From: Ben Widawsky <benjamin.widawsky@intel.com>
To: Intel GFX <intel-gfx@lists.freedesktop.org>
Cc: Ben Widawsky <ben@bwidawsk.net>,
Ben Widawsky <benjamin.widawsky@intel.com>
Subject: [PATCH 56/56] drm/i915/userptr: Mirror GPU addr at ioctl (HACK/POC)
Date: Fri, 9 May 2014 20:59:51 -0700 [thread overview]
Message-ID: <1399694391-3935-57-git-send-email-benjamin.widawsky@intel.com> (raw)
In-Reply-To: <1399694391-3935-1-git-send-email-benjamin.widawsky@intel.com>
This is needed for the proof of concept work that will allow mirrored
GPU addressing via the existing userptr interface. Part of the hack
involves passing the context ID to the ioctl in order to get a VM.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
---
drivers/gpu/drm/i915/i915_gem_userptr.c | 120 +++++++++++++++++++++++++-------
include/uapi/drm/i915_drm.h | 7 +-
2 files changed, 98 insertions(+), 29 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_gem_userptr.c b/drivers/gpu/drm/i915/i915_gem_userptr.c
index 5da37cc..795ea3e 100644
--- a/drivers/gpu/drm/i915/i915_gem_userptr.c
+++ b/drivers/gpu/drm/i915/i915_gem_userptr.c
@@ -224,10 +224,6 @@ i915_mmu_notifier_add(struct i915_mmu_notifier *mmu,
* remove the objects from the interval tree) before we do
* the check for overlapping objects.
*/
- ret = i915_mutex_lock_interruptible(mmu->dev);
- if (ret)
- return ret;
-
i915_gem_retire_requests(mmu->dev);
/* Disallow overlapping userptr objects */
@@ -253,7 +249,6 @@ i915_mmu_notifier_add(struct i915_mmu_notifier *mmu,
ret = 0;
}
spin_unlock(&mmu->lock);
- mutex_unlock(&mmu->dev->struct_mutex);
return ret;
}
@@ -283,19 +278,12 @@ i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
down_write(&obj->userptr.mm->mmap_sem);
- ret = i915_mutex_lock_interruptible(obj->base.dev);
- if (ret == 0) {
- mmu = i915_mmu_notifier_get(obj->base.dev, obj->userptr.mm);
- if (!IS_ERR(mmu))
- mmu->count++; /* preemptive add to act as a refcount */
- else
- ret = PTR_ERR(mmu);
- mutex_unlock(&obj->base.dev->struct_mutex);
- }
+ mmu = i915_mmu_notifier_get(obj->base.dev, obj->userptr.mm);
+ if (!IS_ERR(mmu))
+ mmu->count++; /* preemptive add to act as a refcount */
+ else
+ ret = PTR_ERR(mmu);
up_write(&obj->userptr.mm->mmap_sem);
- if (ret)
- return ret;
-
mn = kzalloc(sizeof(*mn), GFP_KERNEL);
if (mn == NULL) {
ret = -ENOMEM;
@@ -588,6 +576,52 @@ i915_gem_userptr_release(struct drm_i915_gem_object *obj)
}
}
+/* Carve out the address space for later use */
+static int i915_gem_userptr_reserve_vma(struct drm_i915_gem_object *obj,
+ struct i915_address_space *vm,
+ uint64_t offset,
+ uint64_t size)
+{
+ struct i915_vma *vma;
+ int ret;
+
+ vma = i915_gem_obj_to_vma(obj, vm);
+ if (vma)
+ return -ENXIO;
+
+ vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
+ if (!vma)
+ return PTR_ERR(vma);
+
+ BUG_ON(!drm_mm_initialized(&vm->mm));
+
+ if (vma->uptr) {
+ DRM_INFO("Already had a userptr\n");
+ return 0;
+ }
+ if (vma->node.allocated) {
+ DRM_INFO("Node was previously allocated\n");
+ return -EBUSY;
+ }
+
+ vma->node.start = offset;
+ vma->node.size = size;
+ vma->node.color = 0;
+ ret = drm_mm_reserve_node(&vm->mm, &vma->node);
+ if (ret) {
+ /* There are two reasons this can fail.
