drm/i915: Added write-enable pte bit support
Subsystems (2)
PATCH v3 1/1 drm/i915: Added write-enable pte bit support
From: Akash Goel <[email protected]>
This adds support for a write-enable bit in the entry of GTT. This is handled via a read-only flag in the GEM buffer object which is then used to see how to set the bit when writing the GTT entries. Currently by default the Batch buffer & Ring buffers are marked as read only.
v2: Moved the pte override code for read-only bit to 'byt_pte_encode'. (Chris)
Fixed the issue of leaving 'gt_old_ro' as unused. (Chris)
v3: Removed the 'gt_old_ro' field, now setting RO bit only for Ring Buffers(Daniel).
--- drivers/gpu/drm/i915/i915_drv.h | 6 ++++++ drivers/gpu/drm/i915/i915_gem_gtt.c | 14 ++++++++++++++ drivers/gpu/drm/i915/intel_ringbuffer.c | 3 +++ 3 files changed, 23 insertions(+)
drivers/gpu/drm/i915/i915_drv.h +6 -0
@@ -1699,6 +1699,12 @@ struct drm_i915_gem_object { unsigned int pin_display:1;/* +* Is the object to be mapped as read-only to the GPU +* Only honoured if hardware has relevant pte bit +*/ + unsigned long gt_ro:1; + +/* * Is the GPU currently using a fence to access this buffer, */ unsigned int pending_fenced_gpu_access:1;
drivers/gpu/drm/i915/i915_gem_gtt.c +14 -0
@@ -151,6 +151,7 @@ static gen6_gtt_pte_t ivb_pte_encode(dma_addr_t addr,#define BYT_PTE_WRITEABLE (1 << 1)#define BYT_PTE_SNOOPED_BY_CPU_CACHES (1 << 2) +#define BYT_PTE_READ_ONLY (1 << 31) static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr, enum i915_cache_level level,
@@ -167,6 +168,10 @@ static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr,if (level !=I915_CACHE_NONE ) pte |=BYT_PTE_SNOOPED_BY_CPU_CACHES ; +/* Handle read-only request */ +if (level &BYT_PTE_READ_ONLY ) + pte &= ~BYT_PTE_WRITEABLE ; +return pte; }
@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;
@@ -999,6 +1005,10 @@ ppgtt_bind_vma(struct i915_vma *vma,WARN_ON (flags); +if (IS_VALLEYVIEW (vma->vm->dev)) +if (vma->obj->gt_ro) + cache_level |=BYT_PTE_READ_ONLY ; + vma->vm->insert_entries (vma->vm, vma->obj->pages, entry, cache_level); }
@@ -1336,6 +1346,10 @@ static void ggtt_bind_vma(struct i915_vma *vma,struct drm_i915_gem_object *obj = vma->obj;const unsigned long entry = vma->node.start >>PAGE_SHIFT ; +if (IS_VALLEYVIEW (dev)) +if (obj->gt_ro) + cache_level |=BYT_PTE_READ_ONLY ; +/* If there is no aliasing PPGTT, or the caller needs a global mapping, * or we have a global mapping already but the cacheability flags have * changed, set the global PTEs.
drivers/gpu/drm/i915/intel_ringbuffer.c +3 -0
@@ -1354,6 +1354,9 @@ static int intel_init_ring_buffer(struct drm_device *dev, goto err_hws; } +/* mark ring buffers as read-only from GPU side by default */ + obj->gt_ro =1 ; + ring->obj = obj; ret =i915_gem_obj_ggtt_pin (obj,PAGE_SIZE , true, false);
Signature · 1 line
1.8.5.2
Hi Chris,
Gentle reminder, Please can you review this patch, have address the review comments.
Best Regards
Akash
-----Original Message-----
From: Goel, Akash
Sent: Tuesday, February 11, 2014 2:19 PM
To: [email protected]
Cc: Goel, Akash
Subject: [PATCH v3] drm/i915: Added write-enable pte bit support
From: Akash Goel <[email protected]>
This adds support for a write-enable bit in the entry of GTT.
This is handled via a read-only flag in the GEM buffer object which is then used to see how to set the bit when writing the GTT entries. Currently by default the Batch buffer & Ring buffers are marked as read only.
v2: Moved the pte override code for read-only bit to 'byt_pte_encode'. (Chris)
Fixed the issue of leaving 'gt_old_ro' as unused. (Chris)
v3: Removed the 'gt_old_ro' field, now setting RO bit only for Ring Buffers(Daniel).
--- drivers/gpu/drm/i915/i915_drv.h | 6 ++++++ drivers/gpu/drm/i915/i915_gem_gtt.c | 14 ++++++++++++++ drivers/gpu/drm/i915/intel_ringbuffer.c | 3 +++ 3 files changed, 23 insertions(+)
drivers/gpu/drm/i915/i915_drv.h +6 -0
@@ -1699,6 +1699,12 @@ struct drm_i915_gem_object { unsigned int pin_display:1;/* +* Is the object to be mapped as read-only to the GPU +* Only honoured if hardware has relevant pte bit +*/ + unsigned long gt_ro:1; + +/* * Is the GPU currently using a fence to access this buffer, */ unsigned int pending_fenced_gpu_access:1; diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
index 6e858e1..1e02d44 100644
drivers/gpu/drm/i915/i915_gem_gtt.c +14 -0
@@ -151,6 +151,7 @@ static gen6_gtt_pte_t ivb_pte_encode(dma_addr_t addr,#define BYT_PTE_WRITEABLE (1 << 1)#define BYT_PTE_SNOOPED_BY_CPU_CACHES (1 << 2) +#define BYT_PTE_READ_ONLY (1 << 31) static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr, enum i915_cache_level level,
@@ -167,6 +168,10 @@ static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr,if (level !=I915_CACHE_NONE ) pte |=BYT_PTE_SNOOPED_BY_CPU_CACHES ; +/* Handle read-only request */ +if (level &BYT_PTE_READ_ONLY ) + pte &= ~BYT_PTE_WRITEABLE ; +return pte; }
@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;
@@ -999,6 +1005,10 @@ ppgtt_bind_vma(struct i915_vma *vma,WARN_ON (flags); +if (IS_VALLEYVIEW (vma->vm->dev)) +if (vma->obj->gt_ro) + cache_level |=BYT_PTE_READ_ONLY ; + vma->vm->insert_entries (vma->vm, vma->obj->pages, entry, cache_level); }
@@ -1336,6 +1346,10 @@ static void ggtt_bind_vma(struct i915_vma *vma,struct drm_i915_gem_object *obj = vma->obj;const unsigned long entry = vma->node.start >>PAGE_SHIFT ; +if (IS_VALLEYVIEW (dev)) +if (obj->gt_ro) + cache_level |=BYT_PTE_READ_ONLY ; +/* If there is no aliasing PPGTT, or the caller needs a global mapping, * or we have a global mapping already but the cacheability flags have * changed, set the global PTEs.
drivers/gpu/drm/i915/intel_ringbuffer.c +3 -0
@@ -1354,6 +1354,9 @@ static int intel_init_ring_buffer(struct drm_device *dev, goto err_hws; } +/* mark ring buffers as read-only from GPU side by default */ + obj->gt_ro =1 ; + ring->obj = obj; ret =i915_gem_obj_ggtt_pin (obj,PAGE_SIZE , true, false);
--
1.8.5.2
drivers/gpu/drm/i915/i915_gem_gtt.c static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines · On Tue, 11 Feb 2014 14:19:03 +0530 [email protected] wrote:
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;
Some extra whitespace here.
Otherwise:
Might be good to expose this as a param too, so userspace could use it for stuff that shouldn't be written...
Signature · 1 line
Jesse Barnes, Intel Open Source Technology Center
Quoted 14 lines · Jesse Barnes
drivers/gpu/drm/i915/i915_gem_gtt.c static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines · On Tue, 11 Feb 2014 14:19:03 +0530 [email protected] wrote:
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;Some extra whitespace here.
Quoted 2 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>
Hm, looking at the patch again encoding this into the cache_level enum is fraught with fun. And due to IS_VLV aliasing chv this will blow up on chv very likely. My old idea was to eventually add a pte_flags param all over for this stuff with additional bits.
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
Quoted 15 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 14 lines · Jesse Barnes
drivers/gpu/drm/i915/i915_gem_gtt.c static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines · On Tue, 11 Feb 2014 14:19:03 +0530 [email protected] wrote:
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;Some extra whitespace here.
Quoted 7 lines
Quoted 2 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.
That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Ack on the cache_level abuse though; it's an implementation detail that should be changed if/when this is properly exposed to userspace (or maybe we shouldn't bother and just use CHV/BDW stuff going forward, supporting a proper mprotect ioctl).
Signature · 1 line
Jesse Barnes, Intel Open Source Technology Center
Quoted 21 lines · Jesse Barnes
Quoted 18 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 16 lines · Jesse Barnes
drivers/gpu/drm/i915/i915_gem_gtt.c static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines · On Tue, 11 Feb 2014 14:19:03 +0530 [email protected] wrote:
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;Some extra whitespace here.
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>
Quoted 7 lines
Quoted 4 lines
Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Yes, but we have a switch on the cache_level in the gen8 pte encode function. Since the bit31 gets or'ed in for VLV, it'll hit also chv and wreak havoc in that specific switch - we'll hit the default case when we don't expect to.
cache_level = functional behaviour relevant for the kernel's clflushing code
Quoted 4 lines
Ack on the
cache_levelabuse though; it's an implementation detail that should be changed if/when this is properly exposed to userspace (or maybe we shouldn't bother and just use CHV/BDW stuff going forward, supporting a proper mprotect ioctl).
Not talking about exposing to userspace, but adding pte_flags to the low-level platform pte encode functions to prevent the above bugs. I'm ok with keeping obj->gt_ro.
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
Quoted 8 lines · Daniel Vetter
Quoted 7 lines · Jesse Barnes
Quoted 4 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 3 lines · Jesse Barnes
On Tue, 11 Feb 2014 14:19:03 +0530
[email protected] wrote:
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_gtt.c static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;Some extra whitespace here.
Thanks, will remove this.
Quoted 12 lines
Quoted 11 lines
Quoted 8 lines
Quoted 3 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Quoted 8 lines
Yes, but we have a switch on the
cache_levelin the gen8 pte encode function. Since the bit31 gets or'ed in for VLV, it'll hit also chv and wreak havoc in that specific switch - we'll hit the default case when we don't expect to.
cache_level= functional behaviour relevant for the kernel's clflushing code
As the 'IS_VALLEYVIEW' check will alias with CHV also, can I just update the condition here, to include 'GEN7' also (IS_GEN7) in the check.
