On Tue, Sep 26, 2023 at 06:23:39PM +0200, Petr Tesarik wrote:
> When CONFIG_SWIOTLB_DYNAMIC=y, devices which do not use the software IO TLB
> can avoid swiotlb lookup. A flag is added by commit 1395706a1490 ("swiotlb:
> search the software IO TLB only if the device makes use of it"), the flag
> is correctly set, but it is then never checked. Add the actual check here.
>
> Note that this code is an alternative to the default pool check, not an
> additional check, because:
>
> 1. swiotlb_find_pool() also searches the default pool;
> 2. if dma_uses_io_tlb is false, the default swiotlb pool is not used.
>
> Tested in a KVM guest against a QEMU RAM-backed SATA disk over virtio and
> *not* using software IO TLB, this patch increases IOPS by approx 2% for
> 4-way parallel I/O.
>
> The write memory barrier in swiotlb_dyn_alloc() is not needed, because a
> newly allocated pool must always be observed by swiotlb_find_slots() before
> an address from that pool is passed to is_swiotlb_buffer().
>
> Correctness was verified using the following litmus test:
[...]
> Fixes: 1395706a1490 ("swiotlb: search the software IO TLB only if the device makes use of it")
> Reported-by: Jonathan Corbet <[email protected]>
> Closes: https://lore.kernel.org/linux-iommu/[email protected]/
> Signed-off-by: Petr Tesarik <[email protected]>
Thanks for the update.
Reviewed-by: Catalin Marinas <[email protected]>
> diff --git a/include/linux/swiotlb.h b/include/linux/swiotlb.h
> index b4536626f8ff..93b400d9be91 100644
> --- a/include/linux/swiotlb.h
> +++ b/include/linux/swiotlb.h
> @@ -172,14 +172,22 @@ static inline bool is_swiotlb_buffer(struct device *dev, phys_addr_t paddr)
> if (!mem)
> return false;
>
> - if (IS_ENABLED(CONFIG_SWIOTLB_DYNAMIC)) {
> - /* Pairs with smp_wmb() in swiotlb_find_slots() and
> - * swiotlb_dyn_alloc(), which modify the RCU lists.
> - */
> - smp_rmb();
> - return swiotlb_find_pool(dev, paddr);
> - }
> +#ifdef CONFIG_SWIOTLB_DYNAMIC
> + /* All SWIOTLB buffer addresses must have been returned by
> + * swiotlb_tbl_map_single() and passed to a device driver.
> + * If a SWIOTLB address is checked on another CPU, then it was
> + * presumably loaded by the device driver from an unspecified private
> + * data structure. Make sure that this load is ordered before reading
> + * dev->dma_uses_io_tlb here and mem->pools in swiotlb_find_pool().
> + *
> + * This barrier pairs with smp_mb() in swiotlb_find_slots().
> + */
Nitpick. The official multi-line comment style is:
/*
* Text.
*/
i.e. it starts with an empty /* line.
> @@ -1152,9 +1149,25 @@ static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr,
> spin_unlock_irqrestore(&dev->dma_io_tlb_lock, flags);
>
> found:
> - dev->dma_uses_io_tlb = true;
> - /* Pairs with smp_rmb() in is_swiotlb_buffer() */
> - smp_wmb();
> + WRITE_ONCE(dev->dma_uses_io_tlb, true);
> +
> + /* The general barrier orders reads and writes against a presumed store
> + * of the SWIOTLB buffer address by a device driver (to a driver private
> + * data structure). It serves two purposes.
Same here.
--
Catalin
On Tue, 26 Sep 2023 19:17:29 +0100
Catalin Marinas <[email protected]> wrote:
> On Tue, Sep 26, 2023 at 06:23:39PM +0200, Petr Tesarik wrote:
> > When CONFIG_SWIOTLB_DYNAMIC=y, devices which do not use the software IO TLB
> > can avoid swiotlb lookup. A flag is added by commit 1395706a1490 ("swiotlb:
> > search the software IO TLB only if the device makes use of it"), the flag
> > is correctly set, but it is then never checked. Add the actual check here.
> >
> > Note that this code is an alternative to the default pool check, not an
> > additional check, because:
> >
> > 1. swiotlb_find_pool() also searches the default pool;
> > 2. if dma_uses_io_tlb is false, the default swiotlb pool is not used.
> >
> > Tested in a KVM guest against a QEMU RAM-backed SATA disk over virtio and
> > *not* using software IO TLB, this patch increases IOPS by approx 2% for
> > 4-way parallel I/O.
> >
> > The write memory barrier in swiotlb_dyn_alloc() is not needed, because a
> > newly allocated pool must always be observed by swiotlb_find_slots() before
> > an address from that pool is passed to is_swiotlb_buffer().
> >
> > Correctness was verified using the following litmus test:
> [...]
> > Fixes: 1395706a1490 ("swiotlb: search the software IO TLB only if the device makes use of it")
> > Reported-by: Jonathan Corbet <[email protected]>
> > Closes: https://lore.kernel.org/linux-iommu/[email protected]/
> > Signed-off-by: Petr Tesarik <[email protected]>
>
> Thanks for the update.
>
> Reviewed-by: Catalin Marinas <[email protected]>
>
> > diff --git a/include/linux/swiotlb.h b/include/linux/swiotlb.h
> > index b4536626f8ff..93b400d9be91 100644
> > --- a/include/linux/swiotlb.h
> > +++ b/include/linux/swiotlb.h
> > @@ -172,14 +172,22 @@ static inline bool is_swiotlb_buffer(struct device *dev, phys_addr_t paddr)
> > if (!mem)
> > return false;
> >
> > - if (IS_ENABLED(CONFIG_SWIOTLB_DYNAMIC)) {
> > - /* Pairs with smp_wmb() in swiotlb_find_slots() and
> > - * swiotlb_dyn_alloc(), which modify the RCU lists.
> > - */
> > - smp_rmb();
> > - return swiotlb_find_pool(dev, paddr);
> > - }
> > +#ifdef CONFIG_SWIOTLB_DYNAMIC
> > + /* All SWIOTLB buffer addresses must have been returned by
> > + * swiotlb_tbl_map_single() and passed to a device driver.
> > + * If a SWIOTLB address is checked on another CPU, then it was
> > + * presumably loaded by the device driver from an unspecified private
> > + * data structure. Make sure that this load is ordered before reading
> > + * dev->dma_uses_io_tlb here and mem->pools in swiotlb_find_pool().
> > + *
> > + * This barrier pairs with smp_mb() in swiotlb_find_slots().
> > + */
>
> Nitpick. The official multi-line comment style is:
>
> /*
> * Text.
> */
>
> i.e. it starts with an empty /* line.
Right! I should add it to scripts/checkpatch.pl.
> > @@ -1152,9 +1149,25 @@ static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr,
> > spin_unlock_irqrestore(&dev->dma_io_tlb_lock, flags);
> >
> > found:
> > - dev->dma_uses_io_tlb = true;
> > - /* Pairs with smp_rmb() in is_swiotlb_buffer() */
> > - smp_wmb();
> > + WRITE_ONCE(dev->dma_uses_io_tlb, true);
> > +
> > + /* The general barrier orders reads and writes against a presumed store
> > + * of the SWIOTLB buffer address by a device driver (to a driver private
> > + * data structure). It serves two purposes.
>
> Same here.
If that's the only issue... ;-)
Petr T