+ * 1. The user is using a mix of relocs and userptr, and a reloc
+ * won.
+ * TODO: handle better.
+ */
+ return ret;
+ }
+
+ vma->uptr = 1;
+
+ return 0;
+}
+
static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
.get_pages = i915_gem_userptr_get_pages,
.put_pages = i915_gem_userptr_put_pages,
@@ -630,37 +664,62 @@ static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
int
i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
{
- struct drm_i915_private *dev_priv = dev->dev_private;
+ struct drm_i915_file_private *file_priv = file->driver_priv;
struct drm_i915_gem_userptr *args = data;
struct drm_i915_gem_object *obj;
+ struct i915_hw_context *ctx;
+ struct i915_address_space *vm;
int ret;
u32 handle;
+ ret = i915_mutex_lock_interruptible(dev);
+ if (ret)
+ return ret;
+
+#define goto_err(__err) do { \
+ ret = (__err); \
+ goto out; \
+} while (0)
+
+ ctx = i915_gem_context_get(file_priv, args->ctx_id);
+ if (IS_ERR(ctx))
+ goto_err(PTR_ERR(ctx));
+
+ /* i915_gem_context_reference(ctx); */
+
if (args->flags & ~(I915_USERPTR_READ_ONLY |
+ I915_USERPTR_GPU_MIRROR |
I915_USERPTR_UNSYNCHRONIZED))
- return -EINVAL;
+ goto_err(-EINVAL);
if (offset_in_page(args->user_ptr | args->user_size))
- return -EINVAL;
-
- if (args->user_size > dev_priv->gtt.base.total)
- return -E2BIG;
+ goto_err(-EINVAL);
if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
(char __user *)(unsigned long)args->user_ptr, args->user_size))
- return -EFAULT;
+ goto_err(-EFAULT);
if (args->flags & I915_USERPTR_READ_ONLY) {
/* On almost all of the current hw, we cannot tell the GPU that a
* page is readonly, so this is just a placeholder in the uAPI.
*/
- return -ENODEV;
+ goto_err(-ENODEV);
+ }
+
+ vm = ctx->vm;
+ if (args->user_size > vm->total)
+ goto_err(-E2BIG);
+
+ if (args->flags & I915_USERPTR_GPU_MIRROR) {
+ if (!HAS_48B_PPGTT(dev))
+ goto_err(-ENODEV);
}
/* Allocate the new object */
obj = i915_gem_object_alloc(dev);
if (obj == NULL)
- return -ENOMEM;
+ goto_err(-ENOMEM);
+#undef goto_err
drm_gem_private_object_init(dev, &obj->base, args->user_size);
i915_gem_object_init(obj, &i915_gem_userptr_ops);
@@ -680,9 +739,16 @@ i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file
ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
if (ret == 0)
ret = drm_gem_handle_create(file, &obj->base, &handle);
+ if (ret == 0 && args->flags & I915_USERPTR_GPU_MIRROR) {
+ ret = i915_gem_userptr_reserve_vma(obj, vm, args->user_ptr, args->user_size);
+ if (ret)
+ DRM_DEBUG_DRIVER("Failed to reserve GPU mirror %d\n", ret);
+ }
/* drop reference from allocate - handle holds it now */
- drm_gem_object_unreference_unlocked(&obj->base);
+ drm_gem_object_unreference(&obj->base);
+out:
+ mutex_unlock(&dev->struct_mutex);
if (ret)
return ret;
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index dd98463..addab22 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -1056,15 +1056,18 @@ struct drm_i915_reset_stats {
struct drm_i915_gem_userptr {
__u64 user_ptr;
__u64 user_size;
+ __u32 ctx_id;
__u32 flags;
-#define I915_USERPTR_READ_ONLY 0x1
-#define I915_USERPTR_UNSYNCHRONIZED 0x80000000
+#define I915_USERPTR_READ_ONLY (1<<0)
+#define I915_USERPTR_GPU_MIRROR (1<<1)
+#define I915_USERPTR_UNSYNCHRONIZED (1<<31)
/**
* Returned handle for the object.
*
* Object handles are nonzero.
*/
__u32 handle;
+ __u32 pad;
};
#endif /* _UAPI_I915_DRM_H_ */
--
1.9.2
next prev parent reply other threads:[~2014-05-10 4:03 UTC|newest]
Thread overview: 58+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-05-10 3:58 [PATCH 00/56] [RFCish] Dynamic page table alloc, 64b, and GPU/CPU mirror Ben Widawsky
2014-05-10 3:58 ` [PATCH 01/56] drm/i915: Fix flush before context switch comment Ben Widawsky
2014-05-10 3:58 ` [PATCH 02/56] Revert "drm/i915: Drop I915_PARAM_HAS_FULL_PPGTT again" Ben Widawsky
2014-05-10 3:58 ` [PATCH 03/56] drm/i915: Prevent signals from interrupting close() Ben Widawsky
2014-05-10 3:58 ` [PATCH 04/56] drm/i915: Wrap VMA binding Ben Widawsky
2014-05-10 3:59 ` [PATCH 05/56] drm/i915: Make pin global flags explicit Ben Widawsky
2014-05-10 3:59 ` [PATCH 06/56] drm/i915: Split out aliasing binds Ben Widawsky
2014-05-10 3:59 ` [PATCH 07/56] drm/i915: fix gtt_total_entries() Ben Widawsky
2014-05-10 3:59 ` [PATCH 08/56] drm/i915: Rename to GEN8_LEGACY_PDPES Ben Widawsky
2014-05-10 3:59 ` [PATCH 09/56] drm/i915: Split out verbose PPGTT dumping Ben Widawsky
2014-05-10 3:59 ` [PATCH 10/56] drm/i915: s/pd/pdpe, s/pt/pde Ben Widawsky
2014-05-10 3:59 ` [PATCH 11/56] drm/i915: rename map/unmap to dma_map/unmap Ben Widawsky
2014-05-10 3:59 ` [PATCH 12/56] drm/i915: Setup less PPGTT on failed pagedir Ben Widawsky
2014-05-10 3:59 ` [PATCH 13/56] drm/i915: clean up PPGTT init error path Ben Widawsky
2014-05-10 3:59 ` [PATCH 14/56] drm/i915: Un-hardcode number of page directories Ben Widawsky
2014-05-10 3:59 ` [PATCH 15/56] drm/i915: Make gen6_write_pdes gen6_map_page_tables Ben Widawsky
2014-05-10 3:59 ` [PATCH 16/56] drm/i915: Range clearing is PPGTT agnostic Ben Widawsky
2014-05-10 3:59 ` [PATCH 17/56] drm/i915: Page table helpers, and define renames Ben Widawsky
2014-05-10 3:59 ` [PATCH 18/56] drm/i915: construct page table abstractions Ben Widawsky
2014-05-10 3:59 ` [PATCH 19/56] drm/i915: Complete page table structures Ben Widawsky
2014-05-10 3:59 ` [PATCH 20/56] drm/i915: Create page table allocators Ben Widawsky
2014-05-10 3:59 ` [PATCH 21/56] drm/i915: Generalize GEN6 mapping Ben Widawsky
2014-05-10 3:59 ` [PATCH 22/56] drm/i915: Clean up pagetable DMA map & unmap Ben Widawsky
2014-05-10 3:59 ` [PATCH 23/56] drm/i915: Always dma map page table allocations Ben Widawsky
2014-05-10 3:59 ` [PATCH 24/56] drm/i915: Consolidate dma mappings Ben Widawsky
2014-05-10 3:59 ` [PATCH 25/56] drm/i915: Always dma map page directory allocations Ben Widawsky
2014-05-10 3:59 ` [PATCH 26/56] drm/i915: Track GEN6 page table usage Ben Widawsky
2014-05-10 3:59 ` [PATCH 27/56] drm/i915: Extract context switch skip logic Ben Widawsky
2014-05-10 3:59 ` [PATCH 28/56] drm/i915: Force pd restore when PDEs change, gen6-7 Ben Widawsky
