| Message ID | 20190717111147.t776zlyhdqyl5dhc@gondor.apana.org.au (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Herbert Xu |
| Headers | show |
| Series | padata: Replace delayed timer with immediate workqueue in padata_reorder | expand |
On Wed, Jul 17, 2019 at 07:11:47PM +0800, Herbert Xu wrote: > On Tue, Jul 16, 2019 at 12:32:53PM -0400, Daniel Jordan wrote: > > Testing padata with the tcrypt module on a 5.2 kernel... > > Thanks for the patch! > > And here is an incremental patch to get rid of the timer that > appears to be an attempt at fixing a problem related to this. Nice, +1 for getting rid of the timer. > diff --git a/kernel/padata.c b/kernel/padata.c > index 15a8ad63f4ff..b5dfc21e976f 100644 > --- a/kernel/padata.c > +++ b/kernel/padata.c > @@ -165,23 +165,12 @@ EXPORT_SYMBOL(padata_do_parallel); > */ > static struct padata_priv *padata_get_next(struct parallel_data *pd) > { > - int cpu, num_cpus; > - unsigned int next_nr, next_index; > struct padata_parallel_queue *next_queue; > struct padata_priv *padata; > struct padata_list *reorder; > + int cpu = pd->cpu; > > - num_cpus = cpumask_weight(pd->cpumask.pcpu); > - > - /* > - * Calculate the percpu reorder queue and the sequence > - * number of the next object. > - */ > - next_nr = pd->processed; > - next_index = next_nr % num_cpus; > - cpu = padata_index_to_cpu(pd, next_index); > next_queue = per_cpu_ptr(pd->pqueue, cpu); > - > reorder = &next_queue->reorder; > > spin_lock(&reorder->lock); > @@ -192,7 +181,8 @@ static struct padata_priv *padata_get_next(struct parallel_data *pd) > list_del_init(&padata->list); > atomic_dec(&pd->reorder_objects); > > - pd->processed++; > + pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, 0, > + false); We'll crash when cpumask_next_wrap returns nr_cpumask_bits and later try to get the corresponding per-cpu queue. This handles that as well as the case where there's only 1 CPU in the parallel mask: diff --git a/kernel/padata.c b/kernel/padata.c index b5dfc21e976f..ab352839df04 100644 --- a/kernel/padata.c +++ b/kernel/padata.c @@ -181,8 +181,10 @@ static struct padata_priv *padata_get_next(struct parallel_data *pd) list_del_init(&padata->list); atomic_dec(&pd->reorder_objects); - pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, 0, - false); + if (cpumask_weight(pd->cpumask.pcpu) > 1) { + pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, cpu, + false); + } spin_unlock(&reorder->lock); goto out; Haven't finished looking at the patch, but have to run somewhere for now, will pick it up later today.
On Wed, Jul 17, 2019 at 07:11:47PM +0800, Herbert Xu wrote: > Note that we don't bother removing the work queue in > padata_flush_queues because the whole premise is broken. You > cannot flush async crypto requests so it makes no sense to even > try. A subsequent patch will fix it by replacing it with a ref > counting scheme. Interested to see what happens with the ref counting. You mean you don't bother removing the serial workqueue flushing, right, not the parallel? > @@ -122,10 +117,10 @@ struct padata_cpumask { > * @reorder_objects: Number of objects waiting in the reorder queues. > * @refcnt: Number of objects holding a reference on this parallel_data. > * @max_seq_nr: Maximal used sequence number. > + * @cpu: Next CPU to be processed. Maybe something more specific... @cpu: CPU of the next reorder queue to process. > static struct padata_priv *padata_get_next(struct parallel_data *pd) > { > - int cpu, num_cpus; > - unsigned int next_nr, next_index; > struct padata_parallel_queue *next_queue; > struct padata_priv *padata; > struct