* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
[not found] ` <562E5AE4.9070001@arm.com>
@ 2015-10-30 1:17 ` Daniel Kurtz
2015-10-30 14:09 ` Joerg Roedel
` (2 more replies)
0 siblings, 3 replies; 15+ messages in thread
From: Daniel Kurtz @ 2015-10-30 1:17 UTC (permalink / raw)
To: Robin Murphy, Pawel Osciak
Cc: Yong Wu, Joerg Roedel, Will Deacon, Catalin Marinas,
open list:IOMMU DRIVERS, linux-arm-kernel, thunder.leizhen,
Yingjoe Chen, laurent.pinchart+renesas, Thierry Reding,
Lin PoChun, Bobby Batacharia (via Google Docs),
linux-media, Marek Szyprowski, Kyungmin Park
+linux-media & VIDEOBUF2 FRAMEWORK maintainers since this is about the
v4l2-contig's usage of the DMA API.
Hi Robin,
On Tue, Oct 27, 2015 at 12:55 AM, Robin Murphy <robin.murphy@arm.com> wrote:
> On 26/10/15 13:44, Yong Wu wrote:
>>
>> On Thu, 2015-10-01 at 20:13 +0100, Robin Murphy wrote:
>> [...]
>>>
>>> +/*
>>> + * The DMA API client is passing in a scatterlist which could describe
>>> + * any old buffer layout, but the IOMMU API requires everything to be
>>> + * aligned to IOMMU pages. Hence the need for this complicated bit of
>>> + * impedance-matching, to be able to hand off a suitably-aligned list,
>>> + * but still preserve the original offsets and sizes for the caller.
>>> + */
>>> +int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
>>> + int nents, int prot)
>>> +{
>>> + struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
>>> + struct iova_domain *iovad = domain->iova_cookie;
>>> + struct iova *iova;
>>> + struct scatterlist *s, *prev = NULL;
>>> + dma_addr_t dma_addr;
>>> + size_t iova_len = 0;
>>> + int i;
>>> +
>>> + /*
>>> + * Work out how much IOVA space we need, and align the segments
>>> to
>>> + * IOVA granules for the IOMMU driver to handle. With some clever
>>> + * trickery we can modify the list in-place, but reversibly, by
>>> + * hiding the original data in the as-yet-unused DMA fields.
>>> + */
>>> + for_each_sg(sg, s, nents, i) {
>>> + size_t s_offset = iova_offset(iovad, s->offset);
>>> + size_t s_length = s->length;
>>> +
>>> + sg_dma_address(s) = s->offset;
>>> + sg_dma_len(s) = s_length;
>>> + s->offset -= s_offset;
>>> + s_length = iova_align(iovad, s_length + s_offset);
>>> + s->length = s_length;
>>> +
>>> + /*
>>> + * The simple way to avoid the rare case of a segment
>>> + * crossing the boundary mask is to pad the previous one
>>> + * to end at a naturally-aligned IOVA for this one's
>>> size,
>>> + * at the cost of potentially over-allocating a little.
>>> + */
>>> + if (prev) {
>>> + size_t pad_len = roundup_pow_of_two(s_length);
>>> +
>>> + pad_len = (pad_len - iova_len) & (pad_len - 1);
>>> + prev->length += pad_len;
>>
>>
>> Hi Robin,
>> While our v4l2 testing, It seems that we met a problem here.
>> Here we update prev->length again, Do we need update
>> sg_dma_len(prev) again too?
>>
>> Some function like vb2_dc_get_contiguous_size[1] always get
>> sg_dma_len(s) to compare instead of s->length. so it may break
>> unexpectedly while sg_dma_len(s) is not same with s->length.
>
>
> This is just tweaking the faked-up length that we hand off to iommu_map_sg()
> (see also the iova_align() above), to trick it into bumping this segment up
> to a suitable starting IOVA. The real length at this point is stashed in
> sg_dma_len(s), and will be copied back into s->length in __finalise_sg(), so
> both will hold the same true length once we return to the caller.
>
> Yes, it does mean that if you have a list where the segment lengths are page
> aligned but not monotonically decreasing, e.g. {64k, 16k, 64k}, then you'll
> still end up with a gap between the second and third segments, but that's
> fine because the DMA API offers no guarantees about what the resulting DMA
> addresses will be (consider the no-IOMMU case where they would each just be
> "mapped" to their physical address). If that breaks v4l, then it's probably
> v4l's DMA API use that needs looking at (again).
Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
memory pages into a contiguous block in device memory address space.
This would allow passing a dma mapped buffer to device dma using just
a device address and length.
IIUC, the change above breaks this model by inserting gaps in how the
buffer is mapped to device memory, such that the buffer is no longer
contiguous in dma address space.
Here is the code in question from
drivers/media/v4l2-core/videobuf2-dma-contig.c :
static unsigned long vb2_dc_get_contiguous_size(struct sg_table *sgt)
{
struct scatterlist *s;
dma_addr_t expected = sg_dma_address(sgt->sgl);
unsigned int i;
unsigned long size = 0;
for_each_sg(sgt->sgl, s, sgt->nents, i) {
if (sg_dma_address(s) != expected)
break;
expected = sg_dma_address(s) + sg_dma_len(s);
size += sg_dma_len(s);
}
return size;
}
static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
unsigned long size, enum dma_data_direction dma_dir)
{
struct vb2_dc_conf *conf = alloc_ctx;
struct vb2_dc_buf *buf;
struct frame_vector *vec;
unsigned long offset;
int n_pages, i;
int ret = 0;
struct sg_table *sgt;
unsigned long contig_size;
unsigned long dma_align = dma_get_cache_alignment();
DEFINE_DMA_ATTRS(attrs);
dma_set_attr(DMA_ATTR_SKIP_CPU_SYNC, &attrs);
buf = kzalloc(sizeof *buf, GFP_KERNEL);
buf->dma_dir = dma_dir;
offset = vaddr & ~PAGE_MASK;
vec = vb2_create_framevec(vaddr, size, dma_dir == DMA_FROM_DEVICE);
buf->vec = vec;
n_pages = frame_vector_count(vec);
sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
ret = sg_alloc_table_from_pages(sgt, frame_vector_pages(vec), n_pages,
offset, size, GFP_KERNEL);
sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
buf->dma_dir, &attrs);
contig_size = vb2_dc_get_contiguous_size(sgt);
if (contig_size < size) {
<<<=== if the original buffer had sg entries that were not
aligned on the "natural" alignment for their size, the new arm64 iommu
core code inserts a 'gap' in the iommu mapping, which causes
vb2_dc_get_contiguous_size() to exit early (and return a smaller size
than expected).
pr_err("contiguous mapping is too small %lu/%lu\n",
contig_size, size);
ret = -EFAULT;
goto fail_map_sg;
}
So, is the videobuf2-dma-contig.c based on an incorrect assumption
about how the DMA API is supposed to work?
Is it even possible to map a "contiguous-in-iova-range" mapping for a
buffer given as an sg_table with an arbitrary set of pages?
Thanks for helping to move this forward.
-Dan
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-10-30 1:17 ` [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping Daniel Kurtz
@ 2015-10-30 14:09 ` Joerg Roedel
2015-10-30 14:27 ` Robin Murphy
2015-11-17 12:02 ` Marek Szyprowski
2 siblings, 0 replies; 15+ messages in thread
From: Joerg Roedel @ 2015-10-30 14:09 UTC (permalink / raw)
To: Daniel Kurtz
Cc: Robin Murphy, Pawel Osciak, Yong Wu, Will Deacon,
Catalin Marinas, open list:IOMMU DRIVERS, linux-arm-kernel,
thunder.leizhen, Yingjoe Chen, laurent.pinchart+renesas,
Thierry Reding, Lin PoChun, Bobby Batacharia (via Google Docs),
linux-media, Marek Szyprowski, Kyungmin Park
On Fri, Oct 30, 2015 at 09:17:52AM +0800, Daniel Kurtz wrote:
> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
> memory pages into a contiguous block in device memory address space.
> This would allow passing a dma mapped buffer to device dma using just
> a device address and length.
If you are speaking of the dma_map_sg interface, than there is absolutly
no guarantee from the API side that the buffers you pass in will end up
mapped contiguously.
IOMMU drivers handle this differently, and when there is no IOMMU at all
there is also no way to map these buffers together.
> So, is the videobuf2-dma-contig.c based on an incorrect assumption
> about how the DMA API is supposed to work?
If it makes the above assumption, then yes.
Joerg
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-10-30 1:17 ` [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping Daniel Kurtz
2015-10-30 14:09 ` Joerg Roedel
@ 2015-10-30 14:27 ` Robin Murphy
2015-11-02 13:11 ` Daniel Kurtz
2015-11-17 12:02 ` Marek Szyprowski
2 siblings, 1 reply; 15+ messages in thread
From: Robin Murphy @ 2015-10-30 14:27 UTC (permalink / raw)
To: Daniel Kurtz, Pawel Osciak
Cc: Yong Wu, Joerg Roedel, Will Deacon, Catalin Marinas,
open list:IOMMU DRIVERS, linux-arm-kernel, thunder.leizhen,
Yingjoe Chen, laurent.pinchart+renesas, Thierry Reding,
Lin PoChun, Bobby Batacharia (via Google Docs),
linux-media, Marek Szyprowski, Kyungmin Park
Hi Dan,
On 30/10/15 01:17, Daniel Kurtz wrote:
> +linux-media & VIDEOBUF2 FRAMEWORK maintainers since this is about the
> v4l2-contig's usage of the DMA API.
>
> Hi Robin,
>
> On Tue, Oct 27, 2015 at 12:55 AM, Robin Murphy <robin.murphy@arm.com> wrote:
>> On 26/10/15 13:44, Yong Wu wrote:
>>>
>>> On Thu, 2015-10-01 at 20:13 +0100, Robin Murphy wrote:
>>> [...]
>>>>
>>>> +/*
>>>> + * The DMA API client is passing in a scatterlist which could describe
>>>> + * any old buffer layout, but the IOMMU API requires everything to be
>>>> + * aligned to IOMMU pages. Hence the need for this complicated bit of
>>>> + * impedance-matching, to be able to hand off a suitably-aligned list,
>>>> + * but still preserve the original offsets and sizes for the caller.
>>>> + */
>>>> +int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
>>>> + int nents, int prot)
>>>> +{
>>>> + struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
>>>> + struct iova_domain *iovad = domain->iova_cookie;
>>>> + struct iova *iova;
>>>> + struct scatterlist *s, *prev = NULL;
>>>> + dma_addr_t dma_addr;
>>>> + size_t iova_len = 0;
>>>> + int i;
>>>> +
>>>> + /*
>>>> + * Work out how much IOVA space we need, and align the segments
>>>> to
>>>> + * IOVA granules for the IOMMU driver to handle. With some clever
>>>> + * trickery we can modify the list in-place, but reversibly, by
>>>> + * hiding the original data in the as-yet-unused DMA fields.
>>>> + */
>>>> + for_each_sg(sg, s, nents, i) {
>>>> + size_t s_offset = iova_offset(iovad, s->offset);
>>>> + size_t s_length = s->length;
>>>> +
>>>> + sg_dma_address(s) = s->offset;
>>>> + sg_dma_len(s) = s_length;
>>>> + s->offset -= s_offset;
>>>> + s_length = iova_align(iovad, s_length + s_offset);
>>>> + s->length = s_length;
>>>> +
>>>> + /*
>>>> + * The simple way to avoid the rare case of a segment
>>>> + * crossing the boundary mask is to pad the previous one
>>>> + * to end at a naturally-aligned IOVA for this one's
>>>> size,
>>>> + * at the cost of potentially over-allocating a little.
>>>> + */
>>>> + if (prev) {
>>>> + size_t pad_len = roundup_pow_of_two(s_length);
>>>> +
>>>> + pad_len = (pad_len - iova_len) & (pad_len - 1);
>>>> + prev->length += pad_len;
>>>
>>>
>>> Hi Robin,
>>> While our v4l2 testing, It seems that we met a problem here.
>>> Here we update prev->length again, Do we need update
>>> sg_dma_len(prev) again too?
>>>
>>> Some function like vb2_dc_get_contiguous_size[1] always get
>>> sg_dma_len(s) to compare instead of s->length. so it may break
>>> unexpectedly while sg_dma_len(s) is not same with s->length.
>>
>>
>> This is just tweaking the faked-up length that we hand off to iommu_map_sg()
>> (see also the iova_align() above), to trick it into bumping this segment up
>> to a suitable starting IOVA. The real length at this point is stashed in
>> sg_dma_len(s), and will be copied back into s->length in __finalise_sg(), so
>> both will hold the same true length once we return to the caller.
>>
>> Yes, it does mean that if you have a list where the segment lengths are page
>> aligned but not monotonically decreasing, e.g. {64k, 16k, 64k}, then you'll
>> still end up with a gap between the second and third segments, but that's
>> fine because the DMA API offers no guarantees about what the resulting DMA
>> addresses will be (consider the no-IOMMU case where they would each just be
>> "mapped" to their physical address). If that breaks v4l, then it's probably
>> v4l's DMA API use that needs looking at (again).
>
> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
> memory pages into a contiguous block in device memory address space.
> This would allow passing a dma mapped buffer to device dma using just
> a device address and length.
