > core sends another bio with the rest of the data. > > > > > > Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> > > > Cc: stable@vger.kernel.org # v3.16+ > > > > Tested-by: James Johnston <johnstonj.public@codenest.com> > > > > I tested this patch by: > > > > 1. Building v4.7-rc1 from Torvalds git repo. Confirmed that original bug > > still occurs on Ubuntu 15.10. > > > > 2. Applying your patch to v4.7-rc1. My kill sequence no longer works, > > and the writeback cache is now successfully flushed to disk, and the > > cache can be detached from the backing device. > > > > 3. To check data integrity, copied 250 MB of /dev/urandom to some file > > on main volume. Then, dd copy this file to /dev/bcache0. Then, > > detached the cache device from the backing device. Then, rebooted. > > Then, dd copy /dev/bcache0 to another file on main volume. Then, > > diff the files and confirm no changes. > > > > So it looks like it works, based on this admittedly brief testing. Thanks! > > > > Best regards, > > > > James Johnston > > > > > > -- > > dm-devel mailing list > > dm-devel@redhat.com > > https://www.redhat.com/mailman/listinfo/dm-devel > > > > Patch > > > > Index: linux-4.6/drivers/md/dm-crypt.c > > =================================================================== > > --- linux-4.6.orig/drivers/md/dm-crypt.c > > +++ linux-4.6/drivers/md/dm-crypt.c > > @@ -2137,6 +2137,10 @@ static int crypt_map(struct dm_target *t > > struct dm_crypt_io *io; > > struct crypt_config *cc = ti->private; > > > > + if (unlikely(bio->bi_iter.bi_size > BIO_MAX_SIZE) && > > + (bio->bi_rw & (REQ_FLUSH | REQ_DISCARD | REQ_WRITE)) == REQ_WRITE) > > + dm_accept_partial_bio(bio, BIO_MAX_SIZE >> SECTOR_SHIFT); > > + > > /* > > * If bio is REQ_FLUSH or REQ_DISCARD, just bypass crypt queues. > > * - for REQ_FLUSH device-mapper core ensures that no IO is in-flight > > > > > > > > -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Sat, 13 Aug 2016, Mike Snitzer wrote: > [top-posting just because others went wild with it] > > I don't have a strong opinion but I just assumed the local dm-crypt > workaround wasn't the way forward. I didn't stage it because Christoph > disagreed with it: > https://lkml.org/lkml/2016/6/1/456 > https://lkml.org/lkml/2016/6/1/477 > > Also, this would appear to be a more generic fix: > "block: make sure big bio is splitted into at most 256 bvecs > https://lkml.org/lkml/2016/8/12/154 > (but Christoph disagrees there too, so the way forward isn't clear) > > Mike On Wed, Jun 01 2016 at 9:44am -0400, Christoph Hellwig <hch@infradead.org> wrote: > > > be dm-crypt.c. Maybe you've identified some indirect use of > > > BIO_MAX_SIZE? > > > > I mean the recently introduced BIO_MAX_SIZE in -next tree: > > > https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/drivers/md/dm-crypt.c?id=4ed89c97b0706477b822ea2182827640c0cec486 > > The crazy bcache bios striking back once again. I really think it's > harmful having a _MAX value and then having a minor driver > reinterpreting it and sending larger ones. Until we can lift the > maximum limit in general nad have common code exercise it we really need > to stop bcache from sending these instead of littering the tree with > workarounds. The bio_kmalloc function allocates bios with up to 1024 vector entries (as opposed to bio_alloc and bio_alloc_bioset that has a limit of 256 vector entries). Device mapper is using bio_alloc_bioset with a bio set, so it is limited to 256 vector entries, but other kernel users may use bio_kmalloc and create larger bios. So, if you don't want bios with more than 256 vector entries to exist, you should impose this limit in bio_kmalloc (and fix all the callers that use it). Mikulas -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Mon, 15 Aug 2016, Mikulas Patocka wrote: > On Sat, 13 Aug 2016, Mike Snitzer wrote: > > > [top-posting just because others went wild with it] > > > > I don't have a strong opinion but I just assumed the local dm-crypt > > workaround wasn't the way forward. I didn't stage it because Christoph > > disagreed with it: > > https://lkml.org/lkml/2016/6/1/456 > > https://lkml.org/lkml/2016/6/1/477 > > > > Also, this would appear to be a more generic fix: > > "block: make sure big bio is splitted into at most 256 bvecs > > https://lkml.org/lkml/2016/8/12/154 > > (but Christoph disagrees there too, so the way forward isn't clear) > > > > Mike > > > On Wed, Jun 01 2016 at 9:44am -0400, Christoph Hellwig <hch@infradead.org> wrote: > > > > > be dm-crypt.c. Maybe you've identified some indirect use of > > > > BIO_MAX_SIZE? > > > > > > I mean the recently introduced BIO_MAX_SIZE in -next tree: > > > > > https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/drivers/md/dm-crypt.c?id=4ed89c97b0706477b822ea2182827640c0cec486 > > > > The crazy bcache bios striking back once again. I really think it's > > harmful having a _MAX value and then having a minor driver > > reinterpreting it and sending larger ones. Until we can lift the > > maximum limit in general nad have common code exercise it we really need > > to stop bcache from sending these instead of littering the tree with > > workarounds. > > The bio_kmalloc function allocates bios with up to 1024 vector entries (as > opposed to bio_alloc and bio_alloc_bioset that has a limit of 256 vector > entries). > > Device mapper is using bio_alloc_bioset with a bio set, so it is limited > to 256 vector entries, but other kernel users may use bio_kmalloc and > create larger bios. > > So, if you don't want bios with more than 256 vector entries to exist, you > should impose this limit in bio_kmalloc (and fix all the callers that use > it). FYI, Kent Overstreet notes this about bcache from the other thread here: https://lkml.org/lkml/2016/8/15/620 [paste] >> bcache originally had workaround code to split too-large bios when it >> first went upstream - that was dropped only after the patches to make >> generic_make_request() handle arbitrary size bios went in. So to do what >> you're suggesting would mean reverting that bcache patch and bringing that >> code back, which from my perspective would be a step in the wrong >> direction. I just want to get this over and done with. >> >> re: interactions with other drivers - bio_clone() has already been changed >> to only clone biovecs that are live for current bi_iter, so there >> shouldn't be any safety issues. A driver would have to be intentionally >> doing its own open coded bio cloning that clones all of bi_io_vec, not >> just the active ones - but if they're doing that, they're already broken >> because a driver isn't allowed to look at bi_vcnt if it isn't a bio that >> it owns - bi_vcnt is 0 on bios that don't own their biovec (i.e. that were >> created by bio_clone_fast). >> >> And the cloning and bi_vcnt usage stuff I audited very thoroughly back >> when I was working on immutable biovecs and such back in the day, and I >> had to do a fair amount of cleanup/refactoring before that stuff could go >> in. [/paste] They are making progress in the patch-v3 thread, so perhaps this can be fixed for now in generic_make_request(). -- Eric Wheeler -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Thu, 18 Aug 2016, Eric Wheeler wrote: > > On Wed, Jun 01 2016 at 9:44am -0400, Christoph Hellwig <hch@infradead.org> wrote: > > > > > > > be dm-crypt.c. Maybe you've identified some indirect use of > > > > > BIO_MAX_SIZE? > > > > > > > > I mean the recently introduced BIO_MAX_SIZE in -next tree: > > > > > > > https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/drivers/md/dm-crypt.c?id=4ed89c97b0706477b822ea2182827640c0cec486 > > > > > > The crazy bcache bios striking back once again. I really think it's > > > harmful having a _MAX value and then having a minor driver > > > reinterpreting it and sending larger ones. Until we can lift the > > > maximum limit in general nad have common code exercise it we really need > > > to stop bcache from sending these instead of littering the tree with > > > workarounds. > > > > The bio_kmalloc function allocates bios with up to 1024 vector entries (as > > opposed to bio_alloc and bio_alloc_bioset that has a limit of 256 vector > > entries). > > > > Device mapper is using bio_alloc_bioset with a bio set, so it is limited > > to 256 vector entries, but other kernel users may use bio_kmalloc and > > create larger bios. > > > > So, if you don't want bios with more than 256 vector entries to exist, you > > should impose this limit in bio_kmalloc (and fix all the callers that use > > it). > > FYI, Kent Overstreet notes this about bcache from the other thread here: > https://lkml.org/lkml/2016/8/15/620 > > [paste] > >> bcache originally had workaround code to split too-large bios when it > >> first went upstream - that was dropped only after the patches to make > >> generic_make_request() handle arbitrary size bios went in. So to do what > >> you're suggesting would mean reverting that bcache