Quoted 9 lines
Quoted 4 lines
Ack on the
cache_levelabuse though; it's an implementation detail that should be changed if/when this is properly exposed to userspace (or maybe we shouldn't bother and just use CHV/BDW stuff going forward, supporting a proper mprotect ioctl).Not talking about exposing to userspace, but adding
pte_flagsto the low-level platform pte encode functions to prevent the above bugs. I'm ok with keepingobj->gt_ro.
-Daniel
Quoted 39 lines · Akash Goel
Quoted 8 lines · Daniel Vetter
Quoted 7 lines · Jesse Barnes
Quoted 4 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 3 lines · Jesse Barnes
On Tue, 11 Feb 2014 14:19:03 +0530
[email protected] wrote:Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_gtt.c static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), cache_level, true); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;Some extra whitespace here.
Thanks, will remove this.
Quoted 17 lines
Quoted 11 lines
Quoted 8 lines
Quoted 3 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Yes, but we have a switch on the
cache_levelin the gen8 pte encode function. Since the bit31 gets or'ed in for VLV, it'll hit also chv and wreak havoc in that specific switch - we'll hit the default case when we don't expect to.
Quoted 6 lines
Quoted 3 lines
cache_level= functional behaviour relevant for the kernel's clflushing codeAs the '
IS_VALLEYVIEW' check will alias with CHV also, can I just update the condition here, to include 'GEN7' also (IS_GEN7) in the check.
Adding new random bits to an enum which is used all over the place (and not just in the pte encoding functions) makes the code much harder to read. Also the code that deals with enum cache_level is all about cache coherency, which has rather tricky logic.
Hence I expect this choice to cause further bugs down the road, which is why I don't really like it. My apologies for not spotting this earlier. -Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
Quoted 47 lines · Daniel Vetter
Quoted 39 lines · Akash Goel
Quoted 8 lines · Daniel Vetter
Quoted 7 lines · Jesse Barnes
Quoted 4 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 3 lines · Jesse Barnes
On Tue, 11 Feb 2014 14:19:03 +0530
[email protected] wrote:Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Review static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines
(quoted diff) +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode(sg_page_iter_dma_address(&sg_iter), cache_level, true); + if (++act_pte == I915_PPGTT_PT_ENTRIES) { kunmap_atomic(pt_vaddr); pt_vaddr = NULL;Some extra whitespace here.
Thanks, will remove this.
Quoted 17 lines
Quoted 11 lines
Quoted 8 lines
Quoted 3 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Yes, but we have a switch on the
cache_levelin the gen8 pte encode function. Since the bit31 gets or'ed in for VLV, it'll hit also chv and wreak havoc in that specific switch - we'll hit the default case when we don't expect to.Quoted 6 lines
Quoted 3 lines
cache_level= functional behaviour relevant for the kernel's clflushing codeAs the '
IS_VALLEYVIEW' check will alias with CHV also, can I just update the condition here, to include 'GEN7' also (IS_GEN7) in the check.
Quoted 8 lines
Adding new random bits to an enum which is used all over the place (and not just in the pte encoding functions) makes the code much harder to read. Also the code that deals with enum
cache_levelis all about cache coherency, which has rather tricky logic.Hence I expect this choice to cause further bugs down the road, which is why I don't really like it. My apologies for not spotting this earlier. -Daniel
Hi Daniel,
I understand your concern, but actually (as of now) this bit is only VLV specific. As per an earlier suggestion from Chris, I decided to overload the cache_level enum itself, in lieu of adding a new parameter to all the respective 'xxx_insert_entries' & 'xxx_pte_encode' functions.
And this is being done just before calling the 'xxx_insert_entries' function, which simply passes the flag as it is to 'xxx_pte_encode' function. So there may not be any risk here, if we use the appropriate VLV check.
Best Regards
Akash
Quoted 64 lines · Akash Goel
Quoted 47 lines · Daniel Vetter
Quoted 39 lines · Akash Goel
Quoted 8 lines · Daniel Vetter
Quoted 7 lines · Jesse Barnes
Quoted 4 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 3 lines · Jesse Barnes
On Tue, 11 Feb 2014 14:19:03 +0530
[email protected] wrote:Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Review static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines
(quoted diff) +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode(sg_page_iter_dma_address(&sg_iter), cache_level, true); + if (++act_pte == I915_PPGTT_PT_ENTRIES) { kunmap_atomic(pt_vaddr); pt_vaddr = NULL;Some extra whitespace here.
Thanks, will remove this.
Quoted 17 lines
Quoted 11 lines
Quoted 8 lines
Quoted 3 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Yes, but we have a switch on the
cache_levelin the gen8 pte encode function. Since the bit31 gets or'ed in for VLV, it'll hit also chv and wreak havoc in that specific switch - we'll hit the default case when we don't expect to.Quoted 6 lines
Quoted 3 lines
cache_level= functional behaviour relevant for the kernel's clflushing codeAs the '
IS_VALLEYVIEW' check will alias with CHV also, can I just update the condition here, to include 'GEN7' also (IS_GEN7) in the check.Quoted 8 lines
Adding new random bits to an enum which is used all over the place (and not just in the pte encoding functions) makes the code much harder to read. Also the code that deals with enum
cache_levelis all about cache coherency, which has rather tricky logic.Hence I expect this choice to cause further bugs down the road, which is why I don't really like it. My apologies for not spotting this earlier. -Daniel
Hi Daniel,
I understand your concern, but actually (as of now) this bit is only VLV specific. As per an earlier suggestion from Chris, I decided to overload the
cache_levelenum itself, in lieu of adding a new parameter to all the respective 'xxx_insert_entries' & 'xxx_pte_encode' functions.
