diff mbox series

[v3] vhost: fix physical address mapping

Message ID 20211110060641.7666-1-xuan.ding@intel.com (mailing list archive)
State Superseded, archived
Delegated to: Maxime Coquelin
Headers
Series [v3] vhost: fix physical address mapping |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/github-robot: build success github build: passed
ci/Intel-compilation success Compilation OK
ci/intel-Testing success Testing PASS
ci/iol-mellanox-Performance success Performance Testing PASS
ci/iol-broadcom-Functional success Functional Testing PASS
ci/iol-broadcom-Performance success Performance Testing PASS
ci/iol-x86_64-compile-testing success Testing PASS
ci/iol-x86_64-unit-testing success Testing PASS
ci/iol-aarch64-unit-testing success Testing PASS
ci/iol-aarch64-compile-testing success Testing PASS
ci/iol-intel-Functional success Functional Testing PASS
ci/iol-intel-Performance success Performance Testing PASS

Commit Message

Ding, Xuan Nov. 10, 2021, 6:06 a.m. UTC 
  When choosing IOVA as PA mode, IOVA is likely to be discontinuous,
which requires page by page mapping for DMA devices. To be consistent,
this patch implements page by page mapping instead of mapping at the
region granularity for both IOVA as VA and PA mode.

Fixes: 7c61fa08b716 ("vhost: enable IOMMU for async vhost")

Signed-off-by: Xuan Ding <xuan.ding@intel.com>
---

v3:
* Fix commit title.

v2:
* Fix a format issue.
---
 lib/vhost/vhost.h      |   1 +
 lib/vhost/vhost_user.c | 105 ++++++++++++++++++++---------------------
 2 files changed, 53 insertions(+), 53 deletions(-)

Comments

Chenbo Xia Nov. 15, 2021, 7:20 a.m. UTC | #1 
Hi Xuan,

> -----Original Message-----
> From: Ding, Xuan <xuan.ding@intel.com>
> Sent: Wednesday, November 10, 2021 2:07 PM
> To: maxime.coquelin@redhat.com; Xia, Chenbo <chenbo.xia@intel.com>
> Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Wang, YuanX
> <yuanx.wang@intel.com>; He, Xingguang <xingguang.he@intel.com>; Ding, Xuan
> <xuan.ding@intel.com>
> Subject: [PATCH v3] vhost: fix physical address mapping
> 
> When choosing IOVA as PA mode, IOVA is likely to be discontinuous,
> which requires page by page mapping for DMA devices. To be consistent,
> this patch implements page by page mapping instead of mapping at the
> region granularity for both IOVA as VA and PA mode.
> 
> Fixes: 7c61fa08b716 ("vhost: enable IOMMU for async vhost")
> 
> Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> ---
> 
> v3:
> * Fix commit title.
> 
> v2:
> * Fix a format issue.
> ---
>  lib/vhost/vhost.h      |   1 +
>  lib/vhost/vhost_user.c | 105 ++++++++++++++++++++---------------------
>  2 files changed, 53 insertions(+), 53 deletions(-)
> 
> diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h
> index 7085e0885c..d246538ca5 100644
> --- a/lib/vhost/vhost.h
> +++ b/lib/vhost/vhost.h
> @@ -355,6 +355,7 @@ struct vring_packed_desc_event {
>  struct guest_page {
>  	uint64_t guest_phys_addr;
>  	uint64_t host_phys_addr;
> +	uint64_t host_user_addr;
>  	uint64_t size;
>  };
> 
> diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
> index a781346c4d..37cdedda3c 100644
> --- a/lib/vhost/vhost_user.c
> +++ b/lib/vhost/vhost_user.c
> @@ -144,52 +144,55 @@ get_blk_size(int fd)
>  }
> 
>  static int
> -async_dma_map(struct rte_vhost_mem_region *region, bool do_map)
> +async_dma_map(struct virtio_net *dev, bool do_map)
>  {
> -	uint64_t host_iova;
>  	int ret = 0;
> -
> -	host_iova = rte_mem_virt2iova((void *)(uintptr_t)region->host_user_addr);
> +	uint32_t i;
> +	struct guest_page *page;
>  	if (do_map) {
> -		/* Add mapped region into the default container of DPDK. */
> -		ret = rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
> -						 region->host_user_addr,
> -						 host_iova,
> -						 region->size);
> -		if (ret) {
> -			/*
> -			 * DMA device may bind with kernel driver, in this case,
> -			 * we don't need to program IOMMU manually. However, if no
> -			 * device is bound with vfio/uio in DPDK, and vfio kernel
> -			 * module is loaded, the API will still be called and return
> -			 * with ENODEV/ENOSUP.
> -			 *
> -			 * DPDK vfio only returns ENODEV/ENOSUP in very similar
> -			 * situations(vfio either unsupported, or supported
> -			 * but no devices found). Either way, no mappings could be
> -			 * performed. We treat it as normal case in async path.
> -			 */
> -			if (rte_errno == ENODEV || rte_errno == ENOTSUP)
> +		for (i = 0; i < dev->nr_guest_pages; i++) {
> +			page = &dev->guest_pages[i];
> +			ret =
> rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
> +							 page->host_user_addr,
> +							 page->host_phys_addr,
> +							 page->size);
> +			if (ret) {
> +				/*
> +				 * DMA device may bind with kernel driver, in this
> case,
> +				 * we don't need to program IOMMU manually. However,
> if no
> +				 * device is bound with vfio/uio in DPDK, and vfio
> kernel
> +				 * module is loaded, the API will still be called and
> return
> +				 * with ENODEV/ENOSUP.
> +				 *
> +				 * DPDK vfio only returns ENODEV/ENOSUP in very
> similar
> +				 * situations(vfio either unsupported, or supported
> +				 * but no devices found). Either way, no mappings
> could be
> +				 * performed. We treat it as normal case in async path.
> +				 */
> +				if (rte_errno == ENODEV || rte_errno == ENOTSUP)
> +					return 0;

I don't think this logic is good enough to only include the case of kernel driver +
unneeded mapping. Could also be vfio driver + incorrect mapping. It's not good to
assume ENODEV and ENOTSUP only comes from DPDK, it could be from kernel.

> +
> +				VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
> +				/* DMA mapping errors won't stop
> VHST_USER_SET_MEM_TABLE. */
>  				return 0;

I understand this function covers many cases and it's difficult to differentiate,
So you don't check the return value but use the log here to inform users.

I suggest to use a WARNING log (since this could fail with kernel driver case but
it's actually correct) and print the errno info for users.

Note: this is only a workaround, not a perfect solution. But since vhost with dmadev
is in progress and most likely vhost lib will aware of dmadev id. The problem could
be solved later (some dmadev api could be used to know VA/PA mode and kernel/user driver?)

> -
> -			VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
> -			/* DMA mapping errors won't stop VHST_USER_SET_MEM_TABLE. */
> -			return 0;
> +			}
>  		}
> 
>  	} else {
> -		/* Remove mapped region from the default container of DPDK. */
> -		ret = rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
> -						   region->host_user_addr,
> -						   host_iova,
> -						   region->size);
> -		if (ret) {
> -			/* like DMA map, ignore the kernel driver case when unmap.
> */
> -			if (rte_errno == EINVAL)
> -				return 0;
> +		for (i = 0; i < dev->nr_guest_pages; i++) {
> +			page = &dev->guest_pages[i];
> +			ret =
> rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
> +							   page->host_user_addr,
> +							   page->host_phys_addr,
> +							   page->size);
> +			if (ret) {
> +				/* like DMA map, ignore the kernel driver case when
> unmap. */
> +				if (rte_errno == EINVAL)
> +					return 0;
> 
> -			VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
> -			return ret;
> +				VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
> +				return ret;

Same here.

And since you don't check return value, you can just don't return anything and
return type can be void

Thanks,
Chenbo


> +			}
>  		}
>  	}
> 
> @@ -205,12 +208,12 @@ free_mem_region(struct virtio_net *dev)
>  	if (!dev || !dev->mem)
>  		return;
> 
> +	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> +		async_dma_map(dev, false);
> +
>  	for (i = 0; i < dev->mem->nregions; i++) {
>  		reg = &dev->mem->regions[i];
>  		if (reg->host_user_addr) {
> -			if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> -				async_dma_map(reg, false);
> -
>  			munmap(reg->mmap_addr, reg->mmap_size);
>  			close(reg->fd);
>  		}
> @@ -978,7 +981,7 @@ vhost_user_set_vring_base(struct virtio_net **pdev,
> 
>  static int
>  add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
> -		   uint64_t host_phys_addr, uint64_t size)
> +		   uint64_t host_phys_addr, uint64_t host_user_addr, uint64_t size)
>  {
>  	struct guest_page *page, *last_page;
>  	struct guest_page *old_pages;
> @@ -1009,6 +1012,7 @@ add_one_guest_page(struct virtio_net *dev, uint64_t
> guest_phys_addr,
>  	page = &dev->guest_pages[dev->nr_guest_pages++];
>  	page->guest_phys_addr = guest_phys_addr;
>  	page->host_phys_addr  = host_phys_addr;
> +	page->host_user_addr = host_user_addr;
>  	page->size = size;
> 
>  	return 0;
> @@ -1028,7 +1032,8 @@ add_guest_pages(struct virtio_net *dev, struct
> rte_vhost_mem_region *reg,
>  	size = page_size - (guest_phys_addr & (page_size - 1));
>  	size = RTE_MIN(size, reg_size);
> 
> -	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
> +	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
> +			       host_user_addr, size) < 0)
>  		return -1;
> 
>  	host_user_addr  += size;
> @@ -1040,7 +1045,7 @@ add_guest_pages(struct virtio_net *dev, struct
> rte_vhost_mem_region *reg,
>  		host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
>  						  host_user_addr);
>  		if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
> -				size) < 0)
> +				       host_user_addr, size) < 0)
>  			return -1;
> 
>  		host_user_addr  += size;
> @@ -1215,7 +1220,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
>  	uint64_t mmap_size;
>  	uint64_t alignment;
>  	int populate;
> -	int ret;
> 
>  	/* Check for memory_size + mmap_offset overflow */
>  	if (mmap_offset >= -region->size) {
> @@ -1274,14 +1278,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
>  			VHOST_LOG_CONFIG(ERR, "adding guest pages to region
> failed.\n");
>  			return -1;
>  		}
> -
> -		if (rte_vfio_is_enabled("vfio")) {
> -			ret = async_dma_map(region, true);
> -			if (ret) {
> -				VHOST_LOG_CONFIG(ERR, "Configure IOMMU for DMA engine
> failed\n");
> -				return -1;
> -			}
> -		}
>  	}
> 
>  	VHOST_LOG_CONFIG(INFO,
> @@ -1420,6 +1416,9 @@ vhost_user_set_mem_table(struct virtio_net **pdev,
> struct VhostUserMsg *msg,
>  		dev->mem->nregions++;
>  	}
> 
> +	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> +		async_dma_map(dev, true);
> +
>  	if (vhost_user_postcopy_register(dev, main_fd, msg) < 0)
>  		goto free_mem_table;
> 
> --
> 2.17.1
Ding, Xuan Nov. 15, 2021, 8:13 a.m. UTC | #2 
Hi Chenbo,

