Hi, Slava Another question suddenly occurred to me, in order to keep the order that rebuilding global cache before updating ”dev_gen“, the wmb should be before updating "dev_gen" rather than after it. Otherwise, in the out-of-order platforms, current order cannot be kept.
Thus, we should change the code as: a) rebuild global cache; b) rte_smp_wmb(); c) updating dev_gen Best Regards Feifei > -----邮件原件----- > 发件人: Feifei Wang > 发送时间: 2021年4月20日 13:54 > 收件人: Slava Ovsiienko <viachesl...@nvidia.com>; Matan Azrad > <ma...@nvidia.com>; Shahaf Shuler <shah...@nvidia.com> > 抄送: dev@dpdk.org; nd <n...@arm.com>; sta...@dpdk.org; Ruifeng Wang > <ruifeng.w...@arm.com>; nd <n...@arm.com>; nd <n...@arm.com> > 主题: 回复: [PATCH v1 3/4] net/mlx5: fix rebuild bug for Memory Region > cache > > Hi, Slava > > Thanks very much for your explanation. > > I can understand the app can wait all mbufs are returned to the memory pool, > and then it can free this mbufs, I agree with this. > > As a result, I will remove the bug fix patch from this series and just replace > the smp barrier with C11 thread fence. Thanks very much for your patient > explanation again. > > Best Regards > Feifei > > > -----邮件原件----- > > 发件人: Slava Ovsiienko <viachesl...@nvidia.com> > > 发送时间: 2021年4月20日 2:51 > > 收件人: Feifei Wang <feifei.wa...@arm.com>; Matan Azrad > > <ma...@nvidia.com>; Shahaf Shuler <shah...@nvidia.com> > > 抄送: dev@dpdk.org; nd <n...@arm.com>; sta...@dpdk.org; Ruifeng Wang > > <ruifeng.w...@arm.com>; nd <n...@arm.com> > > 主题: RE: [PATCH v1 3/4] net/mlx5: fix rebuild bug for Memory Region > > cache > > > > Hi, Feifei > > > > Please, see below > > > > .... > > > > > > Hi, Feifei > > > > > > > > Sorry, I do not follow what this patch fixes. Do we have some > > > > issue/bug with MR cache in practice? > > > > > > This patch fixes the bug which is based on logical deduction, and it > > > doesn't actually happen. > > > > > > > > > > > Each Tx queue has its own dedicated "local" cache for MRs to > > > > convert buffer address in mbufs being transmitted to LKeys > > > > (HW-related entity > > > > handle) and the "global" cache for all MR registered on the device. > > > > > > > > AFAIK, how conversion happens in datapath: > > > > - check the local queue cache flush request > > > > - lookup in local cache > > > > - if not found: > > > > - acquire lock for global cache read access > > > > - lookup in global cache > > > > - release lock for global cache > > > > > > > > How cache update on memory freeing/unregistering happens: > > > > - acquire lock for global cache write access > > > > - [a] remove relevant MRs from the global cache > > > > - [b] set local caches flush request > > > > - free global cache lock > > > > > > > > If I understand correctly, your patch swaps [a] and [b], and local > > > > caches flush is requested earlier. What problem does it solve? > > > > It is not supposed there are in datapath some mbufs referencing to > > > > the memory being freed. Application must ensure this and must not > > > > allocate new mbufs from this memory regions being freed. Hence, > > > > the lookups for these MRs in caches should not occur. > > > > > > For your first point that, application can take charge of preventing > > > MR freed memory being allocated to data path. > > > > > > Does it means that If there is an emergency of MR fragment, such as > > > hotplug, the application must inform thedata path in advance, and > > > this memory will not be allocated, and then the control path will > > > free this memory? If application can do like this, I agree that this bug > cannot happen. > > > > Actually, this is the only correct way for application to operate. > > Let's suppose we have some memory area that application wants to free. > > ALL references to this area must be removed. If we have some mbufs > > allocated from this area, it means that we have memory pool created there. > > > > What application should do: > > - notify all its components/agents the memory area is going to be > > freed > > - all components/agents free the mbufs they might own > > - PMD might not support freeing for some mbufs (for example being sent > > and awaiting for completion), so app should just wait > > - wait till all mbufs are returned to the memory pool (by monitoring > > available obj == pool size) > > > > Otherwise - it is dangerous to free the memory. There are just some > > mbufs still allocated, it is regardless to buf address to MR > > translation. We just can't free the memory - the mapping will be > > destroyed and might cause the segmentation fault by SW or some HW > > issues on DMA access to unmapped memory. It is very generic safety > > approach - do not free the memory that is still in use. Hence, at the > > moment of freeing and unregistering the MR, there MUST BE NO any > mbufs in flight referencing to the addresses being freed. > > No translation to MR being invalidated can happen. > > > > > > > > > For other side, the cache flush has negative effect - the local > > > > cache is getting empty and can't provide translation for other > > > > valid (not being removed) MRs, and the translation has to look up > > > > in the global cache, that is locked now for rebuilding, this > > > > causes the delays in datapatch > > > on acquiring global cache lock. > > > > So, I see some potential performance impact. > > > > > > If above assumption is true, we can go to your second point. I think > > > this is a problem of the tradeoff between cache coherence and > > performance. > > > > > > I can understand your meaning that though global cache has been > > > changed, we should keep the valid MR in local cache as long as > > > possible to ensure the fast searching speed. > > > In the meanwhile, the local cache can be rebuilt later to reduce its > > > waiting time for acquiring the global cache lock. > > > > > > However, this mechanism just ensures the performance unchanged for > > > the first few mbufs. > > > During the next mbufs lkey searching after 'dev_gen' updated, it is > > > still necessary to update the local cache. And the performance can > > > firstly reduce and then returns. Thus, no matter whether there is > > > this patch or not, the performance will jitter in a certain period of > > > time. > > > > Local cache should be updated to remove MRs no longer valid. But we > > just flush the entire cache. > > Let's suppose we have valid MR0, MR1, and not valid MRX in local cache. > > And there are traffic in the datapath for MR0 and MR1, and no traffic > > for MRX anymore. > > > > 1) If we do as you propose: > > a) take a lock > > b) request flush local cache first - all MR0, MR1, MRX will be removed > > on translation in datapath > > c) update global cache, > > d) free lock > > All the traffic for valid MR0, MR1 ALWAYS will be blocked on lock > > taken for cache update since point b) till point d). > > > > 2) If we do as it is implemented now: > > a) take a lock > > b) update global cache > > c) request flush local cache > > d) free lock > > The traffic MIGHT be locked ONLY for MRs non-existing in local cache > > (not happens for MR0 and MR1, must not happen for MRX), and > > probability should be minor. And lock might happen since c) till d) - > > quite short period of time > > > > Summary, the difference between 1) and 2) > > > > Lock probability: > > - 1) lock ALWAYS happen for ANY MR translation after b), > > 2) lock MIGHT happen, for cache miss ONLY, after c) > > > > Lock duration: > > - 1) lock since b) till d), > > 2) lock since c) till d), that seems to be much shorter. > > > > > > > > Finally, in conclusion, I tend to think that the bottom layer can do > > > more things to ensure the correct execution of the program, which > > > may have a negative impact on the performance in a short time, but > > > in the long run, the performance will eventually come back. > > > Furthermore, maybe we should pay attention to the performance in the > > > stable period, and try our best to ensure the correctness of the > > > program in case of > > emergencies. > > > > If we have some mbufs still allocated in memory being freed - there is > > nothing to say about correctness, it is totally incorrect. In my > > opinion, we should not think how to mitigate this incorrect behavior, > > we should not encourage application developers to follow the wrong > approaches. > > > > With best regards, > > Slava > > > > > > > > Best Regards > > > Feifei > > > > > > > With best regards, > > > > Slava > > > > > > > > > -----Original Message----- > > > > > From: Feifei Wang <feifei.wa...@arm.com> > > > > > Sent: Thursday, March 18, 2021 9:19 > > > > > To: Matan Azrad <ma...@nvidia.com>; Shahaf Shuler > > > > > <shah...@nvidia.com>; Slava Ovsiienko <viachesl...@nvidia.com>; > > > > > Yongseok Koh <ys...@mellanox.com> > > > > > Cc: dev@dpdk.org; n...@arm.com; Feifei Wang > > <feifei.wa...@arm.com>; > > > > > sta...@dpdk.org; Ruifeng Wang <ruifeng.w...@arm.com> > > > > > Subject: [PATCH v1 3/4] net/mlx5: fix rebuild bug for Memory > > > > > Region cache > > > > > > > > > > 'dev_gen' is a variable to inform other cores to flush their > > > > > local cache when global cache is rebuilt. > > > > > > > > > > However, if 'dev_gen' is updated after global cache is rebuilt, > > > > > other cores may load a wrong memory region lkey value from old > > > > > local > > > cache. > > > > > > > > > > Timeslot main core worker core > > > > > 1 rebuild global cache > > > > > 2 load unchanged dev_gen > > > > > 3 update dev_gen > > > > > 4 look up old local cache > > > > > > > > > > From the example above, we can see that though global cache is > > > > > rebuilt, due to that dev_gen is not updated, the worker core may > > > > > look up old cache table and receive a wrong memory region lkey value. > > > > > > > > > > To fix this, updating 'dev_gen' should be moved before > > > > > rebuilding global cache to inform worker cores to flush their > > > > > local cache when global cache start rebuilding. And wmb can > > > > > ensure the sequence of this > > > > process. > > > > > > > > > > Fixes: 974f1e7ef146 ("net/mlx5: add new memory region support") > > > > > Cc: sta...@dpdk.org > > > > > > > > > > Suggested-by: Ruifeng Wang <ruifeng.w...@arm.com> > > > > > Signed-off-by: Feifei Wang <feifei.wa...@arm.com> > > > > > Reviewed-by: Ruifeng Wang <ruifeng.w...@arm.com> > > > > > --- > > > > > drivers/net/mlx5/mlx5_mr.c | 37 > > > > > +++++++++++++++++-------------------- > > > > > 1 file changed, 17 insertions(+), 20 deletions(-) > > > > > > > > > > diff --git a/drivers/net/mlx5/mlx5_mr.c > > > > > b/drivers/net/mlx5/mlx5_mr.c index > > > > > da4e91fc2..7ce1d3e64 100644 > > > > > --- a/drivers/net/mlx5/mlx5_mr.c > > > > > +++ b/drivers/net/mlx5/mlx5_mr.c > > > > > @@ -103,20 +103,18 @@ mlx5_mr_mem_event_free_cb(struct > > > > > mlx5_dev_ctx_shared *sh, > > > > > rebuild = 1; > > > > > } > > > > > if (rebuild) { > > > > > - mlx5_mr_rebuild_cache(&sh->share_cache); > > > > > + ++sh->share_cache.dev_gen; > > > > > + DEBUG("broadcasting local cache flush, gen=%d", > > > > > + sh->share_cache.dev_gen); > > > > > + > > > > > /* > > > > > * Flush local caches by propagating invalidation > across cores. > > > > > - * rte_smp_wmb() is enough to synchronize this > event. If > > > > > one of > > > > > - * freed memsegs is seen by other core, that means > the > > > > > memseg > > > > > - * has been allocated by allocator, which will come > after this > > > > > - * free call. Therefore, this store instruction > (incrementing > > > > > - * generation below) will be guaranteed to be seen > by other > > > > > core > > > > > - * before the core sees the newly allocated memory. > > > > > + * rte_smp_wmb() is to keep the order that dev_gen > > > > > updated before > > > > > + * rebuilding global cache. Therefore, other core can > flush > > > > > their > > > > > + * local cache on time. > > > > > */ > > > > > - ++sh->share_cache.dev_gen; > > > > > - DEBUG("broadcasting local cache flush, gen=%d", > > > > > - sh->share_cache.dev_gen); > > > > > rte_smp_wmb(); > > > > > + mlx5_mr_rebuild_cache(&sh->share_cache); > > > > > } > > > > > rte_rwlock_write_unlock(&sh->share_cache.rwlock); > > > > > } > > > > > @@ -407,20 +405,19 @@ mlx5_dma_unmap(struct rte_pci_device > > > *pdev, > > > > void > > > > > *addr, > > > > > mlx5_mr_free(mr, sh->share_cache.dereg_mr_cb); > > > > > DEBUG("port %u remove MR(%p) from list", dev->data- > >port_id, > > > > > (void *)mr); > > > > > - mlx5_mr_rebuild_cache(&sh->share_cache); > > > > > + > > > > > + ++sh->share_cache.dev_gen; > > > > > + DEBUG("broadcasting local cache flush, gen=%d", > > > > > + sh->share_cache.dev_gen); > > > > > + > > > > > /* > > > > > * Flush local caches by propagating invalidation across cores. > > > > > - * rte_smp_wmb() is enough to synchronize this event. If > one of > > > > > - * freed memsegs is seen by other core, that means the > memseg > > > > > - * has been allocated by allocator, which will come after this > > > > > - * free call. Therefore, this store instruction (incrementing > > > > > - * generation below) will be guaranteed to be seen by other > core > > > > > - * before the core sees the newly allocated memory. > > > > > + * rte_smp_wmb() is to keep the order that dev_gen > updated > > > > > before > > > > > + * rebuilding global cache. Therefore, other core can flush > their > > > > > + * local cache on time. > > > > > */ > > > > > - ++sh->share_cache.dev_gen; > > > > > - DEBUG("broadcasting local cache flush, gen=%d", > > > > > - sh->share_cache.dev_gen); > > > > > rte_smp_wmb(); > > > > > + mlx5_mr_rebuild_cache(&sh->share_cache); > > > > > rte_rwlock_read_unlock(&sh->share_cache.rwlock); > > > > > return 0; > > > > > } > > > > > -- > > > > > 2.25.1