Since commit 295525e29a5b ("virtio_net: merge dma
operations when filling mergeable buffers"), VDUSE device
require support for DMA's .sync_single_for_cpu() operation
as the memory is non-coherent between the device and CPU
because of the use of a bounce buffer.
This patch implements both .sync_single_for_cpu() and
sync_single_for_device() callbacks, and also skip bounce
buffer copies during DMA map and unmap operations if the
DMA_ATTR_SKIP_CPU_SYNC attribute is set to avoid extra
copies of the same buffer.
Signed-off-by: Maxime Coquelin <[email protected]>
---
drivers/vdpa/vdpa_user/iova_domain.c | 27 ++++++++++++++++++++++++---
drivers/vdpa/vdpa_user/iova_domain.h | 8 ++++++++
drivers/vdpa/vdpa_user/vduse_dev.c | 22 ++++++++++++++++++++++
3 files changed, 54 insertions(+), 3 deletions(-)
diff --git a/drivers/vdpa/vdpa_user/iova_domain.c b/drivers/vdpa/vdpa_user/iova_domain.c
index 5e4a77b9bae6..791d38d6284c 100644
--- a/drivers/vdpa/vdpa_user/iova_domain.c
+++ b/drivers/vdpa/vdpa_user/iova_domain.c
@@ -373,6 +373,26 @@ static void vduse_domain_free_iova(struct iova_domain *iovad,
free_iova_fast(iovad, iova >> shift, iova_len);
}
+void vduse_domain_sync_single_for_device(struct vduse_iova_domain *domain,
+ dma_addr_t dma_addr, size_t size,
+ enum dma_data_direction dir)
+{
+ read_lock(&domain->bounce_lock);
+ if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
+ vduse_domain_bounce(domain, dma_addr, size, DMA_TO_DEVICE);
+ read_unlock(&domain->bounce_lock);
+}
+
+void vduse_domain_sync_single_for_cpu(struct vduse_iova_domain *domain,
+ dma_addr_t dma_addr, size_t size,
+ enum dma_data_direction dir)
+{
+ read_lock(&domain->bounce_lock);
+ if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
+ vduse_domain_bounce(domain, dma_addr, size, DMA_FROM_DEVICE);
+ read_unlock(&domain->bounce_lock);
+}
+
dma_addr_t vduse_domain_map_page(struct vduse_iova_domain *domain,
struct page *page, unsigned long offset,
size_t size, enum dma_data_direction dir,
@@ -393,7 +413,8 @@ dma_addr_t vduse_domain_map_page(struct vduse_iova_domain *domain,
if (vduse_domain_map_bounce_page(domain, (u64)iova, (u64)size, pa))
goto err_unlock;
- if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
+ if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
+ (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
vduse_domain_bounce(domain, iova, size, DMA_TO_DEVICE);
read_unlock(&domain->bounce_lock);
@@ -411,9 +432,9 @@ void vduse_domain_unmap_page(struct vduse_iova_domain *domain,
enum dma_data_direction dir, unsigned long attrs)
{
struct iova_domain *iovad = &domain->stream_iovad;
-
read_lock(&domain->bounce_lock);
- if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
+ if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
+ (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
vduse_domain_bounce(domain, dma_addr, size, DMA_FROM_DEVICE);
vduse_domain_unmap_bounce_page(domain, (u64)dma_addr, (u64)size);
diff --git a/drivers/vdpa/vdpa_user/iova_domain.h b/drivers/vdpa/vdpa_user/iova_domain.h
index 173e979b84a9..f92f22a7267d 100644
--- a/drivers/vdpa/vdpa_user/iova_domain.h
+++ b/drivers/vdpa/vdpa_user/iova_domain.h
@@ -44,6 +44,14 @@ int vduse_domain_set_map(struct vduse_iova_domain *domain,
void vduse_domain_clear_map(struct vduse_iova_domain *domain,
struct vhost_iotlb *iotlb);
+void vduse_domain_sync_single_for_device(struct vduse_iova_domain *domain,
+ dma_addr_t dma_addr, size_t size,
+ enum dma_data_direction dir);
+
+void vduse_domain_sync_single_for_cpu(struct vduse_iova_domain *domain,
+ dma_addr_t dma_addr, size_t size,
+ enum dma_data_direction dir);
+
dma_addr_t vduse_domain_map_page(struct vduse_iova_domain *domain,
struct page *page, unsigned long offset,
size_t size, enum dma_data_direction dir,
diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
index 1d24da79c399..75354ce186a1 100644
--- a/drivers/vdpa/vdpa_user/vduse_dev.c
+++ b/drivers/vdpa/vdpa_user/vduse_dev.c
@@ -798,6 +798,26 @@ static const struct vdpa_config_ops vduse_vdpa_config_ops = {
.free = vduse_vdpa_free,
};
+static void vduse_dev_sync_single_for_device(struct device *dev,
+ dma_addr_t dma_addr, size_t size,
+ enum dma_data_direction dir)
+{
+ struct vduse_dev *vdev = dev_to_vduse(dev);
+ struct vduse_iova_domain *domain = vdev->domain;
+
+ vduse_domain_sync_single_for_device(domain, dma_addr, size, dir);
+}
+
+static void vduse_dev_sync_single_for_cpu(struct device *dev,
+ dma_addr_t dma_addr, size_t size,
+ enum dma_data_direction dir)
+{
+ struct vduse_dev *vdev = dev_to_vduse(dev);
+ struct vduse_iova_domain *domain = vdev->domain;
+
+ vduse_domain_sync_single_for_cpu(domain, dma_addr, size, dir);
+}
+
static dma_addr_t vduse_dev_map_page(struct device *dev, struct page *page,
unsigned long offset, size_t size,
enum dma_data_direction dir,
@@ -858,6 +878,8 @@ static size_t vduse_dev_max_mapping_size(struct device *dev)
}
static const struct dma_map_ops vduse_dev_dma_ops = {
+ .sync_single_for_device = vduse_dev_sync_single_for_device,
+ .sync_single_for_cpu = vduse_dev_sync_single_for_cpu,
.map_page = vduse_dev_map_page,
.unmap_page = vduse_dev_unmap_page,
.alloc = vduse_dev_alloc_coherent,
--
2.43.0
On Mon, Feb 19, 2024 at 06:06:06PM +0100, Maxime Coquelin wrote:
> Since commit 295525e29a5b ("virtio_net: merge dma
> operations when filling mergeable buffers"), VDUSE device
> require support for DMA's .sync_single_for_cpu() operation
> as the memory is non-coherent between the device and CPU
> because of the use of a bounce buffer.
>
> This patch implements both .sync_single_for_cpu() and
> sync_single_for_device() callbacks, and also skip bounce
> buffer copies during DMA map and unmap operations if the
> DMA_ATTR_SKIP_CPU_SYNC attribute is set to avoid extra
> copies of the same buffer.
vduse really needs to get out of implementing fake DMA operations for
something that is not DMA.
Hello Christoph,
On 2/20/24 10:01, Christoph Hellwig wrote:
> On Mon, Feb 19, 2024 at 06:06:06PM +0100, Maxime Coquelin wrote:
>> Since commit 295525e29a5b ("virtio_net: merge dma
>> operations when filling mergeable buffers"), VDUSE device
>> require support for DMA's .sync_single_for_cpu() operation
>> as the memory is non-coherent between the device and CPU
>> because of the use of a bounce buffer.
>>
>> This patch implements both .sync_single_for_cpu() and
>> sync_single_for_device() callbacks, and also skip bounce
>> buffer copies during DMA map and unmap operations if the
>> DMA_ATTR_SKIP_CPU_SYNC attribute is set to avoid extra
>> copies of the same buffer.
>
> vduse really needs to get out of implementing fake DMA operations for
> something that is not DMA.
>
I wasn't involved when vduse was initially submitted, I don't know if
any alternative was considered at that time.
Do you have something specific in mind we could do to have VDUSE not
implementing DMA ops, knowing other vDPA devices rely on DMA?
Thanks,
Maxime
On Tue, Feb 20, 2024 at 01:01:36AM -0800, Christoph Hellwig wrote:
> On Mon, Feb 19, 2024 at 06:06:06PM +0100, Maxime Coquelin wrote:
> > Since commit 295525e29a5b ("virtio_net: merge dma
> > operations when filling mergeable buffers"), VDUSE device
> > require support for DMA's .sync_single_for_cpu() operation
> > as the memory is non-coherent between the device and CPU
> > because of the use of a bounce buffer.
> >
> > This patch implements both .sync_single_for_cpu() and
> > sync_single_for_device() callbacks, and also skip bounce
> > buffer copies during DMA map and unmap operations if the
> > DMA_ATTR_SKIP_CPU_SYNC attribute is set to avoid extra
> > copies of the same buffer.
>
> vduse really needs to get out of implementing fake DMA operations for
> something that is not DMA.
In a sense ... but on the other hand, the "fake DMA" metaphor seems to
work surprisingly well, like in this instance - internal bounce buffer
looks a bit like non-coherent DMA. A way to make this all prettier
would I guess be to actually wrap all of DMA with virtio wrappers which
would all go if () dma_... else vduse_...; or something to this end. A
lot of work for sure, and is it really worth it? if the only crazy
driver is vduse I'd maybe rather keep the crazy hacks local there ...
--
MST
On Thu, Feb 22, 2024 at 03:29:10PM -0500, Michael S. Tsirkin wrote:
> In a sense ... but on the other hand, the "fake DMA" metaphor seems to
> work surprisingly well, like in this instance - internal bounce buffer
> looks a bit like non-coherent DMA. A way to make this all prettier
> would I guess be to actually wrap all of DMA with virtio wrappers which
> would all go if () dma_... else vduse_...; or something to this end. A
> lot of work for sure, and is it really worth it? if the only crazy
> driver is vduse I'd maybe rather keep the crazy hacks local there ...
Well, vduse is the only driver that does this hack - we had a few more
and we got rid of it. It basically is the only thing preventing us
from doing direct calls into the iommu code and compile out dma_ops
entirely for non-Xen builds on the common architectures.
So yes, I'd really like to see it gone rather sooner than later.