> -----Original Message-----
> From: Xia, Chenbo <chenbo.xia@intel.com>
> Sent: 2021年11月15日 15:21
> To: Ding, Xuan <xuan.ding@intel.com>; maxime.coquelin@redhat.com
> Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Wang, YuanX
> <yuanx.wang@intel.com>; He, Xingguang <xingguang.he@intel.com>
> Subject: RE: [PATCH v3] vhost: fix physical address mapping
> 
> Hi Xuan,
> 
> > -----Original Message-----
> > From: Ding, Xuan <xuan.ding@intel.com>
> > Sent: Wednesday, November 10, 2021 2:07 PM
> > To: maxime.coquelin@redhat.com; Xia, Chenbo <chenbo.xia@intel.com>
> > Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Wang, YuanX
> > <yuanx.wang@intel.com>; He, Xingguang <xingguang.he@intel.com>; Ding,
> > Xuan <xuan.ding@intel.com>
> > Subject: [PATCH v3] vhost: fix physical address mapping
> >
> > When choosing IOVA as PA mode, IOVA is likely to be discontinuous,
> > which requires page by page mapping for DMA devices. To be consistent,
> > this patch implements page by page mapping instead of mapping at the
> > region granularity for both IOVA as VA and PA mode.
> >
> > Fixes: 7c61fa08b716 ("vhost: enable IOMMU for async vhost")
> >
> > Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> > ---
> >
> > v3:
> > * Fix commit title.
> >
> > v2:
> > * Fix a format issue.
> > ---
> >  lib/vhost/vhost.h      |   1 +
> >  lib/vhost/vhost_user.c | 105
> > ++++++++++++++++++++---------------------
> >  2 files changed, 53 insertions(+), 53 deletions(-)
> >
> > diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h index
> > 7085e0885c..d246538ca5 100644
> > --- a/lib/vhost/vhost.h
> > +++ b/lib/vhost/vhost.h
> > @@ -355,6 +355,7 @@ struct vring_packed_desc_event {  struct
> > guest_page {
> >  	uint64_t guest_phys_addr;
> >  	uint64_t host_phys_addr;
> > +	uint64_t host_user_addr;
> >  	uint64_t size;
> >  };
> >
> > diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c index
> > a781346c4d..37cdedda3c 100644
> > --- a/lib/vhost/vhost_user.c
> > +++ b/lib/vhost/vhost_user.c
> > @@ -144,52 +144,55 @@ get_blk_size(int fd)  }
> >
> >  static int
> > -async_dma_map(struct rte_vhost_mem_region *region, bool do_map)
> > +async_dma_map(struct virtio_net *dev, bool do_map)
> >  {
> > -	uint64_t host_iova;
> >  	int ret = 0;
> > -
> > -	host_iova = rte_mem_virt2iova((void *)(uintptr_t)region-
> >host_user_addr);
> > +	uint32_t i;
> > +	struct guest_page *page;
> >  	if (do_map) {
> > -		/* Add mapped region into the default container of DPDK. */
> > -		ret =
> rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > -						 region->host_user_addr,
> > -						 host_iova,
> > -						 region->size);
> > -		if (ret) {
> > -			/*
> > -			 * DMA device may bind with kernel driver, in this case,
> > -			 * we don't need to program IOMMU manually.
> However, if no
> > -			 * device is bound with vfio/uio in DPDK, and vfio kernel
> > -			 * module is loaded, the API will still be called and
> return
> > -			 * with ENODEV/ENOSUP.
> > -			 *
> > -			 * DPDK vfio only returns ENODEV/ENOSUP in very
> similar
> > -			 * situations(vfio either unsupported, or supported
> > -			 * but no devices found). Either way, no mappings could
> be
> > -			 * performed. We treat it as normal case in async path.
> > -			 */
> > -			if (rte_errno == ENODEV || rte_errno == ENOTSUP)
> > +		for (i = 0; i < dev->nr_guest_pages; i++) {
> > +			page = &dev->guest_pages[i];
> > +			ret =
> > rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > +							 page-
> >host_user_addr,
> > +							 page-
> >host_phys_addr,
> > +							 page->size);
> > +			if (ret) {
> > +				/*
> > +				 * DMA device may bind with kernel driver, in
> this
> > case,
> > +				 * we don't need to program IOMMU manually.
> However,
> > if no
> > +				 * device is bound with vfio/uio in DPDK, and
> vfio
> > kernel
> > +				 * module is loaded, the API will still be called
> and
> > return
> > +				 * with ENODEV/ENOSUP.
> > +				 *
> > +				 * DPDK vfio only returns ENODEV/ENOSUP in
> very
> > similar
> > +				 * situations(vfio either unsupported, or
> supported
> > +				 * but no devices found). Either way, no
> mappings
> > could be
> > +				 * performed. We treat it as normal case in
> async path.
> > +				 */
> > +				if (rte_errno == ENODEV || rte_errno ==
> ENOTSUP)
> > +					return 0;
> 
> I don't think this logic is good enough to only include the case of kernel driver +
> unneeded mapping. Could also be vfio driver + incorrect mapping. It's not good
> to assume ENODEV and ENOTSUP only comes from DPDK, it could be from
> kernel.