2014-05-10 3:59 ` [PATCH 29/56] drm/i915: Finish gen6/7 dynamic page table allocation Ben Widawsky
2014-05-10 3:59 ` [PATCH 30/56] drm/i915/bdw: Use dynamic allocation idioms on free Ben Widawsky
2014-05-10 3:59 ` [PATCH 31/56] drm/i915/bdw: pagedirs rework allocation Ben Widawsky
2014-05-10 3:59 ` [PATCH 32/56] drm/i915/bdw: pagetable allocation rework Ben Widawsky
2014-05-10 3:59 ` [PATCH 33/56] drm/i915/bdw: Make the pdp switch a bit less hacky Ben Widawsky
2014-05-10 3:59 ` [PATCH 34/56] drm/i915: num_pd_pages/num_pd_entries isn't useful Ben Widawsky
2014-05-10 3:59 ` [PATCH 35/56] drm/i915: Extract PPGTT param from pagedir alloc Ben Widawsky
2014-05-10 3:59 ` [PATCH 36/56] drm/i915/bdw: Split out mappings Ben Widawsky
2014-05-10 3:59 ` [PATCH 37/56] drm/i915/bdw: begin bitmap tracking Ben Widawsky
2014-05-10 3:59 ` [PATCH 38/56] drm/i915/bdw: Dynamic page table allocations Ben Widawsky
2014-05-10 3:59 ` [PATCH 39/56] drm/i915/bdw: Scratch unused pages Ben Widawsky
2014-05-10 3:59 ` [PATCH 40/56] drm/i915/bdw: Add ppgtt info for dynamic pages Ben Widawsky
2014-05-10 3:59 ` [PATCH 41/56] drm/i915/bdw: Optimize PDP loads Ben Widawsky
2014-05-10 3:59 ` [PATCH 42/56] TESTME: Either drop the last patch or fix it Ben Widawsky
2014-05-10 3:59 ` [PATCH 43/56] drm/i915/bdw: Add dynamic page trace events Ben Widawsky
2014-05-10 3:59 ` [PATCH 44/56] drm/i915/bdw: Make pdp allocation more dynamic Ben Widawsky
2014-05-10 3:59 ` [PATCH 45/56] drm/i915/bdw: Abstract PDP usage Ben Widawsky
2014-05-10 3:59 ` [PATCH 46/56] drm/i915/bdw: implement alloc/teardown for 4lvl Ben Widawsky
2014-05-10 3:59 ` [PATCH 47/56] drm/i915/bdw: 4 level pages tables Ben Widawsky
2014-05-10 3:59 ` [PATCH 48/56] drm/i915: Restructure map vs. insert entries Ben Widawsky
2014-05-10 3:59 ` [PATCH 49/56] drm/i915/bdw: make aliasing PPGTT dynamic Ben Widawsky
2014-05-10 3:59 ` [PATCH 50/56] drm/i915: Expand error state's address width to 64b Ben Widawsky
2014-05-10 3:59 ` [PATCH 51/56] drm/i915/bdw: Flip the 48b switch Ben Widawsky
2014-05-10 3:59 ` [PATCH 52/56] TESTME: GFX_TLB_INVALIDATE_EXPLICIT Ben Widawsky
2014-05-10 3:59 ` [PATCH 53/56] TESTME: Always force invalidate Ben Widawsky
2014-05-10 3:59 ` [PATCH 54/56] drm/i915: Introduce mapping of user pages into video memory (userptr) ioctl Ben Widawsky
2014-05-10 3:59 ` [PATCH 55/56] drm/i915: Track userptr VMAs Ben Widawsky
2014-05-10 3:59 ` Ben Widawsky [this message]
2014-05-11 17:33 ` [PATCH 00/56] [RFCish] Dynamic page table alloc, 64b, and GPU/CPU mirror Daniel Vetter
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=1399694391-3935-57-git-send-email-benjamin.widawsky@intel.com \
--to=benjamin.widawsky@intel.com \
--cc=ben@bwidawsk.net \
--cc=intel-gfx@lists.freedesktop.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox