On Sat, Oct 10, 2020 at 11:32 PM Daniel Vetter daniel.vetter@ffwll.ch wrote:
On Sat, Oct 10, 2020 at 10:27 PM Oded Gabbay oded.gabbay@gmail.com wrote:
On Fri, Oct 9, 2020 at 10:59 AM Daniel Vetter daniel.vetter@ffwll.ch wrote:
All we need are a pages array, pin_user_pages_fast can give us that directly. Plus this avoids the entire raw pfn side of get_vaddr_frames.
Thanks for the patch Daniel.
Signed-off-by: Daniel Vetter daniel.vetter@intel.com Cc: Jason Gunthorpe jgg@ziepe.ca Cc: Andrew Morton akpm@linux-foundation.org Cc: John Hubbard jhubbard@nvidia.com Cc: Jérôme Glisse jglisse@redhat.com Cc: Jan Kara jack@suse.cz Cc: Dan Williams dan.j.williams@intel.com Cc: linux-mm@kvack.org Cc: linux-arm-kernel@lists.infradead.org Cc: linux-samsung-soc@vger.kernel.org Cc: linux-media@vger.kernel.org Cc: Oded Gabbay oded.gabbay@gmail.com Cc: Omer Shpigelman oshpigelman@habana.ai Cc: Ofir Bitton obitton@habana.ai Cc: Tomer Tayar ttayar@habana.ai Cc: Moti Haimovski mhaimovski@habana.ai Cc: Daniel Vetter daniel.vetter@ffwll.ch Cc: Greg Kroah-Hartman gregkh@linuxfoundation.org Cc: Pawel Piskorski ppiskorski@habana.ai -- v2: Use unpin_user_pages_dirty_lock (John)
drivers/misc/habanalabs/Kconfig | 1 - drivers/misc/habanalabs/common/habanalabs.h | 3 +- drivers/misc/habanalabs/common/memory.c | 49 ++++++++------------- 3 files changed, 20 insertions(+), 33 deletions(-)
diff --git a/drivers/misc/habanalabs/Kconfig b/drivers/misc/habanalabs/Kconfig index 8eb5d38c618e..2f04187f7167 100644 --- a/drivers/misc/habanalabs/Kconfig +++ b/drivers/misc/habanalabs/Kconfig @@ -6,7 +6,6 @@ config HABANA_AI tristate "HabanaAI accelerators (habanalabs)" depends on PCI && HAS_IOMEM
select FRAME_VECTOR select DMA_SHARED_BUFFER select GENERIC_ALLOCATOR select HWMONdiff --git a/drivers/misc/habanalabs/common/habanalabs.h b/drivers/misc/habanalabs/common/habanalabs.h index edbd627b29d2..c1b3ad613b15 100644 --- a/drivers/misc/habanalabs/common/habanalabs.h +++ b/drivers/misc/habanalabs/common/habanalabs.h @@ -881,7 +881,8 @@ struct hl_ctx_mgr { struct hl_userptr { enum vm_type_t vm_type; /* must be first */ struct list_head job_node;
struct frame_vector *vec;
struct page **pages;unsigned int npages;Can you please update the kerneldoc comment section of this structure according to your changes ?
Apologies I missed the nice kerneldoc. I'll fix that in the next round.