padata_list *reorder; > + int cpu = pd->cpu; > > - num_cpus = cpumask_weight(pd->cpumask.pcpu); > - > - /* > - * Calculate the percpu reorder queue and the sequence > - * number of the next object. > - */ > - next_nr = pd->processed; > - next_index = next_nr % num_cpus; > - cpu = padata_index_to_cpu(pd, next_index); After this patch padata_index_to_cpu has only one caller, so it doesn't need to be a function anymore. > @@ -246,7 +237,6 @@ static void padata_reorder(struct parallel_data *pd) > * so exit immediately. > */ > if (PTR_ERR(padata) == -ENODATA) { > - del_timer(&pd->timer); > spin_unlock_bh(&pd->lock); > return; > } > @@ -265,70 +255,29 @@ static void padata_reorder(struct parallel_data *pd) > > /* > * The next object that needs serialization might have arrived to > - * the reorder queues in the meantime, we will be called again > - * from the timer function if no one else cares for it. > + * the reorder queues in the meantime. > * > - * Ensure reorder_objects is read after pd->lock is dropped so we see > - * an increment from another task in padata_do_serial. Pairs with > + * Ensure reorder queue is read after pd->lock is dropped so we see > + * new objects from another task in padata_do_serial. Pairs with > * smp_mb__after_atomic in padata_do_serial. > */ > smp_mb(); > - if (atomic_read(&pd->reorder_objects) > - && !(pinst->flags & PADATA_RESET)) > - mod_timer(&pd->timer, jiffies + HZ); > - else > - del_timer(&pd->timer); > > - return; > + next_queue = per_cpu_ptr(pd->pqueue, pd->cpu); > + if (!list_empty(&next_queue->reorder.list)) > + queue_work(pinst->wq, &pd->reorder_work); It's possible that the work gets queued when it doesn't need to be when another task adds a job to the reorder queue but hasn't grabbed pd->lock yet, but I can't think of a way around it...and it does no harm anyway. > @@ -376,9 +325,8 @@ void padata_do_serial(struct padata_priv *padata) > > cpu = get_cpu(); > > - /* We need to run on the same CPU padata_do_parallel(.., padata, ..) > - * was called on -- or, at least, enqueue the padata object into the > - * correct per-cpu queue. > + /* We need to enqueue the padata object into the correct > + * per-cpu queue. > */ > if (cpu != padata->cpu) { > reorder_via_wq = 1; reorder_via_wq and I think get_cpu/put_cpu can go away now that we're always using padata->cpu to get the parallel queue and then running padata_reorder in the current task. Maybe Steffen can check my reasoning on the get_cpu thing. It looks like that was added in the original padata commit to keep 'cpu' stable for getting the parallel queue but is no longer needed because we just use padata->cpu. > @@ -388,12 +336,12 @@ void padata_do_serial(struct padata_priv *padata) > pqueue = per_cpu_ptr(pd->pqueue, cpu); > > spin_lock(&pqueue->reorder.lock); > - atomic_inc(&pd->reorder_objects); > list_add_tail(&padata->list, &pqueue->reorder.list); > + atomic_inc(&pd->reorder_objects); Why switch the lines? Seems ok to not do this. > @@ -538,8 +479,6 @@ static void padata_flush_queues(struct parallel_data *pd) > flush_work(&pqueue->work); > } > > - del_timer_sync(&pd->timer); > - > if (atomic_read(&pd->reorder_objects)) > padata_reorder(pd); I think we can do away with reorder_objects entirely by checking pd->cpu's reorder queue here. It's racy to read pd->cpu without pd->lock, but it doesn't matter. If there are objects left to process and no other task is in padata_reorder, this path will notice that, and if there's another task in padata_reorder changing pd->cpu from under us, that task will finish the reordering so this path doesn't have to.
On Wed, Jul 17, 2019 at 07:21:36PM -0400, Daniel Jordan wrote: > > > @@ -388,12 +336,12 @@ void padata_do_serial(struct padata_priv *padata) > > pqueue = per_cpu_ptr(pd->pqueue, cpu); > > > > spin_lock(&pqueue->reorder.lock); > > - atomic_inc(&pd->reorder_objects); > > list_add_tail(&padata->list, &pqueue->reorder.list); > > + atomic_inc(&pd->reorder_objects); > > Why switch the lines? Seems ok to not do this. This is crucial because otherwise the memory barrier won't apply to the list insertion. With this patch, we are now using the list insertion as the indicator, rather than reorder_objects. > > @@ -538,8 +479,6 @@ static void padata_flush_queues(struct parallel_data *pd) > > flush_work(&pqueue->work); > > } > > > > - del_timer_sync(&pd->timer); > > - > > > if (atomic_read(&pd->reorder_objects)) > > padata_reorder(pd); > > I think we can do away with reorder_objects entirely by checking pd->cpu's > reorder queue here. As I said this will probably disappear altogether since we can't guarantee that padata_reorder will actually do anything if the jobs are stuck in async crypto processing. Thanks,
On Wed, Jul 17, 2019 at 02:32:27PM -0400, Daniel Jordan wrote: > > We'll crash when cpumask_next_wrap returns nr_cpumask_bits and later try to get > the corresponding per-cpu queue. The whole point of cpumask_next_wrap is to wrap around to the beginning when it hits nr_cpumask_bits. So it cannot return nr_cpumask_bits. Cheers,
On Thu, Jul 18, 2019 at 11:30:08AM +0800, Herbert Xu wrote: > On Wed, Jul 17, 2019 at 07:21:36PM -0400, Daniel Jordan wrote: > > > > > @@ -388,12 +336,12 @@ void padata_do_serial(struct padata_priv *padata) > > > pqueue = per_cpu_ptr(pd->pqueue, cpu); > > > > > > spin_lock(&pqueue->reorder.lock); > > > - atomic_inc(&pd->reorder_objects); > > > list_add_tail(&padata->list, &pqueue->reorder.list); > > > + atomic_inc(&pd->reorder_objects); > > > > Why switch the lines? Seems ok to not do this. > > This is crucial because otherwise the memory barrier won't apply > to the list insertion. With this patch, we are now using the list > insertion as the indicator, rather than reorder_objects. Which memory barrier do you mean? I think you're referring to the one that atomic_inc might provide? If so, the memory model maintainers can correct me here, but my understanding is that RMW atomic ops that don't return values are unordered, so switching the lines has no effect. Besides, the smp_mb__after_atomic is what orders the list insertion with the trylock of pd->lock. > > > @@ -538,8 +479,6 @@ static void padata_flush_queues(struct parallel_data *pd) > > > flush_work(&pqueue->work); > > > } > > > > > > - del_timer_sync(&pd->timer); > > > - > > > > > if (atomic_read(&pd->reorder_objects)) > > > padata_reorder(pd); > > > > I think we can do away with reorder_objects entirely by checking pd->cpu's > > reorder queue here. > > As I said this will probably disappear altogether since we can't > guarantee that padata_reorder will actually do anything if the > jobs are stuck in async crypto processing. Ok, makes sense.
On Thu, Jul 18, 2019 at 11:31:31AM +0800, Herbert Xu wrote: > On Wed, Jul 17, 2019 at 02:32:27PM -0400, Daniel Jordan wrote: > > > > We'll crash when cpumask_next_wrap returns nr_cpumask_bits and later try to get > > the corresponding per-cpu queue. > > The whole point of cpumask_next_wrap is to wrap around to the > beginning when it hits nr_cpumask_bits. So it cannot return > nr_cpumask_bits. That's what I expected when I first saw it too, but nr_cpumask_bits is returned to signal the end of the iteration. The patch always passes 0 for the 'start' argument, so when cpumask_next_wrap is called with the last cpu in the mask, the end-of-iteration case is triggered. To reassure you and myself :) I ran it and got the expected crash. Passing pd->cpu for the start argument instead avoids that problem, but the one-cpu-in-mask case still needs handling because cpumask_next_wrap always signals end of iteration for that, hence the cpumask_weight check.
On Thu, Jul 18, 2019 at 10:25:15AM -0400, Daniel Jordan wrote: > > Which memory barrier do you mean? I think you're referring to the one that > atomic_inc might provide? If so, the memory model maintainers can correct me > here, but my understanding is that RMW atomic ops that don't return values are > unordered, so switching the lines has no effect. > > Besides, the smp_mb__after_atomic is what orders the list insertion with the > trylock of pd->lock. The primitive smp_mb__after_atomic only provides a barrier when used in conjunction with atomic_inc (and similar atomic ops). The actual barrier may either be in smp_mb__after_atomic or the atomic op itself (which is the case on x86). Since we need the barrier to occur after the list insertion we must move both of these after the list_add_tail. Cheers,
On Thu, Jul 18, 2019 at 10:27:30AM -0400, Daniel Jordan wrote: > > That's what I expected when I first saw it too, but nr_cpumask_bits is returned > to signal the end of the iteration. The patch always passes 0 for the 'start' > argument, so when cpumask_next_wrap is called with the last cpu in the mask, > the end-of-iteration case is triggered. To reassure you and myself :) I ran it > and got the expected crash. > > Passing pd->cpu for the start argument instead avoids that problem, but the > one-cpu-in-mask case still needs handling because cpumask_next_wrap always > signals end of iteration for that, hence the cpumask_weight check. My bad. I should have set start to -1 to make it do the right thing. Thanks,
On Thu, Jul 18, 2019 at 10:49:50PM +0800, Herbert Xu wrote: > On Thu, Jul 18, 2019 at 10:25:15AM -0400, Daniel Jordan wrote: > > > > Which memory barrier do you mean? I think you're referring to the one that > > atomic_inc might provide? If so, the memory model maintainers can correct me > > here, but my understanding is that RMW atomic ops that don't return values are > > unordered, so switching the lines has no effect. > > > > Besides, the smp_mb__after_atomic is what orders the list insertion with the > > trylock of pd->lock. > > The primitive smp_mb__after_atomic only provides a barrier when > used in conjunction with atomic_inc (and similar atomic ops). > > The actual barrier may either be in smp_mb__after_atomic or the > atomic op itself (which is the case on x86). Since we need the > barrier to occur after the list insertion we must move both of > these after the list_add_tail. Yes, my mistake! Thanks for clarifying that.
On Thu, Jul 18, 2019 at 10:56:34PM +0800, Herbert Xu wrote: > On Thu, Jul 18, 2019 at 10:27:30AM -0400, Daniel Jordan wrote: > > > > That's what I expected when I first saw it too, but nr_cpumask_bits is returned > > to signal the end of the iteration. The patch always passes 0 for the 'start' > > argument, so when cpumask_next_wrap is called with the last cpu in the mask, > > the end-of-iteration case is triggered. To reassure you and myself :) I ran it > > and got the expected crash. > > > > Passing pd->cpu for the start argument instead avoids that problem, but the > > one-cpu-in-mask case still needs handling because cpumask_next_wrap always > > signals end of iteration for that, hence the cpumask_weight check. > > My bad. I should have set start to -1 to make it do the right thing. Oh, you're right, that's nicer, just noticed other callers do it that way as well.
diff --git a/include/linux/padata.h b/include/linux/padata.h index 5d13d25da2c8..d803397a28f7 100644 --- a/include/linux/padata.h +++ b/include/linux/padata.h @@ -24,7 +24,6 @@ #include <linux/workqueue.h> #include <linux/spinlock.h> #include <linux/list.h> -#include <linux/timer.h> #include <linux/notifier.h> #include <linux/kobject.h> @@ -85,18 +84,14 @@ struct padata_serial_queue { * @serial: List to wait for serialization after reordering. * @pwork: work struct for parallelization. * @swork: work struct for serialization. - * @pd: Backpointer to the internal control structure. * @work: work struct for parallelization. - * @reorder_work: work struct for reordering. * @num_obj: Number of objects that are processed by this cpu. * @cpu_index: Index of the cpu. */ struct padata_parallel_queue { struct padata_list parallel; struct padata_list reorder; - struct parallel_data *pd; struct work_struct work; - struct work_struct reorder_work; atomic_t num_obj; int cpu_index; }; @@ -122,10 +117,10 @@ struct padata_cpumask { * @reorder_objects: Number of objects waiting in the reorder queues. * @refcnt: Number of objects holding a reference on this parallel_data. * @max_seq_nr: Maximal used sequence number. + * @cpu: Next CPU to be processed. * @cpumask: The cpumasks in use for parallel and serial workers. + * @reorder_work: work struct for reordering. * @lock: Reorder lock. - * @processed: Number of already processed objects. - * @timer: Reorder timer. */ struct parallel_data { struct padata_instance *pinst; @@ -134,10 +129,10 @@ struct parallel_data { atomic_t reorder_objects; atomic_t refcnt; atomic_t seq_nr; + int cpu; struct padata_cpumask cpumask; + struct work_struct reorder_work; spinlock_t lock ____cacheline_aligned; - unsigned int processed; - struct timer_list timer; }; /** diff --git a/kernel/padata.c b/kernel/padata.c index 15a8ad63f4ff..b5dfc21e976f 100644 --- a/kernel/padata.c +++ b/kernel/padata.c @@ -165,23 +165,12 @@ EXPORT_SYMBOL(padata_do_parallel); */ static struct padata_priv *padata_get_next(struct parallel_data *pd) { - int cpu, num_cpus; - unsigned int next_nr, next_index; struct padata_parallel_queue *next_queue; struct padata_priv *padata; struct padata_list *reorder; + int cpu = pd->cpu; - num_cpus = cpumask_weight(pd->cpumask.pcpu); - - /* - * Calculate the percpu reorder queue and the sequence - * number of the next object. - */ - next_nr = pd->processed; - next_index = next_nr % num_cpus; - cpu = padata_index_to_cpu(pd, next_index); next_queue = per_cpu_ptr(pd->pqueue, cpu); - reorder = &next_queue->reorder; spin_lock(&reorder->lock); @@ -192,7 +181,8 @@ static struct padata_priv *padata_get_next(struct parallel_data *pd) list_del_init(&padata->list); atomic_dec(&pd->reorder_objects); - pd->processed++; + pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, 0, + false); spin_unlock(&reorder->lock); goto out; @@ -215,6 +205,7 @@ static void padata_reorder(struct parallel_data *pd) struct padata_priv *padata; struct padata_serial_queue *squeue; struct padata_instance *pinst = pd->pinst; + struct padata_parallel_queue *next_queue; /* * We need to ensure that only one cpu can work on dequeueing of @@ -246,7 +237,6 @@ static void padata_reorder(struct parallel_data *pd) * so exit immediately. */ if (PTR_ERR(padata) == -ENODATA) { - del_timer(&pd->timer); spin_unlock_bh(&pd->lock); return; } @@ -265,70 +255,29 @@ static void padata_reorder(struct parallel_data *pd) /* * The next object that needs serialization might have arrived to - * the reorder queues in the meantime, we will be called again - * from the timer function if no one else cares for it. + * the reorder queues in the meantime. * - * Ensure reorder_objects is read after pd->lock is dropped so we see - * an increment from another task in padata_do_serial. Pairs with + * Ensure reorder queue is read after pd->lock is dropped so we see + * new objects from another task in padata_do_serial. Pairs with * smp_mb__after_atomic in padata_do_serial. */ smp_mb(); - if (atomic_read(&pd->reorder_objects) - && !(pinst->flags & PADATA_RESET)) - mod_timer(&pd->timer, jiffies + HZ); - else - del_timer(&pd->timer); - return; + next_queue = per_cpu_ptr(pd->pqueue, pd->cpu); + if (!list_empty(&next_queue->reorder.list)) + queue_work(pinst->wq, &pd->reorder_work); } static void invoke_padata_reorder(struct work_struct *work) { - struct padata_parallel_queue *pqueue; struct parallel_data *pd; local_bh_disable(); - pqueue = container_of(work, struct padata_parallel_queue, reorder_work); - pd = pqueue->pd; + pd = container_of(work, struct parallel_data, reorder_work); padata_reorder(pd); local_bh_enable(); } -static void padata_reorder_timer(struct timer_list *t) -{ - struct parallel_data *pd = from_timer(pd, t, timer); - unsigned int weight; - int target_cpu, cpu; - - cpu = get_cpu(); - - /* We don't lock pd here to not interfere with parallel processing - * padata_reorder() calls on other CPUs. We just need any CPU out of - * the cpumask.pcpu set. It would be nice if it's the right one but - * it doesn't matter if we're off to the next one by using an outdated - * pd->processed value. - */ - weight = cpumask_weight(pd->cpumask.pcpu); - target_cpu = padata_index_to_cpu(pd, pd->processed % weight); - - /* ensure to call the reorder callback on the correct CPU */ - if (cpu != target_cpu) { - struct padata_parallel_queue *pqueue; - struct padata_instance *pinst; - - /* The timer function is serialized wrt itself -- no locking - * needed. - */ - pinst = pd->pinst; - pqueue = per_cpu_ptr(pd->pqueue, target_cpu); - queue_work_on(target_cpu, pinst->wq, &pqueue->reorder_work); - } else { - padata_reorder(pd); - } - - put_cpu(); -} - static void padata_serial_worker(struct work_struct *serial_work) { struct padata_serial_queue *squeue; @@ -376,9 +325,8 @@ void padata_do_serial(struct padata_priv *padata) cpu = get_cpu(); - /* We need to run on the same CPU padata_do_parallel(.., padata, ..) - * was called on -- or, at least, enqueue the padata object into the - * correct per-cpu queue. + /* We need to enqueue the padata object into the correct + * per-cpu queue. */ if (cpu != padata->cpu) { reorder_via_wq = 1; @@ -388,12 +336,12 @@ void padata_do_serial(struct padata_priv *padata) pqueue = per_cpu_ptr(pd->pqueue, cpu); spin_lock(&pqueue->reorder.lock); - atomic_inc(&pd->reorder_objects); list_add_tail(&padata->list, &pqueue->reorder.list); + atomic_inc(&pd->reorder_objects); spin_unlock(&pqueue->reorder.lock); /* - * Ensure the atomic_inc of reorder_objects above is ordered correctly + * Ensure the addition to the reorder list is ordered correctly * with the trylock of pd->lock in padata_reorder. Pairs with smp_mb * in padata_reorder. */ @@ -401,13 +349,7 @@ void padata_do_serial(struct padata_priv *padata) put_cpu(); - /* If we're running on the wrong CPU, call padata_reorder() via a - * kernel worker. - */ - if (reorder_via_wq) - queue_work_on(cpu, pd->pinst->wq, &pqueue->reorder_work); - else - padata_reorder(pd); + padata_reorder(pd); } EXPORT_SYMBOL(padata_do_serial); @@ -463,14 +405,12 @@ static void padata_init_pqueues(struct parallel_data *pd) continue; } - pqueue->pd = pd; pqueue->cpu_index = cpu_index; cpu_index++; __padata_list_init(&pqueue->reorder); __padata_list_init(&pqueue->parallel); INIT_WORK(&pqueue->work, padata_parallel_worker); - INIT_WORK(&pqueue->reorder_work, invoke_padata_reorder); atomic_set(&pqueue->num_obj, 0); } } @@ -498,12 +438,13 @@ static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst, padata_init_pqueues(pd); padata_init_squeues(pd); - timer_setup(&pd->timer, padata_reorder_timer, 0); atomic_set(&pd->seq_nr, -1); atomic_set(&pd->reorder_objects, 0); atomic_set(&pd->refcnt, 0); pd->pinst = pinst; spin_lock_init(&pd->lock); + pd->cpu = cpumask_first(pcpumask); + INIT_WORK(&pd->reorder_work, invoke_padata_reorder); return pd; @@ -538,8 +479,6 @@ static void padata_flush_queues(struct parallel_data *pd) flush_work(&pqueue->work); } - del_timer_sync(&pd->timer); - if (atomic_read(&pd->reorder_objects)) padata_reorder(pd);