patch and bringing that > >> code back, which from my perspective would be a step in the wrong > >> direction. I just want to get this over and done with. > >> > >> re: interactions with other drivers - bio_clone() has already been changed > >> to only clone biovecs that are live for current bi_iter, so there > >> shouldn't be any safety issues. A driver would have to be intentionally > >> doing its own open coded bio cloning that clones all of bi_io_vec, not > >> just the active ones - but if they're doing that, they're already broken > >> because a driver isn't allowed to look at bi_vcnt if it isn't a bio that > >> it owns - bi_vcnt is 0 on bios that don't own their biovec (i.e. that were > >> created by bio_clone_fast). > >> > >> And the cloning and bi_vcnt usage stuff I audited very thoroughly back > >> when I was working on immutable biovecs and such back in the day, and I > >> had to do a fair amount of cleanup/refactoring before that stuff could go > >> in. > [/paste] > > They are making progress in the patch-v3 thread, so perhaps this can be > fixed for now in generic_make_request(). > > -- > Eric Wheeler Device mapper can't split the bio in generic_make_request - it frees the md->queue->bio_split bioset, to save one kernel thread per device. Device mapper uses its own splitting mechanism. So what is the final decision? - should device mapper split the big bio or should bcache not submit big bios? I think splitting big bios in the device mapper is better - simply because it is much less code than reworking bcache to split bios internally. BTW. In the device mapper, we have a layer dm-io, that was created to work around bio size limitations - it accepts unlimited I/O request and splits it to several bios. When bio size limitations are gone, we could simplify dm-io too. Mikulas -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On 08/25/2016 12:34 PM, Mikulas Patocka wrote: > > > On Thu, 18 Aug 2016, Eric Wheeler wrote: > >>> On Wed, Jun 01 2016 at 9:44am -0400, Christoph Hellwig <hch@infradead.org> wrote: >>> >>>>>> be dm-crypt.c. Maybe you've identified some indirect use of >>>>>> BIO_MAX_SIZE? >>>>> >>>>> I mean the recently introduced BIO_MAX_SIZE in -next tree: >>>>> >>>> https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/drivers/md/dm-crypt.c?id=4ed89c97b0706477b822ea2182827640c0cec486 >>>> >>>> The crazy bcache bios striking back once again. I really think it's >>>> harmful having a _MAX value and then having a minor driver >>>> reinterpreting it and sending larger ones. Until we can lift the >>>> maximum limit in general nad have common code exercise it we really need >>>> to stop bcache from sending these instead of littering the tree with >>>> workarounds. >>> >>> The bio_kmalloc function allocates bios with up to 1024 vector entries (as >>> opposed to bio_alloc and bio_alloc_bioset that has a limit of 256 vector >>> entries). >>> >>> Device mapper is using bio_alloc_bioset with a bio set, so it is limited >>> to 256 vector entries, but other kernel users may use bio_kmalloc and >>> create larger bios. >>> >>> So, if you don't want bios with more than 256 vector entries to exist, you >>> should impose this limit in bio_kmalloc (and fix all the callers that use >>> it). >> >> FYI, Kent Overstreet notes this about bcache from the other thread here: >> https://lkml.org/lkml/2016/8/15/620 >> >> [paste] >>>> bcache originally had workaround code to split too-large bios when it >>>> first went upstream - that was dropped only after the patches to make >>>> generic_make_request() handle arbitrary size bios went in. So to do what >>>> you're suggesting would mean reverting that bcache patch and bringing that >>>> code back, which from my perspective would be a step in the wrong >>>> direction. I just want to get this over and done with. >>>> >>>> re: interactions with other drivers - bio_clone() has already been changed >>>> to only clone biovecs that are live for current bi_iter, so there >>>> shouldn't be any safety issues. A driver would have to be intentionally >>>> doing its own open coded bio cloning that clones all of bi_io_vec, not >>>> just the active ones - but if they're doing that, they're already broken >>>> because a driver isn't allowed to look at bi_vcnt if it isn't a bio that >>>> it owns - bi_vcnt is 0 on bios that don't own their biovec (i.e. that were >>>> created by bio_clone_fast). >>>> >>>> And the cloning and bi_vcnt usage stuff I audited very thoroughly