| Message ID | 20210818120642.165-2-xieyongji@bytedance.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | Introduce VDUSE - vDPA Device in Userspace | expand |
| Context | Check | Description |
|---|---|---|
| netdev/tree_selection | success | Not a local patch |
在 2021/8/18 下午8:06, Xie Yongji 写道: > Export alloc_iova_fast() and free_iova_fast() so that > some modules can make use of the per-CPU cache to get > rid of rbtree spinlock in alloc_iova() and free_iova() > during IOVA allocation. > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com> Acked-by: Jason Wang <jasowang@redhat.com> (If we need respin, I'd suggest to put the numbers you measured here) Thanks > --- > drivers/iommu/iova.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c > index b6cf5f16123b..3941ed6bb99b 100644 > --- a/drivers/iommu/iova.c > +++ b/drivers/iommu/iova.c > @@ -521,6 +521,7 @@ alloc_iova_fast(struct iova_domain *iovad, unsigned long size, > > return new_iova->pfn_lo; > } > +EXPORT_SYMBOL_GPL(alloc_iova_fast); > > /** > * free_iova_fast - free iova pfn range into rcache > @@ -538,6 +539,7 @@ free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size) > > free_iova(iovad, pfn); > } > +EXPORT_SYMBOL_GPL(free_iova_fast); > > #define fq_ring_for_each(i, fq) \ > for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
On Wed, Aug 18, 2021 at 08:06:31PM +0800, Xie Yongji wrote: > Export alloc_iova_fast() and free_iova_fast() so that > some modules can make use of the per-CPU cache to get > rid of rbtree spinlock in alloc_iova() and free_iova() > during IOVA allocation. > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com> This needs ack from iommu maintainers. Guys? > --- > drivers/iommu/iova.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c > index b6cf5f16123b..3941ed6bb99b 100644 > --- a/drivers/iommu/iova.c > +++ b/drivers/iommu/iova.c > @@ -521,6 +521,7 @@ alloc_iova_fast(struct iova_domain *iovad, unsigned long size, > > return new_iova->pfn_lo; > } > +EXPORT_SYMBOL_GPL(alloc_iova_fast); > > /** > * free_iova_fast - free iova pfn range into rcache > @@ -538,6 +539,7 @@ free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size) > > free_iova(iovad, pfn); > } > +EXPORT_SYMBOL_GPL(free_iova_fast); > > #define fq_ring_for_each(i, fq) \ > for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE) > -- > 2.11.0
On Tue, Aug 24, 2021 at 02:08:33PM -0400, Michael S. Tsirkin wrote: > On Wed, Aug 18, 2021 at 08:06:31PM +0800, Xie Yongji wrote: > > Export alloc_iova_fast() and free_iova_fast() so that > > some modules can make use of the per-CPU cache to get > > rid of rbtree spinlock in alloc_iova() and free_iova() > > during IOVA allocation. > > > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com> > > > This needs ack from iommu maintainers. Guys? Looks fine to me: Acked-by: Will Deacon <will@kernel.org> Will
On 25/08/2021 10:55, Will Deacon wrote: > On Tue, Aug 24, 2021 at 02:08:33PM -0400, Michael S. Tsirkin wrote: >> On Wed, Aug 18, 2021 at 08:06:31PM +0800, Xie Yongji wrote: >>> Export alloc_iova_fast() and free_iova_fast() so that >>> some modules can make use of the per-CPU cache to get >>> rid of rbtree spinlock in alloc_iova() and free_iova() >>> during IOVA allocation. >>> >>> Signed-off-by: Xie Yongji <xieyongji@bytedance.com> >> >> >> This needs ack from iommu maintainers. Guys? > > Looks fine to me: > > Acked-by: Will Deacon <will@kernel.org> > > Will > _______________________________________________ > iommu mailing list > iommu@lists.linux-foundation.org > https://lists.linuxfoundation.org/mailman/listinfo/iommu > . > JFYI, There was a preliminary discussion to move the iova rcache code (which the iova fast alloc and free functions are based on) out of the iova code and maybe into dma-iommu (being the only user). There was other motivation. https://lore.kernel.org/linux-iommu/83de3911-145d-77c8-17c1-981e4ff825d3@arm.com/ Having more users complicates that... Thanks, John
On Wed, Aug 25, 2021 at 6:35 PM John Garry <john.garry@huawei.com> wrote: > > On 25/08/2021 10:55, Will Deacon wrote: > > On Tue, Aug 24, 2021 at 02:08:33PM -0400, Michael S. Tsirkin wrote: > >> On Wed, Aug 18, 2021 at 08:06:31PM +0800, Xie Yongji wrote: > >>> Export alloc_iova_fast() and free_iova_fast() so that > >>> some modules can make use of the per-CPU cache to get > >>> rid of rbtree spinlock in alloc_iova() and free_iova() > >>> during IOVA allocation. > >>> > >>> Signed-off-by: Xie Yongji <xieyongji@bytedance.com> > >> > >> > >> This needs ack from iommu maintainers. Guys? > > > > Looks fine to me: > > > > Acked-by: Will Deacon <will@kernel.org> > > > > Will > > _______________________________________________ > > iommu mailing list > > iommu@lists.linux-foundation.org > > https://lists.linuxfoundation.org/mailman/listinfo/iommu > > . > > > > JFYI, There was a preliminary discussion to move the iova rcache code > (which the iova fast alloc and free functions are based on) out of the > iova code and maybe into dma-iommu (being the only user). There was > other motivation. > Would it be better to move the code into ./lib as a general library? > https://lore.kernel.org/linux-iommu/83de3911-145d-77c8-17c1-981e4ff825d3@arm.com/ > > Having more users complicates that... > Do we have some plan for this work? From our test [1], iova_alloc_fast() is much better than iova_alloc(). So I'd like to use it as much as possible [1] https://lore.kernel.org/kvm/CACycT3steXFeg7NRbWpo2J59dpYcumzcvM2zcPJAVe40-EvvEg@mail.gmail.com/ Thanks, Yongji
On 25/08/2021 13:17, Yongji Xie wrote: >> JFYI, There was a preliminary discussion to move the iova rcache code >> (which the iova fast alloc and free functions are based on) out of the >> iova code and maybe into dma-iommu (being the only user). There was >> other motivation. >> > Would it be better to move the code into ./lib as a general library? For a start we/I think that the rcache could be removed from the IOVA code, but prob should stay in drivers/iommu. I had another IOVA issue to solve, which complicates things. No solid plans. Need to talk to Robin more. > >> https://lore.kernel.org/linux-iommu/83de3911-145d-77c8-17c1-981e4ff825d3@arm.com/ >> >> Having more users complicates that... >> > Do we have some plan for this work? From our test [1], > iova_alloc_fast() is much better than iova_alloc(). So I'd like to use > it as much as possible > > [1]https://lore.kernel.org/kvm/CACycT3steXFeg7NRbWpo2J59dpYcumzcvM2zcPJAVe40-EvvEg@mail.gmail.com/ Well if you're alloc'ing and free'ing IOVAs a lot then I can imagine it is. Thanks, John
diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c index b6cf5f16123b..3941ed6bb99b 100644 --- a/drivers/iommu/iova.c +++ b/drivers/iommu/iova.c @@ -521,6 +521,7 @@ alloc_iova_fast(struct iova_domain *iovad, unsigned long size, return new_iova->pfn_lo; } +EXPORT_SYMBOL_GPL(alloc_iova_fast); /** * free_iova_fast - free iova pfn range into rcache @@ -538,6 +539,7 @@ free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size) free_iova(iovad, pfn); } +EXPORT_SYMBOL_GPL(free_iova_fast); #define fq_ring_for_each(i, fq) \ for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
Export alloc_iova_fast() and free_iova_fast() so that some modules can make use of the per-CPU cache to get rid of rbtree spinlock in alloc_iova() and free_iova() during IOVA allocation. Signed-off-by: Xie Yongji <xieyongji@bytedance.com> --- drivers/iommu/iova.c | 2 ++ 1 file changed, 2 insertions(+)