Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
Hi Robin, On Mon, Jul 30, 2018 at 12:40 PM, Ganapatrao Kulkarni wrote: > On Fri, Jul 27, 2018 at 9:48 PM, Robin Murphy wrote: >> On 27/07/18 13:56, Ganapatrao Kulkarni wrote: >> [...] > > did you get any chance to look in to this issue? > i am waiting for your suggestion/patch for this issue! I got as far as [1], but I wasn't sure how much I liked it, since it still seems a little invasive for such a specific case (plus I can't remember if it's actually been debugged or not). I think in the end I started wondering whether it's even worth bothering with the 32-bit optimisation for PCIe devices - 4 extra bytes worth of TLP is surely a lot less significant than every transaction taking up to 50% more bus cycles was for legacy PCI. >>> >>> >>> how about tracking previous attempt to get 32bit range iova and avoid >>> further attempts, if it was failed. Later Resume attempts once >>> replenish happens. >>> Created patch for the same [2] >> >> >> Ooh, that's a much neater implementation of essentially the same concept - >> now why couldn't I think of that? :) >> >> Looks like it should be possible to make it entirely self-contained too, >> since alloc_iova() is in a position to both test and update the flag based >> on the limit_pfn passed in. > > is below patch much better? testing with this diff looks ok, shall i send formal patch with your Acked-by? > > diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c > index 83fe262..abb15d6 100644 > --- a/drivers/iommu/iova.c > +++ b/drivers/iommu/iova.c > @@ -56,6 +56,7 @@ init_iova_domain(struct iova_domain *iovad, unsigned > long granule, > iovad->granule = granule; > iovad->start_pfn = start_pfn; > iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad)); > +iovad->free_32bit_pfns = true; > iovad->flush_cb = NULL; > iovad->fq = NULL; > iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR; > @@ -139,8 +140,10 @@ __cached_rbnode_delete_update(struct iova_domain > *iovad, struct iova *free) > > cached_iova = rb_entry(iovad->cached32_node, struct iova, node); > if (free->pfn_hi < iovad->dma_32bit_pfn && > -free->pfn_lo >= cached_iova->pfn_lo) > +free->pfn_lo >= cached_iova->pfn_lo) { > iovad->cached32_node = rb_next(>node); > +iovad->free_32bit_pfns = true; > +} > > cached_iova = rb_entry(iovad->cached_node, struct iova, node); > if (free->pfn_lo >= cached_iova->pfn_lo) > @@ -290,6 +293,10 @@ alloc_iova(struct iova_domain *iovad, unsigned long size, > struct iova *new_iova; > int ret; > > +if (limit_pfn < iovad->dma_32bit_pfn && > +!iovad->free_32bit_pfns) > +return NULL; > + > new_iova = alloc_iova_mem(); > if (!new_iova) > return NULL; > @@ -299,6 +306,8 @@ alloc_iova(struct iova_domain *iovad, unsigned long size, > > if (ret) { > free_iova_mem(new_iova); > +if (limit_pfn < iovad->dma_32bit_pfn) > +iovad->free_32bit_pfns = false; > return NULL; > } > > diff --git a/include/linux/iova.h b/include/linux/iova.h > index 928442d..3810ba9 100644 > --- a/include/linux/iova.h > +++ b/include/linux/iova.h > @@ -96,6 +96,7 @@ struct iova_domain { > flush-queues */ > atomic_t fq_timer_on;/* 1 when timer is active, 0 > when not */ > +boolfree_32bit_pfns; > }; > > static inline unsigned long iova_size(struct iova *iova) > -- > 2.9.4 >> >> Robin. >> >> >>> >>> [2] >>> https://github.com/gpkulkarni/linux/commit/e2343a3e1f55cdeb5694103dd354bcb881dc65c3 >>> note, the testing of this patch is in progress. >>> Robin. [1] http://www.linux-arm.org/git?p=linux-rm.git;a=commitdiff;h=a8e0e4af10ebebb3669750e05bf0028e5bd6afe8 >>> >>> >>> thanks >>> Ganapat >>> >> > > thanks > Ganapat thanks Ganapat ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
On Fri, Jul 27, 2018 at 9:48 PM, Robin Murphy wrote: > On 27/07/18 13:56, Ganapatrao Kulkarni wrote: > [...] did you get any chance to look in to this issue? i am waiting for your suggestion/patch for this issue! >>> >>> >>> >>> I got as far as [1], but I wasn't sure how much I liked it, since it >>> still >>> seems a little invasive for such a specific case (plus I can't remember >>> if >>> it's actually been debugged or not). I think in the end I started >>> wondering >>> whether it's even worth bothering with the 32-bit optimisation for PCIe >>> devices - 4 extra bytes worth of TLP is surely a lot less significant >>> than >>> every transaction taking up to 50% more bus cycles was for legacy PCI. >> >> >> how about tracking previous attempt to get 32bit range iova and avoid >> further attempts, if it was failed. Later Resume attempts once >> replenish happens. >> Created patch for the same [2] > > > Ooh, that's a much neater implementation of essentially the same concept - > now why couldn't I think of that? :) > > Looks like it should be possible to make it entirely self-contained too, > since alloc_iova() is in a position to both test and update the flag based > on the limit_pfn passed in. is below patch much better? diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c index 83fe262..abb15d6 100644 --- a/drivers/iommu/iova.c +++ b/drivers/iommu/iova.c @@ -56,6 +56,7 @@ init_iova_domain(struct iova_domain *iovad, unsigned long granule, iovad->granule = granule; iovad->start_pfn = start_pfn; iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad)); +iovad->free_32bit_pfns = true; iovad->flush_cb = NULL; iovad->fq = NULL; iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR; @@ -139,8 +140,10 @@ __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free) cached_iova = rb_entry(iovad->cached32_node, struct iova, node); if (free->pfn_hi < iovad->dma_32bit_pfn && -free->pfn_lo >= cached_iova->pfn_lo) +free->pfn_lo >= cached_iova->pfn_lo) { iovad->cached32_node = rb_next(>node); +iovad->free_32bit_pfns = true; +} cached_iova = rb_entry(iovad->cached_node, struct iova, node); if (free->pfn_lo >= cached_iova->pfn_lo) @@ -290,6 +293,10 @@ alloc_iova(struct iova_domain *iovad, unsigned long size, struct iova *new_iova; int ret; +if (limit_pfn < iovad->dma_32bit_pfn && +!iovad->free_32bit_pfns) +return NULL; + new_iova = alloc_iova_mem(); if (!new_iova) return NULL; @@ -299,6 +306,8 @@ alloc_iova(struct iova_domain *iovad, unsigned long size, if (ret) { free_iova_mem(new_iova); +if (limit_pfn < iovad->dma_32bit_pfn) +iovad->free_32bit_pfns = false; return NULL; } diff --git a/include/linux/iova.h b/include/linux/iova.h index 928442d..3810ba9 100644 --- a/include/linux/iova.h +++ b/include/linux/iova.h @@ -96,6 +96,7 @@ struct iova_domain { flush-queues */ atomic_t fq_timer_on;/* 1 when timer is active, 0 when not */ +boolfree_32bit_pfns; }; static inline unsigned long iova_size(struct iova *iova) -- 2.9.4 > > Robin. > > >> >> [2] >> https://github.com/gpkulkarni/linux/commit/e2343a3e1f55cdeb5694103dd354bcb881dc65c3 >> note, the testing of this patch is in progress. >> >>> >>> Robin. >>> >>> [1] >>> >>> http://www.linux-arm.org/git?p=linux-rm.git;a=commitdiff;h=a8e0e4af10ebebb3669750e05bf0028e5bd6afe8 >> >> >> thanks >> Ganapat >> > thanks Ganapat ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
On 27/07/18 13:56, Ganapatrao Kulkarni wrote: [...] did you get any chance to look in to this issue? i am waiting for your suggestion/patch for this issue! I got as far as [1], but I wasn't sure how much I liked it, since it still seems a little invasive for such a specific case (plus I can't remember if it's actually been debugged or not). I think in the end I started wondering whether it's even worth bothering with the 32-bit optimisation for PCIe devices - 4 extra bytes worth of TLP is surely a lot less significant than every transaction taking up to 50% more bus cycles was for legacy PCI. how about tracking previous attempt to get 32bit range iova and avoid further attempts, if it was failed. Later Resume attempts once replenish happens. Created patch for the same [2] Ooh, that's a much neater implementation of essentially the same concept - now why couldn't I think of that? :) Looks like it should be possible to make it entirely self-contained too, since alloc_iova() is in a position to both test and update the flag based on the limit_pfn passed in. Robin. [2] https://github.com/gpkulkarni/linux/commit/e2343a3e1f55cdeb5694103dd354bcb881dc65c3 note, the testing of this patch is in progress. Robin. [1] http://www.linux-arm.org/git?p=linux-rm.git;a=commitdiff;h=a8e0e4af10ebebb3669750e05bf0028e5bd6afe8 thanks Ganapat ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
Hi Robin, On Wed, Jul 25, 2018 at 7:50 PM, Robin Murphy wrote: > On 12/07/18 08:45, Ganapatrao Kulkarni wrote: >> >> Hi Robin, >> >> >> On Mon, Jun 4, 2018 at 9:36 AM, Ganapatrao Kulkarni >> wrote: >>> >>> ping?? >>> >>> On Mon, May 21, 2018 at 6:45 AM, Ganapatrao Kulkarni >>> wrote: On Thu, Apr 26, 2018 at 3:15 PM, Ganapatrao Kulkarni wrote: > > Hi Robin, > > On Mon, Apr 23, 2018 at 11:11 PM, Ganapatrao Kulkarni > wrote: >> >> On Mon, Apr 23, 2018 at 10:07 PM, Robin Murphy >> wrote: >>> >>> On 19/04/18 18:12, Ganapatrao Kulkarni wrote: The performance drop is observed with long hours iperf testing using 40G cards. This is mainly due to long iterations in finding the free iova range in 32bit address space. In current implementation for 64bit PCI devices, there is always first attempt to allocate iova from 32bit(SAC preferred over DAC) address range. Once we run out 32bit range, there is allocation from higher range, however due to cached32_node optimization it does not suppose to be painful. cached32_node always points to recently allocated 32-bit node. When address range is full, it will be pointing to last allocated node (leaf node), so walking rbtree to find the available range is not expensive affair. However this optimization does not behave well when one of the middle node is freed. In that case cached32_node is updated to point to next iova range. The next iova allocation will consume free range and again update cached32_node to itself. From now on, walking over 32-bit range is more expensive. This patch adds fix to update cached node to leaf node when there are no iova free range left, which avoids unnecessary long iterations. >>> >>> >>> >>> The only trouble with this is that "allocation failed" doesn't >>> uniquely mean >>> "space full". Say that after some time the 32-bit space ends up empty >>> except >>> for one page at 0x1000 and one at 0x8000, then somebody tries to >>> allocate 2GB. If we move the cached node down to the leftmost entry >>> when >>> that fails, all subsequent allocation attempts are now going to fail >>> despite >>> the space being 99.% free! >>> >>> I can see a couple of ways to solve that general problem of free >>> space above >>> the cached node getting lost, but neither of them helps with the case >>> where >>> there is genuinely insufficient space (and if anything would make it >>> even >>> slower). In terms of the optimisation you want here, i.e. fail fast >>> when an >>> allocation cannot possibly succeed, the only reliable idea which >>> comes to >>> mind is free-PFN accounting. I might give that a go myself to see how >>> ugly >>> it looks. >> >> >> did you get any chance to look in to this issue? >> i am waiting for your suggestion/patch for this issue! > > > I got as far as [1], but I wasn't sure how much I liked it, since it still > seems a little invasive for such a specific case (plus I can't remember if > it's actually been debugged or not). I think in the end I started wondering > whether it's even worth bothering with the 32-bit optimisation for PCIe > devices - 4 extra bytes worth of TLP is surely a lot less significant than > every transaction taking up to 50% more bus cycles was for legacy PCI. how about tracking previous attempt to get 32bit range iova and avoid further attempts, if it was failed. Later Resume attempts once replenish happens. Created patch for the same [2] [2] https://github.com/gpkulkarni/linux/commit/e2343a3e1f55cdeb5694103dd354bcb881dc65c3 note, the testing of this patch is in progress. > > Robin. > > [1] > http://www.linux-arm.org/git?p=linux-rm.git;a=commitdiff;h=a8e0e4af10ebebb3669750e05bf0028e5bd6afe8 thanks Ganapat ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
On 12/07/18 08:45, Ganapatrao Kulkarni wrote: Hi Robin, On Mon, Jun 4, 2018 at 9:36 AM, Ganapatrao Kulkarni wrote: ping?? On Mon, May 21, 2018 at 6:45 AM, Ganapatrao Kulkarni wrote: On Thu, Apr 26, 2018 at 3:15 PM, Ganapatrao Kulkarni wrote: Hi Robin, On Mon, Apr 23, 2018 at 11:11 PM, Ganapatrao Kulkarni wrote: On Mon, Apr 23, 2018 at 10:07 PM, Robin Murphy wrote: On 19/04/18 18:12, Ganapatrao Kulkarni wrote: The performance drop is observed with long hours iperf testing using 40G cards. This is mainly due to long iterations in finding the free iova range in 32bit address space. In current implementation for 64bit PCI devices, there is always first attempt to allocate iova from 32bit(SAC preferred over DAC) address range. Once we run out 32bit range, there is allocation from higher range, however due to cached32_node optimization it does not suppose to be painful. cached32_node always points to recently allocated 32-bit node. When address range is full, it will be pointing to last allocated node (leaf node), so walking rbtree to find the available range is not expensive affair. However this optimization does not behave well when one of the middle node is freed. In that case cached32_node is updated to point to next iova range. The next iova allocation will consume free range and again update cached32_node to itself. From now on, walking over 32-bit range is more expensive. This patch adds fix to update cached node to leaf node when there are no iova free range left, which avoids unnecessary long iterations. The only trouble with this is that "allocation failed" doesn't uniquely mean "space full". Say that after some time the 32-bit space ends up empty except for one page at 0x1000 and one at 0x8000, then somebody tries to allocate 2GB. If we move the cached node down to the leftmost entry when that fails, all subsequent allocation attempts are now going to fail despite the space being 99.% free! I can see a couple of ways to solve that general problem of free space above the cached node getting lost, but neither of them helps with the case where there is genuinely insufficient space (and if anything would make it even slower). In terms of the optimisation you want here, i.e. fail fast when an allocation cannot possibly succeed, the only reliable idea which comes to mind is free-PFN accounting. I might give that a go myself to see how ugly it looks. did you get any chance to look in to this issue? i am waiting for your suggestion/patch for this issue! I got as far as [1], but I wasn't sure how much I liked it, since it still seems a little invasive for such a specific case (plus I can't remember if it's actually been debugged or not). I think in the end I started wondering whether it's even worth bothering with the 32-bit optimisation for PCIe devices - 4 extra bytes worth of TLP is surely a lot less significant than every transaction taking up to 50% more bus cycles was for legacy PCI. Robin. [1] http://www.linux-arm.org/git?p=linux-rm.git;a=commitdiff;h=a8e0e4af10ebebb3669750e05bf0028e5bd6afe8 ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