back >>>> when I was working on immutable biovecs and such back in the day, and I >>>> had to do a fair amount of cleanup/refactoring before that stuff could go >>>> in. >> [/paste] >> >> They are making progress in the patch-v3 thread, so perhaps this can be >> fixed for now in generic_make_request(). >> >> -- >> Eric Wheeler > > Device mapper can't split the bio in generic_make_request - it frees the > md->queue->bio_split bioset, to save one kernel thread per device. Device > mapper uses its own splitting mechanism. > > So what is the final decision? - should device mapper split the big bio or > should bcache not submit big bios? > > I think splitting big bios in the device mapper is better - simply because > it is much less code than reworking bcache to split bios internally. > > BTW. In the device mapper, we have a layer dm-io, that was created to work > around bio size limitations - it accepts unlimited I/O request and splits > it to several bios. When bio size limitations are gone, we could simplify > dm-io too. The patch from Ming Lei was applied for 4.8 the other day.
On Thu, Aug 25 2016 at 4:13pm -0400, Jens Axboe <axboe@kernel.dk> wrote: > On 08/25/2016 12:34 PM, Mikulas Patocka wrote: > > > >Device mapper can't split the bio in generic_make_request - it frees the > >md->queue->bio_split bioset, to save one kernel thread per device. Device > >mapper uses its own splitting mechanism. > > > >So what is the final decision? - should device mapper split the big bio or > >should bcache not submit big bios? > > > >I think splitting big bios in the device mapper is better - simply because > >it is much less code than reworking bcache to split bios internally. > > > >BTW. In the device mapper, we have a layer dm-io, that was created to work > >around bio size limitations - it accepts unlimited I/O request and splits > >it to several bios. When bio size limitations are gone, we could simplify > >dm-io too. > > The patch from Ming Lei was applied for 4.8 the other day. See linux-block commit: 4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Fri, 26 Aug 2016, Mike Snitzer wrote: > On Thu, Aug 25 2016 at 4:13pm -0400, > Jens Axboe <axboe@kernel.dk> wrote: > > > On 08/25/2016 12:34 PM, Mikulas Patocka wrote: > > > > > >Device mapper can't split the bio in generic_make_request - it frees the > > >md->queue->bio_split bioset, to save one kernel thread per device. Device > > >mapper uses its own splitting mechanism. > > > > > >So what is the final decision? - should device mapper split the big bio or > > >should bcache not submit big bios? > > > > > >I think splitting big bios in the device mapper is better - simply because > > >it is much less code than reworking bcache to split bios internally. > > > > > >BTW. In the device mapper, we have a layer dm-io, that was created to work > > >around bio size limitations - it accepts unlimited I/O request and splits > > >it to several bios. When bio size limitations are gone, we could simplify > > >dm-io too. > > > > The patch from Ming Lei was applied for 4.8 the other day. > > See linux-block commit: > 4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs > http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c But this patch won't work for device mapper, blk_bio_segment_split is called from blk_queue_split and device mapper doesn't use blk_queue_split (it can't because it frees queue->bio_split). We still need these two patches: https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html Mikulas -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Sat, Aug 27, 2016 at 11:09 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Fri, 26 Aug 2016, Mike Snitzer wrote: > >> On Thu, Aug 25 2016 at 4:13pm -0400, >> Jens Axboe <axboe@kernel.dk> wrote: >> >> > On 08/25/2016 12:34 PM, Mikulas Patocka wrote: >> > > >> > >Device mapper can't split the bio in generic_make_request - it frees the >> > >md->queue->bio_split bioset, to save one kernel thread per device. Device >> > >mapper uses its own splitting mechanism. >> > > >> > >So what is the final decision? - should device mapper split the big bio or >> > >should bcache not submit big bios? >> > > >> > >I think splitting big bios in the device mapper is better - simply because >> > >it is much less code than reworking bcache to split bios internally. >> > > >> > >BTW. In the device mapper, we have a layer dm-io, that was created to work >> > >around bio size limitations - it accepts unlimited I/O request and splits >> > >it to several bios. When bio size limitations are gone, we could simplify >> > >dm-io too. >> > >> > The patch from Ming Lei was applied for 4.8 the other day. >> >> See linux-block commit: >> 4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs >> http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c > > But this patch won't work for device mapper, blk_bio_segment_split is > called from blk_queue_split and device mapper doesn't use blk_queue_split > (it can't because it frees queue->bio_split). > > We still need these two patches: > https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html > https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html About the 2nd patch, it might not be good enough because in theory a small size bio still may include big bvecs, such as, each bvec points to 512byte buffer, so strictly speaking the bvec number should be checked instead of bio size.
On Sat, Aug 27 2016 at 11:09am -0400, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Fri, 26 Aug 2016, Mike Snitzer wrote: > > > On Thu, Aug 25 2016 at 4:13pm -0400, > > Jens Axboe <axboe@kernel.dk> wrote: > > > > > On 08/25/2016 12:34 PM, Mikulas Patocka wrote: > > > > > > > >Device mapper can't split the bio in generic_make_request - it frees the > > > >md->queue->bio_split bioset, to save one kernel thread per device. Device > > > >mapper uses its own splitting mechanism. > > > > > > > >So what is the final decision? - should device mapper split the big bio or > > > >should bcache not submit big bios? > > > > > > > >I think splitting big bios in the device mapper is better - simply because > > > >it is much less code than reworking bcache to split bios internally. > > > > > > > >BTW. In the device mapper, we have a layer dm-io, that was created to work > > > >around bio size limitations - it accepts unlimited I/O request and splits > > > >it to several bios. When bio size limitations are gone, we could simplify > > > >dm-io too. > > > > > > The patch from Ming Lei was applied for 4.8 the other day. > > > > See linux-block commit: > > 4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs > > http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c > > But this patch won't work for device mapper, blk_bio_segment_split is > called from blk_queue_split and device mapper doesn't use blk_queue_split > (it can't because it frees queue->bio_split). > > We still need these two patches: > https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html > https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html I'm left wondering whether it'd be better to revert commit dbba42d8a9e ("dm: eliminate unused "bioset" process for each bio-based DM device") And also look for other areas of DM that could leverage the block core's relatively new splitting capabilities. Otherwise, we'll just be sprinking more BIO_MAX_PAGES-based splitting code in DM targets because DM is so "special" that it doesn't need the block core's splitting (even though it clearly would now benefit). Mike -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Mon, 29 Aug 2016, Ming Lei wrote: > On Sat, Aug 27, 2016 at 11:09 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > > > > On Fri, 26 Aug 2016, Mike Snitzer wrote: > > > >> On Thu, Aug 25 2016 at 4:13pm -0400, > >> Jens Axboe <axboe@kernel.dk> wrote: > >> > >> > On 08/25/2016 12:34 PM, Mikulas Patocka wrote: > >> > > > >> > >Device mapper can't split the bio in generic_make_request - it frees the > >> > >md->queue->bio_split bioset, to save one kernel thread per device. Device > >> > >mapper uses its own splitting mechanism. > >> > > > >> > >So what is the final decision? - should device mapper split the big bio or > >> > >should bcache not submit big bios? > >> > > > >> > >I think splitting big bios in the device mapper is better - simply because > >> > >it is much less code than reworking bcache to split bios internally. > >> > > > >> > >BTW. In the device mapper, we have a layer dm-io, that was created to work > >> > >around bio size limitations - it accepts unlimited I/O request and splits > >> > >it to several bios. When bio size limitations are gone, we could simplify > >> > >dm-io too. > >> > > >> > The patch from Ming Lei was applied for 4.8 the other day. > >> > >> See linux-block commit: > >> 4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs > >> http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c > > > > But this patch won't work for device mapper, blk_bio_segment_split is > > called from blk_queue_split and device mapper doesn't use blk_queue_split > > (it can't because it frees queue->bio_split). > > > > We still need these two patches: > > https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html > > https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html > > About the 2nd patch, it might not be good enough because in theory > a small size bio still may include big bvecs, such as, each bvec points > to 512byte buffer, so strictly speaking the bvec number should > be checked instead of bio size. > > Ming Lei This is not a problem. dm-crypt allocates new pages for the bio and copies (and encrypts) the data from the original location to the new pages. So yes, the original bio may have a long vector with many small fragments, but the newly allocated memory will be allocated in full pages and the outgoing bio's vector will have full pages. Mikulas -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Mon, 29 Aug 2016, Mike Snitzer wrote: > On Sat, Aug 27 2016 at 11:09am -0400, > Mikulas Patocka <mpatocka@redhat.com> wrote: > > > > > > > On Fri, 26 Aug 2016, Mike Snitzer wrote: > > > > > On Thu, Aug 25 2016 at 4:13pm -0400, > > > Jens Axboe <axboe@kernel.dk> wrote: > > > > > > > On 08/25/2016 12:34 PM, Mikulas Patocka wrote: > > > > > > > > > >Device mapper can't split the bio in generic_make_request - it frees the > > > > >md->queue->bio_split bioset, to save one kernel thread per device. Device > > > > >mapper uses its own splitting mechanism. > > > > > > > > > >So what is the final decision? - should device mapper split the big bio or > > > > >should bcache not submit big bios? > > > > > > > > > >I think splitting big bios in the device mapper is better - simply because > > > > >it is much less code than reworking bcache to split bios internally. > > > > > > > > > >BTW. In the device mapper, we have a layer dm-io, that was created to work > > > > >around bio size limitations - it accepts unlimited I/O request and splits > > > > >it to several bios. When bio size limitations are gone, we could simplify > > > > >dm-io too. > > > > > > > > The patch from Ming Lei was applied for 4.8 the other day. > > > > > > See linux-block commit: > > > 4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs > > > http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c > > > > But this patch won't work for device mapper, blk_bio_segment_split is > > called from blk_queue_split and device mapper doesn't use blk_queue_split > > (it can't because it frees queue->bio_split). > > > > We still need these two patches: > > https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html > > https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html > > I'm left wondering whether it'd be better to revert commit dbba42d8a9e > ("dm: eliminate unused "bioset" process for each bio-based DM device") That would create one more process per dm device. There is no need to create another process if the bio can be split in the device mapper. > And also look for other areas of DM that could leverage the block core's > relatively new splitting capabilities. > > Otherwise, we'll just be sprinking more BIO_MAX_PAGES-based splitting > code in DM targets because DM is so "special" that it doesn't need the > block core's splitting (even though it clearly would now benefit). It would be possible to split bios on BIO_MAX_PAGES boundary in the dm core. But I checked all the dm targets and I found problem only in dm-crypt and dm-log-writes. So there is no need to split bio bios for the other targets. > Mike Mikulas -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Tue, Aug 30, 2016 at 5:57 AM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Mon, 29 Aug 2016, Ming Lei wrote: > >> On Sat, Aug 27, 2016 at 11:09 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> > >> > >> > On Fri, 26 Aug 2016, Mike Snitzer wrote: >> > >> >> On Thu, Aug 25 2016 at 4:13pm -0400, >> >> Jens Axboe <axboe@kernel.dk> wrote: >> >> >> >> > On 08/25/2016 12:34 PM, Mikulas Patocka wrote: >> >> > > >> >> > >Device mapper can't split the bio in generic_make_request - it frees the >> >> > >md->queue->bio_split bioset, to save one kernel thread per device. Device >> >> > >mapper uses its own splitting mechanism. >> >> > > >> >> > >So what is the final decision? - should device mapper split the big bio or >> >> > >should bcache not submit big bios? >> >> > > >> >> > >I think splitting big bios in the device mapper is better - simply because >> >> > >it is much less code than reworking bcache to split bios internally. >> >> > > >> >> > >BTW. In the device mapper, we have a layer dm-io, that was created to work >> >> > >around bio size limitations - it accepts unlimited I/O request and splits >> >> > >it to several bios. When bio size limitations are gone, we could simplify >> >> > >dm-io too. >> >> > >> >> > The patch from Ming Lei was applied for 4.8 the other day. >> >> >> >> See linux-block commit: >> >> 4d70dca4eadf2f block: make sure a big bio is split into at most 256 bvecs >> >> http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=4d70dca4eadf2f95abe389116ac02b8439c2d16c >> > >> > But this patch won't work for device mapper, blk_bio_segment_split is >> > called from blk_queue_split and device mapper doesn't use blk_queue_split >> > (it can't because it frees queue->bio_split). >> > >> > We still need these two patches: >> > https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html >> > https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html >> >> About the 2nd patch, it might not be good enough because in theory >> a small size bio still may include big bvecs, such as, each bvec points >> to 512byte buffer, so strictly speaking the bvec number should >> be checked instead of bio size. >> >> Ming Lei > > This is not a problem. I meant the following code in your 2nd patch: + if (unlikely(bio->bi_iter.bi_size > BIO_MAX_SIZE) && + (bio->bi_rw & (REQ_FLUSH | REQ_DISCARD | REQ_WRITE)) == REQ_WRITE) + dm_accept_partial_bio(bio, BIO_MAX_SIZE >> SECTOR_SHIFT); And the check on .bi_size may not work. > > dm-crypt allocates new pages for the bio and copies (and encrypts) the > data from the original location to the new pages. > > So yes, the original bio may have a long vector with many small fragments, > but the newly allocated memory will be allocated in full pages and the > outgoing bio's vector will have full pages. > > Mikulas
On Tue, 30 Aug 2016, Ming Lei wrote: > On Tue, Aug 30, 2016 at 5:57 AM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > > > > On Mon, 29 Aug 2016, Ming Lei wrote: > > > >> On Sat, Aug 27, 2016 at 11:09 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > >> > > >> > But this patch won't work for device mapper, blk_bio_segment_split is > >> > called from blk_queue_split and device mapper doesn't use blk_queue_split > >> > (it can't because it frees queue->bio_split). > >> > > >> > We still need these two patches: > >> > https://www.redhat.com/archives/dm-devel/2016-May/msg00211.html > >> > https://www.redhat.com/archives/dm-devel/2016-May/msg00210.html > >> > >> About the 2nd patch, it might not be good enough because in theory > >> a small size bio still may include big bvecs, such as, each bvec points > >> to 512byte buffer, so strictly speaking the bvec number should > >> be checked instead of bio size. > >> > >> Ming Lei > > > > This is not a problem. > > I meant the following code in your 2nd patch: > > + if (unlikely(bio->bi_iter.bi_size > BIO_MAX_SIZE) && > + (bio->bi_rw & (REQ_FLUSH | REQ_DISCARD | REQ_WRITE)) == REQ_WRITE) > + dm_accept_partial_bio(bio, BIO_MAX_SIZE >> SECTOR_SHIFT); > > And the check on .bi_size may not work. kcryptd_crypt_write_convert calls: crypt_alloc_buffer(io, io->base_bio->bi_iter.bi_size) crypt_alloc_buffer does: unsigned int nr_iovecs = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; clone = bio_alloc_bioset(GFP_NOIO, nr_iovecs, cc->bs); So, if io->base_bio->bi_iter.bi_size <= BIO_MAX_SIZE, then nr_iovecs will be less or equal than BIO_MAX_PAGES and the function bio_alloc_bioset will succeed. (BTW. BIO_MAX_SIZE was removed in the current kernel, we should use (BIO_MAX_PAGES << PAGE_SHIFT) instead). Mikulas -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
Index: linux-4.6/drivers/md/dm-crypt.c =================================================================== --- linux-4.6.orig/drivers/md/dm-crypt.c +++ linux-4.6/drivers/md/dm-crypt.c @@ -2137,6 +2137,10 @@ static int crypt_map(struct dm_target *t struct dm_crypt_io *io; struct crypt_config *cc = ti->private; + if (unlikely(bio->bi_iter.bi_size > BIO_MAX_SIZE) && + (bio->bi_rw & (REQ_FLUSH | REQ_DISCARD | REQ_WRITE)) == REQ_WRITE) + dm_accept_partial_bio(bio, BIO_MAX_SIZE >> SECTOR_SHIFT); + /* * If bio is REQ_FLUSH or REQ_DISCARD, just bypass crypt queues. * - for REQ_FLUSH device-mapper core ensures that no IO is in-flight