struct sg_table *sgt; enum dma_data_direction dir; struct list_head debugfs_list;diff --git a/drivers/misc/habanalabs/common/memory.c b/drivers/misc/habanalabs/common/memory.c index 5ff4688683fd..327b64479f97 100644 --- a/drivers/misc/habanalabs/common/memory.c +++ b/drivers/misc/habanalabs/common/memory.c @@ -1281,45 +1281,41 @@ static int get_user_memory(struct hl_device *hdev, u64 addr, u64 size, return -EFAULT; }
userptr->vec = frame_vector_create(npages);if (!userptr->vec) {
userptr->pages = kvmalloc_array(npages, sizeof(*userptr->pages),GFP_KERNEL);if (!userptr->pages) { dev_err(hdev->dev, "Failed to create frame vector\n"); return -ENOMEM; }
rc = get_vaddr_frames(start, npages, FOLL_FORCE | FOLL_WRITE,userptr->vec);
rc = pin_user_pages_fast(start, npages, FOLL_FORCE | FOLL_WRITE,userptr->pages); if (rc != npages) { dev_err(hdev->dev, "Failed to map host memory, user ptr probably wrong\n"); if (rc < 0)
goto destroy_framevec;
goto destroy_pages;npages = rc; rc = -EFAULT;
goto put_framevec;}if (frame_vector_to_pages(userptr->vec) < 0) {dev_err(hdev->dev,"Failed to translate frame vector to pages\n");rc = -EFAULT;goto put_framevec;
goto put_pages; }userptr->npages = npages; rc = sg_alloc_table_from_pages(userptr->sgt,
frame_vector_pages(userptr->vec),npages, offset, size, GFP_ATOMIC);
userptr->pages,npages, offset, size, GFP_ATOMIC);I think that because the call to kvmalloc_array() is done with GFP_KERNEL, there is no point in using GFP_ATOMIC here. And actually, this path only needs to avoid yielding when using a special debug mode. So I suggest putting here GFP_KERNEL.
Huh, I didn't even notice the GFP_ATOMIC here. This looks indeed strange and GFP_KERNEL should be perfectly fine in a function that also calls pin_user_pages (since that one can allocate and do worse stuff like userspace pagefaults).
But since that GFP_ATOMIC is there already I'll do that in a separate patch.
Ok I read up on your usage of GFP_ATOMIC in habanalabs, and I'm not going to touch this. But I'm pretty sure it's broken.
You seem to have some requirement of not allocating memory with blocking (see hl_cb_alloc()), and that seems to be way you allocate tons of structures with GFP_ATOMIC. There's 2 pretty tough problems with that: - GFP_ATOMIC can fail, even when the system hasn't run out of memory yet. You _must_ have a fallback back to handle allocation failures for these. Quick survey shows you a ton of GFP_ATOMIC callsites, and very little fallback code - I've found none, but I didn't check the failure handlers all going up the possible callchains. - pin_user_pages can allocate memory, so you're breaking your own "no sleeping in these paths" rules.
This isn't going to get fixed with a quick oneliner patch, depending what's needed you're looking at a driver rearchitecture here :-/ Hence I'm not going to touch this in the next patch, but leave it all as-is.
Cheers, Daniel
In the meanwhile, I'll run this patch (coupled with the next patch) in our C/I to make sure there are no regressions.
Excellent. I'll wait with v3 until that's done, just in case you hit a snag I need to fix.
Cheers, Daniel
Thanks, Oded
if (rc < 0) { dev_err(hdev->dev, "failed to create SG table from pages\n");
goto put_framevec;
goto put_pages; } return 0;-put_framevec:
put_vaddr_frames(userptr->vec);-destroy_framevec:
frame_vector_destroy(userptr->vec);+put_pages:
unpin_user_pages(userptr->pages, npages);+destroy_pages:
kvfree(userptr->pages); return rc;}
@@ -1405,8 +1401,6 @@ int hl_pin_host_memory(struct hl_device *hdev, u64 addr, u64 size, */ void hl_unpin_host_memory(struct hl_device *hdev, struct hl_userptr *userptr) {
struct page **pages;hl_debugfs_remove_userptr(hdev, userptr); if (userptr->dma_mapped)@@ -1414,15 +1408,8 @@ void hl_unpin_host_memory(struct hl_device *hdev, struct hl_userptr *userptr) userptr->sgt->nents, userptr->dir);
pages = frame_vector_pages(userptr->vec);if (!IS_ERR(pages)) {int i;for (i = 0; i < frame_vector_count(userptr->vec); i++)set_page_dirty_lock(pages[i]);}put_vaddr_frames(userptr->vec);frame_vector_destroy(userptr->vec);
unpin_user_pages_dirty_lock(userptr->pages, userptr->npages, true);kvfree(userptr->pages); list_del(&userptr->job_node);-- 2.28.0
-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch