On Wed, Mar 8, 2023 at 2:56 AM Hou Tao <hou...@huaweicloud.com> wrote: > > Hi, > > On 3/7/2023 10:47 PM, Mike Snitzer wrote: > > On Mon, Mar 06 2023 at 9:12P -0500, > > Hou Tao <hou...@huaweicloud.com> wrote: > > > >> Hi, > >> > >> On 3/7/2023 3:31 AM, Mike Snitzer wrote: > >>> On Mon, Mar 06 2023 at 8:49P -0500, > >>> Hou Tao <hou...@huaweicloud.com> wrote: > >>> > >>>> From: Hou Tao <hout...@huawei.com> > >>>> > >>>> When neither no_read_workqueue nor no_write_workqueue are enabled, > >>>> tasklet_trylock() in crypt_dec_pending() may still return false due to > >>>> an uninitialized state, and dm-crypt will do io completion in io_queue > >>>> instead of current context unnecessarily. > >>> Have you actually experienced this? > >> Yes. I had written a bpftrace script to check the completion context of > >> blkdev_bio_end_io_simple() when doing direct io read on dm-crypt device. > >> The > >> expected context should be unbound workers of crypt_queue, but sometimes > >> the > >> context is the bound worker of io_queue. > > OK, thanks for clarifying. Curious to know the circumstance (I > > thought per-bio-data is zero'd -- but it may be I'm mistaken). > The circumstance is just a normal qemu VM running the vanilla kernel for test > purpose. According to the implementation of bio_alloc_bioset(), the front pad > of > bio is not initialized and only bio itself is initialized. AFAIK if > CONFIG_INIT_ON_ALLOC_DEFAULT_ON is enabled, per-bio-data may be zeroed. > > > > I won't be marking this commit for stable@ but if others feel > > differently please let me know and I'll do so. (We can always propose > > it to stable@, after the fact, even if the commit header doesn't Cc > > stable@) > > > >>>> Fix it by initializing io->tasklet in crypt_io_init(). > >>> Really would rather avoid always calling tasklet_init(). But I can > >>> optimize it away with a later patch. > >> My first though was "io->tasklet.state = 0", but it may be fragile because > >> it > >> operated on the internal status of tasklet, so I switch to tasklet_init(). > > Yes, I looked into it and came up with the same hack.. and I too felt > > it was too fragile due to open-coding direct access to the tasklet's > > members. > > > > I have a patch I just staged that staged that uses jump_labels to > > optimize this code. If you might review/test/verify it works well for > > you that'd be appreciated: > > https://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-6.3&id=ae75a25bd83f7c541240449d2fff3a44433e506b > > > > It builds on your patch, which I added a comment to: > > https://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-6.3&id=d9fe0a98a2e0a1cf585e8a6555afb33be968bd13 > Thanks for the comments. It is fine to me. > > > > From: Mike Snitzer <snit...@kernel.org> > > Date: Mon, 6 Mar 2023 15:58:33 -0500 > > Subject: [PATCH] dm crypt: conditionally enable code needed for tasklet > > usecases > > > > Use jump_label to limit the need for branching, and tasklet_init(), > > unless either of the optional "no_read_workqueue" and/or > > "no_write_workqueue" features are used. > > > > Signed-off-by: Mike Snitzer <snit...@kernel.org> > > --- > > drivers/md/dm-crypt.c | 35 +++++++++++++++++++++++++++-------- > > 1 file changed, 27 insertions(+), 8 deletions(-) > > > > diff --git a/drivers/md/dm-crypt.c b/drivers/md/dm-crypt.c > > index 641457e72603..2d0309ca07f5 100644 > > --- a/drivers/md/dm-crypt.c > > +++ b/drivers/md/dm-crypt.c > > @@ -40,6 +40,7 @@ > > #include <keys/user-type.h> > > #include <keys/encrypted-type.h> > > #include <keys/trusted-type.h> > > +#include <linux/jump_label.h> > > > > #include <linux/device-mapper.h> > > > > @@ -85,6 +86,8 @@ struct dm_crypt_io { > > struct rb_node rb_node; > > } CRYPTO_MINALIGN_ATTR; > > > > +static DEFINE_STATIC_KEY_FALSE(use_tasklet_enabled); > > + > > struct dm_crypt_request { > > struct convert_context *ctx; > > struct scatterlist sg_in[4]; > > @@ -1730,12 +1733,15 @@ static void crypt_io_init(struct dm_crypt_io *io, > > struct crypt_config *cc, > > io->sector = sector; > > io->error = 0; > > io->ctx.r.req = NULL; > > - /* > > - * tasklet_init() here to ensure crypt_dec_pending()'s > > - * tasklet_trylock() doesn't incorrectly return false > > - * even when tasklet isn't in use. > > - */ > > - tasklet_init(&io->tasklet, kcryptd_crypt_tasklet, (unsigned > > long)&io->work); > > + if (static_branch_unlikely(&use_tasklet_enabled)) { > > + /* > > + * tasklet_init() here to ensure crypt_dec_pending()'s > > + * tasklet_trylock() doesn't incorrectly return false > > + * even when tasklet isn't in use. > > + */ > > + tasklet_init(&io->tasklet, kcryptd_crypt_tasklet, > > + (unsigned long)&io->work); > > + } > > io->integrity_metadata = NULL; > > io->integrity_metadata_from_pool = false; > > atomic_set(&io->io_pending, 0); > > @@ -1775,6 +1781,10 @@ static void crypt_dec_pending(struct dm_crypt_io *io) > > kfree(io->integrity_metadata); > > > > base_bio->bi_status = error; > > + if (!static_branch_unlikely(&use_tasklet_enabled)) { > > + bio_endio(base_bio); > > + return; > > + } > Because use_tasklet_enabled can be enabled concurrently, so I think it is > still > possible that crypt_dec_pending will try-lock an unitialized tasklet if > use_tasklet_enabled is enabled when invoking crypt_dec_pending().
Perhaps instead we can just pass an additional flag from tasklet_schedule to indicate to the function that we're running in a tasklet. I originally have chosen the tasklet_trylock/unlock hack to avoid passing an extra flag. But unitialized memory makes sense as well as the desire to avoid calling tasklet_init unconditionally. So an extra member in dm_crypt_io might be the most straightforward here. Ignat > > /* > > * If we are running this function from our tasklet, > > @@ -2232,8 +2242,9 @@ static void kcryptd_queue_crypt(struct dm_crypt_io > > *io) > > { > > struct crypt_config *cc = io->cc; > > > > - if ((bio_data_dir(io->base_bio) == READ && > > test_bit(DM_CRYPT_NO_READ_WORKQUEUE, &cc->flags)) || > > - (bio_data_dir(io->base_bio) == WRITE && > > test_bit(DM_CRYPT_NO_WRITE_WORKQUEUE, &cc->flags))) { > > + if (static_branch_unlikely(&use_tasklet_enabled) && > > + ((bio_data_dir(io->base_bio) == READ && > > test_bit(DM_CRYPT_NO_READ_WORKQUEUE, &cc->flags)) || > > + (bio_data_dir(io->base_bio) == WRITE && > > test_bit(DM_CRYPT_NO_WRITE_WORKQUEUE, &cc->flags)))) { > > /* > > * in_hardirq(): Crypto API's skcipher_walk_first() refuses > > to work in hard IRQ context. > > * irqs_disabled(): the kernel may run some IO completion > > from the idle thread, but > > @@ -2746,6 +2757,10 @@ static void crypt_dtr(struct dm_target *ti) > > crypt_calculate_pages_per_client(); > > spin_unlock(&dm_crypt_clients_lock); > > > > + if (test_bit(DM_CRYPT_NO_READ_WORKQUEUE, &cc->flags) || > > + test_bit(DM_CRYPT_NO_WRITE_WORKQUEUE, &cc->flags)) > > + static_branch_dec(&use_tasklet_enabled); > > + > > dm_audit_log_dtr(DM_MSG_PREFIX, ti, 1); > > } > > > > @@ -3375,6 +3390,10 @@ static int crypt_ctr(struct dm_target *ti, unsigned > > int argc, char **argv) > > ti->limit_swap_bios = true; > > ti->accounts_remapped_io = true; > > > > + if (test_bit(DM_CRYPT_NO_READ_WORKQUEUE, &cc->flags) || > > + test_bit(DM_CRYPT_NO_WRITE_WORKQUEUE, &cc->flags)) > > + static_branch_inc(&use_tasklet_enabled); > > + > > dm_audit_log_ctr(DM_MSG_PREFIX, ti, 1); > > return 0; > > > -- dm-devel mailing list dm-devel@redhat.com https://listman.redhat.com/mailman/listinfo/dm-devel