Hi Tomasz,
On Thursday 22 March 2012 16:58:27 Tomasz Stanislawski wrote:
> On 03/22/2012 03:42 PM, Laurent Pinchart wrote:
> > On Thursday 22 March 2012 14:36:33 Tomasz Stanislawski wrote:
> >> On 03/22/2012 11:50 AM, Laurent Pinchart wrote:
> >>> On Thursday 22 March 2012 11:02:23 Laurent Pinchart wrote:
[snip]
> >>>> static void *vb2_dc_alloc(void *alloc_ctx, unsigned long size)
> >>>> {
> >>>>
> >>>> struct device *dev = alloc_ctx;
> >>>> struct vb2_dc_buf *buf;
> >>>>
> >>>> + int ret;
> >>>> + int
> >>>> n_pages;
> >>>>
> >>>> buf = kzalloc(sizeof *buf, GFP_KERNEL);
> >>>> if (!buf)
> >>>>
> >>>> return ERR_PTR(-ENOMEM);
> >>>>
> >>>> - buf->vaddr = dma_alloc_coherent(dev, size, &buf->dma_addr,
> >
> > GFP_KERNEL);
> >
> >>>> + buf->dev = dev;
> >>>> + buf->size = size;
> >>>> + buf->vaddr = dma_alloc_coherent(buf->dev, buf->size, &buf->dma_addr,
> >>>> + GFP_KERNEL);
> >>>> +
> >>>> + ret = -ENOMEM;
> >>>>
> >>>> if (!buf->vaddr) {
> >>>>
> >>>> - dev_err(dev, "dma_alloc_coherent of size %ld failed\n", size);
> >>>> - kfree(buf);
> >>>> - return ERR_PTR(-ENOMEM);
> >>>> + dev_err(dev, "dma_alloc_coherent of size %ld failed\n",
> >>>> + size);
> >>>> + goto fail_buf;
> >>>>
> >>>> }
> >>>>
> >>>> - buf->dev = dev;
> >>>> - buf->size = size;
> >>>> + WARN_ON((unsigned long)buf->vaddr & ~PAGE_MASK);
> >>>> + WARN_ON(buf->dma_addr & ~PAGE_MASK);
> >>>> +
> >>>> + n_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
> >>>> +
> >>>> + pages = kmalloc(n_pages * sizeof pages[0], GFP_KERNEL);
> >>>> + if (!pages) {
> >>>> + printk(KERN_ERR "failed to alloc page table\n");
> >>>> + goto fail_dma;
> >>>> + }
> >>>> +
> >>>> + ret = dma_get_pages(dev, buf->vaddr, buf->dma_addr, pages, n_pages);
> >>>
> >>> As the only purpose of this is to retrieve a list of pages that will be
> >>> used to create a single-entry sgt, wouldn't it be possible to shortcut
> >>> the code and get the physical address of the buffer directly ?
> >>
> >> The physical address should not be used since they are meaningless in a
> >> context of different devices. It seams that only the list of pages is
> >> more-or-less portable between different drivers.
> >
> > The pages are physically contiguous. The physical address of the first
> > page is thus all you need.
>
> No. DMA-CONTIG buffers do not have to be physically contiguous. Please refer
> below.
>
> > struct page and physical addresses can be used interchangeably in this
> > case if I'm not mistaken. If you want to go with pages, you could use the
> > first page only instead of the physical buffer address.
>
> Ok. There are bus addresses, physical addresses, DMA addresses and PFNs.
> As I understand PFNs and 'struct page' can be interchanged, at least in one
> direction. The PFNs are used to create a bus address, I mean addresses that
> are recognized by a RAM chip. So a list of PFNs seams to be the most
> portable way of describing the memory, isn't it?
>
> >> The physical address is already present in buf->dma_addr, but it is only
> >> valid if the device has no MMU. Notice that vb2-dma-contig possess no
> >> knowledge if MMU is present for a given device.
> >
> > That's why buf->dma_addr can't be considered as a physical address. It's
> > only useful in the device context.
>
> ok
>
> >> The sg list is not going to be single-entry if the device is provided
> >> with its own MMU.
> >
> > There's something I don't get then. vb2-dma-contig deals with physically
> > contiguous buffers. The buffer is backed by physically contiguous pages,
> > so the sg list should have a single entry.
>
> As I understand dma-contig deal with DMA contiguous buffers, it means
> buffers that are contiguous from device point of view. Therefore those
> buffers do NOT have to be physically contiguous if the device has its own
> IOMMU.
My bad. There was thus a misunderstanding to begin with.
In the light of this new information, and (at least partially) sharing
Daniel's opinion regarding dma_get_pages(), I think what we need here would be
either
- a DMA API call that maps the memory to the importer device instead of
dma_get_pages() + vb2_dc_pages_to_sgt(). The call would take a DMA memory
"cookie" (see the "Minutes from V4L2 update call" mail thread) and a pointer
to the importer device.
- a DMA API call to retrieve a scatter list suitable to be passed to
dma_map_sg(). This would be similar to dma_get_pages() +
vb2_dc_pages_to_sgt().
We also need to figure out whether the mapping call should be in the exporter
or importer.
> >>>> + if (ret < 0) {
> >>>> + printk(KERN_ERR "failed to get buffer pages from DMA API\n");
> >>>> + goto fail_pages;
> >>>> + }
> >>>> + if (ret != n_pages) {
> >>>> + ret = -EFAULT;
> >>>> + printk(KERN_ERR "failed to get all pages from DMA API\n");
> >>>> + goto fail_pages;
> >>>> + }
> >>>> +
> >>>> + buf->sgt_base = vb2_dc_pages_to_sgt(pages, n_pages, 0, 0);
> >>>> + if (IS_ERR(buf->sgt_base)) {
> >>>> + ret = PTR_ERR(buf->sgt_base);
> >>>> + printk(KERN_ERR "failed to prepare sg table\n");
> >>>> + goto fail_pages;
> >>>> + }
> >>>
> >>> buf->sgt_base isn't used in this patch. I would move the buf->sgt_base
> >>> creation code to the patch that uses it then, or to its own patch just
> >>> before the patch that uses it.
> >>
> >> Good point. The sgt_base is used by exporter only. Thanks for noticing
> >> it.
> >>
> >>>> +
> >>>> + /* pages are no longer needed */
> >>>> + kfree(pages);
> >>>>
> >>>> buf->handler.refcount = &buf->refcount;
> >>>> buf->handler.put = vb2_dc_put;
> >
> > [snip]
> >
> >>>> /*********************************************/
> >>>> /* callbacks for USERPTR buffers */
> >>>> /*********************************************/
> >>>>
> >>>> +static inline int vma_is_io(struct vm_area_struct *vma)
> >>>> +{
> >>>> + return !!(vma->vm_flags & (VM_IO | VM_PFNMAP));
> >>>
> >>> Isn't VM_PFNMAP enough ? Wouldn't it be possible (at least in theory) to
> >>> get a discontinuous physical range with VM_IO ?
> >>
> >> Frankly, I found that that in get_user_pages flags are checked against
> >> (VM_IO | VM_PFNMAP). Probably for noMMU (not no IOMMU) case it is
> >> possible to get vma with VM_IO on and VM_PFNMAP off, isn't it?
> >>
> >> The problem is that this framework should work in both cases so this
> >> check was added just in case :).
> >
> > OK. We can leave it here and deal with problems if they arise :-)
> >
> >>>> +}
> >>>> +
> >>>> +static int vb2_dc_get_pages(unsigned long start, struct page **pages,
> >>>> + int n_pages, struct vm_area_struct **copy_vma, int write)
> >>>> +{
> >>>> + struct vm_area_struct *vma;
> >>>> + int n = 0; /* number of get pages */
> >>>> + int ret = -EFAULT;
> >>>> +
> >>>> + /* entering critical section for mm access */
> >>>> + down_read(¤t->mm->mmap_sem);
> >>>
> >>> This will generate AB-BA deadlock warnings if lockdep is enabled. This
> >>> function is called with the queue lock held, and the mmap() handler
> >>> which
> >>> takes the queue lock is called with current->mm->mmap_sem held.
> >>>
> >>> This is a known issue with videobuf2, not specific to this patch. The
> >>> warning is usually a false positive (which we still need to fix, as it
> >>> worries users), but can become a real issue if an MMAP queue and a
> >>> USERPTR queue are created by a driver with the same queue lock.
