| Message ID | 20170301234319.29584-1-tahsin@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Tejun, > > Ah, indeed, but we can break out allocation of blkg and its > initialization, right? It's a bit more work but then we'd be able to > do something like. > > > retry: > new_blkg = alloc; > lock; > sanity checks; > blkg = blkg_lookup_and_create(..., new_blkg); > if (!blkg) { > unlock; > goto retry; > } I tried doing it based on the sample above but I wasn't happy with the result. The amount of code grew too big. I sent a simplified version that does blkg allocation within blkg_lookup_create(). I think this version is simpler, let me know what you think. On Wed, Mar 1, 2017 at 3:43 PM, Tahsin Erdogan <tahsin@google.com> wrote: > blkg_conf_prep() currently calls blkg_lookup_create() while holding > request queue spinlock. This means allocating memory for struct > blkcg_gq has to be made non-blocking. This causes occasional -ENOMEM > failures in call paths like below: > > pcpu_alloc+0x68f/0x710 > __alloc_percpu_gfp+0xd/0x10 > __percpu_counter_init+0x55/0xc0 > cfq_pd_alloc+0x3b2/0x4e0 > blkg_alloc+0x187/0x230 > blkg_create+0x489/0x670 > blkg_lookup_create+0x9a/0x230 > blkg_conf_prep+0x1fb/0x240 > __cfqg_set_weight_device.isra.105+0x5c/0x180 > cfq_set_weight_on_dfl+0x69/0xc0 > cgroup_file_write+0x39/0x1c0 > kernfs_fop_write+0x13f/0x1d0 > __vfs_write+0x23/0x120 > vfs_write+0xc2/0x1f0 > SyS_write+0x44/0xb0 > entry_SYSCALL_64_fastpath+0x18/0xad > > In the code path above, percpu allocator cannot call vmalloc() due to > queue spinlock. > > A failure in this call path gives grief to tools which are trying to > configure io weights. We see occasional failures happen shortly after > reboots even when system is not under any memory pressure. Machines > with a lot of cpus are more vulnerable to this condition. > > Add a flag to blkg_lookup_create() to indicate whether releasing locks > for memory allocation purposes is okay. > > Suggested-by: Tejun Heo <tj@kernel.org> > Signed-off-by: Tahsin Erdogan <tahsin@google.com> > --- > v2: > Moved blkg creation into blkg_lookup_create() to avoid duplicating > blkg_lookup_create() logic. > > block/blk-cgroup.c | 51 +++++++++++++++++++++++++++++++++++++++------- > include/linux/blk-cgroup.h | 4 ++-- > 2 files changed, 46 insertions(+), 9 deletions(-) > > diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c > index 295e98c2c8cc..afb16e998bf3 100644 > --- a/block/blk-cgroup.c > +++ b/block/blk-cgroup.c > @@ -258,18 +258,22 @@ static struct blkcg_gq *blkg_create(struct blkcg *blkcg, > * blkg_lookup_create - lookup blkg, try to create one if not there > * @blkcg: blkcg of interest > * @q: request_queue of interest > + * @wait_ok: whether blocking for memory allocations is okay > * > * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to > * create one. blkg creation is performed recursively from blkcg_root such > * that all non-root blkg's have access to the parent blkg. This function > * should be called under RCU read lock and @q->queue_lock. > * > + * When @wait_ok is true, rcu and queue locks may be dropped for allocating > + * memory. In this case, the locks will be reacquired on return. > + * > * Returns pointer to the looked up or created blkg on success, ERR_PTR() > * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not > * dead and bypassing, returns ERR_PTR(-EBUSY). > */ > struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, > - struct request_queue *q) > + struct request_queue *q, bool wait_ok) > { > struct blkcg_gq *blkg; > > @@ -300,7 +304,30 @@ struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, > parent = blkcg_parent(parent); > } > > - blkg = blkg_create(pos, q, NULL); > + if (wait_ok) { > + struct blkcg_gq *new_blkg; > + > + spin_unlock_irq(q->queue_lock); > + rcu_read_unlock(); > + > + new_blkg = blkg_alloc(pos, q, GFP_KERNEL); > + > + rcu_read_lock(); > + spin_lock_irq(q->queue_lock); > + > + if (unlikely(!new_blkg)) > + return ERR_PTR(-ENOMEM); > + > + if (unlikely(blk_queue_bypass(q))) { > + blkg_free(new_blkg); > + return ERR_PTR(blk_queue_dying(q) ? > + -ENODEV : -EBUSY); > + } > + > + blkg = blkg_create(pos, q, new_blkg); > + } else > + blkg = blkg_create(pos, q, NULL); > + > if (pos == blkcg || IS_ERR(blkg)) > return blkg; > } > @@ -789,6 +816,7 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, > { > struct gendisk *disk; > struct blkcg_gq *blkg; > + struct request_queue *q; > struct module *owner; > unsigned int major, minor; > int key_len, part, ret; > @@ -812,18 +840,27 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, > return -ENODEV; > } > > + q = disk->queue; > + > rcu_read_lock(); > - spin_lock_irq(disk->queue->queue_lock); > + spin_lock_irq(q->queue_lock); > > - if (blkcg_policy_enabled(disk->queue, pol)) > - blkg = blkg_lookup_create(blkcg, disk->queue); > - else > + if (blkcg_policy_enabled(q, pol)) { > + blkg = blkg_lookup_create(blkcg, q, true /* wait_ok */); > + > + /* > + * blkg_lookup_create() may have dropped and reacquired the > + * queue lock. Check policy enabled state again. > + */ > + if (!IS_ERR(blkg) && unlikely(!blkcg_policy_enabled(q, pol))) > + blkg = ERR_PTR(-EOPNOTSUPP); > + } else > blkg = ERR_PTR(-EOPNOTSUPP); > > if (IS_ERR(blkg)) { > ret = PTR_ERR(blkg); > rcu_read_unlock(); > - spin_unlock_irq(disk->queue->queue_lock); > + spin_unlock_irq(q->queue_lock); > owner = disk->fops->owner; > put_disk(disk); > module_put(owner); > diff --git a/include/linux/blk-cgroup.h b/include/linux/blk-cgroup.h > index 01b62e7bac74..78067dd59c91 100644 > --- a/include/linux/blk-cgroup.h > +++ b/include/linux/blk-cgroup.h > @@ -172,7 +172,7 @@ extern struct cgroup_subsys_state * const blkcg_root_css; > struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg, > struct request_queue *q, bool update_hint); > struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, > - struct request_queue *q); > + struct request_queue *q, bool wait_ok); > int blkcg_init_queue(struct request_queue *q); > void blkcg_drain_queue(struct request_queue *q); > void blkcg_exit_queue(struct request_queue *q); > @@ -694,7 +694,7 @@ static inline bool blkcg_bio_issue_check(struct request_queue *q, > blkg = blkg_lookup(blkcg, q); > if (unlikely(!blkg)) { > spin_lock_irq(q->queue_lock); > - blkg = blkg_lookup_create(blkcg, q); > + blkg = blkg_lookup_create(blkcg, q, false /* wait_ok */); > if (IS_ERR(blkg)) > blkg = NULL; > spin_unlock_irq(q->queue_lock); > -- > 2.12.0.rc1.440.g5b76565f74-goog >
Hello, Tahsin. On Wed, Mar 01, 2017 at 03:43:19PM -0800, Tahsin Erdogan wrote: > @@ -258,18 +258,22 @@ static struct blkcg_gq *blkg_create(struct blkcg *blkcg, > struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, > - struct request_queue *q) > + struct request_queue *q, bool wait_ok) I'm okay with this direction but it probably would be better if the parameter is gfp_mask and we branch on __GFP_WAIT in the function. > { > struct blkcg_gq *blkg; > > @@ -300,7 +304,30 @@ struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, > parent = blkcg_parent(parent); > } > > - blkg = blkg_create(pos, q, NULL); > + if (wait_ok) { > + struct blkcg_gq *new_blkg; > + > + spin_unlock_irq(q->queue_lock); > + rcu_read_unlock(); > + > + new_blkg = blkg_alloc(pos, q, GFP_KERNEL); > + > + rcu_read_lock(); > + spin_lock_irq(q->queue_lock); > + > + if (unlikely(!new_blkg)) > + return ERR_PTR(-ENOMEM); > + > + if (unlikely(blk_queue_bypass(q))) { > + blkg_free(new_blkg); > + return ERR_PTR(blk_queue_dying(q) ? > + -ENODEV : -EBUSY); > + } > + > + blkg = blkg_create(pos, q, new_blkg); > + } else > + blkg = blkg_create(pos, q, NULL); So, while I'm okay with the approach, now we're creating a hybrid approach where we have both pre-allocation and allocation mode altering mechanisms. If we're going to take this route, I think the right thing to do is passing down @gfp_mask all the way down to blkg_create(). > @@ -789,6 +816,7 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, > { > struct gendisk *disk; > struct blkcg_gq *blkg; > + struct request_queue *q; > struct module *owner; > unsigned int major, minor; > int key_len, part, ret; > @@ -812,18 +840,27 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, > return -ENODEV; > } > > + q = disk->queue; > + > rcu_read_lock(); > - spin_lock_irq(disk->queue->queue_lock); > + spin_lock_irq(q->queue_lock); > > - if (blkcg_policy_enabled(disk->queue, pol)) > - blkg = blkg_lookup_create(blkcg, disk->queue); > - else > + if (blkcg_policy_enabled(q, pol)) { > + blkg = blkg_lookup_create(blkcg, q, true /* wait_ok */); > + > + /* > + * blkg_lookup_create() may have dropped and reacquired the > + * queue lock. Check policy enabled state again. > + */ > + if (!IS_ERR(blkg) && unlikely(!blkcg_policy_enabled(q, pol))) > + blkg = ERR_PTR(-EOPNOTSUPP); And let blkg_create() verify these conditions after releasing and regrabbing the lock. This also means that the init path can simply pass in GFP_KERNEL. Thanks.
On Thu, Mar 2, 2017 at 11:32 AM, Tejun Heo <tj@kernel.org> wrote: > Hello, Tahsin. > > On Wed, Mar 01, 2017 at 03:43:19PM -0800, Tahsin Erdogan wrote: >> @@ -258,18 +258,22 @@ static struct blkcg_gq *blkg_create(struct blkcg *blkcg, >> struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, >> - struct request_queue *q) >> + struct request_queue *q, bool wait_ok) > > I'm okay with this direction but it probably would be better if the > parameter is gfp_mask and we branch on __GFP_WAIT in the function. Agreed, gfp mask is better as a parameter. > >> { >> struct blkcg_gq *blkg; >> >> @@ -300,7 +304,30 @@ struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, >> parent = blkcg_parent(parent); >> } >> >> - blkg = blkg_create(pos, q, NULL); >> + if (wait_ok) { >> + struct blkcg_gq *new_blkg; >> + >> + spin_unlock_irq(q->queue_lock); >> + rcu_read_unlock(); >> + >> + new_blkg = blkg_alloc(pos, q, GFP_KERNEL); >> + >> + rcu_read_lock(); >> + spin_lock_irq(q->queue_lock); >> + >> + if (unlikely(!new_blkg)) >> + return ERR_PTR(-ENOMEM); >> + >> + if (unlikely(blk_queue_bypass(q))) { >> + blkg_free(new_blkg); >> + return ERR_PTR(blk_queue_dying(q) ? >> + -ENODEV : -EBUSY); >> + } >> + >> + blkg = blkg_create(pos, q, new_blkg); >> + } else >> + blkg = blkg_create(pos, q, NULL); > > So, while I'm okay with the approach, now we're creating a hybrid > approach where we have both pre-allocation and allocation mode > altering mechanisms. If we're going to take this route, I think the > right thing to do is passing down @gfp_mask all the way down to > blkg_create(). > >> @@ -789,6 +816,7 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, >> { >> struct gendisk *disk; >> struct blkcg_gq *blkg; >> + struct request_queue *q; >> struct module *owner; >> unsigned int major, minor; >> int key_len, part, ret; >> @@ -812,18 +840,27 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, >> return -ENODEV; >> } >> >> + q = disk->queue; >> + >> rcu_read_lock(); >> - spin_lock_irq(disk->queue->queue_lock); >> + spin_lock_irq(q->queue_lock); >> >> - if (blkcg_policy_enabled(disk->queue, pol)) >> - blkg = blkg_lookup_create(blkcg, disk->queue); >> - else >> + if (blkcg_policy_enabled(q, pol)) { >> + blkg = blkg_lookup_create(blkcg, q, true /* wait_ok */); >> + >> + /* >> + * blkg_lookup_create() may have dropped and reacquired the >> + * queue lock. Check policy enabled state again. >> + */ >> + if (!IS_ERR(blkg) && unlikely(!blkcg_policy_enabled(q, pol))) >> + blkg = ERR_PTR(-EOPNOTSUPP); > > And let blkg_create() verify these conditions after releasing and > regrabbing the lock. > > This also means that the init path can simply pass in GFP_KERNEL. I tried that approach, but I encountered two issues that complicate things: 1) Pushing down blk_queue_bypass(q) check in blkg_create() doesn't quite work because when blkcg_init_queue() calls blkg_create(), the queue is still in bypassing mode. 2) Pushing down blkcg_policy_enabled() doesn't work well either, because blkcg_init_queue() doesn't have a policy to pass down. We could let it pass a NULL parameter but that would make blkg_create more ugly.
Hello, Tahsin. On Thu, Mar 02, 2017 at 02:33:11PM -0800, Tahsin Erdogan wrote: > > And let blkg_create() verify these conditions after releasing and > > regrabbing the lock. > > > > This also means that the init path can simply pass in GFP_KERNEL. > > I tried that approach, but I encountered two issues that complicate things: > > 1) Pushing down blk_queue_bypass(q) check in blkg_create() doesn't > quite work because when blkcg_init_queue() calls blkg_create(), the > queue is still in bypassing mode. > > 2) Pushing down blkcg_policy_enabled() doesn't work well either, > because blkcg_init_queue() doesn't have a policy to pass down. We > could let it pass a NULL parameter but that would make blkg_create > more ugly. I see. It kinda really bothers me that we'll have two different modes for non-atomic allocations. Can't we bind both to the policy parameter? Skip the checks if policy is NULL? Thanks.
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c index 295e98c2c8cc..afb16e998bf3 100644 --- a/block/blk-cgroup.c +++ b/block/blk-cgroup.c @@ -258,18 +258,22 @@ static struct blkcg_gq *blkg_create(struct blkcg *blkcg, * blkg_lookup_create - lookup blkg, try to create one if not there * @blkcg: blkcg of interest * @q: request_queue of interest + * @wait_ok: whether blocking for memory allocations is okay * * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to * create one. blkg creation is performed recursively from blkcg_root such * that all non-root blkg's have access to the parent blkg. This function * should be called under RCU read lock and @q->queue_lock. * + * When @wait_ok is true, rcu and queue locks may be dropped for allocating + * memory. In this case, the locks will be reacquired on return. + * * Returns pointer to the looked up or created blkg on success, ERR_PTR() * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not * dead and bypassing, returns ERR_PTR(-EBUSY). */ struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, - struct request_queue *q) + struct request_queue *q, bool wait_ok) { struct blkcg_gq *blkg; @@ -300,7 +304,30 @@ struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, parent = blkcg_parent(parent); } - blkg = blkg_create(pos, q, NULL); + if (wait_ok) { + struct blkcg_gq *new_blkg; + + spin_unlock_irq(q->queue_lock); + rcu_read_unlock(); + + new_blkg = blkg_alloc(pos, q, GFP_KERNEL); + + rcu_read_lock(); + spin_lock_irq(q->queue_lock); + + if (unlikely(!new_blkg)) + return ERR_PTR(-ENOMEM); + + if (unlikely(blk_queue_bypass(q))) { + blkg_free(new_blkg); + return ERR_PTR(blk_queue_dying(q) ? + -ENODEV : -EBUSY); + } + + blkg = blkg_create(pos, q, new_blkg); + } else + blkg = blkg_create(pos, q, NULL); + if (pos == blkcg || IS_ERR(blkg)) return blkg; } @@ -789,6 +816,7 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, { struct gendisk *disk; struct blkcg_gq *blkg; + struct request_queue *q; struct module *owner; unsigned int major, minor; int key_len, part, ret; @@ -812,18 +840,27 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, return -ENODEV; } + q = disk->queue; + rcu_read_lock(); - spin_lock_irq(disk->queue->queue_lock); + spin_lock_irq(q->queue_lock); - if (blkcg_policy_enabled(disk->queue, pol)) - blkg = blkg_lookup_create(blkcg, disk->queue); - else + if (blkcg_policy_enabled(q, pol)) { + blkg = blkg_lookup_create(blkcg, q, true /* wait_ok */); + + /* + * blkg_lookup_create() may have dropped and reacquired the + * queue lock. Check policy enabled state again. + */ + if (!IS_ERR(blkg) && unlikely(!blkcg_policy_enabled(q, pol))) + blkg = ERR_PTR(-EOPNOTSUPP); + } else blkg = ERR_PTR(-EOPNOTSUPP); if (IS_ERR(blkg)) { ret = PTR_ERR(blkg); rcu_read_unlock(); - spin_unlock_irq(disk->queue->queue_lock); + spin_unlock_irq(q->queue_lock); owner = disk->fops->owner; put_disk(disk); module_put(owner); diff --git a/include/linux/blk-cgroup.h b/include/linux/blk-cgroup.h index 01b62e7bac74..78067dd59c91 100644 --- a/include/linux/blk-cgroup.h +++ b/include/linux/blk-cgroup.h @@ -172,7 +172,7 @@ extern struct cgroup_subsys_state * const blkcg_root_css; struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg, struct request_queue *q, bool update_hint); struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, - struct request_queue *q); + struct request_queue *q, bool wait_ok); int blkcg_init_queue(struct request_queue *q); void blkcg_drain_queue(struct request_queue *q); void blkcg_exit_queue(struct request_queue *q); @@ -694,7 +694,7 @@ static inline bool blkcg_bio_issue_check(struct request_queue *q, blkg = blkg_lookup(blkcg, q); if (unlikely(!blkg)) { spin_lock_irq(q->queue_lock); - blkg = blkg_lookup_create(blkcg, q); + blkg = blkg_lookup_create(blkcg, q, false /* wait_ok */); if (IS_ERR(blkg)) blkg = NULL; spin_unlock_irq(q->queue_lock);
blkg_conf_prep() currently calls blkg_lookup_create() while holding request queue spinlock. This means allocating memory for struct blkcg_gq has to be made non-blocking. This causes occasional -ENOMEM failures in call paths like below: pcpu_alloc+0x68f/0x710 __alloc_percpu_gfp+0xd/0x10 __percpu_counter_init+0x55/0xc0 cfq_pd_alloc+0x3b2/0x4e0 blkg_alloc+0x187/0x230 blkg_create+0x489/0x670 blkg_lookup_create+0x9a/0x230 blkg_conf_prep+0x1fb/0x240 __cfqg_set_weight_device.isra.105+0x5c/0x180 cfq_set_weight_on_dfl+0x69/0xc0 cgroup_file_write+0x39/0x1c0 kernfs_fop_write+0x13f/0x1d0 __vfs_write+0x23/0x120 vfs_write+0xc2/0x1f0 SyS_write+0x44/0xb0 entry_SYSCALL_64_fastpath+0x18/0xad In the code path above, percpu allocator cannot call vmalloc() due to queue spinlock. A failure in this call path gives grief to tools which are trying to configure io weights. We see occasional failures happen shortly after reboots even when system is not under any memory pressure. Machines with a lot of cpus are more vulnerable to this condition. Add a flag to blkg_lookup_create() to indicate whether releasing locks for memory allocation purposes is okay. Suggested-by: Tejun Heo <tj@kernel.org> Signed-off-by: Tahsin Erdogan <tahsin@google.com> --- v2: Moved blkg creation into blkg_lookup_create() to avoid duplicating blkg_lookup_create() logic. block/blk-cgroup.c | 51 +++++++++++++++++++++++++++++++++++++++------- include/linux/blk-cgroup.h | 4 ++-- 2 files changed, 46 insertions(+), 9 deletions(-)