On Tue, May 05, 2020 at 04:04:42AM +0530, Ritesh Harjani wrote: > Hello Paul, > > On 5/2/20 12:01 AM, Paul E. McKenney wrote: > > On Fri, May 01, 2020 at 11:59:44AM +0530, Ritesh Harjani wrote: > > > There could be a race in function ext4_mb_discard_group_preallocations() > > > where the 1st thread may iterate through group's bb_prealloc_list and > > > remove all the PAs and add to function's local list head. > > > Now if the 2nd thread comes in to discard the group preallocations, > > > it will see that the group->bb_prealloc_list is empty and will return 0. > > > > > > Consider for a case where we have less number of groups > > > (for e.g. just group 0), > > > this may even return an -ENOSPC error from ext4_mb_new_blocks() > > > (where we call for ext4_mb_discard_group_preallocations()). > > > But that is wrong, since 2nd thread should have waited for 1st thread > > > to release all the PAs and should have retried for allocation. > > > Since 1st thread was anyway going to discard the PAs. > > > > > > The algorithm using this percpu seq counter goes below: > > > 1. We sample the percpu discard_pa_seq counter before trying for block > > > allocation in ext4_mb_new_blocks(). > > > 2. We increment this percpu discard_pa_seq counter when we succeed in > > > allocating blocks and hence while adding the remaining blocks in group's > > > bb_prealloc_list (ext4_mb_new_inode_pa/ext4_mb_new_group_pa). > > > 3. We also increment this percpu seq counter when we successfully identify > > > that the bb_prealloc_list is not empty and hence proceed for discarding > > > of those PAs inside ext4_mb_discard_group_preallocations(). > > > > > > Now to make sure that the regular fast path of block allocation is not > > > affected, as a small optimization we only sample the percpu seq counter > > > on that cpu. Only when the block allocation fails and when freed blocks > > > found were 0, that is when we sample percpu seq counter for all cpus using > > > below function ext4_get_discard_pa_seq_sum(). This happens after making > > > sure that all the PAs on grp->bb_prealloc_list got freed. > > > > > > TO CHECK: Though in here rcu_barrier only happens in ENOSPC path. > > > ================================================================= > > > On rcu_barrier() - How expensive it can be? > > > Does that mean that every thread who is coming and waiting on > > > rcu_barrier() will actually check whether call_rcu() has completed by > > > checking that on every cpu? So will this be a O(n*m) operation? > > > (n = no. of threads, m = no. of cpus). > > > Or are there some sort of optimization in using rcu_barrier()? > > > > This first part assumes that a single task is executing a series of > > rcu_barrier() calls sequentially. > > > > Yes, the rcu_barrier() function makes heavy use of batching optimizations > > and asynchrony. So from an overhead perspective, it queues an RCU > > callback on each CPU that has at least one callback already queued. > > It does this queuing in such a way to avoid the need for any additional > > grace periods. Of course, each callback actually queued will eventually > > be invoked (function call through a pointer), and the callback will > > atomically decrement a counter and wake up the rcu_barrier() task if > > that counter reaches zero. So the CPU overhead is O(M), where M is the > > number of CPUs, with negligible overhead except for those CPUs that have > > at least one RCU callback queued. > > > > So the CPU overhead of rcu_barrier() is quite small, at least when > > taken on a per-CPU basis. > > > > However, the latency can be a bit longer than than that of an RCU > > grace period. In normal operation, this will normally range from a few > > milliseconds to a few hundred milliseconds. > > > > And the caller of rcu_barrier() can also do batching. For example, > > the following: > > > > make_call_rcu_stop(a); > > rcu_barrier(); > > depend_on_callbacks_finished(a); > > make_call_rcu_stop(b); > > rcu_barrier(); > > depend_on_callbacks_finished(b); > > ... > > make_call_rcu_stop(z); > > rcu_barrier(); > > depend_on_callbacks_finished(z); > > > > Can be safely transformed into this: > > > > make_call_rcu_stop(a); > > make_call_rcu_stop(b); > > ... > > make_call_rcu_stop(z); > > rcu_barrier(); > > depend_on_callbacks_finished(a); > > depend_on_callbacks_finished(b); > > ... > > depend_on_callbacks_finished(z); > > > > That single rcu_barrier() call can cover all 26 instances. > > Give or take possible software-engineering and maintainability > > issues, of course. > > > > But what if a large number of tasks are concurrently executing > > rcu_barrier()? > > > > These will also be batched. The first task will run the rcu_barrier() > > machinery. Any rcu_barrier() calls that show up before the first one > > completes will share a single second run of the rcu_barrier() machinery. > > And any rcu_barrier() calls that show up before this second run of > > the machinery completes will be handled by a single third run of this > > machinery. And so on. > > Thanks for explaining that. So I was doing more reading on rcu_barrier() > code and I think what it essentially does, is depicted from line 1 - 12 > below. > > So AFAIU, > when there are multiple threads calling rcu_barrier(), what will happen > mostly is that 1st thread which calls this func will get the mutex_lock > (line 1) and will make sure that all calls to call_rcu() already queued > on different cpus are completed before this rcu_barrier() returns. > And how exactly this happens is that rcu_barrier() runs it's own > rcu_barrier_func() on all of those cpus which have call_rcu() queued up. > rcu_barrier_func() while running on each of those cpus then queues up > a call_rcu() (on it's tail of call_rcu queue) with > rcu_barrier_callback() function. This call_rcu queued on tail will run, > after all of previous call_rcu() functions from those per cpu queues are > completed. This rcu_barrier_callback() will mostly run after a grace > period (like how in general call_rcu() works). > > And meanwhile all other threads will sleep while trying to acquire > that mutex lock. So at this time there could be something useful which > these cpus should be able to do (while waiting on mutex). > Now, when these other threads will be able to acquire the mutex_lock() > to proceed inside rcu_barrier(), in case if there are no more call_rcu() > queued in (since it should have been taken care by previous thread), > (checked in line 5 below) then nothing else will be done by > rcu_barrier(), and it will simply exit. True enough. However, on a busy system it is very likely that there will be additional callbacks enqueued somewhere by that time, and so the next task to acquire the mutex will have to wait. However, the third task might well be able to skip the whole process based on the ->barrier_sequence counter in the rcu_state structure. (The ->barrier_sequence manipulations are not shown in your steps below, but they lead to the "if (rcu_seq_done" early exit just after the mutex is acquired. > Is the above understanding correct? Reasonably close. > Instead if I use a spinlock. Then in the worst case what will happen is > that most of the threads will just spin on this lock while not being > able to do anything useful. Also in case if it's heavy multi-threaded > and if we hit this scenario, then this could also result into high cpu > usage and sluggish performance since all threads are waiting on > spinlock. It can also result in deadlock because something holding a pure spinlock will need to do a wait_for_completion() while holding that spinlock. And of course, lockdep will complain bitterly, as well it should. ;-) You also don't get to drop the lock before doing the wait_for_completion() without adding some sort of synchronization preventing the next task from using the rcu_data structure's ->barrier_head before the current task is done with them. > My main reason to understand rcu_barrier() functionality is to weigh > my options of which one is a better approach to be deployed here. > Hence all of these queries :) Why not just try it and then measure the actual effect on performance? Also, given actual measurements, there might be things I could do to speed this up, for example, applying funnel locking to decrease memory contention on ->barrier_mutex, should that turn out to be a real problem. But it all depends on whether there is an actual performance problem, and if so, what aspect of the overall algorithm is in trouble. Thanx, Paul > 1. mutex_lock(&rcu_state.barrier_mutex); > 2. init_completion(&rcu_state.barrier_completion); > 3. atomic_set(&rcu_state.barrier_cpu_count, 1); > > 4. for_each_possible_cpu(cpu) { > 5. if_we_have_any_call_rcu_pending_on_this_cpu { > 6. smp_call_function_single(cpu, rcu_barrier_func, NULL, 1) > 7. } > 8. } > > 9. if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) > 10. complete(&rcu_state.barrier_completion); > > 11. wait_for_completion(&rcu_state.barrier_completion); > > 12. mutex_unlock(&rcu_state.barrier_mutex); > > > Thanks > -ritesh > > > > > > --- > > > Note: The other method [1] also used to check for grp->bb_free next time > > > if we couldn't discard anything. But it had below concerns due to which > > > we cannot use that approach. > > > 1. But one suspicion with that was that if grp->bb_free is non-zero for some > > > reason (not sure), then we may result in that loop indefinitely but still > > > won't be able to satisfy any request. > > > 2. To check for grp->bb_free all threads were to wait on grp's spin_lock > > > which might result in increased cpu usage. > > > 3. In case of group's PA allocation (i.e. when file's size is < 1MB for > > > 64K blocksize), there is still a case where ENOSPC could be returned. > > > This happens when the grp->bb_free is set to 0 but those blocks are > > > actually not yet added to PA. Yes, this could actually happen since > > > reducing grp->bb_free and adding those extra blocks in > > > bb_prealloc_list are not done atomically. Hence the race. > > > > > > [1]: https://patchwork.ozlabs.org/project/linux-ext4/patch/533ac1f5b19c520b08f8c99aec5baf8729185714.1586954511.git.riteshh@xxxxxxxxxxxxx/ > > > > > > Signed-off-by: Ritesh Harjani <riteshh@xxxxxxxxxxxxx> > > > --- > > > fs/ext4/mballoc.c | 65 +++++++++++++++++++++++++++++++++++++++++++---- > > > 1 file changed, 60 insertions(+), 5 deletions(-) > > > > > > diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c > > > index a742e51e33b8..6bb08bb3c0ce 100644 > > > --- a/fs/ext4/mballoc.c > > > +++ b/fs/ext4/mballoc.c > > > @@ -357,6 +357,35 @@ static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, > > > static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, > > > ext4_group_t group); > > > +/* > > > + * The algorithm using this percpu seq counter goes below: > > > + * 1. We sample the percpu discard_pa_seq counter before trying for block > > > + * allocation in ext4_mb_new_blocks(). > > > + * 2. We increment this percpu discard_pa_seq counter when we succeed in > > > + * allocating blocks and hence while adding the remaining blocks in group's > > > + * bb_prealloc_list (ext4_mb_new_inode_pa/ext4_mb_new_group_pa). > > > + * 3. We also increment this percpu seq counter when we successfully identify > > > + * that the bb_prealloc_list is not empty and hence proceed for discarding > > > + * of those PAs inside ext4_mb_discard_group_preallocations(). > > > + * > > > + * Now to make sure that the regular fast path of block allocation is not > > > + * affected, as a small optimization we only sample the percpu seq counter > > > + * on that cpu. Only when the block allocation fails and when freed blocks > > > + * found were 0, that is when we sample percpu seq counter for all cpus using > > > + * below function ext4_get_discard_pa_seq_sum(). This happens after making > > > + * sure that all the PAs on grp->bb_prealloc_list got freed. > > > + */ > > > +DEFINE_PER_CPU(u64, discard_pa_seq); > > > +static inline u64 ext4_get_discard_pa_seq_sum(void) > > > +{ > > > + int __cpu; > > > + u64 __seq = 0; > > > + > > > + for_each_possible_cpu(__cpu) > > > + __seq += per_cpu(discard_pa_seq, __cpu); > > > + return __seq; > > > +} > > > + > > > static inline void *mb_correct_addr_and_bit(int *bit, void *addr) > > > { > > > #if BITS_PER_LONG == 64 > > > @@ -3730,6 +3759,7 @@ ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) > > > pa->pa_inode = ac->ac_inode; > > > ext4_lock_group(sb, ac->ac_b_ex.fe_group); > > > + this_cpu_inc(discard_pa_seq); > > > list_add(&pa->pa_group_list, &grp->bb_prealloc_list); > > > ext4_unlock_group(sb, ac->ac_b_ex.fe_group); > > > @@ -3791,6 +3821,7 @@ ext4_mb_new_group_pa(struct ext4_allocation_context *ac) > > > pa->pa_inode = NULL; > > > ext4_lock_group(sb, ac->ac_b_ex.fe_group); > > > + this_cpu_inc(discard_pa_seq); > > > list_add(&pa->pa_group_list, &grp->bb_prealloc_list); > > > ext4_unlock_group(sb, ac->ac_b_ex.fe_group); > > > @@ -3943,6 +3974,7 @@ ext4_mb_discard_group_preallocations(struct super_block *sb, > > > INIT_LIST_HEAD(&list); > > > repeat: > > > ext4_lock_group(sb, group); > > > + this_cpu_inc(discard_pa_seq); > > > list_for_each_entry_safe(pa, tmp, > > > &grp->bb_prealloc_list, pa_group_list) { > > > spin_lock(&pa->pa_lock); > > > @@ -4487,14 +4519,35 @@ static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) > > > } > > > static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, > > > - struct ext4_allocation_context *ac) > > > + struct ext4_allocation_context *ac, u64 *seq) > > > { > > > int freed; > > > + u64 seq_retry = 0; > > > + bool ret = false; > > > freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); > > > - if (freed) > > > - return true; > > > - return false; > > > + if (freed) { > > > + ret = true; > > > + goto out_dbg; > > > + } > > > + /* > > > + * Unless it is ensured that PAs are actually freed, we may hit > > > + * a ENOSPC error since the next time seq may match while the PA blocks > > > + * are still getting freed in ext4_mb_release_inode/group_pa(). > > > + * So, rcu_barrier() here is to make sure that any call_rcu queued in > > > + * ext4_mb_discard_group_preallocations() is completed before we > > > + * proceed further to retry for block allocation. > > > + */ > > > + rcu_barrier(); > > > + seq_retry = ext4_get_discard_pa_seq_sum(); > > > + if (seq_retry != *seq) { > > > + *seq = seq_retry; > > > + ret = true; > > > + } > > > + > > > +out_dbg: > > > + mb_debug(1, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); > > > + return ret; > > > } > > > /* > > > @@ -4511,6 +4564,7 @@ ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, > > > ext4_fsblk_t block = 0; > > > unsigned int inquota = 0; > > > unsigned int reserv_clstrs = 0; > > > + u64 seq; > > > might_sleep(); > > > sb = ar->inode->i_sb; > > > @@ -4572,6 +4626,7 @@ ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, > > > } > > > ac->ac_op = EXT4_MB_HISTORY_PREALLOC; > > > + seq = *this_cpu_ptr(&discard_pa_seq); > > > if (!ext4_mb_use_preallocated(ac)) { > > > ac->ac_op = EXT4_MB_HISTORY_ALLOC; > > > ext4_mb_normalize_request(ac, ar); > > > @@ -4603,7 +4658,7 @@ ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, > > > ar->len = ac->ac_b_ex.fe_len; > > > } > > > } else { > > > - if (ext4_mb_discard_preallocations_should_retry(sb, ac)) > > > + if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) > > > goto repeat; > > > *errp = -ENOSPC; > > > } > > > -- > > > 2.21.0 > > >