> >>
> >> Good point. Do you know any good solution to this problem?
> >
> > http://patchwork.linuxtv.org/patch/8455/
> >
> > It seems QBUF is safe, but PREPAREBUF isn't (both call __buf_prepare,
> > which end up calling the memops get_userptr operation).
> >
> > I'll post a patch to fix it for PREPAREBUF. If I'm not mistaken, you can
> > drop the down_read/up_read here.
>
> ok. Thanks for the link.
>
> >>>> + vma = find_vma(current->mm, start);
> >>>> + if (!vma) {
> >>>> + printk(KERN_ERR "no vma for address %lu\n", start);
> >>>> + goto cleanup;
> >>>> + }
> >>>> +
> >>>> + if (vma_is_io(vma)) {
> >>>> + unsigned long pfn;
> >>>> +
> >>>> + if (vma->vm_end - start < n_pages * PAGE_SIZE) {
> >>>> + printk(KERN_ERR "vma is too small\n");
> >>>> + goto cleanup;
> >>>> + }
> >>>> +
> >>>> + for (n = 0; n < n_pages; ++n, start += PAGE_SIZE) {
> >>>> + ret = follow_pfn(vma, start, &pfn);
> >>>> + if (ret) {
> >>>> + printk(KERN_ERR "no page for address %lu\n",
> >>>> + start);
> >>>> + goto cleanup;
> >>>> + }
> >>>> + pages[n] = pfn_to_page(pfn);
> >>>> + get_page(pages[n]);
> >>>
> >>> This worries me. When the VM_PFNMAP flag is set, the memory pages are
> >>> not backed by a struct page. Creating a struct page pointer out of it
> >>> can be an acceptable hack (for instance to store a page in an
> >>> scatterlist with sg_set_page() and then retrieve its physical address
> >>> with sg_phys()), but you should not expect the struct page to be valid
> >>> for anything else. Calling get_page() on it will likely crash.
> >>
> >> You are completetly right. This is the corner case where list of pages is
> >> not a portable way of describing the memory.
> >> Maybe pfn_valid should be used to check validity of the page (pfn)
> >> before getting it?
> >
> > I think you should just drop the get_page() call. There's no page, so
> > there's no need to get a reference count to it.
>
> The problem is that get_user_pages does call get_page. Not calling get_page
> will break the symmetry between PFNMAP and non-PFNMAP buffers. Maybe
> checking page validity before get_page/put_page is enough?
PFNMAP and non-PFNMAP buffers are inherently different, so I don't see a
problem in handling them differently. We will likely run into an issue though,
with hardware such as the OMAP TILER, where the memory isn't backed by normal
memory (and thus no struct page is present), but for which the target must be
pinned somehow (in the case of the tiler, that would be a tiler mapping). I
don't think we have an API to ask the kernel to pin a memory range regardless
of how the memory is handled (system memory, reserved memory with PFNMAP,
device mapping such as with the tiler, ...). This is an open issue. One
possible solution is to deprecate USERPTR support for that kind of memory and
use dma-buf instead.
> > The VM_PFNMAP flag is mostly used with memory out of the kernel
> > allocator's control if I'm not mistaken. The main use case I've seen is
> > memory reserved at boot time and use as a frame buffer for instance. In
> > that case the pages can't go away, as there no page in the first place.
> >
> > This won't fix the DMA SG problem though (see below).
> >
> >>>> + }
> >>>> + } else {
> >>>> + n = get_user_pages(current, current->mm, start & PAGE_MASK,
> >>>> + n_pages, write, 1, pages, NULL);
> >>>> + if (n != n_pages) {
> >>>> + printk(KERN_ERR "got only %d of %d user pages\n",
> >>>> + n, n_pages);
> >>>> + goto cleanup;
> >>>> + }
> >>>> + }
> >>>> +
> >>>> + *copy_vma = vb2_get_vma(vma);
> >>>> + if (!*copy_vma) {
> >>>> + printk(KERN_ERR "failed to copy vma\n");
> >>>> + ret = -ENOMEM;
> >>>> + goto cleanup;
> >>>> + }
> >>>
> >>> Do we really need to make a copy of the VMA ? The only reason why we
> >>> store a pointer to it is to check the flags in vb2_dc_put_userptr(). We
> >>> could store the flags instead and avoid vb2_get_dma()/vb2_put_dma()
> >>> calls altogether.
> >>
> >> I remember that there was a very good reason of copying this vma
> >> structure.
> >> You caught me on 'cargo-cult' programming.
> >> I will do some reverse engineering and try to answer it soon.
> >
> > OK :-) I'm not copying the VMA in the OMAP3 ISP driver, which is why this
> > caught my eyes. If you find the reason why copying it is needed, please
> > add a comment to the code.
>
> The reason of copying vma was that 'struct vma' has no reference counters.
> Therefore it could be deleted after mm lock is freed, ending with freeing
> its all pages belonging to vma. To prevent it, a copy of vma is created.
> Notice that inside vb2_get_vma the callback open is called for original
> vma, preventing memory from being released. On vb2_put_vma the
> complementary close is called.
Feel free to prove me wrong, but I think get_user_pages() is enough to prevent
the pages from being freed, even if the VMA is deleted.
However, there's one subtle issue that we will need to deal with when we will
implement cache management. It took me a lot of time to debug and fix it when
I was working on the OMAP3 ISP driver, so I'll explain it in the hope that
someone will find it later when dealing with the same problems :-)
When a buffer is queued, the OMAP3 ISP driver cleans up the cache using the
userspace mapping addresses (for USERPTR buffers). This might be a bad thing,
but that's the way we currently implement that.
A prior get_user_pages() call will make sure pages are not freed, but due to
optimization in the lockless memory management algorithms the userspace
mapping can disappear: the kernel might consider that a page can be freed,
remove its userspace mapping, and then find out that the page is locked. It
will then move on without restoring the userspace mapping, which will be
restored when the next page fault occurs.
When cleaning the cache using the userspace mapping addresses, any page for
which the userspace mapping has been removed will trigger a page fault. The
page fault handler (do_page_fault() in arm/arch/mm/fault.c) will try to
read_lock mmap_sem. If it fails, it will check if the page fault occured in
userspace context, or from a known exception location. As neither condition is
true, it will panic.
The solution I found to this issue was to lock the VMA. This ensured that the
userspace mapping would stay in place. See isp_video_buffer_lock_vma() in
drivers/media/video/omap3isp/ispqueue.c. You could use a similar approach here
if you want to ensure that userspace mappings are not removed, but once again
I don't think that's needed (until we get to cache handling) as
get_user_pages() will ensure that the pages are not freed.
> >>>> +
> >>>> + /* leaving critical section for mm access */
> >>>> + up_read(¤t->mm->mmap_sem);
> >>>> +
> >>>> + return 0;
> >>>> +
> >>>> +cleanup:
> >>>> + up_read(¤t->mm->mmap_sem);
> >>>> +
> >>>> + /* putting user pages if used, can be done wothout the lock */
> >>>> + while (n)
> >>>> + put_page(pages[--n]);
> >>>> +
> >>>> + return ret;
> >>>> +}
> >>>> +
> >>>>
> >>>> static void *vb2_dc_get_userptr(void *alloc_ctx, unsigned long vaddr,
> >>>>
> >>>> - unsigned long size, int write)
> >>>> + unsigned long size, int write)
> >>>>
> >>>> {
> >>>>
> >>>> struct vb2_dc_buf *buf;
> >>>>
> >>>> - struct vm_area_struct *vma;
> >>>> - dma_addr_t dma_addr = 0;
> >>>> - int ret;
> >>>> + unsigned long start, end, offset, offset2;
> >>>> + struct page **pages;
> >>>> + int n_pages;
> >>>> + int ret = 0;
> >>>> + struct sg_table *sgt;
> >>>> + unsigned long contig_size;
> >>>>
> >>>> buf = kzalloc(sizeof *buf, GFP_KERNEL);
> >>>> if (!buf)
> >>>>
> >>>> return ERR_PTR(-ENOMEM);
> >>>>
> >>>> - ret = vb2_get_contig_userptr(vaddr, size, &vma, &dma_addr);
> >>>> + buf->dev = alloc_ctx;
> >>>> + buf->dma_dir = write ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
> >>>> +
> >>>> + start = (unsigned long)vaddr & PAGE_MASK;
> >>>> + offset = (unsigned long)vaddr & ~PAGE_MASK;
> >>>> + end = PAGE_ALIGN((unsigned long)vaddr + size);
> >>>> + offset2 = end - (unsigned long)vaddr - size;
> >>>> + n_pages = (end - start) >> PAGE_SHIFT;
> >>>> +
> >>>> + pages = kmalloc(n_pages * sizeof pages[0], GFP_KERNEL);
> >>>> + if (!pages) {
> >>>> + ret = -ENOMEM;
> >>>> + printk(KERN_ERR "failed to allocate pages table\n");
> >>>> + goto fail_buf;
> >>>> + }
> >>>> +
> >>>> + /* extract page list from userspace mapping */
> >>>> + ret = vb2_dc_get_pages(start, pages, n_pages, &buf->vma, write);
> >>>>
> >>>> if (ret) {
> >>>>
> >>>> - printk(KERN_ERR "Failed acquiring VMA for vaddr 0x%08lx\n",
> >>>> - vaddr);
> >>>> - kfree(buf);
> >>>> - return ERR_PTR(ret);
> >>>> + printk(KERN_ERR "failed to get user pages\n");
> >>>> + goto fail_pages;
> >>>> + }
> >>>> +
> >>>> + sgt = vb2_dc_pages_to_sgt(pages, n_pages, offset, offset2);
> >>>> + if (!sgt) {
> >>>> + printk(KERN_ERR "failed to create scatterlist table\n");
> >>>> + ret = -ENOMEM;
> >>>> + goto fail_get_pages;
> >>>>
> >>>> }
> >>>
> >>> This looks overly complex to me. You create a multi-chunk sgt out of the
> >>> user pointer address and map it completely, and then check if it starts
> >>> with a big enough contiguous chunk.
> >>
> >> Notice that vb2_dc_pages_to_sgt does compress contiguous ranges of pfns
> >> (pages). So if the memory is contiguous, then sigle chunk sglist is
> >> produced. The memory used to store pages list is just temporary. It is
> >> freed after the sglist is created.
> >
> > That's exactly my point. The memory needs to be contiguous to be usable.
> > If it isn't, vb2-dma-contig will only use the first contiguous chunk. We
> > could thus simplify the code by hardcoding the single-chunk assumption.
> > vb2-dma-contig would walk user user pages list (or the PFN, depending on
> > the VMA flags) and stop at the first discontinuity. It would then create
> > a single-entry sg list and operate on that, without mapping or otherwise
> > touching the rest of the VMA, which is unusable to the device anyway.
>
> [see above]
>
> >>> Why don't you create an sgt with a single continuous
> >>> chunk then ? In the VM_PFNMAP case you could check whether the area is
> >>> contiguous when you follow the PFNs, stop at the first discontinuity,
> >>> and create an sgt with a single element right there. You would then need
> >>> to call vb2_dc_pages_to_sgt() in the normal case only, and stop at the
> >>> first discontinuity as well.
> >>
> >> Discontinuity of pfns is not a problem if a device has own IOMMU. It is
> >> not known for vb2-dma-contig if mapping this multi-chunk sglist will
> >> succeed until calling and checking a result of dma_map_sg.
> >
> > If the device has an IOMMU it won't need contiguous memory. Shouldn't it
> > then use vb2-dma-sg instead ?
>
> No. The vb2-dma-sg is used for devices that can deal with memory accessed
> from multiple chunks. For example a device with attached DMA engine that
> can handle a scattergather list.
>
> >> Why bothering if both VM_PFNMAP and non-VM_PFNMAP are handled in the same
> >> way after list of pages is obtained? Trating them the same way allows to
> >> reuse code and simplify the program flow.
> >>
> >> The DMA framework does not provide any way to force single chunk mapping
> >> in sg. If the device is capable of mapping discontinous pages into a
> >> single chunk the DMA framework will probably do merge the pages. The
> >> documentation encourages to merge the list but it is not obligatory.
> >>
> >> The reason is that if 'struct scatterlist' contains no dma_length field,
> >> what is controlled by CONFIG_NEED_SG_DMA_LENGTH macro, then field length
> >> is used instead. Chunk marging cannot be done in such a case.