The code here is a workaround to be compatible with kernel driver, since we can't
know the device driver in vhost for now. In fact, the appropriate approach for justifying
whether perform mapping here should base on the device driver.

> 
> > +
> > +				VHOST_LOG_CONFIG(ERR, "DMA engine map
> failed\n");
> > +				/* DMA mapping errors won't stop
> > VHST_USER_SET_MEM_TABLE. */
> >  				return 0;
> 
> I understand this function covers many cases and it's difficult to differentiate, So
> you don't check the return value but use the log here to inform users.
> 
> I suggest to use a WARNING log (since this could fail with kernel driver case but
> it's actually correct) and print the errno info for users.

Thanks for your suggestion. Accordingly, I will update the doc to inform user should
take which action according to different warning message.

> 
> Note: this is only a workaround, not a perfect solution. But since vhost with
> dmadev is in progress and most likely vhost lib will aware of dmadev id. The
> problem could be solved later (some dmadev api could be used to know VA/PA
> mode and kernel/user driver?)
 
With rte_iova_mode(), we can know current IOVA mode in DPDK. It will not affect
vfio mapping because bound a device to vfio always requires mapping. For dmadev,
it is possible to know current driver, since the integration for async vhost with dmadev
is in progress, so let's use this solution for now.

> 
> > -
> > -			VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
> > -			/* DMA mapping errors won't stop
> VHST_USER_SET_MEM_TABLE. */
> > -			return 0;
> > +			}
> >  		}
> >
> >  	} else {
> > -		/* Remove mapped region from the default container of DPDK.
> */
> > -		ret =
> rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > -						   region->host_user_addr,
> > -						   host_iova,
> > -						   region->size);
> > -		if (ret) {
> > -			/* like DMA map, ignore the kernel driver case when
> unmap.
> > */
> > -			if (rte_errno == EINVAL)
> > -				return 0;
> > +		for (i = 0; i < dev->nr_guest_pages; i++) {
> > +			page = &dev->guest_pages[i];
> > +			ret =
> > rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > +							   page-
> >host_user_addr,
> > +							   page-
> >host_phys_addr,
> > +							   page->size);
> > +			if (ret) {
> > +				/* like DMA map, ignore the kernel driver case
> when
> > unmap. */
> > +				if (rte_errno == EINVAL)
> > +					return 0;
> >
> > -			VHOST_LOG_CONFIG(ERR, "DMA engine unmap
> failed\n");
> > -			return ret;
> > +				VHOST_LOG_CONFIG(ERR, "DMA engine
> unmap failed\n");
> > +				return ret;
> 
> Same here.
> 
> And since you don't check return value, you can just don't return anything and
> return type can be void

Good catch, pls see next version.