Quoted 5 lines
And this is being done just before calling the '
xxx_insert_entries' function, which simply passes the flag as it is to 'xxx_pte_encode' function. So there may not be any risk here, if we use the appropriate VLV check.
Hi Daniel,
Please can you provide your inputs here.
Best regards
Akash
Quoted 6 lines
Best Regards
Akash
Quoted 65 lines · Akash Goel
Quoted 64 lines · Akash Goel
Quoted 47 lines · Daniel Vetter
Quoted 39 lines · Akash Goel
Quoted 8 lines · Daniel Vetter
Quoted 7 lines · Jesse Barnes
Quoted 4 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 3 lines · Jesse Barnes
On Tue, 11 Feb 2014 14:19:03 +0530
[email protected] wrote:Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Review static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines
(quoted diff) +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode(sg_page_iter_dma_address(&sg_iter), cache_level, true); + if (++act_pte == I915_PPGTT_PT_ENTRIES) { kunmap_atomic(pt_vaddr); pt_vaddr = NULL;Some extra whitespace here.
Thanks, will remove this.
Quoted 17 lines
Quoted 11 lines
Quoted 8 lines
Quoted 3 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Yes, but we have a switch on the
cache_levelin the gen8 pte encode function. Since the bit31 gets or'ed in for VLV, it'll hit also chv and wreak havoc in that specific switch - we'll hit the default case when we don't expect to.Quoted 6 lines
Quoted 3 lines
cache_level= functional behaviour relevant for the kernel's clflushing codeAs the '
IS_VALLEYVIEW' check will alias with CHV also, can I just update the condition here, to include 'GEN7' also (IS_GEN7) in the check.Quoted 8 lines
Adding new random bits to an enum which is used all over the place (and not just in the pte encoding functions) makes the code much harder to read. Also the code that deals with enum
cache_levelis all about cache coherency, which has rather tricky logic.Hence I expect this choice to cause further bugs down the road, which is why I don't really like it. My apologies for not spotting this earlier. -Daniel
Hi Daniel,
I understand your concern, but actually (as of now) this bit is only VLV specific. As per an earlier suggestion from Chris, I decided to overload the
cache_levelenum itself, in lieu of adding a new parameter to all the respective 'xxx_insert_entries' & 'xxx_pte_encode' functions.
Quoted 6 lines
Quoted 4 lines
And this is being done just before calling the '
xxx_insert_entries' function, which simply passes the flag as it is to 'xxx_pte_encode' function. So there may not be any risk here, if we use the appropriate VLV check.Please can you provide your inputs here.
I guess you've run into a case where Chris&I disagree. I still think wiring up a flags parameter to all the pte encode functions is the sane option to pick here.
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
Quoted 71 lines · Daniel Vetter
Quoted 70 lines · Akash Goel
Quoted 68 lines · Akash Goel
Quoted 47 lines · Daniel Vetter
Quoted 39 lines · Akash Goel
Quoted 8 lines · Daniel Vetter
Quoted 7 lines · Jesse Barnes
Quoted 4 lines · On Wed, 14 May 2014 00:30:34 +0200 Daniel Vetter <[email protected]> wrote:
Quoted 3 lines · Jesse Barnes
On Tue, 11 Feb 2014 14:19:03 +0530
[email protected] wrote:Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Review static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, · line 821
Quoted 8 lines
(quoted diff) +1
-0@@ -810,6 +815,7 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode(sg_page_iter_dma_address(&sg_iter), cache_level, true); + if (++act_pte == I915_PPGTT_PT_ENTRIES) { kunmap_atomic(pt_vaddr); pt_vaddr = NULL;Some extra whitespace here.
Thanks, will remove this.
Quoted 17 lines
Quoted 11 lines
Quoted 8 lines
Quoted 3 lines
Otherwise:
Reviewed-by Jesse Barnes <[email protected]>Hm, looking at the patch again encoding this into the
cache_levelenum is fraught with fun. And due toIS_VLValiasing chv this will blow up on chv very likely. My old idea was to eventually add apte_flagsparam all over for this stuff with additional bits.That works too; and yeah CHV is all different here isn't it. But won't it go through the gen8 paths anyway?