Hi Robin, On Mon, Jun 4, 2018 at 9:36 AM, Ganapatrao Kulkarni wrote: > ping?? > > On Mon, May 21, 2018 at 6:45 AM, Ganapatrao Kulkarni > wrote: >> On Thu, Apr 26, 2018 at 3:15 PM, Ganapatrao Kulkarni >> wrote: >>> Hi Robin, >>> >>> On Mon, Apr 23, 2018 at 11:11 PM, Ganapatrao Kulkarni >>> wrote: On Mon, Apr 23, 2018 at 10:07 PM, Robin Murphy wrote: > On 19/04/18 18:12, Ganapatrao Kulkarni wrote: >> >> The performance drop is observed with long hours iperf testing using 40G >> cards. This is mainly due to long iterations in finding the free iova >> range in 32bit address space. >> >> In current implementation for 64bit PCI devices, there is always first >> attempt to allocate iova from 32bit(SAC preferred over DAC) address >> range. Once we run out 32bit range, there is allocation from higher >> range, >> however due to cached32_node optimization it does not suppose to be >> painful. cached32_node always points to recently allocated 32-bit node. >> When address range is full, it will be pointing to last allocated node >> (leaf node), so walking rbtree to find the available range is not >> expensive affair. However this optimization does not behave well when >> one of the middle node is freed. In that case cached32_node is updated >> to point to next iova range. The next iova allocation will consume free >> range and again update cached32_node to itself. From now on, walking >> over 32-bit range is more expensive. >> >> This patch adds fix to update cached node to leaf node when there are no >> iova free range left, which avoids unnecessary long iterations. > > > The only trouble with this is that "allocation failed" doesn't uniquely > mean > "space full". Say that after some time the 32-bit space ends up empty > except > for one page at 0x1000 and one at 0x8000, then somebody tries to > allocate 2GB. If we move the cached node down to the leftmost entry when > that fails, all subsequent allocation attempts are now going to fail > despite > the space being 99.% free! > > I can see a couple of ways to solve that general problem of free space > above > the cached node getting lost, but neither of them helps with the case > where > there is genuinely insufficient space (and if anything would make it even > slower). In terms of the optimisation you want here, i.e. fail fast when > an > allocation cannot possibly succeed, the only reliable idea which comes to > mind is free-PFN accounting. I might give that a go myself to see how ugly > it looks. did you get any chance to look in to this issue? i am waiting for your suggestion/patch for this issue! >>> >>> For this testing, dual port intel 40G card(XL710) used and both ports >>> were connected in loop-back. Ran iperf server and clients on both >>> ports(used NAT to route packets out on intended ports).There were 10 >>> iperf clients invoked every 60 seconds in loop for hours for each >>> port. Initially the performance on both ports is seen close to line >>> rate, however after test ran about 4 to 6 hours, the performance >>> started dropping to very low (to few hundred Mbps) on both >>> connections. >>> >>> IMO, this is common bug and should happen on any other platforms too >>> and needs to be fixed at the earliest. >>> Please let me know if you have better way to fix this, i am happy to >>> test your patch! >> >> any update on this issue? >>> i see 2 problems in current implementation, 1. We don't replenish the 32 bits range, until first attempt of second allocation(64 bit) fails. 2. Having per cpu cache might not yield good hit on platforms with more number of CPUs. however irrespective of current issues, It makes sense to update cached node as done in this patch , when there is failure to get iova range using current cached pointer which is forcing for the unnecessary time consuming do-while iterations until any replenish happens! thanks Ganapat > > Robin. > > >> Signed-off-by: Ganapatrao Kulkarni >> --- >> drivers/iommu/iova.c | 6 ++ >> 1 file changed, 6 insertions(+) >> >> diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c >> index 83fe262..e6ee2ea 100644 >> --- a/drivers/iommu/iova.c >> +++ b/drivers/iommu/iova.c >> @@ -201,6 +201,12 @@ static int __alloc_and_insert_iova_range(struct >> iova_domain *iovad, >> } while (curr && new_pfn <= curr_iova->pfn_hi); >> if (limit_pfn < size || new_pfn < iovad->start_pfn) { >> + /* No more cached node points to free hole, update to >> leaf >> node. >> +*/ >> + struct iova *prev_iova; >> + >> + prev_iova = rb_entry(prev, struct iova, node);
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
ping?? On Mon, May 21, 2018 at 6:45 AM, Ganapatrao Kulkarni wrote: > On Thu, Apr 26, 2018 at 3:15 PM, Ganapatrao Kulkarni > wrote: >> Hi Robin, >> >> On Mon, Apr 23, 2018 at 11:11 PM, Ganapatrao Kulkarni >> wrote: >>> On Mon, Apr 23, 2018 at 10:07 PM, Robin Murphy wrote: On 19/04/18 18:12, Ganapatrao Kulkarni wrote: > > The performance drop is observed with long hours iperf testing using 40G > cards. This is mainly due to long iterations in finding the free iova > range in 32bit address space. > > In current implementation for 64bit PCI devices, there is always first > attempt to allocate iova from 32bit(SAC preferred over DAC) address > range. Once we run out 32bit range, there is allocation from higher range, > however due to cached32_node optimization it does not suppose to be > painful. cached32_node always points to