Thanks,
Xuan

> 
> Thanks,
> Chenbo
> 
> 
> > +			}
> >  		}
> >  	}
> >
> > @@ -205,12 +208,12 @@ free_mem_region(struct virtio_net *dev)
> >  	if (!dev || !dev->mem)
> >  		return;
> >
> > +	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> > +		async_dma_map(dev, false);
> > +
> >  	for (i = 0; i < dev->mem->nregions; i++) {
> >  		reg = &dev->mem->regions[i];
> >  		if (reg->host_user_addr) {
> > -			if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> > -				async_dma_map(reg, false);
> > -
> >  			munmap(reg->mmap_addr, reg->mmap_size);
> >  			close(reg->fd);
> >  		}
> > @@ -978,7 +981,7 @@ vhost_user_set_vring_base(struct virtio_net
> > **pdev,
> >
> >  static int
> >  add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
> > -		   uint64_t host_phys_addr, uint64_t size)
> > +		   uint64_t host_phys_addr, uint64_t host_user_addr, uint64_t
> size)
> >  {
> >  	struct guest_page *page, *last_page;
> >  	struct guest_page *old_pages;
> > @@ -1009,6 +1012,7 @@ add_one_guest_page(struct virtio_net *dev,
> > uint64_t guest_phys_addr,
> >  	page = &dev->guest_pages[dev->nr_guest_pages++];
> >  	page->guest_phys_addr = guest_phys_addr;
> >  	page->host_phys_addr  = host_phys_addr;
> > +	page->host_user_addr = host_user_addr;
> >  	page->size = size;
> >
> >  	return 0;
> > @@ -1028,7 +1032,8 @@ add_guest_pages(struct virtio_net *dev, struct
> > rte_vhost_mem_region *reg,
> >  	size = page_size - (guest_phys_addr & (page_size - 1));
> >  	size = RTE_MIN(size, reg_size);
> >
> > -	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) <
> 0)
> > +	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
> > +			       host_user_addr, size) < 0)
> >  		return -1;
> >
> >  	host_user_addr  += size;
> > @@ -1040,7 +1045,7 @@ add_guest_pages(struct virtio_net *dev, struct
> > rte_vhost_mem_region *reg,
> >  		host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
> >  						  host_user_addr);
> >  		if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
> > -				size) < 0)
> > +				       host_user_addr, size) < 0)
> >  			return -1;
> >
> >  		host_user_addr  += size;
> > @@ -1215,7 +1220,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
> >  	uint64_t mmap_size;
> >  	uint64_t alignment;
> >  	int populate;
> > -	int ret;
> >
> >  	/* Check for memory_size + mmap_offset overflow */
> >  	if (mmap_offset >= -region->size) {
> > @@ -1274,14 +1278,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
> >  			VHOST_LOG_CONFIG(ERR, "adding guest pages to
> region failed.\n");
> >  			return -1;
> >  		}
> > -
> > -		if (rte_vfio_is_enabled("vfio")) {
> > -			ret = async_dma_map(region, true);
> > -			if (ret) {
> > -				VHOST_LOG_CONFIG(ERR, "Configure IOMMU
> for DMA engine
> > failed\n");
> > -				return -1;
> > -			}
> > -		}
> >  	}
> >
> >  	VHOST_LOG_CONFIG(INFO,
> > @@ -1420,6 +1416,9 @@ vhost_user_set_mem_table(struct virtio_net
> > **pdev, struct VhostUserMsg *msg,
> >  		dev->mem->nregions++;
> >  	}
> >
> > +	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> > +		async_dma_map(dev, true);
> > +
> >  	if (vhost_user_postcopy_register(dev, main_fd, msg) < 0)
> >  		goto free_mem_table;
> >
> > --
> > 2.17.1
Chenbo Xia Nov. 15, 2021, 12:11 p.m. UTC | #3 
> -----Original Message-----
> From: Ding, Xuan <xuan.ding@intel.com>
> Sent: Monday, November 15, 2021 4:13 PM
> To: Xia, Chenbo <chenbo.xia@intel.com>; maxime.coquelin@redhat.com
> Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Wang, YuanX
> <yuanx.wang@intel.com>; He, Xingguang <xingguang.he@intel.com>; Burakov,
> Anatoly <anatoly.burakov@intel.com>
> Subject: RE: [PATCH v3] vhost: fix physical address mapping
> 
> Hi Chenbo,
> 
> > -----Original Message-----
> > From: Xia, Chenbo <chenbo.xia@intel.com>
> > Sent: 2021年11月15日 15:21
> > To: Ding, Xuan <xuan.ding@intel.com>; maxime.coquelin@redhat.com
> > Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Wang, YuanX
> > <yuanx.wang@intel.com>; He, Xingguang <xingguang.he@intel.com>
> > Subject: RE: [PATCH v3] vhost: fix physical address mapping
> >
> > Hi Xuan,
> >
> > > -----Original Message-----
> > > From: Ding, Xuan <xuan.ding@intel.com>
> > > Sent: Wednesday, November 10, 2021 2:07 PM
> > > To: maxime.coquelin@redhat.com; Xia, Chenbo <chenbo.xia@intel.com>
> > > Cc: dev@dpdk.org; Hu, Jiayu <jiayu.hu@intel.com>; Wang, YuanX
> > > <yuanx.wang@intel.com>; He, Xingguang <xingguang.he@intel.com>; Ding,
> > > Xuan <xuan.ding@intel.com>
> > > Subject: [PATCH v3] vhost: fix physical address mapping
> > >
> > > When choosing IOVA as PA mode, IOVA is likely to be discontinuous,
> > > which requires page by page mapping for DMA devices. To be consistent,
> > > this patch implements page by page mapping instead of mapping at the
> > > region granularity for both IOVA as VA and PA mode.
> > >
> > > Fixes: 7c61fa08b716 ("vhost: enable IOMMU for async vhost")
> > >
> > > Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> > > ---
> > >
> > > v3:
> > > * Fix commit title.
> > >
> > > v2:
> > > * Fix a format issue.
> > > ---
> > >  lib/vhost/vhost.h      |   1 +
> > >  lib/vhost/vhost_user.c | 105
> > > ++++++++++++++++++++---------------------
> > >  2 files changed, 53 insertions(+), 53 deletions(-)
> > >
> > > diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h index
> > > 7085e0885c..d246538ca5 100644
> > > --- a/lib/vhost/vhost.h
> > > +++ b/lib/vhost/vhost.h
> > > @@ -355,6 +355,7 @@ struct vring_packed_desc_event {  struct
> > > guest_page {
> > >  	uint64_t guest_phys_addr;
> > >  	uint64_t host_phys_addr;
> > > +	uint64_t host_user_addr;
> > >  	uint64_t size;
> > >  };
> > >
> > > diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c index
> > > a781346c4d..37cdedda3c 100644
> > > --- a/lib/vhost/vhost_user.c
> > > +++ b/lib/vhost/vhost_user.c
> > > @@ -144,52 +144,55 @@ get_blk_size(int fd)  }
> > >
> > >  static int
> > > -async_dma_map(struct rte_vhost_mem_region *region, bool do_map)
> > > +async_dma_map(struct virtio_net *dev, bool do_map)
> > >  {
> > > -	uint64_t host_iova;
> > >  	int ret = 0;
> > > -
> > > -	host_iova = rte_mem_virt2iova((void *)(uintptr_t)region-
> > >host_user_addr);
> > > +	uint32_t i;
> > > +	struct guest_page *page;
> > >  	if (do_map) {
> > > -		/* Add mapped region into the default container of DPDK. */
> > > -		ret =
> > rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > > -						 region->host_user_addr,
> > > -						 host_iova,
> > > -						 region->size);
> > > -		if (ret) {
> > > -			/*
> > > -			 * DMA device may bind with kernel driver, in this case,
> > > -			 * we don't need to program IOMMU manually.
> > However, if no
> > > -			 * device is bound with vfio/uio in DPDK, and vfio kernel
> > > -			 * module is loaded, the API will still be called and
> > return
> > > -			 * with ENODEV/ENOSUP.
> > > -			 *
> > > -			 * DPDK vfio only returns ENODEV/ENOSUP in very
> > similar
> > > -			 * situations(vfio either unsupported, or supported
> > > -			 * but no devices found). Either way, no mappings could
> > be
> > > -			 * performed. We treat it as normal case in async path.
> > > -			 */
> > > -			if (rte_errno == ENODEV || rte_errno == ENOTSUP)
> > > +		for (i = 0; i < dev->nr_guest_pages; i++) {
> > > +			page = &dev->guest_pages[i];
> > > +			ret =
> > > rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > > +							 page-
> > >host_user_addr,
> > > +							 page-
> > >host_phys_addr,
> > > +							 page->size);
> > > +			if (ret) {
> > > +				/*
> > > +				 * DMA device may bind with kernel driver, in
> > this
> > > case,
> > > +				 * we don't need to program IOMMU manually.
> > However,
> > > if no
> > > +				 * device is bound with vfio/uio in DPDK, and
> > vfio
> > > kernel
> > > +				 * module is loaded, the API will still be called
> > and
> > > return
> > > +				 * with ENODEV/ENOSUP.
> > > +				 *
> > > +				 * DPDK vfio only returns ENODEV/ENOSUP in
> > very
> > > similar
> > > +				 * situations(vfio either unsupported, or
> > supported
> > > +				 * but no devices found). Either way, no
> > mappings
> > > could be
> > > +				 * performed. We treat it as normal case in
> > async path.
> > > +				 */
> > > +				if (rte_errno == ENODEV || rte_errno ==
> > ENOTSUP)
> > > +					return 0;
> >
> > I don't think this logic is good enough to only include the case of kernel
> driver +
> > unneeded mapping. Could also be vfio driver + incorrect mapping. It's not
> good
> > to assume ENODEV and ENOTSUP only comes from DPDK, it could be from
> > kernel.
> 
> The code here is a workaround to be compatible with kernel driver, since we
> can't
> know the device driver in vhost for now. In fact, the appropriate approach for
> justifying
> whether perform mapping here should base on the device driver.

OK. But since this code does not consider the case of FreeBSD, checking ENODEV 
is enough.

> 
> >
> > > +
> > > +				VHOST_LOG_CONFIG(ERR, "DMA engine map
> > failed\n");
> > > +				/* DMA mapping errors won't stop
> > > VHST_USER_SET_MEM_TABLE. */

VHST -> VHOST

> > >  				return 0;
> >
> > I understand this function covers many cases and it's difficult to
> differentiate, So
> > you don't check the return value but use the log here to inform users.
> >
> > I suggest to use a WARNING log (since this could fail with kernel driver
> case but
> > it's actually correct) and print the errno info for users.
> 
> Thanks for your suggestion. Accordingly, I will update the doc to inform user
> should
> take which action according to different warning message.

If ENODEV is used to filter the error but correct case (using kernel driver),
you can just use ERR log here. But return type should be void.

Thanks,
Chenbo

> 
> >
> > Note: this is only a workaround, not a perfect solution. But since vhost
> with
> > dmadev is in progress and most likely vhost lib will aware of dmadev id. The
> > problem could be solved later (some dmadev api could be used to know VA/PA
> > mode and kernel/user driver?)
> 
> With rte_iova_mode(), we can know current IOVA mode in DPDK. It will not
> affect
> vfio mapping because bound a device to vfio always requires mapping. For
> dmadev,
> it is possible to know current driver, since the integration for async vhost
> with dmadev
> is in progress, so let's use this solution for now.
> 
> >
> > > -
> > > -			VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
> > > -			/* DMA mapping errors won't stop
> > VHST_USER_SET_MEM_TABLE. */
> > > -			return 0;
> > > +			}
> > >  		}
> > >
> > >  	} else {
> > > -		/* Remove mapped region from the default container of DPDK.
> > */
> > > -		ret =
> > rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > > -						   region->host_user_addr,
> > > -						   host_iova,
> > > -						   region->size);
> > > -		if (ret) {
> > > -			/* like DMA map, ignore the kernel driver case when
> > unmap.
> > > */
> > > -			if (rte_errno == EINVAL)
> > > -				return 0;
> > > +		for (i = 0; i < dev->nr_guest_pages; i++) {
> > > +			page = &dev->guest_pages[i];
> > > +			ret =
> > > rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
> > > +							   page-
> > >host_user_addr,
> > > +							   page-
> > >host_phys_addr,
> > > +							   page->size);
> > > +			if (ret) {
> > > +				/* like DMA map, ignore the kernel driver case
> > when
> > > unmap. */
> > > +				if (rte_errno == EINVAL)
> > > +					return 0;
> > >
> > > -			VHOST_LOG_CONFIG(ERR, "DMA engine unmap
> > failed\n");
> > > -			return ret;
> > > +				VHOST_LOG_CONFIG(ERR, "DMA engine
> > unmap failed\n");
> > > +				return ret;
> >
> > Same here.
> >
> > And since you don't check return value, you can just don't return anything
> and
> > return type can be void
> 
> Good catch, pls see next version.
> 
> Thanks,
> Xuan
> 
> >
> > Thanks,
> > Chenbo
> >
> >
> > > +			}
> > >  		}
> > >  	}
> > >
> > > @@ -205,12 +208,12 @@ free_mem_region(struct virtio_net *dev)
> > >  	if (!dev || !dev->mem)
> > >  		return;
> > >
> > > +	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> > > +		async_dma_map(dev, false);
> > > +
> > >  	for (i = 0; i < dev->mem->nregions; i++) {
> > >  		reg = &dev->mem->regions[i];
> > >  		if (reg->host_user_addr) {
> > > -			if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> > > -				async_dma_map(reg, false);
> > > -
> > >  			munmap(reg->mmap_addr, reg->mmap_size);
> > >  			close(reg->fd);
> > >  		}
> > > @@ -978,7 +981,7 @@ vhost_user_set_vring_base(struct virtio_net
> > > **pdev,
> > >
> > >  static int
> > >  add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
> > > -		   uint64_t host_phys_addr, uint64_t size)
> > > +		   uint64_t host_phys_addr, uint64_t host_user_addr, uint64_t
> > size)
> > >  {
> > >  	struct guest_page *page, *last_page;
> > >  	struct guest_page *old_pages;
> > > @@ -1009,6 +1012,7 @@ add_one_guest_page(struct virtio_net *dev,
> > > uint64_t guest_phys_addr,
> > >  	page = &dev->guest_pages[dev->nr_guest_pages++];
> > >  	page->guest_phys_addr = guest_phys_addr;
> > >  	page->host_phys_addr  = host_phys_addr;
> > > +	page->host_user_addr = host_user_addr;
> > >  	page->size = size;
> > >
> > >  	return 0;
> > > @@ -1028,7 +1032,8 @@ add_guest_pages(struct virtio_net *dev, struct
> > > rte_vhost_mem_region *reg,
> > >  	size = page_size - (guest_phys_addr & (page_size - 1));
> > >  	size = RTE_MIN(size, reg_size);
> > >
> > > -	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) <
> > 0)
> > > +	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
> > > +			       host_user_addr, size) < 0)
> > >  		return -1;
> > >
> > >  	host_user_addr  += size;
> > > @@ -1040,7 +1045,7 @@ add_guest_pages(struct virtio_net *dev, struct
> > > rte_vhost_mem_region *reg,
> > >  		host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
> > >  						  host_user_addr);
> > >  		if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
> > > -				size) < 0)
> > > +				       host_user_addr, size) < 0)
> > >  			return -1;
> > >
> > >  		host_user_addr  += size;
> > > @@ -1215,7 +1220,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
> > >  	uint64_t mmap_size;
> > >  	uint64_t alignment;
> > >  	int populate;
> > > -	int ret;
> > >
> > >  	/* Check for memory_size + mmap_offset overflow */
> > >  	if (mmap_offset >= -region->size) {
> > > @@ -1274,14 +1278,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
> > >  			VHOST_LOG_CONFIG(ERR, "adding guest pages to
> > region failed.\n");
> > >  			return -1;
> > >  		}
> > > -
> > > -		if (rte_vfio_is_enabled("vfio")) {
> > > -			ret = async_dma_map(region, true);
> > > -			if (ret) {
> > > -				VHOST_LOG_CONFIG(ERR, "Configure IOMMU
> > for DMA engine
> > > failed\n");
> > > -				return -1;
> > > -			}
> > > -		}
> > >  	}
> > >
> > >  	VHOST_LOG_CONFIG(INFO,
> > > @@ -1420,6 +1416,9 @@ vhost_user_set_mem_table(struct virtio_net
> > > **pdev, struct VhostUserMsg *msg,
> > >  		dev->mem->nregions++;
> > >  	}
> > >
> > > +	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
> > > +		async_dma_map(dev, true);
> > > +
> > >  	if (vhost_user_postcopy_register(dev, main_fd, msg) < 0)
> > >  		goto free_mem_table;
> > >
> > > --
> > > 2.17.1
diff mbox series

Patch

diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h
index 7085e0885c..d246538ca5 100644
--- a/lib/vhost/vhost.h
+++ b/lib/vhost/vhost.h
@@ -355,6 +355,7 @@  struct vring_packed_desc_event {
 struct guest_page {
 	uint64_t guest_phys_addr;
 	uint64_t host_phys_addr;
+	uint64_t host_user_addr;
 	uint64_t size;
 };
 
diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
index a781346c4d..37cdedda3c 100644
--- a/lib/vhost/vhost_user.c
+++ b/lib/vhost/vhost_user.c
@@ -144,52 +144,55 @@  get_blk_size(int fd)
 }
 
 static int
-async_dma_map(struct rte_vhost_mem_region *region, bool do_map)
+async_dma_map(struct virtio_net *dev, bool do_map)
 {
-	uint64_t host_iova;
 	int ret = 0;
-
-	host_iova = rte_mem_virt2iova((void *)(uintptr_t)region->host_user_addr);
+	uint32_t i;
+	struct guest_page *page;
 	if (do_map) {
-		/* Add mapped region into the default container of DPDK. */
-		ret = rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
-						 region->host_user_addr,
-						 host_iova,
-						 region->size);
-		if (ret) {
-			/*
-			 * DMA device may bind with kernel driver, in this case,
-			 * we don't need to program IOMMU manually. However, if no
-			 * device is bound with vfio/uio in DPDK, and vfio kernel
-			 * module is loaded, the API will still be called and return
-			 * with ENODEV/ENOSUP.
-			 *
-			 * DPDK vfio only returns ENODEV/ENOSUP in very similar
-			 * situations(vfio either unsupported, or supported
-			 * but no devices found). Either way, no mappings could be
-			 * performed. We treat it as normal case in async path.
-			 */
-			if (rte_errno == ENODEV || rte_errno == ENOTSUP)
+		for (i = 0; i < dev->nr_guest_pages; i++) {
+			page = &dev->guest_pages[i];
+			ret = rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
+							 page->host_user_addr,
+							 page->host_phys_addr,
+							 page->size);
+			if (ret) {
+				/*
+				 * DMA device may bind with kernel driver, in this case,
+				 * we don't need to program IOMMU manually. However, if no
+				 * device is bound with vfio/uio in DPDK, and vfio kernel
+				 * module is loaded, the API will still be called and return
+				 * with ENODEV/ENOSUP.
+				 *
+				 * DPDK vfio only returns ENODEV/ENOSUP in very similar
+				 * situations(vfio either unsupported, or supported
+				 * but no devices found). Either way, no mappings could be
+				 * performed. We treat it as normal case in async path.
+				 */
+				if (rte_errno == ENODEV || rte_errno == ENOTSUP)
+					return 0;
+
+				VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
+				/* DMA mapping errors won't stop VHST_USER_SET_MEM_TABLE. */
 				return 0;
-
-			VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
-			/* DMA mapping errors won't stop VHST_USER_SET_MEM_TABLE. */
-			return 0;
+			}
 		}
 
 	} else {
-		/* Remove mapped region from the default container of DPDK. */
-		ret = rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
-						   region->host_user_addr,
-						   host_iova,
-						   region->size);
-		if (ret) {
-			/* like DMA map, ignore the kernel driver case when unmap. */
-			if (rte_errno == EINVAL)
-				return 0;
+		for (i = 0; i < dev->nr_guest_pages; i++) {
+			page = &dev->guest_pages[i];
+			ret = rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
+							   page->host_user_addr,
+							   page->host_phys_addr,
+							   page->size);
+			if (ret) {
+				/* like DMA map, ignore the kernel driver case when unmap. */
+				if (rte_errno == EINVAL)
+					return 0;
 
-			VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
-			return ret;
+				VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
+				return ret;
+			}
 		}
 	}
 
@@ -205,12 +208,12 @@  free_mem_region(struct virtio_net *dev)
 	if (!dev || !dev->mem)
 		return;
 
+	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
+		async_dma_map(dev, false);
+
 	for (i = 0; i < dev->mem->nregions; i++) {
 		reg = &dev->mem->regions[i];
 		if (reg->host_user_addr) {
-			if (dev->async_copy && rte_vfio_is_enabled("vfio"))
-				async_dma_map(reg, false);
-
 			munmap(reg->mmap_addr, reg->mmap_size);
 			close(reg->fd);
 		}
@@ -978,7 +981,7 @@  vhost_user_set_vring_base(struct virtio_net **pdev,
 
 static int
 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
-		   uint64_t host_phys_addr, uint64_t size)
+		   uint64_t host_phys_addr, uint64_t host_user_addr, uint64_t size)
 {
 	struct guest_page *page, *last_page;
 	struct guest_page *old_pages;
@@ -1009,6 +1012,7 @@  add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 	page = &dev->guest_pages[dev->nr_guest_pages++];
 	page->guest_phys_addr = guest_phys_addr;
 	page->host_phys_addr  = host_phys_addr;
+	page->host_user_addr = host_user_addr;
 	page->size = size;
 
 	return 0;
@@ -1028,7 +1032,8 @@  add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 	size = page_size - (guest_phys_addr & (page_size - 1));
 	size = RTE_MIN(size, reg_size);
 
-	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
+	if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
+			       host_user_addr, size) < 0)
 		return -1;
 
 	host_user_addr  += size;
@@ -1040,7 +1045,7 @@  add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 		host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 						  host_user_addr);
 		if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
-				size) < 0)
+				       host_user_addr, size) < 0)
 			return -1;
 
 		host_user_addr  += size;
@@ -1215,7 +1220,6 @@  vhost_user_mmap_region(struct virtio_net *dev,
 	uint64_t mmap_size;
 	uint64_t alignment;
 	int populate;
-	int ret;
 
 	/* Check for memory_size + mmap_offset overflow */
 	if (mmap_offset >= -region->size) {
@@ -1274,14 +1278,6 @@  vhost_user_mmap_region(struct virtio_net *dev,
 			VHOST_LOG_CONFIG(ERR, "adding guest pages to region failed.\n");
 			return -1;
 		}
-
-		if (rte_vfio_is_enabled("vfio")) {
-			ret = async_dma_map(region, true);
-			if (ret) {
-				VHOST_LOG_CONFIG(ERR, "Configure IOMMU for DMA engine failed\n");
-				return -1;
-			}
-		}
 	}
 
 	VHOST_LOG_CONFIG(INFO,
@@ -1420,6 +1416,9 @@  vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		dev->mem->nregions++;
 	}
 
+	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
+		async_dma_map(dev, true);
+
 	if (vhost_user_postcopy_register(dev, main_fd, msg) < 0)
 		goto free_mem_table;