[v2,1/1] vfio: add page-by-page mapping API

  Currently, there is no way to map memory for DMA in a way that allows
unmapping it partially later, because some IOMMU's do not support
partial unmapping. There is a workaround of mapping all of these
segments separately, but this is inconvenient and silly, so this
commit adds a proper API that does it.

This commit relies on earlier infrastructure that was built out to
support "chunking", as the concept of "chunks" is essentially the same
as page size.

Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
---

Notes:
    v2:
    - Added checks for page alignment of supplied values
    - Added rte_errno values (similar to patch 103165 [1])
    
    [1] http://patches.dpdk.org/project/dpdk/patch/e8c5e7ba089e2283c3cd67e4529e52fe49390eb9.1635428963.git.anatoly.burakov@intel.com/

 lib/eal/freebsd/eal.c      |  11 ++++
 lib/eal/include/rte_vfio.h |  33 ++++++++++++
 lib/eal/linux/eal_vfio.c   | 108 +++++++++++++++++++++++++++++++------
 lib/eal/version.map        |   1 +
 lib/eal/windows/eal.c      |  11 ++++
 5 files changed, 148 insertions(+), 16 deletions(-)
  

On 28-Oct-21 3:09 PM, Anatoly Burakov wrote:
> Currently, there is no way to map memory for DMA in a way that allows
> unmapping it partially later, because some IOMMU's do not support
> partial unmapping. There is a workaround of mapping all of these
> segments separately, but this is inconvenient and silly, so this
> commit adds a proper API that does it.
> 
> This commit relies on earlier infrastructure that was built out to
> support "chunking", as the concept of "chunks" is essentially the same
> as page size.
> 
> Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
> ---
> 
> Notes:
>      v2:
>      - Added checks for page alignment of supplied values
>      - Added rte_errno values (similar to patch 103165 [1])
>      
>      [1] http://patches.dpdk.org/project/dpdk/patch/e8c5e7ba089e2283c3cd67e4529e52fe49390eb9.1635428963.git.anatoly.burakov@intel.com/

I think this API also needs VA and IOVA addresses in an array, so that 
it's possible to map IOVA-discontiguous segments. This is too late for 
this release, so let's postpone it till 22.02.
  

Hello Anatoly,

On Tue, Nov 2, 2021 at 4:54 PM Burakov, Anatoly
<anatoly.burakov@intel.com> wrote:
>
> On 28-Oct-21 3:09 PM, Anatoly Burakov wrote:
> > Currently, there is no way to map memory for DMA in a way that allows
> > unmapping it partially later, because some IOMMU's do not support
> > partial unmapping. There is a workaround of mapping all of these
> > segments separately, but this is inconvenient and silly, so this
> > commit adds a proper API that does it.
> >
> > This commit relies on earlier infrastructure that was built out to
> > support "chunking", as the concept of "chunks" is essentially the same
> > as page size.
> >
> > Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
> > ---
> >
> > Notes:
> >      v2:
> >      - Added checks for page alignment of supplied values
> >      - Added rte_errno values (similar to patch 103165 [1])
> >
> >      [1] http://patches.dpdk.org/project/dpdk/patch/e8c5e7ba089e2283c3cd67e4529e52fe49390eb9.1635428963.git.anatoly.burakov@intel.com/
>
> I think this API also needs VA and IOVA addresses in an array, so that
> it's possible to map IOVA-discontiguous segments. This is too late for
> this release, so let's postpone it till 22.02.

It seems this patch fell through the cracks.
Any update?

Thanks.
  

Anatoly Burakov <anatoly.burakov@intel.com> writes:

> Currently, there is no way to map memory for DMA in a way that allows
> unmapping it partially later, because some IOMMU's do not support
> partial unmapping. There is a workaround of mapping all of these
> segments separately, but this is inconvenient and silly, so this
> commit adds a proper API that does it.
>
> This commit relies on earlier infrastructure that was built out to
> support "chunking", as the concept of "chunks" is essentially the same
> as page size.
>
> Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
> ---
>
> Notes:
>     v2:
>     - Added checks for page alignment of supplied values
>     - Added rte_errno values (similar to patch 103165 [1])
>     
>     [1] http://patches.dpdk.org/project/dpdk/patch/e8c5e7ba089e2283c3cd67e4529e52fe49390eb9.1635428963.git.anatoly.burakov@intel.com/
>
>  lib/eal/freebsd/eal.c      |  11 ++++
>  lib/eal/include/rte_vfio.h |  33 ++++++++++++
>  lib/eal/linux/eal_vfio.c   | 108 +++++++++++++++++++++++++++++++------
>  lib/eal/version.map        |   1 +
>  lib/eal/windows/eal.c      |  11 ++++
>  5 files changed, 148 insertions(+), 16 deletions(-)
>
> diff --git a/lib/eal/freebsd/eal.c b/lib/eal/freebsd/eal.c
> index 9935356ed4..5ca5fa8e11 100644
> --- a/lib/eal/freebsd/eal.c
> +++ b/lib/eal/freebsd/eal.c
> @@ -1090,6 +1090,17 @@ rte_vfio_container_dma_map(__rte_unused int container_fd,
>  	return -1;
>  }
>  
> +int
> +rte_vfio_container_dma_map_paged(__rte_unused int container_fd,
> +		__rte_unused uint64_t vaddr,
> +		__rte_unused uint64_t iova,
> +		__rte_unused uint64_t len,
> +		__rte_unused uint64_t pagesz)
> +{
> +	rte_errno = ENOTSUP;
> +	return -1;
> +}
> +
>  int
>  rte_vfio_container_dma_unmap(__rte_unused int container_fd,
>  			__rte_unused uint64_t vaddr,
> diff --git a/lib/eal/include/rte_vfio.h b/lib/eal/include/rte_vfio.h
> index 2d90b36480..6afae2ccce 100644
> --- a/lib/eal/include/rte_vfio.h
> +++ b/lib/eal/include/rte_vfio.h
> @@ -17,6 +17,8 @@ extern "C" {
>  #include <stdbool.h>
>  #include <stdint.h>
>  
> +#include <rte_compat.h>
> +
>  /*
>   * determine if VFIO is present on the system
>   */
> @@ -331,6 +333,37 @@ int
>  rte_vfio_container_dma_map(int container_fd, uint64_t vaddr,
>  		uint64_t iova, uint64_t len);
>  
> +/**
> + * @warning
> + * @b EXPERIMENTAL: this API may change, or be removed, without prior notice
> + *
> + * Perform DMA mapping for devices in a container, mapping memory page-by-page.
> + *
> + * @param container_fd
> + *   the specified container fd. Use RTE_VFIO_DEFAULT_CONTAINER_FD to
> + *   use the default container.
> + *
> + * @param vaddr
> + *   Starting virtual address of memory to be mapped.
> + *
> + * @param iova
> + *   Starting IOVA address of memory to be mapped.
> + *
> + * @param len
> + *   Length of memory segment being mapped.
> + *
> + * @param pagesz
> + *   Page size of the underlying memory.
> + *
> + * @return
> + *    0 if successful
> + *   <0 if failed
> + */
> +__rte_experimental
> +int
> +rte_vfio_container_dma_map_paged(int container_fd, uint64_t vaddr,
> +		uint64_t iova, uint64_t len, uint64_t pagesz);
> +
>  /**
>   * Perform DMA unmapping for devices in a container.
>   *
> diff --git a/lib/eal/linux/eal_vfio.c b/lib/eal/linux/eal_vfio.c
> index aa2087a2da..0f5b6c14af 100644
> --- a/lib/eal/linux/eal_vfio.c
> +++ b/lib/eal/linux/eal_vfio.c
> @@ -1872,11 +1872,12 @@ vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
>  
>  static int
>  container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
> -		uint64_t len)
> +		uint64_t len, uint64_t pagesz)
>  {
>  	struct user_mem_map *new_map;
>  	struct user_mem_maps *user_mem_maps;
>  	bool has_partial_unmap;
> +	uint64_t chunk_size;
>  	int ret = 0;
>  
>  	user_mem_maps = &vfio_cfg->mem_maps;
> @@ -1887,19 +1888,37 @@ container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
>  		ret = -1;
>  		goto out;
>  	}
> -	/* map the entry */
> -	if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
> -		/* technically, this will fail if there are currently no devices
> -		 * plugged in, even if a device were added later, this mapping
> -		 * might have succeeded. however, since we cannot verify if this
> -		 * is a valid mapping without having a device attached, consider
> -		 * this to be unsupported, because we can't just store any old
> -		 * mapping and pollute list of active mappings willy-nilly.
> -		 */
> -		RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
> -		ret = -1;
> -		goto out;
> +
> +	/* technically, mapping will fail if there are currently no devices
> +	 * plugged in, even if a device were added later, this mapping might
> +	 * have succeeded. however, since we cannot verify if this is a valid
> +	 * mapping without having a device attached, consider this to be
> +	 * unsupported, because we can't just store any old mapping and pollute
> +	 * list of active mappings willy-nilly.
> +	 */
> +
> +	/* if page size was not specified, map the entire segment in one go */
> +	if (pagesz == 0) {
> +		if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
> +			RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
> +			ret = -1;
> +			goto out;
> +		}
> +	} else {
> +		/* otherwise, do mappings page-by-page */
> +		uint64_t offset;
> +
> +		for (offset = 0; offset < len; offset += pagesz) {
> +			uint64_t va = vaddr + offset;
> +			uint64_t io = iova + offset;
> +			if (vfio_dma_mem_map(vfio_cfg, va, io, pagesz, 1)) {
> +				RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
> +				ret = -1;
> +				goto out;
> +			}
> +		}
>  	}
> +
>  	/* do we have partial unmap support? */
>  	has_partial_unmap = vfio_cfg->vfio_iommu_type->partial_unmap;
>  
> @@ -1908,8 +1927,18 @@ container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
>  	new_map->addr = vaddr;
>  	new_map->iova = iova;
>  	new_map->len = len;
> -	/* for IOMMU types supporting partial unmap, we don't need chunking */
> -	new_map->chunk = has_partial_unmap ? 0 : len;
> +
> +	/*
> +	 * Chunking essentially serves largely the same purpose as page sizes,
> +	 * so for the purposes of this calculation, we treat them as the same.
> +	 * The reason we have page sizes is because we want to map things in a
> +	 * way that allows us to partially unmap later. Therefore, when IOMMU
> +	 * supports partial unmap, page size is irrelevant and can be ignored.
> +	 * For IOMMU that don't support partial unmap, page size is equivalent
> +	 * to chunk size.
> +	 */
> +	chunk_size = pagesz == 0 ? len : pagesz;
> +	new_map->chunk = has_partial_unmap ? 0 : chunk_size;
>  
>  	compact_user_maps(user_mem_maps);
>  out:
> @@ -2179,7 +2208,44 @@ rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
>  		return -1;
>  	}
>  
> -	return container_dma_map(vfio_cfg, vaddr, iova, len);
> +	/* not having page size means we map entire segment */
> +	return container_dma_map(vfio_cfg, vaddr, iova, len, 0);
> +}
> +
> +int
> +rte_vfio_container_dma_map_paged(int container_fd, uint64_t vaddr,
> +		uint64_t iova, uint64_t len, uint64_t pagesz)
> +{
> +	struct vfio_config *vfio_cfg;
> +	bool va_aligned, iova_aligned, len_aligned, pagesz_aligned;
> +
> +	/* avoid divide by zero first */
> +	if (len == 0 || pagesz == 0) {
> +		rte_errno = EINVAL;
> +		return -1;
> +	}
> +	/*
> +	 * for direct maps, we leave it up to the kernel to decide whether the
> +	 * parameters are valid, but since we are going to map page-by-page and
> +	 * are expecting certain guarantees from the parameters we receive, we
> +	 * are going to check those here.
> +	 */
> +	pagesz_aligned = rte_align64pow2(pagesz) == pagesz;
> +	len_aligned = (len % pagesz) == 0;
> +	va_aligned = RTE_ALIGN(vaddr, pagesz) == pagesz;
> +	iova_aligned = RTE_ALIGN(iova, pagesz) == pagesz;
> +	if (!pagesz_aligned || !len_aligned || !va_aligned || !iova_aligned) {
> +		rte_errno = EINVAL;
> +		return -1;
> +	}
> +
> +	vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
> +	if (vfio_cfg == NULL) {
> +		RTE_LOG(ERR, EAL, "Invalid VFIO container fd\n");
> +		return -1;
> +	}
> +
> +	return container_dma_map(vfio_cfg, vaddr, iova, len, pagesz);
>  }
>  
>  int
> @@ -2299,6 +2365,16 @@ rte_vfio_container_dma_map(__rte_unused int container_fd,
>  	return -1;
>  }
>  
> +int
> +rte_vfio_container_dma_map_paged(__rte_unused int container_fd,
> +		__rte_unused uint64_t vaddr,
> +		__rte_unused uint64_t iova,
> +		__rte_unused uint64_t len,
> +		__rte_unused uint64_t pagesz)
> +{
> +	return -1;
> +}
> +
>  int
>  rte_vfio_container_dma_unmap(__rte_unused int container_fd,
>  		__rte_unused uint64_t vaddr,
> diff --git a/lib/eal/version.map b/lib/eal/version.map
> index 9d43655b66..82a34c2014 100644
> --- a/lib/eal/version.map
> +++ b/lib/eal/version.map
> @@ -420,6 +420,7 @@ EXPERIMENTAL {
>  	rte_intr_instance_free;
>  	rte_intr_type_get;
>  	rte_intr_type_set;

Pls anontate with the target release in a comment. 

> +	rte_vfio_container_dma_map_paged;
>  };
>  
>  INTERNAL {
> diff --git a/lib/eal/windows/eal.c b/lib/eal/windows/eal.c
> index f7ce1b6671..93771cb230 100644
> --- a/lib/eal/windows/eal.c
> +++ b/lib/eal/windows/eal.c
> @@ -484,6 +484,17 @@ rte_vfio_container_dma_map(__rte_unused int container_fd,
>  	return -1;
>  }
>  
> +int
> +rte_vfio_container_dma_map_paged(__rte_unused int container_fd,
> +		__rte_unused uint64_t vaddr,
> +		__rte_unused uint64_t iova,
> +		__rte_unused uint64_t len,
> +		__rte_unused uint64_t pagesz)
> +{
> +	rte_errno = ENOTSUP;
> +	return -1;
> +}
> +
>  int
>  rte_vfio_container_dma_unmap(__rte_unused int container_fd,
>  			__rte_unused uint64_t vaddr,
  

On Thu, 2 Jun 2022 10:23:07 +0200
David Marchand <david.marchand@redhat.com> wrote:

> Hello Anatoly,
> 
> On Tue, Nov 2, 2021 at 4:54 PM Burakov, Anatoly
> <anatoly.burakov@intel.com> wrote:
> >
> > On 28-Oct-21 3:09 PM, Anatoly Burakov wrote:  
> > > Currently, there is no way to map memory for DMA in a way that allows
> > > unmapping it partially later, because some IOMMU's do not support
> > > partial unmapping. There is a workaround of mapping all of these
> > > segments separately, but this is inconvenient and silly, so this
> > > commit adds a proper API that does it.
> > >
> > > This commit relies on earlier infrastructure that was built out to
> > > support "chunking", as the concept of "chunks" is essentially the same
> > > as page size.
> > >
> > > Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
> > > ---
> > >
> > > Notes:
> > >      v2:
> > >      - Added checks for page alignment of supplied values
> > >      - Added rte_errno values (similar to patch 103165 [1])
> > >
> > >      [1] http://patches.dpdk.org/project/dpdk/patch/e8c5e7ba089e2283c3cd67e4529e52fe49390eb9.1635428963.git.anatoly.burakov@intel.com/  
> >
> > I think this API also needs VA and IOVA addresses in an array, so that
> > it's possible to map IOVA-discontiguous segments. This is too late for
> > this release, so let's postpone it till 22.02.  
> 
> It seems this patch fell through the cracks.
> Any update?
> 
> Thanks.

It is not clear what problem this is trying to solve.
Is it for some hardware platform that is not available?
Any updates on this patch or should it just be dropped?
  

On 7/4/2023 12:48 AM, Stephen Hemminger wrote:
> On Thu, 2 Jun 2022 10:23:07 +0200
> David Marchand <david.marchand@redhat.com> wrote:
> 
>> Hello Anatoly,
>>
>> On Tue, Nov 2, 2021 at 4:54 PM Burakov, Anatoly
>> <anatoly.burakov@intel.com> wrote:
>>>
>>> On 28-Oct-21 3:09 PM, Anatoly Burakov wrote:
>>>> Currently, there is no way to map memory for DMA in a way that allows
>>>> unmapping it partially later, because some IOMMU's do not support
>>>> partial unmapping. There is a workaround of mapping all of these
>>>> segments separately, but this is inconvenient and silly, so this
>>>> commit adds a proper API that does it.
>>>>
>>>> This commit relies on earlier infrastructure that was built out to
>>>> support "chunking", as the concept of "chunks" is essentially the same
>>>> as page size.
>>>>
>>>> Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
>>>> ---
>>>>
>>>> Notes:
>>>>       v2:
>>>>       - Added checks for page alignment of supplied values
>>>>       - Added rte_errno values (similar to patch 103165 [1])
>>>>
>>>>       [1] http://patches.dpdk.org/project/dpdk/patch/e8c5e7ba089e2283c3cd67e4529e52fe49390eb9.1635428963.git.anatoly.burakov@intel.com/
>>>
>>> I think this API also needs VA and IOVA addresses in an array, so that
>>> it's possible to map IOVA-discontiguous segments. This is too late for
>>> this release, so let's postpone it till 22.02.
>>
>> It seems this patch fell through the cracks.
>> Any update?
>>
>> Thanks.
> 
> It is not clear what problem this is trying to solve.
> Is it for some hardware platform that is not available?
> Any updates on this patch or should it just be dropped?
> 

This should probably be dropped, because I believe there was already 
another patch that makes all mappings as page-by-page.