Yes, but we have a switch on the
cache_levelin the gen8 pte encode function. Since the bit31 gets or'ed in for VLV, it'll hit also chv and wreak havoc in that specific switch - we'll hit the default case when we don't expect to.Quoted 6 lines
Quoted 3 lines
cache_level= functional behaviour relevant for the kernel's clflushing codeAs the '
IS_VALLEYVIEW' check will alias with CHV also, can I just update the condition here, to include 'GEN7' also (IS_GEN7) in the check.Quoted 8 lines
Adding new random bits to an enum which is used all over the place (and not just in the pte encoding functions) makes the code much harder to read. Also the code that deals with enum
cache_levelis all about cache coherency, which has rather tricky logic.Hence I expect this choice to cause further bugs down the road, which is why I don't really like it. My apologies for not spotting this earlier. -Daniel
Hi Daniel,
I understand your concern, but actually (as of now) this bit is only VLV specific. As per an earlier suggestion from Chris, I decided to overload the
cache_levelenum itself, in lieu of adding a new parameter to all the respective 'xxx_insert_entries' & 'xxx_pte_encode' functions.And this is being done just before calling the '
xxx_insert_entries' function, which simply passes the flag as it is to 'xxx_pte_encode' function. So there may not be any risk here, if we use the appropriate VLV check.
Quoted 5 lines
Quoted 1 line
Please can you provide your inputs here.
I guess you've run into a case where Chris&I disagree. I still think wiring up a flags parameter to all the pte encode functions is the sane option to pick here.
Hi Daniel,
Yes I just followed Chris's suggestion, which I also felt was reasonable. But it's fine, will implement it as per your suggestion only. I hope that will be acceptable to all.
Best regards
Akash
Quoted 1 line
-Daniel
PATCH v4 1/1 drm/i915: Added write-enable pte bit support
In reply to Daniel VetterFrom: Akash Goel <[email protected]>
This adds support for a write-enable bit in the entry of GTT. This is handled via a read-only flag in the GEM buffer object which is then used to see how to set the bit when writing the GTT entries. Currently by default the Batch buffer & Ring buffers are marked as read only.
v2: Moved the pte override code for read-only bit to 'byt_pte_encode'. (Chris)
Fixed the issue of leaving 'gt_old_ro' as unused. (Chris)
v3: Removed the 'gt_old_ro' field, now setting RO bit only for Ring Buffers(Daniel).
v4: Added a new 'flags' parameter to all the pte(gen6) encode & insert_entries functions,
in lieu of overloading the cache_level enum (Daniel).
--- drivers/gpu/drm/i915/i915_drv.h | 6 +++++ drivers/gpu/drm/i915/i915_gem_gtt.c | 48 ++++++++++++++++++++------------- drivers/gpu/drm/i915/i915_gem_gtt.h | 5 ++-- drivers/gpu/drm/i915/intel_ringbuffer.c | 3 +++ 4 files changed, 41 insertions(+), 21 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +6 -0
@@ -1658,6 +1658,12 @@ struct drm_i915_gem_object { unsigned int pin_display:1;/* +* Is the object to be mapped as read-only to the GPU +* Only honoured if hardware has relevant pte bit +*/ + unsigned long gt_ro:1; + +/* * Is the GPU currently using a fence to access this buffer, */ unsigned int pending_fenced_gpu_access:1;
drivers/gpu/drm/i915/i915_gem_gtt.c +29 -19
169 lines not on this page · raw message
drivers/gpu/drm/i915/i915_gem_gtt.h +3 -2
@@ -197,15 +197,16 @@ struct i915_address_space {/* FIXME: Need a more generic return type */ gen6_gtt_pte_t (*pte_encode)(dma_addr_t addr,enum i915_cache_level level, -bool valid);/* Create a valid PTE */ +bool valid,u32 flags);/* Create a valid PTE */ void (*clear_range)(struct i915_address_space *vm,uint64_t start,uint64_t length,bool use_scratch); +#define PTE_READ_ONLY (1<<2)void (*insert_entries)(struct i915_address_space *vm,struct sg_table *st,uint64_t start, -enum i915_cache_level cache_level); +enum i915_cache_level cache_level,u32 flags);void (*cleanup)(struct i915_address_space *vm); };
drivers/gpu/drm/i915/intel_ringbuffer.c +3 -0
@@ -1397,6 +1397,9 @@ static int allocate_ring_buffer(struct intel_engine_cs *ring)if (obj ==NULL )return -ENOMEM ; +/* mark ring buffers as read-only from GPU side by default */ + obj->gt_ro =1 ; + ret =i915_gem_obj_ggtt_pin (obj,PAGE_SIZE ,PIN_MAPPABLE );if (ret) goto err_unref;
Signature · 1 line
1.9.2
Hi Akash,
Quoted 146 lines · On Mon, 2014-06-09 at 08:36 +0530, [email protected] wrote:
From: Akash Goel <[email protected]>
This adds support for a write-enable bit in the entry of GTT. This is handled via a read-only flag in the GEM buffer object which is then used to see how to set the bit when writing the GTT entries. Currently by default the Batch buffer & Ring buffers are marked as read only.