Not at all. The streaming DMA API (dma_map_* and friends) has two
responsibilities: performing any necessary cache maintenance to ensure
the device will correctly see data from the CPU, and the CPU will
correctly see data from the device; and working out an address for that
buffer from the device's point of view to actually hand off to the
hardware (which is perfectly well allowed to fail).
Consider SWIOTLB's implementation - segments which already lie at
physical addresses within the device's DMA mask just get passed through,
while those that lie outside it get mapped into the bounce buffer, but
still as individual allocations (arch code just handles cache
maintenance on the resulting physical addresses and can apply any
hard-wired DMA offset for the device concerned).
> IIUC, the change above breaks this model by inserting gaps in how the
> buffer is mapped to device memory, such that the buffer is no longer
> contiguous in dma address space.
Even the existing arch/arm IOMMU DMA code which I guess this implicitly
relies on doesn't guarantee that behaviour - if the mapping happens to
reach one of the segment length/boundary limits it won't just leave a
gap, it'll start an entirely new IOVA allocation which could well start
at a wildly different address[0].
> Here is the code in question from
> drivers/media/v4l2-core/videobuf2-dma-contig.c :
>
> static unsigned long vb2_dc_get_contiguous_size(struct sg_table *sgt)
> {
> struct scatterlist *s;
> dma_addr_t expected = sg_dma_address(sgt->sgl);
> unsigned int i;
> unsigned long size = 0;
>
> for_each_sg(sgt->sgl, s, sgt->nents, i) {
> if (sg_dma_address(s) != expected)
> break;
> expected = sg_dma_address(s) + sg_dma_len(s);
> size += sg_dma_len(s);
> }
> return size;
> }
>
>
> static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
> unsigned long size, enum dma_data_direction dma_dir)
> {
> struct vb2_dc_conf *conf = alloc_ctx;
> struct vb2_dc_buf *buf;
> struct frame_vector *vec;
> unsigned long offset;
> int n_pages, i;
> int ret = 0;
> struct sg_table *sgt;
> unsigned long contig_size;
> unsigned long dma_align = dma_get_cache_alignment();
> DEFINE_DMA_ATTRS(attrs);
>
> dma_set_attr(DMA_ATTR_SKIP_CPU_SYNC, &attrs);
>
> buf = kzalloc(sizeof *buf, GFP_KERNEL);
> buf->dma_dir = dma_dir;
>
> offset = vaddr & ~PAGE_MASK;
> vec = vb2_create_framevec(vaddr, size, dma_dir == DMA_FROM_DEVICE);
> buf->vec = vec;
> n_pages = frame_vector_count(vec);
>
> sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
>
> ret = sg_alloc_table_from_pages(sgt, frame_vector_pages(vec), n_pages,
> offset, size, GFP_KERNEL);
>
> sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
> buf->dma_dir, &attrs);
>
> contig_size = vb2_dc_get_contiguous_size(sgt);
(as an aside, it's rather unintuitive that the handling of the
dma_map_sg call actually failing is entirely implicit here)
> if (contig_size < size) {
>
> <<<=== if the original buffer had sg entries that were not
> aligned on the "natural" alignment for their size, the new arm64 iommu
> core code inserts a 'gap' in the iommu mapping, which causes
> vb2_dc_get_contiguous_size() to exit early (and return a smaller size
> than expected).
>
> pr_err("contiguous mapping is too small %lu/%lu\n",
> contig_size, size);
> ret = -EFAULT;
> goto fail_map_sg;
> }
>
>
> So, is the videobuf2-dma-contig.c based on an incorrect assumption
> about how the DMA API is supposed to work?
> Is it even possible to map a "contiguous-in-iova-range" mapping for a
> buffer given as an sg_table with an arbitrary set of pages?
From the Streaming DMA mappings section of Documentation/DMA-API.txt:
Note also that the above constraints on physical contiguity and
dma_mask may not apply if the platform has an IOMMU (a device which
maps an I/O DMA address to a physical memory address). However, to be
portable, device driver writers may *not* assume that such an IOMMU
exists.
There's not strictly any harm in using the DMA API this way and *hoping*
you get what you want, as long as you're happy for it to fail pretty
much 100% of the time on some systems, and still in a minority of corner
cases on any system. However, if there's a real dependency on IOMMUs and
tight control of IOVA allocation here, then the DMA API isn't really the
right tool for the job, and maybe it's time to start looking to how to
better fit these multimedia-subsystem-type use cases into the IOMMU API
- as far as I understand it there's at least some conceptual overlap
with the HSA PASID stuff being prototyped in PCI/x86-land at the moment,
so it could be an apposite time to try and bang out some common
requirements.
Robin.
[0]:http://article.gmane.org/gmane.linux.kernel.iommu/11185
>
> Thanks for helping to move this forward.
>
> -Dan
>
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-10-30 14:27 ` Robin Murphy
@ 2015-11-02 13:11 ` Daniel Kurtz
2015-11-02 13:43 ` Tomasz Figa
0 siblings, 1 reply; 15+ messages in thread
From: Daniel Kurtz @ 2015-11-02 13:11 UTC (permalink / raw)
To: Robin Murphy
Cc: Lin PoChun, linux-arm-kernel, Yingjoe Chen, Will Deacon,
linux-media, Thierry Reding, open list:IOMMU DRIVERS,
Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
laurent.pinchart+renesas, Joerg Roedel, thunder.leizhen,
Catalin Marinas, Tomasz Figa, Russell King, linux-mediatek
+Tomasz, so he can reply to the thread
+Marek and Russell as recommended by Tomasz
On Oct 30, 2015 22:27, "Robin Murphy" <robin.murphy@arm.com> wrote:
>
> Hi Dan,
>
> On 30/10/15 01:17, Daniel Kurtz wrote:
>>
>> +linux-media & VIDEOBUF2 FRAMEWORK maintainers since this is about the
>> v4l2-contig's usage of the DMA API.
>>
>> Hi Robin,
>>
>> On Tue, Oct 27, 2015 at 12:55 AM, Robin Murphy <robin.murphy@arm.com> wrote:
>>>
>>> On 26/10/15 13:44, Yong Wu wrote:
>>>>
>>>>
>>>> On Thu, 2015-10-01 at 20:13 +0100, Robin Murphy wrote:
>>>> [...]
>>>>>
>>>>>
>>>>> +/*
>>>>> + * The DMA API client is passing in a scatterlist which could describe
>>>>> + * any old buffer layout, but the IOMMU API requires everything to be
>>>>> + * aligned to IOMMU pages. Hence the need for this complicated bit of
>>>>> + * impedance-matching, to be able to hand off a suitably-aligned list,
>>>>> + * but still preserve the original offsets and sizes for the caller.
>>>>> + */
>>>>> +int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
>>>>> + int nents, int prot)
>>>>> +{
>>>>> + struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
>>>>> + struct iova_domain *iovad = domain->iova_cookie;
>>>>> + struct iova *iova;
>>>>> + struct scatterlist *s, *prev = NULL;
>>>>> + dma_addr_t dma_addr;
>>>>> + size_t iova_len = 0;
>>>>> + int i;
>>>>> +
>>>>> + /*
>>>>> + * Work out how much IOVA space we need, and align the segments
>>>>> to
>>>>> + * IOVA granules for the IOMMU driver to handle. With some clever
>>>>> + * trickery we can modify the list in-place, but reversibly, by
>>>>> + * hiding the original data in the as-yet-unused DMA fields.
>>>>> + */
>>>>> + for_each_sg(sg, s, nents, i) {
>>>>> + size_t s_offset = iova_offset(iovad, s->offset);
>>>>> + size_t s_length = s->length;
>>>>> +
>>>>> + sg_dma_address(s) = s->offset;
>>>>> + sg_dma_len(s) = s_length;
>>>>> + s->offset -= s_offset;
>>>>> + s_length = iova_align(iovad, s_length + s_offset);
>>>>> + s->length = s_length;
>>>>> +
>>>>> + /*
>>>>> + * The simple way to avoid the rare case of a segment
>>>>> + * crossing the boundary mask is to pad the previous one
>>>>> + * to end at a naturally-aligned IOVA for this one's
>>>>> size,
>>>>> + * at the cost of potentially over-allocating a little.
>>>>> + */
>>>>> + if (prev) {
>>>>> + size_t pad_len = roundup_pow_of_two(s_length);
>>>>> +
>>>>> + pad_len = (pad_len - iova_len) & (pad_len - 1);
>>>>> + prev->length += pad_len;
>>>>
>>>>
>>>>
>>>> Hi Robin,
>>>> While our v4l2 testing, It seems that we met a problem here.
>>>> Here we update prev->length again, Do we need update
>>>> sg_dma_len(prev) again too?
>>>>
>>>> Some function like vb2_dc_get_contiguous_size[1] always get
>>>> sg_dma_len(s) to compare instead of s->length. so it may break
>>>> unexpectedly while sg_dma_len(s) is not same with s->length.
>>>
>>>
>>>
>>> This is just tweaking the faked-up length that we hand off to iommu_map_sg()
>>> (see also the iova_align() above), to trick it into bumping this segment up
>>> to a suitable starting IOVA. The real length at this point is stashed in
>>> sg_dma_len(s), and will be copied back into s->length in __finalise_sg(), so
>>> both will hold the same true length once we return to the caller.
>>>
>>> Yes, it does mean that if you have a list where the segment lengths are page
>>> aligned but not monotonically decreasing, e.g. {64k, 16k, 64k}, then you'll
>>> still end up with a gap between the second and third segments, but that's
>>> fine because the DMA API offers no guarantees about what the resulting DMA
>>> addresses will be (consider the no-IOMMU case where they would each just be
>>> "mapped" to their physical address). If that breaks v4l, then it's probably
>>> v4l's DMA API use that needs looking at (again).
>>
>>
>> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
>> memory pages into a contiguous block in device memory address space.
>> This would allow passing a dma mapped buffer to device dma using just
>> a device address and length.
>
>
> Not at all. The streaming DMA API (dma_map_* and friends) has two responsibilities: performing any necessary cache maintenance to ensure the device will correctly see data from the CPU, and the CPU will correctly see data from the device; and working out an address for that buffer from the device's point of view to actually hand off to the hardware (which is perfectly well allowed to fail).
>
> Consider SWIOTLB's implementation - segments which already lie at physical addresses within the device's DMA mask just get passed through, while those that lie outside it get mapped into the bounce buffer, but still as individual allocations (arch code just handles cache maintenance on the resulting physical addresses and can apply any hard-wired DMA offset for the device concerned).
>
>> IIUC, the change above breaks this model by inserting gaps in how the
>> buffer is mapped to device memory, such that the buffer is no longer
>> contiguous in dma address space.
>
>
> Even the existing arch/arm IOMMU DMA code which I guess this implicitly relies on doesn't guarantee that behaviour - if the mapping happens to reach one of the segment length/boundary limits it won't just leave a gap, it'll start an entirely new IOVA allocation which could well start at a wildly different address[0].
>
>> Here is the code in question from
>> drivers/media/v4l2-core/videobuf2-dma-contig.c :
>>
>> static unsigned long vb2_dc_get_contiguous_size(struct sg_table *sgt)
>> {
>> struct scatterlist *s;
>> dma_addr_t expected = sg_dma_address(sgt->sgl);
>> unsigned int i;
>> unsigned long size = 0;
>>
>> for_each_sg(sgt->sgl, s, sgt->nents, i) {
>> if (sg_dma_address(s) != expected)
>> break;
>> expected = sg_dma_address(s) + sg_dma_len(s);
>> size += sg_dma_len(s);
>> }
>> return size;
>> }
>>
>>
>> static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
>> unsigned long size, enum dma_data_direction dma_dir)
>> {
>> struct vb2_dc_conf *conf = alloc_ctx;
>> struct vb2_dc_buf *buf;
>> struct frame_vector *vec;
>> unsigned long offset;
>> int n_pages, i;
>> int ret = 0;
>> struct sg_table *sgt;
>> unsigned long contig_size;
>> unsigned long dma_align = dma_get_cache_alignment();
>> DEFINE_DMA_ATTRS(attrs);
>>
>> dma_set_attr(DMA_ATTR_SKIP_CPU_SYNC, &attrs);
>>
>> buf = kzalloc(sizeof *buf, GFP_KERNEL);
>> buf->dma_dir = dma_dir;
>>
>> offset = vaddr & ~PAGE_MASK;
>> vec = vb2_create_framevec(vaddr, size, dma_dir == DMA_FROM_DEVICE);
>> buf->vec = vec;
>> n_pages = frame_vector_count(vec);
>>
>> sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
>>
>> ret = sg_alloc_table_from_pages(sgt, frame_vector_pages(vec), n_pages,
>> offset, size, GFP_KERNEL);
>>
>> sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
>> buf->dma_dir, &attrs);
>>
>> contig_size = vb2_dc_get_contiguous_size(sgt);
>
>
> (as an aside, it's rather unintuitive that the handling of the dma_map_sg call actually failing is entirely implicit here)
>
>> if (contig_size < size) {
>>
>> <<<=== if the original buffer had sg entries that were not
>> aligned on the "natural" alignment for their size, the new arm64 iommu
>> core code inserts a 'gap' in the iommu mapping, which causes
>> vb2_dc_get_contiguous_size() to exit early (and return a smaller size
>> than expected).