> >> This is the reason why I look for the longest contiguous block.
> >>
> >>>> + /* pages are no longer needed */
> >>>> + kfree(pages);
> >>>> + pages = NULL;
> >>>> +
> >>>> + sgt->nents = dma_map_sg(buf->dev, sgt->sgl, sgt->orig_nents,
> >>>> + buf->dma_dir);
> >>>> + if (sgt->nents <= 0) {
> >>>> + printk(KERN_ERR "failed to map scatterlist\n");
> >>>> + ret = -EIO;
> >>>> + goto fail_sgt;
> >>>> + }
> >>>> +
> >>>> + contig_size = vb2_dc_get_contiguous_size(sgt);
> >>>> + if (contig_size < size) {
> >>>> + printk(KERN_ERR "contiguous mapping is too small %lu/%lu\n",
> >>>> + contig_size, size);
> >>>> + ret = -EFAULT;
> >>>> + goto fail_map_sg;
> >>>> + }
> >>>> +
> >>>> + buf->dma_addr = sg_dma_address(sgt->sgl);
> >>>>
> >>>> buf->size = size;
> >>>>
> >>>> - buf->dma_addr = dma_addr;
> >>>> - buf->vma = vma;
> >>>> + buf->dma_sgt = sgt;
> >>>> +
> >>>> + atomic_inc(&buf->refcount);
> >>>>
> >>>> return buf;
> >>>>
> >>>> +
> >>>> +fail_map_sg:
> >>>> + dma_unmap_sg(buf->dev, sgt->sgl, sgt->nents, buf->dma_dir);
> >>>
> >>> I think this will break in the VM_PFNMAP case on non-coherent
> >>> architectures. arm_dma_unmap_page() will call __dma_page_dev_to_cpu() in
> >>> that case, which can dereference struct page. As explain above, the
> >>> struct page isn't valid with VM_PFNMAP. I haven't check the dma_map_sg()
> >>> and dma_sync_sg_*() calls, but changes are they might break as well.
> >>
> >> It will crash as long it is true that there is no struct page behind
> >> given pfn. In practice, I found that VM_PFNMAP means that one cannot
> >> assume that there is a 'struct page' behind PFNs of a given mapping.
> >> Thoses struct pages really exists for all our drivers. Anyway, I agree
> >> that using those pages is a hack.
> >
> > They don't exist for the memory used as frame buffer on the OMAP3 (or at
> > least didn't exist in the N900 and N9, I haven't checked since). This
> > could become just a bad distant memory when drivers will use CMA.
>
> I see. That is why one should check page validity before calling get_pages.
> Other way would be splitting logic between PFNMAP and non-PFNMAP userspace
> mappings.
The OMAP3 ISP driver splits the logic for the PFNMAP and non-PFNMAP cases and
the code is quite clean (at least in my opinion ;-)). I can give you a quick
tour around its allocator if you want. I would go for that solution here as
well.
> >> It could be avoided if vb2_dc_get_pages returned a list of PFNs. Anyway,
> >> those PFNs have to be transformed to pages to create an sglist. Those
> >> pointers might be accessed somewhere deep inside dma_map_sg internals.
> >>
> >> The quite good solution would be dropping support for VM_PFNMAP mappings
> >> since they cannot be handled reliably.
> >
> > We should either drop VM_PFNMAP support or fix the DMA SG mapping API to
> > support VM_PFNMAP-style memory. I would vote for the former, as that's way
> > simpler and we have no VM_PFNMAP use case right now.
>
> Good point. But this way we have new dependencies on DMA mapping framework.
>
> >>>> +
> >>>> +fail_sgt:
> >>>> + vb2_dc_put_sgtable(sgt, 0);
> >>>> +
> >>>> +fail_get_pages:
> >>>> + while (pages && n_pages)
> >>>> + put_page(pages[--n_pages]);
> >>>> + vb2_put_vma(buf->vma);
> >>>> +
> >>>> +fail_pages:
> >>>> + kfree(pages); /* kfree is NULL-proof */
> >>>> +
> >>>> +fail_buf:
> >>>> + kfree(buf);
> >>>> +
> >>>> + return ERR_PTR(ret);
> >>>>
> >>>> }
--
Regards,
Laurent Pinchart
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/dri-devel