recently allocated 32-bit node. > When address range is full, it will be pointing to last allocated node > (leaf node), so walking rbtree to find the available range is not > expensive affair. However this optimization does not behave well when > one of the middle node is freed. In that case cached32_node is updated > to point to next iova range. The next iova allocation will consume free > range and again update cached32_node to itself. From now on, walking > over 32-bit range is more expensive. > > This patch adds fix to update cached node to leaf node when there are no > iova free range left, which avoids unnecessary long iterations. The only trouble with this is that "allocation failed" doesn't uniquely mean "space full". Say that after some time the 32-bit space ends up empty except for one page at 0x1000 and one at 0x8000, then somebody tries to allocate 2GB. If we move the cached node down to the leftmost entry when that fails, all subsequent allocation attempts are now going to fail despite the space being 99.% free! I can see a couple of ways to solve that general problem of free space above the cached node getting lost, but neither of them helps with the case where there is genuinely insufficient space (and if anything would make it even slower). In terms of the optimisation you want here, i.e. fail fast when an allocation cannot possibly succeed, the only reliable idea which comes to mind is free-PFN accounting. I might give that a go myself to see how ugly it looks. >> >> For this testing, dual port intel 40G card(XL710) used and both ports >> were connected in loop-back. Ran iperf server and clients on both >> ports(used NAT to route packets out on intended ports).There were 10 >> iperf clients invoked every 60 seconds in loop for hours for each >> port. Initially the performance on both ports is seen close to line >> rate, however after test ran about 4 to 6 hours, the performance >> started dropping to very low (to few hundred Mbps) on both >> connections. >> >> IMO, this is common bug and should happen on any other platforms too >> and needs to be fixed at the earliest. >> Please let me know if you have better way to fix this, i am happy to >> test your patch! > > any update on this issue? >> >>> >>> i see 2 problems in current implementation, >>> 1. We don't replenish the 32 bits range, until first attempt of second >>> allocation(64 bit) fails. >>> 2. Having per cpu cache might not yield good hit on platforms with >>> more number of CPUs. >>> >>> however irrespective of current issues, It makes sense to update >>> cached node as done in this patch , when there is failure to get iova >>> range using current cached pointer which is forcing for the >>> unnecessary time consuming do-while iterations until any replenish >>> happens! >>> >>> thanks >>> Ganapat >>> Robin. > Signed-off-by: Ganapatrao Kulkarni > --- > drivers/iommu/iova.c | 6 ++ > 1 file changed, 6 insertions(+) > > diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c > index 83fe262..e6ee2ea 100644 > --- a/drivers/iommu/iova.c > +++ b/drivers/iommu/iova.c > @@ -201,6 +201,12 @@ static int __alloc_and_insert_iova_range(struct > iova_domain *iovad, > } while (curr && new_pfn <= curr_iova->pfn_hi); > if (limit_pfn < size || new_pfn < iovad->start_pfn) { > + /* No more cached node points to free hole, update to leaf > node. > +*/ > + struct iova *prev_iova; > + > + prev_iova = rb_entry(prev, struct iova, node); > + __cached_rbnode_insert_update(iovad, prev_iova); > spin_unlock_irqrestore(>iova_rbtree_lock, flags); > return -ENOMEM; > } > >> >> thanks >> Ganapat ___ iommu mailing list
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
On Thu, Apr 26, 2018 at 3:15 PM, Ganapatrao Kulkarniwrote: > Hi Robin, > > On Mon, Apr 23, 2018 at 11:11 PM, Ganapatrao Kulkarni > wrote: >> On Mon, Apr 23, 2018 at 10:07 PM, Robin Murphy wrote: >>> On 19/04/18 18:12, Ganapatrao Kulkarni wrote: The performance drop is observed with long hours iperf testing using 40G cards. This is mainly due to long iterations in finding the free iova range in 32bit address space. In current implementation for 64bit PCI devices, there is always first attempt to allocate iova from 32bit(SAC preferred over DAC) address range. Once we run out 32bit range, there is allocation from higher range, however due to cached32_node optimization it does not suppose to be painful. cached32_node always points to recently allocated 32-bit node. When address range is full, it will be pointing to last allocated node (leaf node), so walking rbtree to find the available range is not expensive affair. However this optimization does not behave well when one of the middle node is freed. In that case cached32_node is updated to point to next iova range. The next iova allocation will consume free range and again update cached32_node to itself. From now on, walking over 32-bit range is more expensive. This patch adds fix to update cached node to leaf node when there are no iova free range left, which avoids unnecessary long iterations. >>> >>> >>> The only trouble with this is that "allocation failed" doesn't uniquely mean >>> "space full". Say that after some time the 32-bit space ends up empty except >>> for one page at 0x1000 and one at 0x8000, then somebody tries to >>> allocate 2GB. If we move the cached node down to the leftmost entry when >>> that fails, all subsequent allocation attempts are now going to fail despite >>> the space being 99.