>>
>> pr_err("contiguous mapping is too small %lu/%lu\n",
>> contig_size, size);
>> ret = -EFAULT;
>> goto fail_map_sg;
>> }
>>
>>
>> So, is the videobuf2-dma-contig.c based on an incorrect assumption
>> about how the DMA API is supposed to work?
>> Is it even possible to map a "contiguous-in-iova-range" mapping for a
>> buffer given as an sg_table with an arbitrary set of pages?
>
>
> From the Streaming DMA mappings section of Documentation/DMA-API.txt:
>
> Note also that the above constraints on physical contiguity and
> dma_mask may not apply if the platform has an IOMMU (a device which
> maps an I/O DMA address to a physical memory address). However, to be
> portable, device driver writers may *not* assume that such an IOMMU
> exists.
>
> There's not strictly any harm in using the DMA API this way and *hoping* you get what you want, as long as you're happy for it to fail pretty much 100% of the time on some systems, and still in a minority of corner cases on any system. However, if there's a real dependency on IOMMUs and tight control of IOVA allocation here, then the DMA API isn't really the right tool for the job, and maybe it's time to start looking to how to better fit these multimedia-subsystem-type use cases into the IOMMU API - as far as I understand it there's at least some conceptual overlap with the HSA PASID stuff being prototyped in PCI/x86-land at the moment, so it could be an apposite time to try and bang out some common requirements.
>
> Robin.
>
> [0]:http://article.gmane.org/gmane.linux.kernel.iommu/11185
>
>>
>> Thanks for helping to move this forward.
>>
>> -Dan
>>
>
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-02 13:11 ` Daniel Kurtz
@ 2015-11-02 13:43 ` Tomasz Figa
2015-11-03 17:41 ` Robin Murphy
0 siblings, 1 reply; 15+ messages in thread
From: Tomasz Figa @ 2015-11-02 13:43 UTC (permalink / raw)
To: Daniel Kurtz
Cc: Robin Murphy, Lin PoChun, linux-arm-kernel, Yingjoe Chen,
Will Deacon, linux-media, Thierry Reding,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
Laurent Pinchart, Joerg Roedel, thunder.leizhen, Catalin Marinas,
Russell King, linux-mediatek
On Mon, Nov 2, 2015 at 10:11 PM, Daniel Kurtz <djkurtz@chromium.org> wrote:
>
> +Tomasz, so he can reply to the thread
> +Marek and Russell as recommended by Tomasz
>
> On Oct 30, 2015 22:27, "Robin Murphy" <robin.murphy@arm.com> wrote:
> >
> > Hi Dan,
> >
> > On 30/10/15 01:17, Daniel Kurtz wrote:
> >>
> >> +linux-media & VIDEOBUF2 FRAMEWORK maintainers since this is about the
> >> v4l2-contig's usage of the DMA API.
> >>
> >> Hi Robin,
> >>
> >> On Tue, Oct 27, 2015 at 12:55 AM, Robin Murphy <robin.murphy@arm.com> wrote:
> >>>
> >>> On 26/10/15 13:44, Yong Wu wrote:
> >>>>
> >>>>
> >>>> On Thu, 2015-10-01 at 20:13 +0100, Robin Murphy wrote:
> >>>> [...]
> >>>>>
> >>>>>
> >>>>> +/*
> >>>>> + * The DMA API client is passing in a scatterlist which could describe
> >>>>> + * any old buffer layout, but the IOMMU API requires everything to be
> >>>>> + * aligned to IOMMU pages. Hence the need for this complicated bit of
> >>>>> + * impedance-matching, to be able to hand off a suitably-aligned list,
> >>>>> + * but still preserve the original offsets and sizes for the caller.
> >>>>> + */
> >>>>> +int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
> >>>>> + int nents, int prot)
> >>>>> +{
> >>>>> + struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
> >>>>> + struct iova_domain *iovad = domain->iova_cookie;
> >>>>> + struct iova *iova;
> >>>>> + struct scatterlist *s, *prev = NULL;
> >>>>> + dma_addr_t dma_addr;
> >>>>> + size_t iova_len = 0;
> >>>>> + int i;
> >>>>> +
> >>>>> + /*
> >>>>> + * Work out how much IOVA space we need, and align the segments
> >>>>> to
> >>>>> + * IOVA granules for the IOMMU driver to handle. With some clever
> >>>>> + * trickery we can modify the list in-place, but reversibly, by
> >>>>> + * hiding the original data in the as-yet-unused DMA fields.
> >>>>> + */
> >>>>> + for_each_sg(sg, s, nents, i) {
> >>>>> + size_t s_offset = iova_offset(iovad, s->offset);
> >>>>> + size_t s_length = s->length;
> >>>>> +
> >>>>> + sg_dma_address(s) = s->offset;
> >>>>> + sg_dma_len(s) = s_length;
> >>>>> + s->offset -= s_offset;
> >>>>> + s_length = iova_align(iovad, s_length + s_offset);
> >>>>> + s->length = s_length;
> >>>>> +
> >>>>> + /*
> >>>>> + * The simple way to avoid the rare case of a segment
> >>>>> + * crossing the boundary mask is to pad the previous one
> >>>>> + * to end at a naturally-aligned IOVA for this one's
> >>>>> size,
> >>>>> + * at the cost of potentially over-allocating a little.
I'd like to know what is the boundary mask and what hardware imposes
requirements like this. The cost here is not only over-allocating a
little, but making many, many buffers contiguously mappable on the
CPU, unmappable contiguously in IOMMU, which just defeats the purpose
of having an IOMMU, which I believe should be there for simple IP
blocks taking one DMA address to be able to view the buffer the same
way as the CPU.
> >>>>> + */
> >>>>> + if (prev) {
> >>>>> + size_t pad_len = roundup_pow_of_two(s_length);
> >>>>> +
> >>>>> + pad_len = (pad_len - iova_len) & (pad_len - 1);
> >>>>> + prev->length += pad_len;
> >>>>
> >>>>
> >>>>
> >>>> Hi Robin,
> >>>> While our v4l2 testing, It seems that we met a problem here.
> >>>> Here we update prev->length again, Do we need update
> >>>> sg_dma_len(prev) again too?
> >>>>
> >>>> Some function like vb2_dc_get_contiguous_size[1] always get
> >>>> sg_dma_len(s) to compare instead of s->length. so it may break
> >>>> unexpectedly while sg_dma_len(s) is not same with s->length.
> >>>
> >>>
> >>>
> >>> This is just tweaking the faked-up length that we hand off to iommu_map_sg()
> >>> (see also the iova_align() above), to trick it into bumping this segment up
> >>> to a suitable starting IOVA. The real length at this point is stashed in
> >>> sg_dma_len(s), and will be copied back into s->length in __finalise_sg(), so
> >>> both will hold the same true length once we return to the caller.
> >>>
> >>> Yes, it does mean that if you have a list where the segment lengths are page
> >>> aligned but not monotonically decreasing, e.g. {64k, 16k, 64k}, then you'll
> >>> still end up with a gap between the second and third segments, but that's
> >>> fine because the DMA API offers no guarantees about what the resulting DMA
> >>> addresses will be (consider the no-IOMMU case where they would each just be
> >>> "mapped" to their physical address). If that breaks v4l, then it's probably
> >>> v4l's DMA API use that needs looking at (again).
> >>
> >>
> >> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
> >> memory pages into a contiguous block in device memory address space.
> >> This would allow passing a dma mapped buffer to device dma using just
> >> a device address and length.
> >
> >
> > Not at all. The streaming DMA API (dma_map_* and friends) has two responsibilities: performing any necessary cache maintenance to ensure the device will correctly see data from the CPU, and the CPU will correctly see data from the device; and working out an address for that buffer from the device's point of view to actually hand off to the hardware (which is perfectly well allowed to fail).
Agreed. The dma_map_*() API is not guaranteed to return a single
contiguous part of virtual address space for any given SG list.
However it was understood to be able to map buffers contiguously
mappable by the CPU into a single segment and users,
videobuf2-dma-contig in particular, relied on this.
> >
> > Consider SWIOTLB's implementation - segments which already lie at physical addresses within the device's DMA mask just get passed through, while those that lie outside it get mapped into the bounce buffer, but still as individual allocations (arch code just handles cache maintenance on the resulting physical addresses and can apply any hard-wired DMA offset for the device concerned).
And this is fine for vb2-dma-contig, which was made for devices that
require buffers contiguous in its address space. Without IOMMU it will
allow only physically contiguous buffers and fails otherwise, which is
fine, because it's a hardware requirement.
> >
> >> IIUC, the change above breaks this model by inserting gaps in how the
> >> buffer is mapped to device memory, such that the buffer is no longer
> >> contiguous in dma address space.
> >
> >
> > Even the existing arch/arm IOMMU DMA code which I guess this implicitly relies on doesn't guarantee that behaviour - if the mapping happens to reach one of the segment length/boundary limits it won't just leave a gap, it'll start an entirely new IOVA allocation which could well start at a wildly different address[0].
Could you explain segment length/boundary limits and when buffers can
reach them? Sorry, i haven't been following all the discussions, but
I'm not aware of any similar requirements of the IOMMU hardware I
worked with.
> >
> >> Here is the code in question from
> >> drivers/media/v4l2-core/videobuf2-dma-contig.c :
> >>
> >> static unsigned long vb2_dc_get_contiguous_size(struct sg_table *sgt)
> >> {
> >> struct scatterlist *s;
> >> dma_addr_t expected = sg_dma_address(sgt->sgl);
> >> unsigned int i;
> >> unsigned long size = 0;
> >>
> >> for_each_sg(sgt->sgl, s, sgt->nents, i) {
> >> if (sg_dma_address(s) != expected)
> >> break;
> >> expected = sg_dma_address(s) + sg_dma_len(s);
> >> size += sg_dma_len(s);
> >> }
> >> return size;
> >> }
> >>
> >>
> >> static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
> >> unsigned long size, enum dma_data_direction dma_dir)
> >> {
> >> struct vb2_dc_conf *conf = alloc_ctx;
> >> struct vb2_dc_buf *buf;
> >> struct frame_vector *vec;
> >> unsigned long offset;
> >> int n_pages, i;
> >> int ret = 0;
> >> struct sg_table *sgt;
> >> unsigned long contig_size;
> >> unsigned long dma_align = dma_get_cache_alignment();
> >> DEFINE_DMA_ATTRS(attrs);
> >>
> >> dma_set_attr(DMA_ATTR_SKIP_CPU_SYNC, &attrs);
> >>
> >> buf = kzalloc(sizeof *buf, GFP_KERNEL);
> >> buf->dma_dir = dma_dir;
> >>
> >> offset = vaddr & ~PAGE_MASK;
> >> vec = vb2_create_framevec(vaddr, size, dma_dir == DMA_FROM_DEVICE);
> >> buf->vec = vec;
> >> n_pages = frame_vector_count(vec);
> >>
> >> sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
> >>
> >> ret = sg_alloc_table_from_pages(sgt, frame_vector_pages(vec), n_pages,
> >> offset, size, GFP_KERNEL);
> >>
> >> sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
> >> buf->dma_dir, &attrs);
> >>
> >> contig_size = vb2_dc_get_contiguous_size(sgt);
> >
> >
> > (as an aside, it's rather unintuitive that the handling of the dma_map_sg call actually failing is entirely implicit here)
I'm not sure what you mean, please elaborate. The code considers only
the case of contiguously mapping at least the requested size as a
success, because anything else is useless with the hardware.
> >
> >> if (contig_size < size) {
> >>
> >> <<<=== if the original buffer had sg entries that were not
> >> aligned on the "natural" alignment for their size, the new arm64 iommu
> >> core code inserts a 'gap' in the iommu mapping, which causes
> >> vb2_dc_get_contiguous_size() to exit early (and return a smaller size
> >> than expected).
> >>
> >> pr_err("contiguous mapping is too small %lu/%lu\n",
> >> contig_size, size);
> >> ret = -EFAULT;
> >> goto fail_map_sg;
> >> }
> >>
> >>
> >> So, is the videobuf2-dma-contig.c based on an incorrect assumption
> >> about how the DMA API is supposed to work?
> >> Is it even possible to map a "contiguous-in-iova-range" mapping for a
> >> buffer given as an sg_table with an arbitrary set of pages?
> >
> >
> > From the Streaming DMA mappings section of Documentation/DMA-API.txt:
> >
> > Note also that the above constraints on physical contiguity and
> > dma_mask may not apply if the platform has an IOMMU (a device which
> > maps an I/O DMA address to a physical memory address). However, to be
> > portable, device driver writers may *not* assume that such an IOMMU
> > exists.
> >
> > There's not strictly any harm in using the DMA API this way and *hoping* you get what you want, as long as you're happy for it to fail pretty much 100% of the time on some systems, and still in a minority of corner cases on any system.
Could you please elaborate? I'd like to see examples, because I can't
really imagine buffers mappable contiguously on CPU, but not on IOMMU.
Also, as I said, the hardware I worked with didn't suffer from
problems like this.
> > However, if there's a real dependency on IOMMUs and tight control of IOVA allocation here, then the DMA API isn't really the right tool for the job, and maybe it's time to start looking to how to better fit these multimedia-subsystem-type use cases into the IOMMU API - as far as I understand it there's at least some conceptual overlap with the HSA PASID stuff being prototyped in PCI/x86-land at the moment, so it could be an apposite time to try and bang out some common requirements.
The DMA API is actually the only good tool to use here to keep the
videobuf2-dma-contig code away from the knowledge about platform
specific data, e.g. presence of IOMMU. The only thing it knows is that
the target hardware requires a single contiguous buffer and it relies
on the fact that in correct cases the buffer given to it will meet
this requirement (i.e. physically contiguous w/o IOMMU; CPU mappable
with IOMMU).
Best regards,
Tomasz
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-02 13:43 ` Tomasz Figa
@ 2015-11-03 17:41 ` Robin Murphy
2015-11-03 18:40 ` Russell King - ARM Linux
2015-11-04 5:12 ` Tomasz Figa
0 siblings, 2 replies; 15+ messages in thread
From: Robin Murphy @ 2015-11-03 17:41 UTC (permalink / raw)
To: Tomasz Figa, Daniel Kurtz
Cc: Lin PoChun, linux-arm-kernel, Yingjoe Chen, Will Deacon,
linux-media, Thierry Reding, open list:IOMMU DRIVERS,
Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
Laurent Pinchart, Joerg Roedel, thunder.leizhen, Catalin Marinas,
Russell King, linux-mediatek
Hi Tomasz,
On 02/11/15 13:43, Tomasz Figa wrote:
> I'd like to know what is the boundary mask and what hardware imposes
> requirements like this. The cost here is not only over-allocating a
> little, but making many, many buffers contiguously mappable on the
> CPU, unmappable contiguously in IOMMU, which just defeats the purpose
> of having an IOMMU, which I believe should be there for simple IP
> blocks taking one DMA address to be able to view the buffer the same
> way as the CPU.
The expectation with dma_map_sg() is that you're either going to be
iterating over the buffer segments, handing off each address to the
device to process one by one; or you have a scatter-gather-capable
device, in which case you hand off the whole list at once. It's in the
latter case where you have to make sure the list doesn't exceed the
hardware limitations of that device. I believe the original concern was
disk controllers (the introduction of dma_parms seems to originate from
the linux-scsi list), but most scatter-gather engines are going to have
some limit on how much they can handle per entry (IMO the dmaengine
drivers are the easiest example to look at).
Segment boundaries are a little more arcane, but my assumption is that
they relate to the kind of devices whose addressing is not flat but
relative to some separate segment register (The "64-bit" mode of USB
EHCI is one concrete example I can think of) - since you cannot
realistically change the segment register while the device is in the
middle of accessing a single buffer entry, that entry must not fall
across a segment boundary or at some point the device's accesses are
going to overflow the offset address bits and wrap around to bogus
addresses at the bottom of the segment.
Now yes, it will be possible under _most_ circumstances to use an IOMMU
to lay out a list of segments with page-aligned lengths within a single
IOVA allocation whilst still meeting all the necessary constraints. It
just needs some unavoidably complicated calculations - quite likely
significantly more complex than my v5 version of map_sg() that tried to
do that and merge segments but failed to take the initial alignment into
account properly - since there are much simpler ways to enforce just the
_necessary_ behaviour for the DMA API, I put the complicated stuff to
one side for now to prevent it holding up getting the basic functional
support in place.
>>>> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
>>>> memory pages into a contiguous block in device memory address space.
>>>> This would allow passing a dma mapped buffer to device dma using just
>>>> a device address and length.
>>>
>>>
>>> Not at all. The streaming DMA API (dma_map_* and friends) has two responsibilities: performing any necessary cache maintenance to ensure the device will correctly see data from the CPU, and the CPU will correctly see data from the device; and working out an address for that buffer from the device's point of view to actually hand off to the hardware (which is perfectly well allowed to fail).
>
> Agreed. The dma_map_*() API is not guaranteed to return a single
> contiguous part of virtual address space for any given SG list.
> However it was understood to be able to map buffers contiguously
> mappable by the CPU into a single segment and users,
> videobuf2-dma-contig in particular, relied on this.
I don't follow that - _any_ buffer made of page-sized chunks is going to
be mappable contiguously by the CPU; it's clearly impossible for the
streaming DMA API itself to offer such a guarantee, because it's
entirely orthogonal to the presence or otherwise of an IOMMU.
Furthermore, I can't see any existing dma_map_sg implementation (between
arm/64 and x86, at least), that _won't_ break that expectation under
certain conditions (ranging from "relatively pathological" to "always"),
so it still seems questionable to have a dependency on it.
>>> Consider SWIOTLB's implementation - segments which already lie at physical addresses within the device's DMA mask just get passed through, while those that lie outside it get mapped into the bounce buffer, but still as individual allocations (arch code just handles cache maintenance on the resulting physical addresses and can apply any hard-wired DMA offset for the device concerned).
>
> And this is fine for vb2-dma-contig, which was made for devices that
> require buffers contiguous in its address space. Without IOMMU it will
> allow only physically contiguous buffers and fails otherwise, which is
> fine, because it's a hardware requirement.
If it depends on having contiguous-from-the-device's-view DMA buffers
either way, that's a sign it should perhaps be using the coherent DMA
API instead, which _does_ give such a guarantee. I'm well aware of the
"but the noncacheable mappings make userspace access unacceptably slow!"
issue many folks have with that, though, and don't particularly fancy
going off on that tangent here.
>>>> IIUC, the change above breaks this model by inserting gaps in how the
>>>> buffer is mapped to device memory, such that the buffer is no longer
>>>> contiguous in dma address space.
>>>
>>>
>>> Even the existing arch/arm IOMMU DMA code which I guess this implicitly relies on doesn't guarantee that behaviour - if the mapping happens to reach one of the segment length/boundary limits it won't just leave a gap, it'll start an entirely new IOVA allocation which could well start at a wildly different address[0].
>
> Could you explain segment length/boundary limits and when buffers can
> reach them? Sorry, i haven't been following all the discussions, but
> I'm not aware of any similar requirements of the IOMMU hardware I
> worked with.
I hope the explanation at the top makes sense - it's purely about the
requirements of the DMA master device itself, nothing to do with the
IOMMU (or lack of) in the middle. Devices with scatter-gather DMA
limitations exist, therefore the API for scatter-gather DMA is designed
to represent and respect such limitations.
>>>> Here is the code in question from
>>>> drivers/media/v4l2-core/videobuf2-dma-contig.c :
[...]
>>>> static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
>>>> unsigned long size, enum dma_data_direction dma_dir)
>>>> {
[...]
>>>> sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
>>>> buf->dma_dir, &attrs);
>>>>
>>>> contig_size = vb2_dc_get_contiguous_size(sgt);
>>>
>>>
>>> (as an aside, it's rather unintuitive that the handling of the dma_map_sg call actually failing is entirely implicit here)
>
> I'm not sure what you mean, please elaborate. The code considers only
> the case of contiguously mapping at least the requested size as a
> success, because anything else is useless with the hardware.
My bad; having now compared against the actual file I see this is just a
cherry-picking of relevant lines with all the error checking stripped
out. Objection withdrawn ;)
>>>> So, is the videobuf2-dma-contig.c based on an incorrect assumption
>>>> about how the DMA API is supposed to work?
>>>> Is it even possible to map a "contiguous-in-iova-range" mapping for a
>>>> buffer given as an sg_table with an arbitrary set of pages?
>>>
>>>
>>> From the Streaming DMA mappings section of Documentation/DMA-API.txt:
>>>
>>> Note also that the above constraints on physical contiguity and
>>> dma_mask may not apply if the platform has an IOMMU (a device which
>>> maps an I/O DMA address to a physical memory address). However, to be
>>> portable, device driver writers may *not* assume that such an IOMMU
>>> exists.
>>>
>>> There's not strictly any harm in using the DMA API this way and *hoping* you get what you want, as long as you're happy for it to fail pretty much 100% of the time on some systems, and still in a minority of corner cases on any system.
>
> Could you please elaborate? I'd like to see examples, because I can't
> really imagine buffers mappable contiguously on CPU, but not on IOMMU.
> Also, as I said, the hardware I worked with didn't suffer from
> problems like this.
"...device driver writers may *not* assume that such an IOMMU exists."
>>> However, if there's a real dependency on IOMMUs and tight control of IOVA allocation here, then the DMA API isn't really the right tool for the job, and maybe it's time to start looking to how to better fit these multimedia-subsystem-type use cases into the IOMMU API - as far as I understand it there's at least some conceptual overlap with the HSA PASID stuff being prototyped in PCI/x86-land at the moment, so it could be an apposite time to try and bang out some common requirements.
>
> The DMA API is actually the only good tool to use here to keep the
> videobuf2-dma-contig code away from the knowledge about platform
> specific data, e.g. presence of IOMMU. The only thing it knows is that
> the target hardware requires a single contiguous buffer and it relies
> on the fact that in correct cases the buffer given to it will meet
> this requirement (i.e. physically contiguous w/o IOMMU; CPU mappable
> with IOMMU).
As above; the DMA API guarantees only what the DMA API guarantees. An
IOMMU-based implementation of streaming DMA is free to identity-map
pages if it only cares about device isolation; a non-IOMMU
implementation is free to provide streaming DMA remapping via some
elaborate bounce-buffering scheme if it really wants to. GART-type
IOMMUs... let's not even go there.
If v4l needs a guarantee of a single contiguous DMA buffer, then it
needs to use dma_alloc_coherent() for that, not streaming mappings.
Robin.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-03 17:41 ` Robin Murphy
@ 2015-11-03 18:40 ` Russell King - ARM Linux
2015-11-04 5:15 ` Tomasz Figa
2015-11-04 5:12 ` Tomasz Figa
1 sibling, 1 reply; 15+ messages in thread
From: Russell King - ARM Linux @ 2015-11-03 18:40 UTC (permalink / raw)
To: Robin Murphy
Cc: Tomasz Figa, Daniel Kurtz, Lin PoChun, linux-arm-kernel,
Yingjoe Chen, Will Deacon, linux-media, Thierry Reding,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
Laurent Pinchart, Joerg Roedel, thunder.leizhen, Catalin Marinas,
linux-mediatek
On Tue, Nov 03, 2015 at 05:41:24PM +0000, Robin Murphy wrote:
> Hi Tomasz,
>
> On 02/11/15 13:43, Tomasz Figa wrote:
> >Agreed. The dma_map_*() API is not guaranteed to return a single
> >contiguous part of virtual address space for any given SG list.
> >However it was understood to be able to map buffers contiguously
> >mappable by the CPU into a single segment and users,
> >videobuf2-dma-contig in particular, relied on this.
>
> I don't follow that - _any_ buffer made of page-sized chunks is going to be
> mappable contiguously by the CPU; it's clearly impossible for the streaming
> DMA API itself to offer such a guarantee, because it's entirely orthogonal
> to the presence or otherwise of an IOMMU.
Tomasz's use of "virtual address space" above in combination with the
DMA API is really confusing.
dma_map_sg() does *not* construct a CPU view of the passed scatterlist.
The only thing dma_map_sg() might do with virtual addresses is to use
them as a way to achieve cache coherence for one particular view of
that memory, that being the kernel's own lowmem mapping and any kmaps.
It doesn't extend to vmalloc() or userspace mappings of the memory.
If the scatterlist is converted to an array of struct page pointers,
it's possible to map it with vmap(), but it's implementation defined
whether such a mapping will receive cache maintanence as part of the
DMA API or not. (If you have PIPT caches, it will, if they're VIPT
caches, maybe not.)
There is a separate set of calls to deal with the flushing issues for
vmap()'d memory in this case - see flush_kernel_vmap_range() and
invalidate_kernel_vmap_range().
--
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-03 17:41 ` Robin Murphy
2015-11-03 18:40 ` Russell King - ARM Linux
@ 2015-11-04 5:12 ` Tomasz Figa
2015-11-04 9:27 ` Russell King - ARM Linux
2015-11-09 13:11 ` Robin Murphy
1 sibling, 2 replies; 15+ messages in thread
From: Tomasz Figa @ 2015-11-04 5:12 UTC (permalink / raw)
To: Robin Murphy
Cc: Daniel Kurtz, Lin PoChun, linux-arm-kernel, Yingjoe Chen,
Will Deacon, linux-media, Thierry Reding,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
Laurent Pinchart, Joerg Roedel, thunder.leizhen, Catalin Marinas,
Russell King, linux-mediatek
On Wed, Nov 4, 2015 at 2:41 AM, Robin Murphy <robin.murphy@arm.com> wrote:
> Hi Tomasz,
>
> On 02/11/15 13:43, Tomasz Figa wrote:
>>
>> I'd like to know what is the boundary mask and what hardware imposes
>> requirements like this. The cost here is not only over-allocating a
>> little, but making many, many buffers contiguously mappable on the
>> CPU, unmappable contiguously in IOMMU, which just defeats the purpose
>> of having an IOMMU, which I believe should be there for simple IP
>> blocks taking one DMA address to be able to view the buffer the same
>> way as the CPU.
>
>
> The expectation with dma_map_sg() is that you're either going to be
> iterating over the buffer segments, handing off each address to the device
> to process one by one;
My understanding of a scatterlist was that it represents a buffer as a
whole, by joining together its physically discontinuous segments.
I don't see how single segments (layout of which is completely up to
the allocator; often just single pages) would be usable for hardware
that needs to do some work more serious than just writing a byte
stream continuously to subsequent buffers. In case of such simple
devices you don't even need an IOMMU (for means other than protection
and/or getting over address space limitations).
However, IMHO the most important use case of an IOMMU is to make
buffers, which are contiguous in CPU virtual address space (VA),
contiguous in device's address space (IOVA). Your implementation of
dma_map_sg() effectively breaks this ability, so I'm not really
following why it's located under drivers/iommu and supposed to be used
with IOMMU-enabled platforms...
> or you have a scatter-gather-capable device, in which
> case you hand off the whole list at once.
No need for mapping ability of the IOMMU here as well (except for
working around address space issues, as I mentioned above).
> It's in the latter case where you
> have to make sure the list doesn't exceed the hardware limitations of that
> device. I believe the original concern was disk controllers (the
> introduction of dma_parms seems to originate from the linux-scsi list), but
> most scatter-gather engines are going to have some limit on how much they
> can handle per entry (IMO the dmaengine drivers are the easiest example to
> look at).
>
> Segment boundaries are a little more arcane, but my assumption is that they
> relate to the kind of devices whose addressing is not flat but relative to
> some separate segment register (The "64-bit" mode of USB EHCI is one
> concrete example I can think of) - since you cannot realistically change the
> segment register while the device is in the middle of accessing a single
> buffer entry, that entry must not fall across a segment boundary or at some
> point the device's accesses are going to overflow the offset address bits
> and wrap around to bogus addresses at the bottom of the segment.
The two requirements above sound like something really specific to
scatter-gather-capable hardware, which as I pointed above, barely need
an IOMMU (at least its mapping capabilities). We are talking here
about very IOMMU-specific code, though...
Now, while I see that on some systems there might be IOMMU used for
improving protection and working around addressing issues with
SG-capable hardware, the code shouldn't be breaking the majority of
systems with IOMMU used as the only possible way to make physically
discontinuous appear (IO-virtually) continuous to devices incapable of
scatter-gather.
>
> Now yes, it will be possible under _most_ circumstances to use an IOMMU to
> lay out a list of segments with page-aligned lengths within a single IOVA
> allocation whilst still meeting all the necessary constraints. It just needs
> some unavoidably complicated calculations - quite likely significantly more
> complex than my v5 version of map_sg() that tried to do that and merge
> segments but failed to take the initial alignment into account properly -
> since there are much simpler ways to enforce just the _necessary_ behaviour
> for the DMA API, I put the complicated stuff to one side for now to prevent
> it holding up getting the basic functional support in place.
Somehow just whatever currently done in arch/arm/mm/dma-mapping.c was
sufficient and not overly complicated.
See http://lxr.free-electrons.com/source/arch/arm/mm/dma-mapping.c#L1547 .
I can see that the code there at least tries to comply with maximum
segment size constraint. Segment boundary seems to be ignored, though.
However, I'm convinced that in most (if not all) cases where IOMMU
IOVA-contiguous mapping is needed, those two requirements don't exist.
Do we really have to break the good hardware only because the
bad^Wlimited one is broken?
Couldn't we preserve the ARM-like behavior whenever
dma_parms->segment_boundary_mask is set to all 1s and
dma_parms->max_segment_size to UINT_MAX (what currently drivers used
to set) or 0 (sounds more logical for the meaning of "no maximum
given")?
>
>>>>> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
>>>>> memory pages into a contiguous block in device memory address space.
>>>>> This would allow passing a dma mapped buffer to device dma using just
>>>>> a device address and length.
>>>>
>>>>
>>>>
>>>> Not at all. The streaming DMA API (dma_map_* and friends) has two
>>>> responsibilities: performing any necessary cache maintenance to ensure the
>>>> device will correctly see data from the CPU, and the CPU will correctly see
>>>> data from the device; and working out an address for that buffer from the
>>>> device's point of view to actually hand off to the hardware (which is
>>>> perfectly well allowed to fail).
>>
>>
>> Agreed. The dma_map_*() API is not guaranteed to return a single
>> contiguous part of virtual address space for any given SG list.
>> However it was understood to be able to map buffers contiguously
>> mappable by the CPU into a single segment and users,
>> videobuf2-dma-contig in particular, relied on this.
>
>
> I don't follow that - _any_ buffer made of page-sized chunks is going to be
> mappable contiguously by the CPU;'
Yes it is. Actually the last chunk might not even need to be
page-sized. However I believe we can have a scatterlist consisting of
non-page-sized chunks in the middle as well, which is obviously not
mappable in a contiguous way even for the CPU.
> it's clearly impossible for the streaming
> DMA API itself to offer such a guarantee, because it's entirely orthogonal
> to the presence or otherwise of an IOMMU.
But we are talking here about the very IOMMU-specific implementation of DMA API.
>
> Furthermore, I can't see any existing dma_map_sg implementation (between
> arm/64 and x86, at least), that _won't_ break that expectation under certain
> conditions (ranging from "relatively pathological" to "always"), so it still
> seems questionable to have a dependency on it.
The current implementation for arch/arm doesn't break that
expectation. As long as we fit inside the maximum segment size (which
in most, if not all, cases of the hardware that actually requires such
contiguous mapping to be created, is UINT_MAX).
http://lxr.free-electrons.com/source/arch/arm/mm/dma-mapping.c#L1547
>
>>>> Consider SWIOTLB's implementation - segments which already lie at
>>>> physical addresses within the device's DMA mask just get passed through,
>>>> while those that lie outside it get mapped into the bounce buffer, but still
>>>> as individual allocations (arch code just handles cache maintenance on the
>>>> resulting physical addresses and can apply any hard-wired DMA offset for the
>>>> device concerned).
>>
>>
>> And this is fine for vb2-dma-contig, which was made for devices that
>> require buffers contiguous in its address space. Without IOMMU it will
>> allow only physically contiguous buffers and fails otherwise, which is
>> fine, because it's a hardware requirement.
>
>
> If it depends on having contiguous-from-the-device's-view DMA buffers either
> way, that's a sign it should perhaps be using the coherent DMA API instead,
> which _does_ give such a guarantee. I'm well aware of the "but the
> noncacheable mappings make userspace access unacceptably slow!" issue many
> folks have with that, though, and don't particularly fancy going off on that
> tangent here.
The keywords here are DMA-BUF and user pointer. Neither of these cases
can use coherent DMA API, because the buffer is already allocated, so
it just needs to be mapped into another device's (or its IOMMU's)
address space. Obviously we can't guarantee mappability of such
buffers, e.g. in case of importing non-contiguous buffers to a device
without an IOMMU, However we expect the pipelines to be sane
(physically contiguous buffers or both devices IOMMU-enabled), so that
such things won't happen.
>
>>>>> IIUC, the change above breaks this model by inserting gaps in how the
>>>>> buffer is mapped to device memory, such that the buffer is no longer
>>>>> contiguous in dma address space.
>>>>
>>>>
>>>>
>>>> Even the existing arch/arm IOMMU DMA code which I guess this implicitly
>>>> relies on doesn't guarantee that behaviour - if the mapping happens to reach
>>>> one of the segment length/boundary limits it won't just leave a gap, it'll
>>>> start an entirely new IOVA allocation which could well start at a wildly
>>>> different address[0].
>>
>>
>> Could you explain segment length/boundary limits and when buffers can
>> reach them? Sorry, i haven't been following all the discussions, but
>> I'm not aware of any similar requirements of the IOMMU hardware I
>> worked with.
>
>
> I hope the explanation at the top makes sense - it's purely about the
> requirements of the DMA master device itself, nothing to do with the IOMMU
> (or lack of) in the middle. Devices with scatter-gather DMA limitations
> exist, therefore the API for scatter-gather DMA is designed to represent and
> respect such limitations.
Yes, it makes sense, thanks for the explanation. However there also
exist devices with no scatter-gather capability, but behind an IOMMU
without such fancy mapping limitations. I believe we should also
respect the limitation of such setups, which is the lack of support
for multiple IOVA segments.
>>>>> So, is the videobuf2-dma-contig.c based on an incorrect assumption
>>>>> about how the DMA API is supposed to work?
>>>>> Is it even possible to map a "contiguous-in-iova-range" mapping for a
>>>>> buffer given as an sg_table with an arbitrary set of pages?
>>>>
>>>>
>>>>
>>>> From the Streaming DMA mappings section of Documentation/DMA-API.txt:
>>>>
>>>> Note also that the above constraints on physical contiguity and
>>>> dma_mask may not apply if the platform has an IOMMU (a device which
>>>> maps an I/O DMA address to a physical memory address). However, to
>>>> be
>>>> portable, device driver writers may *not* assume that such an IOMMU
>>>> exists.
>>>>
>>>> There's not strictly any harm in using the DMA API this way and *hoping*
>>>> you get what you want, as long as you're happy for it to fail pretty much
>>>> 100% of the time on some systems, and still in a minority of corner cases on
>>>> any system.
>>
>>
>> Could you please elaborate? I'd like to see examples, because I can't
>> really imagine buffers mappable contiguously on CPU, but not on IOMMU.
>> Also, as I said, the hardware I worked with didn't suffer from
>> problems like this.
>
>
> "...device driver writers may *not* assume that such an IOMMU exists."
>
And this is exactly why they _should_ use dma_map_sg(), because it was
supposed to work correctly for both physically contiguous (i.e. 1
segment) buffers and non-IOMMU-enabled devices, as well as with
non-contiguous (i.e. > 1 segment) buffers and IOMMU-enabled devices.
>>>> However, if there's a real dependency on IOMMUs and tight control of
>>>> IOVA allocation here, then the DMA API isn't really the right tool for the
>>>> job, and maybe it's time to start looking to how to better fit these
>>>> multimedia-subsystem-type use cases into the IOMMU API - as far as I
>>>> understand it there's at least some conceptual overlap with the HSA PASID
>>>> stuff being prototyped in PCI/x86-land at the moment, so it could be an
>>>> apposite time to try and bang out some common requirements.
>>
>>
>> The DMA API is actually the only good tool to use here to keep the
>> videobuf2-dma-contig code away from the knowledge about platform
>> specific data, e.g. presence of IOMMU. The only thing it knows is that
>> the target hardware requires a single contiguous buffer and it relies
>> on the fact that in correct cases the buffer given to it will meet
>> this requirement (i.e. physically contiguous w/o IOMMU; CPU mappable
>> with IOMMU).
>
>
> As above; the DMA API guarantees only what the DMA API guarantees. An
> IOMMU-based implementation of streaming DMA is free to identity-map pages if
> it only cares about device isolation; a non-IOMMU implementation is free to
> provide streaming DMA remapping via some elaborate bounce-buffering scheme
I guess this is the area where our understandings of IOMMU-backed DMA
API differ.
> if it really wants to. GART-type IOMMUs... let's not even go there.
I believe that's how IOMMU-based implementation of DMA API was
supposed to work when first implemented for ARM...
>
> If v4l needs a guarantee of a single contiguous DMA buffer, then it needs to
> use dma_alloc_coherent() for that, not streaming mappings.
Except that it can't use it, because the buffers are already allocated
by another entity.
Best regards,
Tomasz
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-03 18:40 ` Russell King - ARM Linux
@ 2015-11-04 5:15 ` Tomasz Figa
2015-11-04 9:10 ` Russell King - ARM Linux
0 siblings, 1 reply; 15+ messages in thread
From: Tomasz Figa @ 2015-11-04 5:15 UTC (permalink / raw)
To: Russell King - ARM Linux
Cc: Robin Murphy, Daniel Kurtz, Lin PoChun, linux-arm-kernel,
Yingjoe Chen, Will Deacon, linux-media, Thierry Reding,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
Laurent Pinchart, Joerg Roedel, thunder.leizhen, Catalin Marinas,
linux-mediatek
On Wed, Nov 4, 2015 at 3:40 AM, Russell King - ARM Linux
<linux@arm.linux.org.uk> wrote:
> On Tue, Nov 03, 2015 at 05:41:24PM +0000, Robin Murphy wrote:
>> Hi Tomasz,
>>
>> On 02/11/15 13:43, Tomasz Figa wrote:
>> >Agreed. The dma_map_*() API is not guaranteed to return a single
>> >contiguous part of virtual address space for any given SG list.
>> >However it was understood to be able to map buffers contiguously
>> >mappable by the CPU into a single segment and users,
>> >videobuf2-dma-contig in particular, relied on this.
>>
>> I don't follow that - _any_ buffer made of page-sized chunks is going to be
>> mappable contiguously by the CPU; it's clearly impossible for the streaming
>> DMA API itself to offer such a guarantee, because it's entirely orthogonal
>> to the presence or otherwise of an IOMMU.
>
> Tomasz's use of "virtual address space" above in combination with the
> DMA API is really confusing.
I suppose I must have mistakenly use "virtual address space" somewhere
instead of "IO virtual address space". I'm sorry for causing
confusion.
The thing being discussed here is mapping of buffers described by
scatterlists into IO virtual address space, i.e. the operation
happening when dma_map_sg() is called for an IOMMU-enabled device.
Best regards,
Tomasz
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-04 5:15 ` Tomasz Figa
@ 2015-11-04 9:10 ` Russell King - ARM Linux
0 siblings, 0 replies; 15+ messages in thread
From: Russell King - ARM Linux @ 2015-11-04 9:10 UTC (permalink / raw)
To: Tomasz Figa
Cc: Robin Murphy, Daniel Kurtz, Lin PoChun, linux-arm-kernel,
Yingjoe Chen, Will Deacon, linux-media, Thierry Reding,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
Laurent Pinchart, Joerg Roedel, thunder.leizhen, Catalin Marinas,
linux-mediatek
On Wed, Nov 04, 2015 at 02:15:41PM +0900, Tomasz Figa wrote:
> On Wed, Nov 4, 2015 at 3:40 AM, Russell King - ARM Linux
> <linux@arm.linux.org.uk> wrote:
> > On Tue, Nov 03, 2015 at 05:41:24PM +0000, Robin Murphy wrote:
> >> Hi Tomasz,
> >>
> >> On 02/11/15 13:43, Tomasz Figa wrote:
> >> >Agreed. The dma_map_*() API is not guaranteed to return a single
> >> >contiguous part of virtual address space for any given SG list.
> >> >However it was understood to be able to map buffers contiguously
> >> >mappable by the CPU into a single segment and users,
> >> >videobuf2-dma-contig in particular, relied on this.
> >>
> >> I don't follow that - _any_ buffer made of page-sized chunks is going to be
> >> mappable contiguously by the CPU; it's clearly impossible for the streaming
> >> DMA API itself to offer such a guarantee, because it's entirely orthogonal
> >> to the presence or otherwise of an IOMMU.
> >
> > Tomasz's use of "virtual address space" above in combination with the
> > DMA API is really confusing.
>
> I suppose I must have mistakenly use "virtual address space" somewhere
> instead of "IO virtual address space". I'm sorry for causing
> confusion.
>
> The thing being discussed here is mapping of buffers described by
> scatterlists into IO virtual address space, i.e. the operation
> happening when dma_map_sg() is called for an IOMMU-enabled device.
... and there, it's perfectly legal for an IOMMU to merge all entries
in a scatterlist into one mapping - so dma_map_sg() would return 1.
What that means is that the scatterlist contains the original number of
entries which describes the CPU view of the buffer list using the original
number of entries, and the DMA device view of the same but using just the
first entry.
In other words, if you're walking a scatterlist, and doing a mixture of
DMA and PIO, you can't assume that if you're at scatterlist entry N for
DMA, you can switch to PIO for entry N and you'll write to the same
memory. (I know that there's badly written drivers in the kernel which
unfortunately do make this assumption, and if they're used in the
presence of an IOMMU, they _will_ be silently data corrupting.)
--
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-04 5:12 ` Tomasz Figa
@ 2015-11-04 9:27 ` Russell King - ARM Linux
2015-11-04 9:48 ` Tomasz Figa
2015-11-09 13:11 ` Robin Murphy
1 sibling, 1 reply; 15+ messages in thread
From: Russell King - ARM Linux @ 2015-11-04 9:27 UTC (permalink / raw)
To: Tomasz Figa
Cc: Robin Murphy, Laurent Pinchart, Pawel Osciak, Catalin Marinas,
Joerg Roedel, Will Deacon, Kyungmin Park, Daniel Kurtz, Yong Wu,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
linux-mediatek, Lin PoChun, thunder.leizhen, Marek Szyprowski,
Yingjoe Chen, Thierry Reding, linux-arm-kernel, linux-media
On Wed, Nov 04, 2015 at 02:12:03PM +0900, Tomasz Figa wrote:
> My understanding of a scatterlist was that it represents a buffer as a
> whole, by joining together its physically discontinuous segments.
Correct, and it may also be scattered in CPU virtual space as well.
> I don't see how single segments (layout of which is completely up to
> the allocator; often just single pages) would be usable for hardware
> that needs to do some work more serious than just writing a byte
> stream continuously to subsequent buffers. In case of such simple
> devices you don't even need an IOMMU (for means other than protection
> and/or getting over address space limitations).
All that's required is that the addresses described in the scatterlist
are accessed as an apparently contiguous series of bytes. They don't
have to be contiguous in any address view, provided the device access
appears to be contiguous. How that is done is really neither here nor
there.
IOMMUs are normally there as an address translator - for example, the
underlying device may not have the capability to address a scatterlist
(eg, because it makes effectively random access) and in order to be
accessible to the device, it needs to be made contiguous in device
address space.
Another scenario is that you have more bits of physical address than
a device can generate itself for DMA purposes, and you need an IOMMU
to create a (possibly scattered) mapping in device address space
within the ability of the device to address.
The requirements here depend on the device behind the IOMMU.
> However, IMHO the most important use case of an IOMMU is to make
> buffers, which are contiguous in CPU virtual address space (VA),
> contiguous in device's address space (IOVA).
No - there is no requirement for CPU virtual contiguous buffers to also
be contiguous in the device address space.
--
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-04 9:27 ` Russell King - ARM Linux
@ 2015-11-04 9:48 ` Tomasz Figa
2015-11-04 10:50 ` Russell King - ARM Linux
0 siblings, 1 reply; 15+ messages in thread
From: Tomasz Figa @ 2015-11-04 9:48 UTC (permalink / raw)
To: Russell King - ARM Linux
Cc: Robin Murphy, Laurent Pinchart, Pawel Osciak, Catalin Marinas,
Joerg Roedel, Will Deacon, Kyungmin Park, Daniel Kurtz, Yong Wu,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
linux-mediatek, Lin PoChun, thunder.leizhen, Marek Szyprowski,
Yingjoe Chen, Thierry Reding, linux-arm-kernel, linux-media
On Wed, Nov 4, 2015 at 6:27 PM, Russell King - ARM Linux
<linux@arm.linux.org.uk> wrote:
> On Wed, Nov 04, 2015 at 02:12:03PM +0900, Tomasz Figa wrote:
>> My understanding of a scatterlist was that it represents a buffer as a
>> whole, by joining together its physically discontinuous segments.
>
> Correct, and it may also be scattered in CPU virtual space as well.
>
>> I don't see how single segments (layout of which is completely up to
>> the allocator; often just single pages) would be usable for hardware
>> that needs to do some work more serious than just writing a byte
>> stream continuously to subsequent buffers. In case of such simple
>> devices you don't even need an IOMMU (for means other than protection
>> and/or getting over address space limitations).
>
> All that's required is that the addresses described in the scatterlist
> are accessed as an apparently contiguous series of bytes. They don't
> have to be contiguous in any address view, provided the device access
> appears to be contiguous. How that is done is really neither here nor
> there.
>
> IOMMUs are normally there as an address translator - for example, the
> underlying device may not have the capability to address a scatterlist
> (eg, because it makes effectively random access) and in order to be
> accessible to the device, it needs to be made contiguous in device
> address space.
>
> Another scenario is that you have more bits of physical address than
> a device can generate itself for DMA purposes, and you need an IOMMU
> to create a (possibly scattered) mapping in device address space
> within the ability of the device to address.
>
> The requirements here depend on the device behind the IOMMU.
I fully agree with you.
The problem is that the code being discussed here breaks the case of
devices that don't have the capability of addressing a scatterlist,
supposedly for the sake of devices that have such capability (but as I
suggested, they both could be happily supported, by distinguishing
special values of DMA max segment size and boundary mask).
>> However, IMHO the most important use case of an IOMMU is to make
>> buffers, which are contiguous in CPU virtual address space (VA),
>> contiguous in device's address space (IOVA).
>
> No - there is no requirement for CPU virtual contiguous buffers to also
> be contiguous in the device address space.
There is no requirement, but shouldn't it be desired for the mapping
code to map them as such? Otherwise, how could the IOMMU use case you
described above (address translator for devices which don't have the
capability to address a scatterlist) be handled properly?
Is the general conclusion now that dma_map_sg() should not be used to
create IOMMU mappings and we should make a step backwards making all
drivers (or frameworks, such as videobuf2) do that manually? That
would be really backwards, because code not aware of IOMMU existence
at all would have to become aware of it.
Best regards,
Tomasz
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-04 9:48 ` Tomasz Figa
@ 2015-11-04 10:50 ` Russell King - ARM Linux
0 siblings, 0 replies; 15+ messages in thread
From: Russell King - ARM Linux @ 2015-11-04 10:50 UTC (permalink / raw)
To: Tomasz Figa
Cc: Robin Murphy, Laurent Pinchart, Pawel Osciak, Catalin Marinas,
Joerg Roedel, Will Deacon, Kyungmin Park, Daniel Kurtz, Yong Wu,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
linux-mediatek, Lin PoChun, thunder.leizhen, Marek Szyprowski,
Yingjoe Chen, Thierry Reding, linux-arm-kernel, linux-media
On Wed, Nov 04, 2015 at 06:48:50PM +0900, Tomasz Figa wrote:
> There is no requirement, but shouldn't it be desired for the mapping
> code to map them as such? Otherwise, how could the IOMMU use case you
> described above (address translator for devices which don't have the
> capability to address a scatterlist) be handled properly?
It's up to the IOMMU code to respect the parameters that the device has
supplied to it via the device_dma_parameters. This doesn't currently
allow a device to say "I want this scatterlist to be mapped as a
contiguous device address", so really if a device has such a requirement,
at the moment the device driver _must_ check the dma_map_sg() return
value and act accordingly.
While it's possible to say "an IOMMU should map as a single contiguous
address" what happens when the IOMMU's device address space becomes
fragmented?
> Is the general conclusion now that dma_map_sg() should not be used to
> create IOMMU mappings and we should make a step backwards making all
> drivers (or frameworks, such as videobuf2) do that manually? That
> would be really backwards, because code not aware of IOMMU existence
> at all would have to become aware of it.
No. The DMA API has always had the responsibility for managing the
IOMMU device, which may well be shared between multiple different
devices.
However, if the IOMMU is part of a device IP block (such as a GPU)
then the decision on whether the DMA API should be used or not is up
to the driver author. If it has special management requirements,
then it's probably appropriate for the device driver to manage it by
itself.
For example, a GPUs MMU may need something inserted into the GPUs
command stream to flush the MMU TLBs. Such cases are inappropriate
to be using the DMA API for IOMMU management.
--
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-11-04 5:12 ` Tomasz Figa
2015-11-04 9:27 ` Russell King - ARM Linux
@ 2015-11-09 13:11 ` Robin Murphy
1 sibling, 0 replies; 15+ messages in thread
From: Robin Murphy @ 2015-11-09 13:11 UTC (permalink / raw)
To: Tomasz Figa
Cc: Daniel Kurtz, Lin PoChun, linux-arm-kernel, Yingjoe Chen,
Will Deacon, linux-media, Thierry Reding,
open list:IOMMU DRIVERS, Bobby Batacharia (via Google Docs),
Kyungmin Park, Marek Szyprowski, Yong Wu, Pawel Osciak,
Laurent Pinchart, Joerg Roedel, thunder.leizhen, Catalin Marinas,
Russell King, linux-mediatek
On 04/11/15 05:12, Tomasz Figa wrote:
> On Wed, Nov 4, 2015 at 2:41 AM, Robin Murphy <robin.murphy@arm.com> wrote:
>> Hi Tomasz,
>>
>> On 02/11/15 13:43, Tomasz Figa wrote:
>>>
>>> I'd like to know what is the boundary mask and what hardware imposes
>>> requirements like this. The cost here is not only over-allocating a
>>> little, but making many, many buffers contiguously mappable on the
>>> CPU, unmappable contiguously in IOMMU, which just defeats the purpose
>>> of having an IOMMU, which I believe should be there for simple IP
>>> blocks taking one DMA address to be able to view the buffer the same
>>> way as the CPU.
>>
>>
>> The expectation with dma_map_sg() is that you're either going to be
>> iterating over the buffer segments, handing off each address to the device
>> to process one by one;
>
> My understanding of a scatterlist was that it represents a buffer as a
> whole, by joining together its physically discontinuous segments.
It can, but there are also cases where a single scatterlist is used to
batch up multiple I/O requests - see the stuff in block/blk-merge.c as
described in section 2.2 of Documentation/biodoc.txt, and AFAICS anyone
could quite happily use the dmaengine API, and possibly others, in the
same way. Ultimately a scatterlist is no more specific than "a list of
blocks of physical memory that each want giving a DMA address".
> I don't see how single segments (layout of which is completely up to
> the allocator; often just single pages) would be usable for hardware
> that needs to do some work more serious than just writing a byte
> stream continuously to subsequent buffers. In case of such simple
> devices you don't even need an IOMMU (for means other than protection
> and/or getting over address space limitations).
>
> However, IMHO the most important use case of an IOMMU is to make
> buffers, which are contiguous in CPU virtual address space (VA),
> contiguous in device's address space (IOVA). Your implementation of
> dma_map_sg() effectively breaks this ability, so I'm not really
> following why it's located under drivers/iommu and supposed to be used
> with IOMMU-enabled platforms...
>
>> or you have a scatter-gather-capable device, in which
>> case you hand off the whole list at once.
>
> No need for mapping ability of the IOMMU here as well (except for
> working around address space issues, as I mentioned above).
Ok, now I'm starting to wonder if you're wilfully choosing to miss the
point. Look at 64-bit systems of any architecture, and those address
space issues are pretty much the primary consideration for including an
IOMMU in the first place (behind virtualisation, which we can forget
about here). Take the Juno board on my desk - most of the peripherals
cannot address 75% of the RAM, and CPU bounce buffers are both not
overly efficient and a limited resource (try using dmatest with
sufficiently large buffers to stress/measure memory bandwidth and watch
it take down the kernel, and that's without any other SWIOTLB
contention). The only one that really cares at all about contiguous
buffers is the HDLCD, but that's perfectly happy when it calls
dma_alloc_coherent() via drm_fb_cma_helper and pulls a contiguous 8MB
framebuffer out of thin air, without even knowing that CMA itself is
disabled and it couldn't natively address 75% of the memory that might
be backing that buffer.
That last point also illustrates that the thing for providing
DMA-contiguous buffers is indeed very good at providing DMA-contiguous
buffers when backed by an IOMMU.
>> It's in the latter case where you
>> have to make sure the list doesn't exceed the hardware limitations of that
>> device. I believe the original concern was disk controllers (the
>> introduction of dma_parms seems to originate from the linux-scsi list), but
>> most scatter-gather engines are going to have some limit on how much they
>> can handle per entry (IMO the dmaengine drivers are the easiest example to
>> look at).
>>
>> Segment boundaries are a little more arcane, but my assumption is that they
>> relate to the kind of devices whose addressing is not flat but relative to
>> some separate segment register (The "64-bit" mode of USB EHCI is one
>> concrete example I can think of) - since you cannot realistically change the
>> segment register while the device is in the middle of accessing a single
>> buffer entry, that entry must not fall across a segment boundary or at some
>> point the device's accesses are going to overflow the offset address bits
>> and wrap around to bogus addresses at the bottom of the segment.
>
> The two requirements above sound like something really specific to
> scatter-gather-capable hardware, which as I pointed above, barely need
> an IOMMU (at least its mapping capabilities). We are talking here
> about very IOMMU-specific code, though...
>
> Now, while I see that on some systems there might be IOMMU used for
> improving protection and working around addressing issues with
> SG-capable hardware, the code shouldn't be breaking the majority of
> systems with IOMMU used as the only possible way to make physically
> discontinuous appear (IO-virtually) continuous to devices incapable of
> scatter-gather.
Unless this majority of systems are all 64-bit ARMv8 ones running code
that works perfectly _with the existing SWIOTLB DMA API implementation_
but not with this implementation, then I disagree that anything is being
broken that wasn't already broken with respect to portability.
Otherwise, please give me the details of any regressions with these
patches relative to SWIOTLB DMA on arm64 so I can look into them.
>> Now yes, it will be possible under _most_ circumstances to use an IOMMU to
>> lay out a list of segments with page-aligned lengths within a single IOVA
>> allocation whilst still meeting all the necessary constraints. It just needs
>> some unavoidably complicated calculations - quite likely significantly more
>> complex than my v5 version of map_sg() that tried to do that and merge
>> segments but failed to take the initial alignment into account properly -
>> since there are much simpler ways to enforce just the _necessary_ behaviour
>> for the DMA API, I put the complicated stuff to one side for now to prevent
>> it holding up getting the basic functional support in place.
>
> Somehow just whatever currently done in arch/arm/mm/dma-mapping.c was
> sufficient and not overly complicated.
>
> See http://lxr.free-electrons.com/source/arch/arm/mm/dma-mapping.c#L1547 .
>
> I can see that the code there at least tries to comply with maximum
> segment size constraint. Segment boundary seems to be ignored, though.
It certainly doesn't map the entire list into a single IOVA allocation
as here (such that everything is laid out in contiguous IOVA pages
_regardless_ of the segment lengths, and unmapping becomes nicely
trivial). That it also is the only implementation which fails to respect
segment boundaries really just implies that it's probably not seen much
use beyond supporting graphics hardware on 32-bit systems, and/or has
just got lucky otherwise.
> However, I'm convinced that in most (if not all) cases where IOMMU
> IOVA-contiguous mapping is needed, those two requirements don't exist.
> Do we really have to break the good hardware only because the
> bad^Wlimited one is broken?
Where "is broken" at least encompasses "is a SATA controller",
presumably. Here's an example I've actually played with:
http://lxr.free-electrons.com/source/drivers/ata/sata_sil24.c#L390
It doesn't seem all that unreasonable that hardware that fundamentally
works in fixed-size blocks of data wants its data aligned to its block
size (or some efficient multiple). Implementing an API which has
guaranteed support for that requirement from the outset necessitates
supporting that requirement. I'm not going to buy the argument that
having some video device DMA into userspace pages is more important than
being able to boot at all (and not corrupting your filesystem).
> Couldn't we preserve the ARM-like behavior whenever
> dma_parms->segment_boundary_mask is set to all 1s and
> dma_parms->max_segment_size to UINT_MAX (what currently drivers used
> to set) or 0 (sounds more logical for the meaning of "no maximum
> given")?
Sure, I was always aiming to ultimately improve on the arch/arm
implementation (i.e. with the single allocation thing), but for a common
general-purpose implementation that's going to be shared by multiple
architectures, correctness comes way before optimisation for one
specific use-case. Thus we start with a baseline version that we know
correctly implements all the required behaviour specified by the DMA
API, then start tweaking it for other considerations later.
FWIW, I've already sketched out such a follow-on patch to start
tightening up map_sg (because exposing any pages to the device more than
absolutely necessary is not what we want in the long run). The thought
that it's likely to be jumped on and used as an excuse to justify bad
code elsewhere does rather sour the idea, though.
>>>>>> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
>>>>>> memory pages into a contiguous block in device memory address space.
>>>>>> This would allow passing a dma mapped buffer to device dma using just
>>>>>> a device address and length.
>>>>>
>>>>>
>>>>>
>>>>> Not at all. The streaming DMA API (dma_map_* and friends) has two
>>>>> responsibilities: performing any necessary cache maintenance to ensure the
>>>>> device will correctly see data from the CPU, and the CPU will correctly see
>>>>> data from the device; and working out an address for that buffer from the
>>>>> device's point of view to actually hand off to the hardware (which is
>>>>> perfectly well allowed to fail).
>>>
>>>
>>> Agreed. The dma_map_*() API is not guaranteed to return a single
>>> contiguous part of virtual address space for any given SG list.
>>> However it was understood to be able to map buffers contiguously
>>> mappable by the CPU into a single segment and users,
>>> videobuf2-dma-contig in particular, relied on this.
>>
>>
>> I don't follow that - _any_ buffer made of page-sized chunks is going to be
>> mappable contiguously by the CPU;'
>
> Yes it is. Actually the last chunk might not even need to be
> page-sized. However I believe we can have a scatterlist consisting of
> non-page-sized chunks in the middle as well, which is obviously not
> mappable in a contiguous way even for the CPU.
>
>> it's clearly impossible for the streaming
>> DMA API itself to offer such a guarantee, because it's entirely orthogonal
>> to the presence or otherwise of an IOMMU.
>
> But we are talking here about the very IOMMU-specific implementation of DMA API.
Exactly, therein lies the problem! The whole point of an API is that we
write code against the provided _interface_, not against some particular
implementation detail. To quote Raymond Chen, "I can't believe I had to
write that".
I fail to see how anyone would be surprised that code which is reliant
on specific non-contractual behaviour of a particular API implementation
is not portable to other implementations of that API.
>> Furthermore, I can't see any existing dma_map_sg implementation (between
>> arm/64 and x86, at least), that _won't_ break that expectation under certain
>> conditions (ranging from "relatively pathological" to "always"), so it still
>> seems questionable to have a dependency on it.
>
> The current implementation for arch/arm doesn't break that
> expectation. As long as we fit inside the maximum segment size (which
> in most, if not all, cases of the hardware that actually requires such
> contiguous mapping to be created, is UINT_MAX).
Well, yes, that just restates my point exactly; outside of certain
conditions you will still get a non-contiguous mapping. Put that exact
code on a 64-bit system, throw a scatterlist describing a "relatively
pathological" 5GB buffer into it, and see what you get out.
> http://lxr.free-electrons.com/source/arch/arm/mm/dma-mapping.c#L1547
>
>>
>>>>> Consider SWIOTLB's implementation - segments which already lie at
>>>>> physical addresses within the device's DMA mask just get passed through,
>>>>> while those that lie outside it get mapped into the bounce buffer, but still
>>>>> as individual allocations (arch code just handles cache maintenance on the
>>>>> resulting physical addresses and can apply any hard-wired DMA offset for the
>>>>> device concerned).
>>>
>>>
>>> And this is fine for vb2-dma-contig, which was made for devices that
>>> require buffers contiguous in its address space. Without IOMMU it will
>>> allow only physically contiguous buffers and fails otherwise, which is
>>> fine, because it's a hardware requirement.
>>
>>
>> If it depends on having contiguous-from-the-device's-view DMA buffers either
>> way, that's a sign it should perhaps be using the coherent DMA API instead,
>> which _does_ give such a guarantee. I'm well aware of the "but the
>> noncacheable mappings make userspace access unacceptably slow!" issue many
>> folks have with that, though, and don't particularly fancy going off on that
>> tangent here.
>
> The keywords here are DMA-BUF and user pointer. Neither of these cases
> can use coherent DMA API, because the buffer is already allocated, so
> it just needs to be mapped into another device's (or its IOMMU's)
> address space. Obviously we can't guarantee mappability of such
> buffers, e.g. in case of importing non-contiguous buffers to a device
> without an IOMMU, However we expect the pipelines to be sane
> (physically contiguous buffers or both devices IOMMU-enabled), so that
> such things won't happen.
The "guarantee to map these scatterlist pages contiguously in IOVA space
if an IOMMU is present" function is named iommu_map_sg(). There is
nothing in the DMA API offering that behaviour. How well does
vb2-dma-contig work with the x86 IOMMUs?
>>>>>> IIUC, the change above breaks this model by inserting gaps in how the
>>>>>> buffer is mapped to device memory, such that the buffer is no longer
>>>>>> contiguous in dma address space.
>>>>>
>>>>>
>>>>>
>>>>> Even the existing arch/arm IOMMU DMA code which I guess this implicitly
>>>>> relies on doesn't guarantee that behaviour - if the mapping happens to reach
>>>>> one of the segment length/boundary limits it won't just leave a gap, it'll
>>>>> start an entirely new IOVA allocation which could well start at a wildly
>>>>> different address[0].
>>>
>>>
>>> Could you explain segment length/boundary limits and when buffers can
>>> reach them? Sorry, i haven't been following all the discussions, but
>>> I'm not aware of any similar requirements of the IOMMU hardware I
>>> worked with.
>>
>>
>> I hope the explanation at the top makes sense - it's purely about the
>> requirements of the DMA master device itself, nothing to do with the IOMMU
>> (or lack of) in the middle. Devices with scatter-gather DMA limitations
>> exist, therefore the API for scatter-gather DMA is designed to represent and
>> respect such limitations.
>
> Yes, it makes sense, thanks for the explanation. However there also
> exist devices with no scatter-gather capability, but behind an IOMMU
> without such fancy mapping limitations. I believe we should also
> respect the limitation of such setups, which is the lack of support
> for multiple IOVA segments.
>
>>>>>> So, is the videobuf2-dma-contig.c based on an incorrect assumption
>>>>>> about how the DMA API is supposed to work?
>>>>>> Is it even possible to map a "contiguous-in-iova-range" mapping for a
>>>>>> buffer given as an sg_table with an arbitrary set of pages?
>>>>>
>>>>>
>>>>>
>>>>> From the Streaming DMA mappings section of Documentation/DMA-API.txt:
>>>>>
>>>>> Note also that the above constraints on physical contiguity and
>>>>> dma_mask may not apply if the platform has an IOMMU (a device which
>>>>> maps an I/O DMA address to a physical memory address). However, to
>>>>> be
>>>>> portable, device driver writers may *not* assume that such an IOMMU
>>>>> exists.
>>>>>
>>>>> There's not strictly any harm in using the DMA API this way and *hoping*
>>>>> you get what you want, as long as you're happy for it to fail pretty much
>>>>> 100% of the time on some systems, and still in a minority of corner cases on
>>>>> any system.
>>>
>>>
>>> Could you please elaborate? I'd like to see examples, because I can't
>>> really imagine buffers mappable contiguously on CPU, but not on IOMMU.
>>> Also, as I said, the hardware I worked with didn't suffer from
>>> problems like this.
>>
>>
>> "...device driver writers may *not* assume that such an IOMMU exists."
>>
>
> And this is exactly why they _should_ use dma_map_sg(), because it was
> supposed to work correctly for both physically contiguous (i.e. 1
> segment) buffers and non-IOMMU-enabled devices, as well as with
> non-contiguous (i.e. > 1 segment) buffers and IOMMU-enabled devices.
Note that the number of segments has nothing to do with whether they are
contiguous (in any address space) or not.
In fact, while I've been thinking about this I realise we have another
misapprehension here: the point of dma_parms is to expose a device's
scatter-gather capabilities to _restrict_ what an IOMMU-based DMA API
implementation can do (see 6b7b65105522) - thus setting fake
"restrictions" for non-scatter-gather hardware in an attempt to force an
implementation into merging segments is entirely backwards.
>>>>> However, if there's a real dependency on IOMMUs and tight control of
>>>>> IOVA allocation here, then the DMA API isn't really the right tool for the
>>>>> job, and maybe it's time to start looking to how to better fit these
>>>>> multimedia-subsystem-type use cases into the IOMMU API - as far as I
>>>>> understand it there's at least some conceptual overlap with the HSA PASID
>>>>> stuff being prototyped in PCI/x86-land at the moment, so it could be an
>>>>> apposite time to try and bang out some common requirements.
>>>
>>>
>>> The DMA API is actually the only good tool to use here to keep the
>>> videobuf2-dma-contig code away from the knowledge about platform
>>> specific data, e.g. presence of IOMMU. The only thing it knows is that
>>> the target hardware requires a single contiguous buffer and it relies
>>> on the fact that in correct cases the buffer given to it will meet
>>> this requirement (i.e. physically contiguous w/o IOMMU; CPU mappable
>>> with IOMMU).
>>
>>
>> As above; the DMA API guarantees only what the DMA API guarantees. An
>> IOMMU-based implementation of streaming DMA is free to identity-map pages if
>> it only cares about device isolation; a non-IOMMU implementation is free to
>> provide streaming DMA remapping via some elaborate bounce-buffering scheme
>
> I guess this is the area where our understandings of IOMMU-backed DMA
> API differ.
The DMA API provides a hardware-independent abstraction of a set of
operations for exposing kernel memory to devices. When someone calls a
DMA API function, they don't get to choose the details of that
abstraction, and they don't get to choose the semantics of those operations.
Of course they can always go ahead and propose adding something to the
API, if they really believe there's something else it needs to offer.
>> if it really wants to. GART-type IOMMUs... let's not even go there.
>
> I believe that's how IOMMU-based implementation of DMA API was
> supposed to work when first implemented for ARM...
>
>> If v4l needs a guarantee of a single contiguous DMA buffer, then it needs to
>> use dma_alloc_coherent() for that, not streaming mappings.
>
> Except that it can't use it, because the buffers are already allocated
> by another entity.
dma_alloc_coherent(...
for_each_sg(..
memcpy(...
Or v4l is rearchitected such that the userspace pages came from
mmap()ing a guaranteed-contiguous DMA buffer in the first place. Or
vb2-dma-contig is rearchitected to use the IOMMU API directly where it
has an IOMMU dependency. Or someone posts a patch to extend the DMA API
with a dma_try_to_map_sg_as_contiguously_as_you_can_manage() operation
that doesn't even necessarily have to depend on an IOMMU... Plenty of
ways to replace incorrect assumptions with reliable ones.
Or to put it another way; Fast, Easy to implement, Correct: pick two.
With the caveat that for upstream, one of the two _must_ be "Correct".
Robin.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping
2015-10-30 1:17 ` [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping Daniel Kurtz
2015-10-30 14:09 ` Joerg Roedel
2015-10-30 14:27 ` Robin Murphy
@ 2015-11-17 12:02 ` Marek Szyprowski
2 siblings, 0 replies; 15+ messages in thread
From: Marek Szyprowski @ 2015-11-17 12:02 UTC (permalink / raw)
To: Daniel Kurtz, Robin Murphy, Pawel Osciak
Cc: Yong Wu, Joerg Roedel, Will Deacon, Catalin Marinas,
open list:IOMMU DRIVERS, linux-arm-kernel, thunder.leizhen,
Yingjoe Chen, laurent.pinchart+renesas, Thierry Reding,
Lin PoChun, Bobby Batacharia (via Google Docs),
linux-media, Kyungmin Park, Tomasz Figa,
Russell King - ARM Linux, Bartlomiej Zolnierkiewicz
Hello,
I'm really sorry do late joining this discussion, but I was terribly
busy with other things.
On 2015-10-30 02:17, Daniel Kurtz wrote:
> +linux-media & VIDEOBUF2 FRAMEWORK maintainers since this is about the
> v4l2-contig's usage of the DMA API.
>
> Hi Robin,
>
> On Tue, Oct 27, 2015 at 12:55 AM, Robin Murphy <robin.murphy@arm.com> wrote:
>> On 26/10/15 13:44, Yong Wu wrote:
>>> On Thu, 2015-10-01 at 20:13 +0100, Robin Murphy wrote:
>>> [...]
>>>> +/*
>>>> + * The DMA API client is passing in a scatterlist which could describe
>>>> + * any old buffer layout, but the IOMMU API requires everything to be
>>>> + * aligned to IOMMU pages. Hence the need for this complicated bit of
>>>> + * impedance-matching, to be able to hand off a suitably-aligned list,
>>>> + * but still preserve the original offsets and sizes for the caller.
>>>> + */
>>>> +int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
>>>> + int nents, int prot)
>>>> +{
>>>> + struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
>>>> + struct iova_domain *iovad = domain->iova_cookie;
>>>> + struct iova *iova;
>>>> + struct scatterlist *s, *prev = NULL;
>>>> + dma_addr_t dma_addr;
>>>> + size_t iova_len = 0;
>>>> + int i;
>>>> +
>>>> + /*
>>>> + * Work out how much IOVA space we need, and align the segments
>>>> to
>>>> + * IOVA granules for the IOMMU driver to handle. With some clever
>>>> + * trickery we can modify the list in-place, but reversibly, by
>>>> + * hiding the original data in the as-yet-unused DMA fields.
>>>> + */
>>>> + for_each_sg(sg, s, nents, i) {
>>>> + size_t s_offset = iova_offset(iovad, s->offset);
>>>> + size_t s_length = s->length;
>>>> +
>>>> + sg_dma_address(s) = s->offset;
>>>> + sg_dma_len(s) = s_length;
>>>> + s->offset -= s_offset;
>>>> + s_length = iova_align(iovad, s_length + s_offset);
>>>> + s->length = s_length;
>>>> +
>>>> + /*
>>>> + * The simple way to avoid the rare case of a segment
>>>> + * crossing the boundary mask is to pad the previous one
>>>> + * to end at a naturally-aligned IOVA for this one's
>>>> size,
>>>> + * at the cost of potentially over-allocating a little.
>>>> + */
>>>> + if (prev) {
>>>> + size_t pad_len = roundup_pow_of_two(s_length);
>>>> +
>>>> + pad_len = (pad_len - iova_len) & (pad_len - 1);
>>>> + prev->length += pad_len;
>>>
>>> Hi Robin,
>>> While our v4l2 testing, It seems that we met a problem here.
>>> Here we update prev->length again, Do we need update
>>> sg_dma_len(prev) again too?
>>>
>>> Some function like vb2_dc_get_contiguous_size[1] always get
>>> sg_dma_len(s) to compare instead of s->length. so it may break
>>> unexpectedly while sg_dma_len(s) is not same with s->length.
>>
>> This is just tweaking the faked-up length that we hand off to iommu_map_sg()
>> (see also the iova_align() above), to trick it into bumping this segment up
>> to a suitable starting IOVA. The real length at this point is stashed in
>> sg_dma_len(s), and will be copied back into s->length in __finalise_sg(), so
>> both will hold the same true length once we return to the caller.
>>
>> Yes, it does mean that if you have a list where the segment lengths are page
>> aligned but not monotonically decreasing, e.g. {64k, 16k, 64k}, then you'll
>> still end up with a gap between the second and third segments, but that's
>> fine because the DMA API offers no guarantees about what the resulting DMA
>> addresses will be (consider the no-IOMMU case where they would each just be
>> "mapped" to their physical address). If that breaks v4l, then it's probably
>> v4l's DMA API use that needs looking at (again).
> Hmm, I thought the DMA API maps a (possibly) non-contiguous set of
> memory pages into a contiguous block in device memory address space.
> This would allow passing a dma mapped buffer to device dma using just
> a device address and length.
> IIUC, the change above breaks this model by inserting gaps in how the
> buffer is mapped to device memory, such that the buffer is no longer
> contiguous in dma address space.
>
> Here is the code in question from
> drivers/media/v4l2-core/videobuf2-dma-contig.c :
>
> static unsigned long vb2_dc_get_contiguous_size(struct sg_table *sgt)
> {
> struct scatterlist *s;
> dma_addr_t expected = sg_dma_address(sgt->sgl);
> unsigned int i;
> unsigned long size = 0;
>
> for_each_sg(sgt->sgl, s, sgt->nents, i) {
> if (sg_dma_address(s) != expected)
> break;
> expected = sg_dma_address(s) + sg_dma_len(s);
> size += sg_dma_len(s);
> }
> return size;
> }
>
>
> static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
> unsigned long size, enum dma_data_direction dma_dir)
> {
> struct vb2_dc_conf *conf = alloc_ctx;
> struct vb2_dc_buf *buf;
> struct frame_vector *vec;
> unsigned long offset;
> int n_pages, i;
> int ret = 0;
> struct sg_table *sgt;
> unsigned long contig_size;
> unsigned long dma_align = dma_get_cache_alignment();
> DEFINE_DMA_ATTRS(attrs);
>
> dma_set_attr(DMA_ATTR_SKIP_CPU_SYNC, &attrs);
>
> buf = kzalloc(sizeof *buf, GFP_KERNEL);
> buf->dma_dir = dma_dir;
>
> offset = vaddr & ~PAGE_MASK;
> vec = vb2_create_framevec(vaddr, size, dma_dir == DMA_FROM_DEVICE);
> buf->vec = vec;
> n_pages = frame_vector_count(vec);
>
> sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
>
> ret = sg_alloc_table_from_pages(sgt, frame_vector_pages(vec), n_pages,
> offset, size, GFP_KERNEL);
>
> sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
> buf->dma_dir, &attrs);
>
> contig_size = vb2_dc_get_contiguous_size(sgt);
> if (contig_size < size) {
>
> <<<=== if the original buffer had sg entries that were not
> aligned on the "natural" alignment for their size, the new arm64 iommu
> core code inserts a 'gap' in the iommu mapping, which causes
> vb2_dc_get_contiguous_size() to exit early (and return a smaller size
> than expected).
>
> pr_err("contiguous mapping is too small %lu/%lu\n",
> contig_size, size);
> ret = -EFAULT;
> goto fail_map_sg;
> }
>
>
> So, is the videobuf2-dma-contig.c based on an incorrect assumption
> about how the DMA API is supposed to work?
> Is it even possible to map a "contiguous-in-iova-range" mapping for a
> buffer given as an sg_table with an arbitrary set of pages?
>
> Thanks for helping to move this forward.
As I'm responsible for the code of both dma-mapping IOMMU integration
code for
ARM and videobuf2-dc subdriver, I would like to share the background behind
them.
This code is a result of our (Samsung R&D Institute Poland) works on
mainlining
drivers for various multimedia devices found in Exynos SoCs. All those
devices
can only process buffers, which are contiguous in the DMA address space.
This
is requirement was one of the fundamental reason for using IOMMU modules.
However it turned out that there was no straightforward way of
integrating it
for our purposes and some extension to core frameworks were needed.
Our initial proposal integrated IOMMU drivers directly to the V4L2
helper code,
as a separate memory managing subdriver for videobuf2:
http://www.spinics.net/lists/linux-media/msg31455.html
Then I've been suggested to use dma-mapping API and hide IOMMU behind it.
Allocating a buffer suitable for DMA with IOMMU mapper enabled was easy.
However creating a contiguous mapping in DMA address space for a buffer
scattered in the physical memory was still a bit tricky. The only
possible way
was to use scatter-list and assume that DMA-mapping will do it right (will
create only one segment for the whole scatter list). I found no other
possibility. This part of the IOMMU and DMA-mapping was especially
problematic
when buffer sharing (dma-buf) has been introduced and it turned out that
there are different ways of interpreting scatter lists done by different
drivers.
From the time I see that I focused mainly on re-using existing
DMA-mapping API,
what shouldn't be the main goal. This resulted in somehow tricky way of
doing
one of the most common operation for existing multimedia devices.
Scatter lists
are also a bit over-engineered for doing a such simple operation like
mapping
scattered memory into single contiguous DMA address space. They waste memory
for storing useless parameters like per-page offset and dma address/len.
Maybe it would be better if something like page vector (or PFN vector to
solve
the problem of mapping buffers that cannot be described by pages) have been
introduced and operations like dma_map_vector() will make thing much
more clear.
I can provide a proof-of-concept code for further discussion if needed.
Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
^ permalink raw reply [flat|nested] 15+ messages in thread
end of thread, other threads:[~2015-11-17 12:02 UTC | newest]
Thread overview: 15+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
[not found] <cover.1443718557.git.robin.murphy@arm.com>
[not found] ` <ab8e1caa40d6da1afa4a49f30242ef4e6e1f17df.1443718557.git.robin.murphy@arm.com>
[not found] ` <1445867094.30736.14.camel@mhfsdcap03>
[not found] ` <562E5AE4.9070001@arm.com>
2015-10-30 1:17 ` [PATCH v6 1/3] iommu: Implement common IOMMU ops for DMA mapping Daniel Kurtz
2015-10-30 14:09 ` Joerg Roedel
2015-10-30 14:27 ` Robin Murphy
2015-11-02 13:11 ` Daniel Kurtz
2015-11-02 13:43 ` Tomasz Figa
2015-11-03 17:41 ` Robin Murphy
2015-11-03 18:40 ` Russell King - ARM Linux
2015-11-04 5:15 ` Tomasz Figa
2015-11-04 9:10 ` Russell King - ARM Linux
2015-11-04 5:12 ` Tomasz Figa
2015-11-04 9:27 ` Russell King - ARM Linux
2015-11-04 9:48 ` Tomasz Figa
2015-11-04 10:50 ` Russell King - ARM Linux
2015-11-09 13:11 ` Robin Murphy
2015-11-17 12:02 ` Marek Szyprowski
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).