v2: Moved the pte override code for read-only bit to '
byt_pte_encode'. (Chris)
Fixed the issue of leaving 'gt_old_ro' as unused. (Chris)v3: Removed the '
gt_old_ro' field, now setting RO bit only for Ring Buffers(Daniel).v4: Added a new 'flags' parameter to all the pte(gen6) encode &
insert_entriesfunctions,
in lieu of overloading thecache_levelenum (Daniel).Signed-off-by Akash Goel <[email protected]>--- drivers/gpu/drm/i915/i915_drv.h | 6 +++++ drivers/gpu/drm/i915/i915_gem_gtt.c | 48 ++++++++++++++++++++------------- drivers/gpu/drm/i915/i915_gem_gtt.h | 5 ++-- drivers/gpu/drm/i915/intel_ringbuffer.c | 3 +++ 4 files changed, 41 insertions(+), 21 deletions(-)drivers/gpu/drm/i915/i915_drv.h +6
-0@@ -1658,6 +1658,12 @@ struct drm_i915_gem_object { unsigned int pin_display:1;/* +* Is the object to be mapped as read-only to the GPU +* Only honoured if hardware has relevant pte bit +*/ + unsigned long gt_ro:1; + +/* * Is the GPU currently using a fence to access this buffer, */ unsigned int pending_fenced_gpu_access:1;drivers/gpu/drm/i915/i915_gem_gtt.c +13
-11@@ -110,7 +110,7 @@ static inline gen8_ppgtt_pde_t gen8_pde_encode(struct drm_device *dev,static gen6_gtt_pte_t snb_pte_encode (dma_addr_t addr,enum i915_cache_level level, -bool valid) +bool valid,u32 unused) {gen6_gtt_pte_t pte = valid ?GEN6_PTE_VALID :0 ; pte |=GEN6_PTE_ADDR_ENCODE (addr);@@ -132,7 +132,7 @@ static gen6_gtt_pte_t snb_pte_encode(dma_addr_t addr,static gen6_gtt_pte_t ivb_pte_encode (dma_addr_t addr,enum i915_cache_level level, -bool valid) +bool valid,u32 unused) {gen6_gtt_pte_t pte = valid ?GEN6_PTE_VALID :0 ; pte |=GEN6_PTE_ADDR_ENCODE (addr);@@ -156,7 +156,7 @@ static gen6_gtt_pte_t ivb_pte_encode(dma_addr_t addr,static gen6_gtt_pte_t byt_pte_encode (dma_addr_t addr,enum i915_cache_level level, -bool valid) +bool valid,u32 flags) {gen6_gtt_pte_t pte = valid ?GEN6_PTE_VALID :0 ; pte |=GEN6_PTE_ADDR_ENCODE (addr);@@ -164,7 +164,8 @@ static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr,/* Mark the page as writeable. Other platforms don't have a * setting for read-only/writable, so this matches that behavior. */ - pte |=BYT_PTE_WRITEABLE ; +if (!(flags &PTE_READ_ONLY )) + pte |=BYT_PTE_WRITEABLE ;if (level !=I915_CACHE_NONE ) pte |=BYT_PTE_SNOOPED_BY_CPU_CACHES ;@@ -174,7 +175,7 @@ static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr,static gen6_gtt_pte_t hsw_pte_encode (dma_addr_t addr,enum i915_cache_level level, -bool valid) +bool valid,u32 unused) {gen6_gtt_pte_t pte = valid ?GEN6_PTE_VALID :0 ; pte |=HSW_PTE_ADDR_ENCODE (addr);@@ -187,7 +188,7 @@ static gen6_gtt_pte_t hsw_pte_encode(dma_addr_t addr,static gen6_gtt_pte_t iris_pte_encode (dma_addr_t addr,enum i915_cache_level level, -bool valid) +bool valid,u32 unused) {gen6_gtt_pte_t pte = valid ?GEN6_PTE_VALID :0 ; pte |=HSW_PTE_ADDR_ENCODE (addr);@@ -301,7 +302,7 @@ static void gen8_ppgtt_clear_range(struct i915_address_space *vm,static void gen8_ppgtt_insert_entries (struct i915_address_space *vm,struct sg_table *pages,uint64_t start, -enum i915_cache_level cache_level) +enum i915_cache_level cache_level,u32 unused) {struct i915_hw_ppgtt *ppgtt = container_of(vm, struct i915_hw_ppgtt, base);@@ -639,7 +640,7 @@ static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)uint32_t pd_entry;int pte, pde; - scratch_pte = vm->pte_encode (vm->scratch.addr,I915_CACHE_LLC , true); + scratch_pte = vm->pte_encode (vm->scratch.addr,I915_CACHE_LLC , true,0 ); pd_addr = (gen6_gtt_pte_t __iomem *)dev_priv->gtt.gsm + ppgtt->pd_offset /sizeof (gen6_gtt_pte_t);@@ -941,7 +942,7 @@ static void gen6_ppgtt_clear_range(struct i915_address_space *vm,unsigned first_pte = first_entry %I915_PPGTT_PT_ENTRIES ;unsigned last_pte, i; - scratch_pte = vm->pte_encode (vm->scratch.addr,I915_CACHE_LLC , true); + scratch_pte = vm->pte_encode (vm->scratch.addr,I915_CACHE_LLC , true,0 );while (num_entries) { last_pte = first_pte + num_entries;@@ -964,7 +965,7 @@ static void gen6_ppgtt_clear_range(struct i915_address_space *vm,static void gen6_ppgtt_insert_entries (struct i915_address_space *vm,struct sg_table *pages,uint64_t start, -enum i915_cache_level cache_level) +enum i915_cache_level cache_level,u32 flags) {struct i915_hw_ppgtt *ppgtt = container_of(vm, struct i915_hw_ppgtt, base);@@ -981,7 +982,8 @@ static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, pt_vaddr[act_pte] = vm->pte_encode (sg_page_iter_dma_address (&sg_iter), - cache_level, true); + cache_level, true, flags); +if (++act_pte ==I915_PPGTT_PT_ENTRIES ) {kunmap_atomic (pt_vaddr); pt_vaddr =NULL ;
drivers/gpu/drm/i915/i915_gem_gtt.c ppgtt_bind_vma(struct i915_vma *vma, · line 1223
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -1218,8 +1220,12 @@ ppgtt_bind_vma(struct i915_vma *vma,enum i915_cache_level cache_level, u32 flags) { + if (IS_VALLEYVIEW(vma->vm->dev))
The above check is redundant, since with the vlv/byt specific pte_encode() we'll get already the right behavior.
Quoted 55 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +9
-6+if (vma->obj->gt_ro) + flags |=PTE_READ_ONLY ; + vma->vm->insert_entries (vma->vm, vma->obj->pages, vma->node.start, - cache_level); + cache_level, flags); }static void ppgtt_unbind_vma (struct i915_vma *vma)@@ -1394,7 +1400,7 @@ static inline void gen8_set_pte(void __iomem *addr, gen8_gtt_pte_t pte)static void gen8_ggtt_insert_entries (struct i915_address_space *vm,struct sg_table *st,uint64_t start, -enum i915_cache_level level) +enum i915_cache_level level,u32 unused) {struct drm_i915_private *dev_priv = vm->dev->dev_private;unsigned first_entry = start >>PAGE_SHIFT ;@@ -1440,7 +1446,7 @@ static void gen8_ggtt_insert_entries(struct i915_address_space *vm,static void gen6_ggtt_insert_entries (struct i915_address_space *vm,struct sg_table *st,uint64_t start, -enum i915_cache_level level) +enum i915_cache_level level,u32 flags) {struct drm_i915_private *dev_priv = vm->dev->dev_private;unsigned first_entry = start >>PAGE_SHIFT ;@@ -1452,7 +1458,7 @@ static void gen6_ggtt_insert_entries(struct i915_address_space *vm,for_each_sg_page (st->sgl, &sg_iter, st->nents,0 ) { addr =sg_page_iter_dma_address (&sg_iter); -iowrite32 (vm->pte_encode (addr, level, true), >t_entries[i]); +iowrite32 (vm->pte_encode (addr, level, true, flags), >t_entries[i]); i++; }@@ -1464,7 +1470,7 @@ static void gen6_ggtt_insert_entries(struct i915_address_space *vm,*/ if (i !=0 )WARN_ON (readl (>t_entries[i-1 ]) != - vm->pte_encode (addr, level, true)); + vm->pte_encode (addr, level, true, flags));/* This next bit makes the above posting read even more important. We * want to flush the TLBs only after we're certain all the PTE updates @@ -1518,7 +1524,7 @@ static void gen6_ggtt_clear_range(struct i915_address_space *vm, first_entry, num_entries, max_entries)) num_entries = max_entries; - scratch_pte = vm->pte_encode (vm->scratch.addr,I915_CACHE_LLC , use_scratch); + scratch_pte = vm->pte_encode (vm->scratch.addr,I915_CACHE_LLC , use_scratch,0 );for (i =0 ; i < num_entries; i++)iowrite32 (scratch_pte, >t_base[i]);
drivers/gpu/drm/i915/i915_gem_gtt.c static void ggtt_bind_vma(struct i915_vma *vma, · line 1576
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +1
-0@@ -1567,6 +1573,10 @@ static void ggtt_bind_vma(struct i915_vma *vma,struct drm_i915_private *dev_priv = dev->dev_private;struct drm_i915_gem_object *obj = vma->obj; +if (IS_VALLEYVIEW (vma->vm->dev))
The check is redundant as above.
Quoted 29 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +5
-2+if (obj->gt_ro) + flags |=PTE_READ_ONLY ; +/* If there is no aliasing PPGTT, or the caller needs a global mapping, * or we have a global mapping already but the cacheability flags have * changed, set the global PTEs. @@ -1583,7 +1593,7 @@ static void ggtt_bind_vma(struct i915_vma *vma, (cache_level != obj->cache_level)) { vma->vm->insert_entries (vma->vm, obj->pages, vma->node.start, - cache_level); + cache_level, flags); obj->has_global_gtt_mapping =1 ; } }@@ -1595,7 +1605,7 @@ static void ggtt_bind_vma(struct i915_vma *vma, appgtt->base.insert_entries (&appgtt->base, vma->obj->pages, vma->node.start, - cache_level); + cache_level, flags); vma->obj->has_aliasing_ppgtt_mapping =1 ; } }drivers/gpu/drm/i915/i915_gem_gtt.h +0
-0
drivers/gpu/drm/i915/i915_gem_gtt.h struct i915_address_space { · line 205
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_gtt.h +2
-1@@ -197,15 +197,16 @@ struct i915_address_space {/* FIXME: Need a more generic return type */ gen6_gtt_pte_t (*pte_encode)(dma_addr_t addr,enum i915_cache_level level, -bool valid);/* Create a valid PTE */ +bool valid,u32 flags);/* Create a valid PTE */ void (*clear_range)(struct i915_address_space *vm,uint64_t start,uint64_t length,bool use_scratch); +#define PTE_READ_ONLY (1<<2)
As this uses the same namespace as GLOBAL_BIND, I would keep their definitions together. Also I couldn't find any other flag besides GLOBAL_BIND, in which case this could be (1 << 1).
With the above fixed the patch looks good to me:
Quoted 22 lines
drivers/gpu/drm/i915/i915_gem_gtt.h +1
-1void (*insert_entries)(struct i915_address_space *vm,struct sg_table *st,uint64_t start, -enum i915_cache_level cache_level); +enum i915_cache_level cache_level,u32 flags);void (*cleanup)(struct i915_address_space *vm); };drivers/gpu/drm/i915/intel_ringbuffer.c +3
-0@@ -1397,6 +1397,9 @@ static int allocate_ring_buffer(struct intel_engine_cs *ring)if (obj ==NULL )return -ENOMEM ; +/* mark ring buffers as read-only from GPU side by default */ + obj->gt_ro =1 ; + ret =i915_gem_obj_ggtt_pin (obj,PAGE_SIZE ,PIN_MAPPABLE );if (ret) goto err_unref;
PATCH v5 1/1 drm/i915: Added write-enable pte bit support
From: Akash Goel <[email protected]>
This adds support for a write-enable bit in the entry of GTT. This is handled via a read-only flag in the GEM buffer object which is then used to see how to set the bit when writing the GTT entries. Currently by default the Batch buffer & Ring buffers are marked as read only.
v2: Moved the pte override code for read-only bit to 'byt_pte_encode'. (Chris)
Fixed the issue of leaving 'gt_old_ro' as unused. (Chris)
v3: Removed the 'gt_old_ro' field, now setting RO bit only for Ring Buffers(Daniel).
v4: Added a new 'flags' parameter to all the pte(gen6) encode & insert_entries functions,
in lieu of overloading the cache_level enum (Daniel).
v5: Removed the superfluous VLV check & changed the definition location of PTE_READ_ONLY flag (Imre)
--- drivers/gpu/drm/i915/i915_drv.h | 6 +++++ drivers/gpu/drm/i915/i915_gem_gtt.c | 48 ++++++++++++++++++++------------- drivers/gpu/drm/i915/i915_gem_gtt.h | 5 ++-- drivers/gpu/drm/i915/intel_ringbuffer.c | 3 +++ 4 files changed, 41 insertions(+), 21 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +6 -0
@@ -1661,6 +1661,12 @@ struct drm_i915_gem_object { unsigned int pin_display:1;/* +* Is the object to be mapped as read-only to the GPU +* Only honoured if hardware has relevant pte bit +*/ + unsigned long gt_ro:1; + +/* * Is the GPU currently using a fence to access this buffer, */ unsigned int pending_fenced_gpu_access:1;
drivers/gpu/drm/i915/i915_gem_gtt.c +29 -19
169 lines not on this page · raw message
drivers/gpu/drm/i915/i915_gem_gtt.h +3 -2
@@ -154,6 +154,7 @@ struct i915_vma {void (*unbind_vma)(struct i915_vma *vma);/* Map an object into an address space with the given cache flags. */ #define GLOBAL_BIND (1<<0) +#define PTE_READ_ONLY (1<<1)void (*bind_vma)(struct i915_vma *vma,enum i915_cache_level cache_level,u32 flags);
@@ -197,7 +198,7 @@ struct i915_address_space {/* FIXME: Need a more generic return type */ gen6_gtt_pte_t (*pte_encode)(dma_addr_t addr,enum i915_cache_level level, -bool valid);/* Create a valid PTE */ +bool valid,u32 flags);/* Create a valid PTE */ void (*clear_range)(struct i915_address_space *vm, uint64_t start, uint64_t length,
@@ -205,7 +206,7 @@ struct i915_address_space {void (*insert_entries)(struct i915_address_space *vm,struct sg_table *st,uint64_t start, -enum i915_cache_level cache_level); +enum i915_cache_level cache_level,u32 flags);void (*cleanup)(struct i915_address_space *vm); };
drivers/gpu/drm/i915/intel_ringbuffer.c +3 -0
@@ -1397,6 +1397,9 @@ static int allocate_ring_buffer(struct intel_engine_cs *ring)if (obj ==NULL )return -ENOMEM ; +/* mark ring buffers as read-only from GPU side by default */ + obj->gt_ro =1 ; + ret =i915_gem_obj_ggtt_pin (obj,PAGE_SIZE ,PIN_MAPPABLE );if (ret) goto err_unref;
Signature · 1 line
1.9.2
Daniel Vetter
AppliedQuoted 12 lines · On Tue, Jun 17, 2014 at 10:59:42AM +0530, [email protected] wrote:
From: Akash Goel <[email protected]>
This adds support for a write-enable bit in the entry of GTT. This is handled via a read-only flag in the GEM buffer object which is then used to see how to set the bit when writing the GTT entries. Currently by default the Batch buffer & Ring buffers are marked as read only.
v2: Moved the pte override code for read-only bit to '
byt_pte_encode'. (Chris)
Fixed the issue of leaving 'gt_old_ro' as unused. (Chris)v3: Removed the '
gt_old_ro' field, now setting RO bit only for Ring Buffers(Daniel).
Quoted 7 lines
v4: Added a new 'flags' parameter to all the pte(gen6) encode &
insert_entriesfunctions,
in lieu of overloading thecache_levelenum (Daniel).v5: Removed the superfluous VLV check & changed the definition location of
PTE_READ_ONLYflag (Imre)Reviewed-by Imre Deak <[email protected]>Signed-off-by Akash Goel <[email protected]>
Queued for -next, thanks for the patch.
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch