Re: [PATCH v6 19/19] swiotlb-xen: instead of bouncing on the swiotlb,pin single pages
From: Stefano Stabellini
Date: Wed Oct 09 2013 - 13:28:42 EST
On Mon, 30 Sep 2013, Konrad Rzeszutek Wilk wrote:
> On Fri, Sep 27, 2013 at 05:10:07PM +0100, Stefano Stabellini wrote:
> > If we are dealing with single page mappings that don't cross page
> > boundaries, we can try to pin the page and get the corresponding mfn,
> > using xen_pin_page. This avoids going through the swiotlb bounce
> > buffer. If xen_pin_page fails (because the underlying mfn doesn't
> > respect the dma_mask) fall back to the swiotlb bounce buffer.
> > Add a ref count to xen_dma_info, so that we can avoid pinnig pages that
> > are already pinned.
> > Use a spinlock to protect accesses, insertions and deletions in the
> > rbtrees.
> >
> > Signed-off-by: Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>
Thanks for the review, however I am dropping this patch because it
wasn't improving performances as I was hoping it would be.
I am taking a new approach now: I am keeping the 1:1 physical to machine
mapping for dom0 and using swiotlb-xen only to handle dma requests
involving foreign grants.
The code is much nicer, and it runs much faster.
> > drivers/xen/swiotlb-xen.c | 152 ++++++++++++++++++++++++++++++++++++++++++---
> > 1 files changed, 143 insertions(+), 9 deletions(-)
> >
> > diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
> > index 022bcaf..6f94285 100644
> > --- a/drivers/xen/swiotlb-xen.c
> > +++ b/drivers/xen/swiotlb-xen.c
> > @@ -57,6 +57,8 @@
> > #define NR_DMA_SEGS ((xen_io_tlb_nslabs + IO_TLB_SEGSIZE - 1) / IO_TLB_SEGSIZE)
> > static char *xen_io_tlb_start, *xen_io_tlb_end;
> > static unsigned long xen_io_tlb_nslabs;
> > +spinlock_t swiotlb_lock;
> > +
> > /*
> > * Quick lookup value of the bus address of the IOTLB.
> > */
> > @@ -79,6 +81,7 @@ struct xen_dma_info {
> > dma_addr_t dma_addr;
> > phys_addr_t phys_addr;
> > size_t size;
> > + atomic_t refs;
> > struct rb_node rbnode_dma;
> > struct rb_node rbnode_phys;
> > };
> > @@ -254,6 +257,48 @@ static dma_addr_t xen_virt_to_bus(void *address)
> > return xen_phys_to_bus_quick(virt_to_phys(address));
> > }
> >
> > +static int xen_pin_dev_page(struct device *dev,
> > + phys_addr_t phys,
> > + dma_addr_t *dev_addr)
>
> Something is odd with your tabs.
> > +{
> > + u64 dma_mask = DMA_BIT_MASK(32);
>
> Why 32?
>
> > + xen_pfn_t in;
> > + struct xen_dma_info *dma_info = xen_get_dma_info_from_phys(phys);
> > +
> > + if (dma_info != NULL) {
> > + atomic_inc(&dma_info->refs);
> > + *dev_addr = dma_info->dma_addr + (phys - dma_info->phys_addr);
> > + return 0;
> > + }
> > +
> > + if (dev && dev->coherent_dma_mask)
> > + dma_mask = dma_alloc_coherent_mask(dev, GFP_KERNEL);
> > +
> > + in = phys >> PAGE_SHIFT;
> > + if (!xen_pin_page(&in, fls64(dma_mask))) {
>
> Why not just make xen_pin_page use an phys address and it can also
> do the appropiate bit shifting in it?
>
> > + *dev_addr = in << PAGE_SHIFT;
> > + dma_info = kzalloc(sizeof(struct xen_dma_info), GFP_NOWAIT);
> > + if (!dma_info) {
> > + pr_warn("cannot allocate xen_dma_info\n");
> > + xen_destroy_contiguous_region(phys & PAGE_MASK, 0);
>
> Perhaps we should add an inline function for that called 'xen_unpin_page' ?
>
> > + return -ENOMEM;
> > + }
> > + dma_info->phys_addr = phys & PAGE_MASK;
> > + dma_info->size = PAGE_SIZE;
> > + dma_info->dma_addr = *dev_addr;
> > + if (xen_dma_add_entry(dma_info)) {
> > + pr_warn("cannot add new entry to bus_to_phys\n");
> > + xen_destroy_contiguous_region(phys & PAGE_MASK, 0);
> > + kfree(dma_info);
> > + return -EFAULT;
> > + }
> > + atomic_set(&dma_info->refs, 1);
> > + *dev_addr += (phys & ~PAGE_MASK);
> > + return 0;
> > + }
>
> Don't you want to the opposite of dma_alloc_coherent_mask ?
>
> > + return -EFAULT;
> > +}
> > +
> > static int check_pages_physically_contiguous(unsigned long pfn,
> > unsigned int offset,
> > size_t length)
> > @@ -434,6 +479,7 @@ retry:
> > rc = 0;
> > } else
> > rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
> > + spin_lock_init(&swiotlb_lock);
> > return rc;
> > error:
> > if (repeat--) {
> > @@ -461,6 +507,7 @@ xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
> > phys_addr_t phys;
> > dma_addr_t dev_addr;
> > struct xen_dma_info *dma_info = NULL;
> > + unsigned long irqflags;
> >
> > /*
> > * Ignore region specifiers - the kernel's ideas of
> > @@ -497,7 +544,7 @@ xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
> > !range_straddles_page_boundary(phys, size))
> > *dma_handle = dev_addr;
> > else {
> > - if (xen_create_contiguous_region(phys, order,
> > + if (xen_create_contiguous_region(phys & PAGE_MASK, order,
> > fls64(dma_mask), dma_handle) != 0) {
> > xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
> > return NULL;
> > @@ -509,15 +556,19 @@ xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
> > xen_destroy_contiguous_region(phys, order);
> > return NULL;
> > }
> > - dma_info->phys_addr = phys;
> > - dma_info->size = size;
> > + dma_info->phys_addr = phys & PAGE_MASK;
> > + dma_info->size = (1U << order) << PAGE_SHIFT;
> > dma_info->dma_addr = *dma_handle;
> > + atomic_set(&dma_info->refs, 1);
> > + spin_lock_irqsave(&swiotlb_lock, irqflags);
> > if (xen_dma_add_entry(dma_info)) {
> > + spin_unlock_irqrestore(&swiotlb_lock, irqflags);
> > pr_warn("cannot add new entry to bus_to_phys\n");
> > xen_destroy_contiguous_region(phys, order);
> > kfree(dma_info);
> > return NULL;
> > }
> > + spin_unlock_irqrestore(&swiotlb_lock, irqflags);
> > }
> > memset(ret, 0, size);
> > return ret;
> > @@ -532,6 +583,7 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
> > phys_addr_t phys;
> > u64 dma_mask = DMA_BIT_MASK(32);
> > struct xen_dma_info *dma_info = NULL;
> > + unsigned long flags;
> >
> > if (dma_release_from_coherent(hwdev, order, vaddr))
> > return;
> > @@ -539,6 +591,7 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
> > if (hwdev && hwdev->coherent_dma_mask)
> > dma_mask = hwdev->coherent_dma_mask;
> >
> > + spin_lock_irqsave(&swiotlb_lock, flags);
> > /* do not use virt_to_phys because on ARM it doesn't return you the
> > * physical address */
> > phys = xen_bus_to_phys(dev_addr);
> > @@ -546,12 +599,16 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
> > if (xen_feature(XENFEAT_auto_translated_physmap) ||
> > (((dev_addr + size - 1 > dma_mask)) ||
> > range_straddles_page_boundary(phys, size))) {
> > - xen_destroy_contiguous_region(phys, order);
> > dma_info = xen_get_dma_info_from_dma(dev_addr);
> > - rb_erase(&dma_info->rbnode, &bus_to_phys);
> > - kfree(dma_info);
> > + if (atomic_dec_and_test(&dma_info->refs)) {
> > + xen_destroy_contiguous_region(phys & PAGE_MASK, order);
> > + rb_erase(&dma_info->rbnode_dma, &bus_to_phys);
> > + rb_erase(&dma_info->rbnode_phys, &phys_to_bus);
> > + kfree(dma_info);
> > + }
>
> If xen_pin_dev_page failed or was not called we would still end up
> calling this. And we would decrement a potentially garbage value? Or not.
> > }
> >
> > + spin_unlock_irqrestore(&swiotlb_lock, flags);
> > xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
> > }
> > EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
> > @@ -583,6 +640,23 @@ dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
> > !range_straddles_page_boundary(phys, size) && !swiotlb_force)
> > return dev_addr;
> >
> > + if (xen_feature(XENFEAT_auto_translated_physmap) &&
> > + size <= PAGE_SIZE &&
> > + !range_straddles_page_boundary(phys, size) &&
> > + !swiotlb_force) {
> > + unsigned long flags;
> > + int rc;
> > +
> > + spin_lock_irqsave(&swiotlb_lock, flags);
> > + rc = xen_pin_dev_page(dev, phys, &dev_addr);
> > + spin_unlock_irqrestore(&swiotlb_lock, flags);
> > +
> > + if (!rc) {
> > + dma_mark_clean(phys_to_virt(phys), size);
> > + return dev_addr;
> > + }
>
> And if there is an rc you should probably do
> dev_warn(.., "RC ..")
>
>
> But more importantly - all of this code adds an extra lock on the X86 side
> which will get -ENOxxx on the xen_pin_dev_page.
>
> I am wondering if it makes sense to make most of this code dependent
> on CONFIG_ARM? As the check for auto-xlat falls flat on X86 + PVH. Thought
> I have no idea what we want to do with PVH and X86 at this point.
>
> > + }
> > +
> > /*
> > * Oh well, have to allocate and map a bounce buffer.
> > * Pass the dma_addr of the first slab in the iotlb buffer as
> > @@ -618,10 +692,37 @@ EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
> > static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
> > size_t size, enum dma_data_direction dir)
> > {
> > - phys_addr_t paddr = xen_bus_to_phys(dev_addr);
> > + struct xen_dma_info *dma_info;
> > + phys_addr_t paddr = DMA_ERROR_CODE;
> > + char *vaddr = NULL;
> > + unsigned long flags;
> >
> > BUG_ON(dir == DMA_NONE);
> >
> > + spin_lock_irqsave(&swiotlb_lock, flags);
> > + dma_info = xen_get_dma_info_from_dma(dev_addr);
> > + if (dma_info != NULL) {
> > + paddr = dma_info->phys_addr + (dev_addr - dma_info->dma_addr);
> > + vaddr = phys_to_virt(paddr);
> > + }
> > +
> > + if (xen_feature(XENFEAT_auto_translated_physmap) &&
> > + paddr != DMA_ERROR_CODE &&
> > + !(vaddr >= xen_io_tlb_start && vaddr < xen_io_tlb_end) &&
> > + !swiotlb_force) {
> > + if (atomic_dec_and_test(&dma_info->refs)) {
> > + xen_destroy_contiguous_region(paddr & PAGE_MASK, 0);
> > + rb_erase(&dma_info->rbnode_dma, &bus_to_phys);
> > + rb_erase(&dma_info->rbnode_phys, &phys_to_bus);
> > + kfree(dma_info);
> > + }
> > + spin_unlock_irqrestore(&swiotlb_lock, flags);
> > + if ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL))
> > + dma_mark_clean(vaddr, size);
> > + return;
> > + }
> > + spin_unlock_irqrestore(&swiotlb_lock, flags);
> > +
> > /* NOTE: We use dev_addr here, not paddr! */
> > if (is_xen_swiotlb_buffer(dev_addr)) {
> > swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
> > @@ -664,9 +765,19 @@ xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
> > enum dma_sync_target target)
> > {
> > phys_addr_t paddr = xen_bus_to_phys(dev_addr);
> > + char *vaddr = phys_to_virt(paddr);
> >
> > BUG_ON(dir == DMA_NONE);
> >
> > + if (xen_feature(XENFEAT_auto_translated_physmap) &&
> > + paddr != DMA_ERROR_CODE &&
> > + size <= PAGE_SIZE &&
> > + !(vaddr >= xen_io_tlb_start && vaddr < xen_io_tlb_end) &&
> > + !range_straddles_page_boundary(paddr, size) && !swiotlb_force) {
> > + dma_mark_clean(vaddr, size);
> > + return;
> > + }
> > +
> > /* NOTE: We use dev_addr here, not paddr! */
> > if (is_xen_swiotlb_buffer(dev_addr)) {
> > swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
> > @@ -717,13 +828,36 @@ xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
> > struct dma_attrs *attrs)
> > {
> > struct scatterlist *sg;
> > - int i;
> > + int i, rc;
> > + u64 dma_mask = DMA_BIT_MASK(32);
> > + unsigned long flags;
> >
> > BUG_ON(dir == DMA_NONE);
> >
> > + if (hwdev && hwdev->coherent_dma_mask)
> > + dma_mask = dma_alloc_coherent_mask(hwdev, GFP_KERNEL);
> > +
> > for_each_sg(sgl, sg, nelems, i) {
> > phys_addr_t paddr = sg_phys(sg);
> > - dma_addr_t dev_addr = xen_phys_to_bus_quick(paddr);
> > + dma_addr_t dev_addr;
> > +
> > + if (xen_feature(XENFEAT_auto_translated_physmap) &&
> > + !range_straddles_page_boundary(paddr, sg->length) &&
> > + sg->length <= PAGE_SIZE &&
> > + !swiotlb_force) {
> > +
> > + spin_lock_irqsave(&swiotlb_lock, flags);
> > + rc = xen_pin_dev_page(hwdev, paddr, &dev_addr);
> > + spin_unlock_irqrestore(&swiotlb_lock, flags);
> > +
> > + if (!rc) {
> > + dma_mark_clean(phys_to_virt(paddr), sg->length);
> > + sg_dma_len(sg) = sg->length;
> > + sg->dma_address = dev_addr;
> > + continue;
> > + }
> > + }
> > + dev_addr = xen_phys_to_bus_quick(paddr);
> >
> > if (swiotlb_force ||
> > xen_feature(XENFEAT_auto_translated_physmap) ||
> > --
> > 1.7.2.5
> >
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/