% free! >>> >>> I can see a couple of ways to solve that general problem of free space above >>> the cached node getting lost, but neither of them helps with the case where >>> there is genuinely insufficient space (and if anything would make it even >>> slower). In terms of the optimisation you want here, i.e. fail fast when an >>> allocation cannot possibly succeed, the only reliable idea which comes to >>> mind is free-PFN accounting. I might give that a go myself to see how ugly >>> it looks. > > For this testing, dual port intel 40G card(XL710) used and both ports > were connected in loop-back. Ran iperf server and clients on both > ports(used NAT to route packets out on intended ports).There were 10 > iperf clients invoked every 60 seconds in loop for hours for each > port. Initially the performance on both ports is seen close to line > rate, however after test ran about 4 to 6 hours, the performance > started dropping to very low (to few hundred Mbps) on both > connections. > > IMO, this is common bug and should happen on any other platforms too > and needs to be fixed at the earliest. > Please let me know if you have better way to fix this, i am happy to > test your patch! any update on this issue? > >> >> i see 2 problems in current implementation, >> 1. We don't replenish the 32 bits range, until first attempt of second >> allocation(64 bit) fails. >> 2. Having per cpu cache might not yield good hit on platforms with >> more number of CPUs. >> >> however irrespective of current issues, It makes sense to update >> cached node as done in this patch , when there is failure to get iova >> range using current cached pointer which is forcing for the >> unnecessary time consuming do-while iterations until any replenish >> happens! >> >> thanks >> Ganapat >> >>> >>> Robin. >>> >>> Signed-off-by: Ganapatrao Kulkarni --- drivers/iommu/iova.c | 6 ++ 1 file changed, 6 insertions(+) diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c index 83fe262..e6ee2ea 100644 --- a/drivers/iommu/iova.c +++ b/drivers/iommu/iova.c @@ -201,6 +201,12 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad, } while (curr && new_pfn <= curr_iova->pfn_hi); if (limit_pfn < size || new_pfn < iovad->start_pfn) { + /* No more cached node points to free hole, update to leaf node. +*/ + struct iova *prev_iova; + + prev_iova = rb_entry(prev, struct iova, node); + __cached_rbnode_insert_update(iovad, prev_iova); spin_unlock_irqrestore(>iova_rbtree_lock, flags); return -ENOMEM; } >>> > > thanks > Ganapat ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
Hi Robin, On Mon, Apr 23, 2018 at 11:11 PM, Ganapatrao Kulkarniwrote: > On Mon, Apr 23, 2018 at 10:07 PM, Robin Murphy wrote: >> On 19/04/18 18:12, Ganapatrao Kulkarni wrote: >>> >>> The performance drop is observed with long hours iperf testing using 40G >>> cards. This is mainly due to long iterations in finding the free iova >>> range in 32bit address space. >>> >>> In current implementation for 64bit PCI devices, there is always first >>> attempt to allocate iova from 32bit(SAC preferred over DAC) address >>> range. Once we run out 32bit range, there is allocation from higher range, >>> however due to cached32_node optimization it does not suppose to be >>> painful. cached32_node always points to recently allocated 32-bit node. >>> When address range is full, it will be pointing to last allocated node >>> (leaf node), so walking rbtree to find the available range is not >>> expensive affair. However this optimization does not behave well when >>> one of the middle node is freed. In that case cached32_node is updated >>> to point to next iova range. The next iova allocation will consume free >>> range and again update cached32_node to itself. From now on, walking >>> over 32-bit range is more expensive. >>> >>> This patch adds fix to update cached node to leaf node when there are no >>> iova free range left, which avoids unnecessary long iterations. >> >> >> The only trouble with this is that "allocation failed" doesn't uniquely mean >> "space full". Say that after some time the 32-bit space ends up empty except >> for one page at 0x1000 and one at 0x8000, then somebody tries to >> allocate 2GB. If we move the cached node down to the leftmost entry when >> that fails, all subsequent allocation attempts are now going to fail despite >> the space being 99.% free! >> >> I can see a couple of ways to solve that general problem of free space above >> the cached node getting lost, but neither of them helps with the case where >> there is genuinely insufficient space (and if anything would make it even >> slower). In terms of the optimisation you want here, i.e. fail fast when an >> allocation cannot possibly succeed, the only reliable idea which comes to >> mind is free-PFN accounting. I might give that a go myself to see how ugly >> it looks. For this testing, dual port intel 40G card(XL710) used and both ports were connected in loop-back. Ran iperf server and clients on both ports(used NAT to route packets out on intended ports).There were 10 iperf clients invoked every 60 seconds in loop for hours for each port. Initially the performance on both ports is seen close to line rate, however after test ran about 4 to 6 hours, the performance started dropping to very low (to few hundred Mbps) on both connections. IMO, this is common bug and should happen on any other platforms too and needs to be fixed at the earliest. Please let me know if you have better way to fix this, i am happy to test your patch! > > i see 2 problems in current implementation, > 1. We don't replenish the 32 bits range, until first attempt of second > allocation(64 bit) fails. > 2. Having per cpu cache might not yield good hit on platforms with > more number of CPUs. > > however irrespective of current issues, It makes sense to update > cached node as done in this patch , when there is failure to get iova > range using current cached pointer which is forcing for the > unnecessary time consuming do-while iterations until any replenish > happens! > > thanks > Ganapat > >> >> Robin. >> >> >>> Signed-off-by: Ganapatrao Kulkarni >>> --- >>> drivers/iommu/iova.c | 6 ++ >>> 1 file changed, 6 insertions(+) >>> >>> diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c >>> index 83fe262..e6ee2ea 100644 >>> --- a/drivers/iommu/iova.c >>> +++ b/drivers/iommu/iova.c >>> @@ -201,6 +201,12 @@ static int __alloc_and_insert_iova_range(struct >>> iova_domain *iovad, >>> } while (curr && new_pfn <= curr_iova->pfn_hi); >>> if (limit_pfn < size || new_pfn < iovad->start_pfn) { >>> + /* No more cached node points to free hole, update to leaf >>> node. >>> +*/ >>> + struct iova *prev_iova; >>> + >>> + prev_iova = rb_entry(prev, struct iova, node); >>> + __cached_rbnode_insert_update(iovad, prev_iova); >>> spin_unlock_irqrestore(>iova_rbtree_lock, flags); >>> return -ENOMEM; >>> } >>> >> thanks Ganapat ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
On Mon, Apr 23, 2018 at 10:07 PM, Robin Murphywrote: > On 19/04/18 18:12, Ganapatrao Kulkarni wrote: >> >> The performance drop is observed with long hours iperf testing using 40G >> cards. This is mainly due to long iterations in finding the free iova >> range in 32bit address space. >> >> In current implementation for 64bit PCI devices, there is always first >> attempt to allocate iova from 32bit(SAC preferred over DAC) address >> range. Once we run out 32bit range, there is allocation from higher range, >> however due to cached32_node optimization it does not suppose to be >> painful. cached32_node always points to recently allocated 32-bit node. >> When address range is full, it will be pointing to last allocated node >> (leaf node), so walking rbtree to find the available range is not >> expensive affair. However this optimization does not behave well when >> one of the middle node is freed. In that case cached32_node is updated >> to point to next iova range. The next iova allocation will consume free >> range and again update cached32_node to itself. From now on, walking >> over 32-bit range is more expensive. >> >> This patch adds fix to update cached node to leaf node when there are no >> iova free range left, which avoids unnecessary long iterations. > > > The only trouble with this is that "allocation failed" doesn't uniquely mean > "space full". Say that after some time the 32-bit space ends up empty except > for one page at 0x1000 and one at 0x8000, then somebody tries to > allocate 2GB. If we move the cached node down to the leftmost entry when > that fails, all subsequent allocation attempts are now going to fail despite > the space being 99.% free! > > I can see a couple of ways to solve that general problem of free space above > the cached node getting lost, but neither of them helps with the case where > there is genuinely insufficient space (and if anything would make it even > slower). In terms of the optimisation you want here, i.e. fail fast when an > allocation cannot possibly succeed, the only reliable idea which comes to > mind is free-PFN accounting. I might give that a go myself to see how ugly > it looks. i see 2 problems in current implementation, 1. We don't replenish the 32 bits range, until first attempt of second allocation(64 bit) fails. 2. Having per cpu cache might not yield good hit on platforms with more number of CPUs. however irrespective of current issues, It makes sense to update cached node as done in this patch , when there is failure to get iova range using current cached pointer which is forcing for the unnecessary time consuming do-while iterations until any replenish happens! thanks Ganapat > > Robin. > > >> Signed-off-by: Ganapatrao Kulkarni >> --- >> drivers/iommu/iova.c | 6 ++ >> 1 file changed, 6 insertions(+) >> >> diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c >> index 83fe262..e6ee2ea 100644 >> --- a/drivers/iommu/iova.c >> +++ b/drivers/iommu/iova.c >> @@ -201,6 +201,12 @@ static int __alloc_and_insert_iova_range(struct >> iova_domain *iovad, >> } while (curr && new_pfn <= curr_iova->pfn_hi); >> if (limit_pfn < size || new_pfn < iovad->start_pfn) { >> + /* No more cached node points to free hole, update to leaf >> node. >> +*/ >> + struct iova *prev_iova; >> + >> + prev_iova = rb_entry(prev, struct iova, node); >> + __cached_rbnode_insert_update(iovad, prev_iova); >> spin_unlock_irqrestore(>iova_rbtree_lock, flags); >> return -ENOMEM; >> } >> > ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
On 19/04/18 18:12, Ganapatrao Kulkarni wrote: The performance drop is observed with long hours iperf testing using 40G cards. This is mainly due to long iterations in finding the free iova range in 32bit address space. In current implementation for 64bit PCI devices, there is always first attempt to allocate iova from 32bit(SAC preferred over DAC) address range. Once we run out 32bit range, there is allocation from higher range, however due to cached32_node optimization it does not suppose to be painful. cached32_node always points to recently allocated 32-bit node. When address range is full, it will be pointing to last allocated node (leaf node), so walking rbtree to find the available range is not expensive affair. However this optimization does not behave well when one of the middle node is freed. In that case cached32_node is updated to point to next iova range. The next iova allocation will consume free range and again update cached32_node to itself. From now on, walking over 32-bit range is more expensive. This patch adds fix to update cached node to leaf node when there are no iova free range left, which avoids unnecessary long iterations. The only trouble with this is that "allocation failed" doesn't uniquely mean "space full". Say that after some time the 32-bit space ends up empty except for one page at 0x1000 and one at 0x8000, then somebody tries to allocate 2GB. If we move the cached node down to the leftmost entry when that fails, all subsequent allocation attempts are now going to fail despite the space being 99.% free! I can see a couple of ways to solve that general problem of free space above the cached node getting lost, but neither of them helps with the case where there is genuinely insufficient space (and if anything would make it even slower). In terms of the optimisation you want here, i.e. fail fast when an allocation cannot possibly succeed, the only reliable idea which comes to mind is free-PFN accounting. I might give that a go myself to see how ugly it looks. Robin. Signed-off-by: Ganapatrao Kulkarni--- drivers/iommu/iova.c | 6 ++ 1 file changed, 6 insertions(+) diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c index 83fe262..e6ee2ea 100644 --- a/drivers/iommu/iova.c +++ b/drivers/iommu/iova.c @@ -201,6 +201,12 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad, } while (curr && new_pfn <= curr_iova->pfn_hi); if (limit_pfn < size || new_pfn < iovad->start_pfn) { + /* No more cached node points to free hole, update to leaf node. +*/ + struct iova *prev_iova; + + prev_iova = rb_entry(prev, struct iova, node); + __cached_rbnode_insert_update(iovad, prev_iova); spin_unlock_irqrestore(>iova_rbtree_lock, flags); return -ENOMEM; } ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
[PATCH] iommu/iova: Update cached node pointer when current node fails to get any free IOVA
The performance drop is observed with long hours iperf testing using 40G cards. This is mainly due to long iterations in finding the free iova range in 32bit address space. In current implementation for 64bit PCI devices, there is always first attempt to allocate iova from 32bit(SAC preferred over DAC) address range. Once we run out 32bit range, there is allocation from higher range, however due to cached32_node optimization it does not suppose to be painful. cached32_node always points to recently allocated 32-bit node. When address range is full, it will be pointing to last allocated node (leaf node), so walking rbtree to find the available range is not expensive affair. However this optimization does not behave well when one of the middle node is freed. In that case cached32_node is updated to point to next iova range. The next iova allocation will consume free range and again update cached32_node to itself. From now on, walking over 32-bit range is more expensive. This patch adds fix to update cached node to leaf node when there are no iova free range left, which avoids unnecessary long iterations. Signed-off-by: Ganapatrao Kulkarni--- drivers/iommu/iova.c | 6 ++ 1 file changed, 6 insertions(+) diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c index 83fe262..e6ee2ea 100644 --- a/drivers/iommu/iova.c +++ b/drivers/iommu/iova.c @@ -201,6 +201,12 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad, } while (curr && new_pfn <= curr_iova->pfn_hi); if (limit_pfn < size || new_pfn < iovad->start_pfn) { + /* No more cached node points to free hole, update to leaf node. +*/ + struct iova *prev_iova; + + prev_iova = rb_entry(prev, struct iova, node); + __cached_rbnode_insert_update(iovad, prev_iova); spin_unlock_irqrestore(>iova_rbtree_lock, flags); return -ENOMEM; } -- 2.9.4 ___ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu