drm/i915: refactor i915_gem_object_pin_map()
Subsystems (2)
PATCH v4 1/4 drm/i915: refactor i915_gem_object_pin_map()
The recently-added i915_gem_object_pin_map() can be further optimised for "small" objects. To facilitate this, and simplify the error paths before adding the new code, this patch pulls out the "mapping" part of the operation (involving local allocations which must be undone before return) into its own subfunction.
The next patch will then insert the new optimisation into the middle of the now-separated subfunction.
This reorganisation will probably not affect the generated code, as the compiler will most likely inline it anyway, but it makes the logical structure a bit clearer and easier to modify.
v2:
Restructure loop-over-pages & error check [Chris Wilson]
v3:
Add page count to debug messages [Chris Wilson] Convert WARN_ON() to GEM_BUG_ON()
v4:
Drop the DEBUG messages [Tvrtko Ursulin]
--- drivers/gpu/drm/i915/i915_gem.c | 54 +++++++++++++++++++++++++---------------- 1 file changed, 33 insertions(+), 21 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +33 -21
@@ -2398,6 +2398,38 @@ static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)return 0 ; } +/* The 'mapping' part of i915_gem_object_pin_map() below */ +static void *i915_gem_object_map (const struct drm_i915_gem_object *obj) +{ +unsigned long n_pages = obj->base.size >>PAGE_SHIFT ; +struct sg_table *sgt = obj->pages; +struct sg_page_iter sg_iter; +struct page **pages; +unsigned long i =0 ; +void *addr; + +/* A single page can always be kmapped */ +if (n_pages ==1 ) +return kmap (sg_page (sgt->sgl)); + + pages =drm_malloc_gfp (n_pages,sizeof (*pages),GFP_TEMPORARY ); +if (pages ==NULL ) +return NULL ; + +for_each_sg_page (sgt->sgl, &sg_iter, sgt->nents,0 ) + pages[i++] =sg_page_iter_page (&sg_iter); + +/* Check that we have the expected number of pages */ +GEM_BUG_ON (i != n_pages); + + addr =vmap (pages, n_pages,0 ,PAGE_KERNEL ); + +drm_free_large (pages); + +return addr; +} + +/* get, pin, and map the pages of the object into kernel space */ void *i915_gem_object_pin_map (struct drm_i915_gem_object *obj) {int ret;
@@ -2411,27 +2443,7 @@ void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj)i915_gem_object_pin_pages (obj);if (obj->mapping ==NULL ) { -struct page **pages; - - pages =NULL ; -if (obj->base.size ==PAGE_SIZE ) - obj->mapping =kmap (sg_page (obj->pages->sgl)); -else - pages =drm_malloc_gfp (obj->base.size >>PAGE_SHIFT , -sizeof (*pages), -GFP_TEMPORARY ); -if (pages !=NULL ) { -struct sg_page_iter sg_iter; -int n; - - n =0 ; -for_each_sg_page (obj->pages->sgl, &sg_iter, - obj->pages->nents,0 ) - pages[n++] =sg_page_iter_page (&sg_iter); - - obj->mapping =vmap (pages, n,0 ,PAGE_KERNEL ); -drm_free_large (pages); - } + obj->mapping =i915_gem_object_map (obj);if (obj->mapping ==NULL ) {i915_gem_object_unpin_pages (obj);return ERR_PTR (-ENOMEM );
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 2/4 drm/i915: optimise i915_gem_object_map() for small objects
We're using this function for ringbuffers and other "small" objects, so it's worth avoiding an extra malloc()/free() cycle if the page array is small enough to put on the stack. Here we've chosen an arbitrary cutoff of 32 (4k) pages, which is big enough for a ringbuffer (4 pages) or a context image (currently up to 22 pages).
v5:
change name of local array [Chris Wilson]
--- drivers/gpu/drm/i915/i915_gem.c | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +10 -5
@@ -2404,7 +2404,8 @@ static void *i915_gem_object_map(const struct drm_i915_gem_object *obj)unsigned long n_pages = obj->base.size >>PAGE_SHIFT ;struct sg_table *sgt = obj->pages;struct sg_page_iter sg_iter; -struct page **pages; +struct page *stack_pages[32 ]; +struct page **pages = stack_pages;unsigned long i =0 ;void *addr;
@@ -2412,9 +2413,12 @@ static void *i915_gem_object_map(const struct drm_i915_gem_object *obj)if (n_pages ==1 )return kmap (sg_page (sgt->sgl)); - pages =drm_malloc_gfp (n_pages,sizeof (*pages),GFP_TEMPORARY ); -if (pages ==NULL ) -return NULL ; +if (n_pages >ARRAY_SIZE (stack_pages)) { +/* Too big for stack -- allocate temporary array instead */ + pages =drm_malloc_gfp (n_pages,sizeof (*pages),GFP_TEMPORARY ); +if (pages ==NULL ) +return NULL ; + }for_each_sg_page (sgt->sgl, &sg_iter, sgt->nents,0 ) pages[i++] =sg_page_iter_page (&sg_iter);
@@ -2424,7 +2428,8 @@ static void *i915_gem_object_map(const struct drm_i915_gem_object *obj) addr =vmap (pages, n_pages,0 ,PAGE_KERNEL ); -drm_free_large (pages); +if (pages != stack_pages) +drm_free_large (pages);return addr; }
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 3/4 Introduce & use new lightweight SGL iterators
In reply to Dave GordonThe existing for_each_sg_page() iterator is somewhat heavyweight, and is limiting i915 driver performance in a few benchmarks. So here we introduce somewhat lighter weight iterators, primarily for use with GEM objects or other case where we need only deal with whole aligned pages.
Unlike the old iterator, the new iterators use an internal state structure which is not intended to be accessed by the caller; instead each takes as a parameter an output variable which is set before each iteration. This makes them particularly simple to use :)
One of the new iterators provides the caller with the DMA address of each page in turn; the other provides the 'struct page' pointer required by many memory management operations.
Various uses of for_each_sg_page() are then converted to the new macros.
--- drivers/gpu/drm/i915/i915_drv.h | 62 +++++++++++++++++++++++++-- drivers/gpu/drm/i915/i915_gem.c | 20 ++++----- drivers/gpu/drm/i915/i915_gem_fence.c | 14 +++--- drivers/gpu/drm/i915/i915_gem_gtt.c | 76 ++++++++++++++++----------------- drivers/gpu/drm/i915/i915_gem_userptr.c | 7 ++- 5 files changed, 116 insertions(+), 63 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +59 -3
@@ -2120,6 +2120,10 @@ struct drm_i915_gem_object_ops {#define INTEL_FRONTBUFFER_ALL_MASK (pipe) \ (0xff << (INTEL_FRONTBUFFER_BITS_PER_PIPE * (pipe))) +void i915_gem_track_fb (struct drm_i915_gem_object *old, +struct drm_i915_gem_object *new, +unsigned frontbuffer_bits); +struct drm_i915_gem_object {struct drm_gem_object base;
@@ -2251,9 +2255,61 @@ struct drm_i915_gem_object { };#define to_intel_bo (x) container_of(x, struct drm_i915_gem_object, base) -void i915_gem_track_fb (struct drm_i915_gem_object *old, -struct drm_i915_gem_object *new, -unsigned frontbuffer_bits); +/* +* New optimised SGL iterator for GEM objects +*/ +struct sgt_iter { +struct scatterlist *sgp; +union { +unsigned long pfn; +unsigned long dma; + } ix; +unsigned int curr; +unsigned int max; +}; + +/* Constructor for new iterator */ +static inline struct sgt_iter +__sgt_iter (struct scatterlist *sgl,bool dma) +{ +struct sgt_iter s = { .sgp = sgl }; + +if (s.sgp) { + s.max = s.curr = s.sgp->offset; + s.max += s.sgp->length; +if (dma) + s.ix.dma =sg_dma_address (s.sgp); +else + s.ix.pfn =page_to_pfn (sg_page (s.sgp)); + } + +return s; +} + +/** +* for_each_sgt_dma - iterate over the DMA addresses of the given sg_table +* @__dmap: DMA address (output) +* @__iter: 'struct sgt_iter' (iterator state, internal) +* @__sgt: sg_table to iterate over (input) +*/ +#define for_each_sgt_dma (__dmap, __iter, __sgt) \ + for ((__iter) = __sgt_iter((__sgt)->sgl, true); \ + ((__dmap) = (__iter).ix.dma + (__iter).curr); \ + (((__iter).curr += PAGE_SIZE) < (__iter).max) || \ + ((__iter) = __sgt_iter(sg_next((__iter).sgp), true), 0)) + +/** +* for_each_sgt_page - iterate over the pages of the given sg_table +* @__pp: page pointer (output) +* @__iter: 'struct sgt_iter' (iterator state, internal) +* @__sgt: sg_table to iterate over (input) +*/ +#define for_each_sgt_page (__pp, __iter, __sgt) \ + for ((__iter) = __sgt_iter((__sgt)->sgl, false); \ + ((__pp) = (__iter).ix.pfn == 0 ? NULL : \ + pfn_to_page((__iter).ix.pfn + ((__iter).curr >> PAGE_SHIFT)));\ + (((__iter).curr += PAGE_SIZE) < (__iter).max) || \ + ((__iter) = __sgt_iter(sg_next((__iter).sgp), false), 0))/** * Request queue structure.
drivers/gpu/drm/i915/i915_gem.c +10 -10
@@ -2165,7 +2165,8 @@ static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)static void i915_gem_object_put_pages_gtt (struct drm_i915_gem_object *obj) { -struct sg_page_iter sg_iter; +struct sgt_iter sgt_iter; +struct page *page;int ret;BUG_ON (obj->madv == __I915_MADV_PURGED);
@@ -2187,9 +2188,7 @@ static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)if (obj->madv ==I915_MADV_DONTNEED ) obj->dirty =0 ; -for_each_sg_page (obj->pages->sgl, &sg_iter, obj->pages->nents,0 ) { -struct page *page =sg_page_iter_page (&sg_iter); - +for_each_sgt_page (page, sgt_iter, obj->pages) {if (obj->dirty)set_page_dirty (page);
@@ -2246,7 +2245,7 @@ static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)struct address_space *mapping;struct sg_table *st;struct scatterlist *sg; -struct sg_page_iter sg_iter; +struct sgt_iter sgt_iter;struct page *page;unsigned long last_pfn =0 ;/* suppress gcc warning */ int ret;
@@ -2343,8 +2342,8 @@ static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj) err_pages:sg_mark_end (sg); -for_each_sg_page (st->sgl, &sg_iter, st->nents,0 ) -put_page (sg_page_iter_page (&sg_iter)); +for_each_sgt_page (page, sgt_iter, st) +put_page (page);sg_free_table (st);kfree (st);
@@ -2403,7 +2402,8 @@ static void *i915_gem_object_map(const struct drm_i915_gem_object *obj) {unsigned long n_pages = obj->base.size >>PAGE_SHIFT ;struct sg_table *sgt = obj->pages; -struct sg_page_iter sg_iter; +struct sgt_iter sgt_iter; +struct page *page;struct page *stack_pages[32 ];struct page **pages = stack_pages;unsigned long i =0 ;
@@ -2420,8 +2420,8 @@ static void *i915_gem_object_map(const struct drm_i915_gem_object *obj)return NULL ; } -for_each_sg_page (sgt->sgl, &sg_iter, sgt->nents,0 ) - pages[i++] =sg_page_iter_page (&sg_iter); +for_each_sgt_page (page, sgt_iter, sgt) + pages[i++] = page;/* Check that we have the expected number of pages */ GEM_BUG_ON (i != n_pages);
drivers/gpu/drm/i915/i915_gem_fence.c +8 -6
@@ -745,15 +745,15 @@ void i915_gem_restore_fences(struct drm_device *dev)void i915_gem_object_do_bit_17_swizzle (struct drm_i915_gem_object *obj) { -struct sg_page_iter sg_iter; +struct sgt_iter sgt_iter; +struct page *page;int i;if (obj->bit_17 ==NULL )return ; i =0 ; -for_each_sg_page (obj->pages->sgl, &sg_iter, obj->pages->nents,0 ) { -struct page *page =sg_page_iter_page (&sg_iter); +for_each_sgt_page (page, sgt_iter, obj->pages) {char new_bit_17 =page_to_phys (page) >>17 ;if ((new_bit_17 &0x1 ) != (test_bit (i, obj->bit_17) !=0 )) {
@@ -775,7 +775,8 @@ void i915_gem_restore_fences(struct drm_device *dev)void i915_gem_object_save_bit_17_swizzle (struct drm_i915_gem_object *obj) { -struct sg_page_iter sg_iter; +struct sgt_iter sgt_iter; +struct page *page;int page_count = obj->base.size >>PAGE_SHIFT ;int i;
@@ -790,8 +791,9 @@ void i915_gem_restore_fences(struct drm_device *dev) } i =0 ; -for_each_sg_page (obj->pages->sgl, &sg_iter, obj->pages->nents,0 ) { -if (page_to_phys (sg_page_iter_page (&sg_iter)) & (1 <<17 )) + +for_each_sgt_page (page, sgt_iter, obj->pages) { +if (page_to_phys (page) & (1 <<17 ))__set_bit (i, obj->bit_17);else __clear_bit (i, obj->bit_17);
drivers/gpu/drm/i915/i915_gem_gtt.c +36 -40
149 lines not on this page · raw message
drivers/gpu/drm/i915/i915_gem_userptr.c +3 -4
@@ -706,7 +706,8 @@ struct get_pages_work {static void i915_gem_userptr_put_pages (struct drm_i915_gem_object *obj) { -struct sg_page_iter sg_iter; +struct sgt_iter sgt_iter; +struct page *page;BUG_ON (obj->userptr.work !=NULL );__i915_gem_userptr_set_active (obj, false);
@@ -716,9 +717,7 @@ struct get_pages_work {i915_gem_gtt_finish_object (obj); -for_each_sg_page (obj->pages->sgl, &sg_iter, obj->pages->nents,0 ) { -struct page *page =sg_page_iter_page (&sg_iter); - +for_each_sgt_page (page, sgt_iter, obj->pages) {if (obj->dirty)set_page_dirty (page);
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4 3/4] Introduce & use new lightweight SGL iterators
Quoted 31 lines · Dave Gordon
The existing
for_each_sg_page()iterator is somewhat heavyweight, and is limiting i915 driver performance in a few benchmarks. So here we introduce somewhat lighter weight iterators, primarily for use with GEM objects or other case where we need only deal with whole aligned pages.Unlike the old iterator, the new iterators use an internal state structure which is not intended to be accessed by the caller; instead each takes as a parameter an output variable which is set before each iteration. This makes them particularly simple to use :)
One of the new iterators provides the caller with the DMA address of each page in turn; the other provides the 'struct page' pointer required by many memory management operations.
Various uses of
for_each_sg_page()are then converted to the new macros.Signed-off-by Dave Gordon <[email protected]>Cc Chris Wilson <[email protected]>--- drivers/gpu/drm/i915/i915_drv.h | 62 +++++++++++++++++++++++++-- drivers/gpu/drm/i915/i915_gem.c | 20 ++++----- drivers/gpu/drm/i915/i915_gem_fence.c | 14 +++--- drivers/gpu/drm/i915/i915_gem_gtt.c | 76 ++++++++++++++++----------------- drivers/gpu/drm/i915/i915_gem_userptr.c | 7 ++- 5 files changed, 116 insertions(+), 63 deletions(-)drivers/gpu/drm/i915/i915_drv.h +0
-0
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object_ops { · line 2126
Quoted 7 lines
drivers/gpu/drm/i915/i915_drv.h +4
-0@@ -2120,6 +2120,10 @@ struct drm_i915_gem_object_ops {#define INTEL_FRONTBUFFER_ALL_MASK (pipe) \ (0xff << (INTEL_FRONTBUFFER_BITS_PER_PIPE * (pipe))) +void i915_gem_track_fb (struct drm_i915_gem_object *old, +struct drm_i915_gem_object *new, +unsigned frontbuffer_bits); +
That's not a much better spot, since this is now before the object it acts upon. One day we'll get our types separated from their actors.
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.h +0
-2struct drm_i915_gem_object {struct drm_gem_object base;@@ -2251,9 +2255,61 @@ struct drm_i915_gem_object { };#define to_intel_bo (x) container_of(x, struct drm_i915_gem_object, base) -void i915_gem_track_fb(struct drm_i915_gem_object *old, - struct drm_i915_gem_object *new,
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object { · line 2266
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.h +9
-1- unsigned frontbuffer_bits); +/* +* New optimised SGL iterator for GEM objects +*/ +struct sgt_iter { +struct scatterlist *sgp; +union { +unsigned long pfn; +unsigned long dma; + } ix;
An anonymous union here would be slightly more pleasant. And we should be using the correct types.
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object { · line 2277
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.h +11
-0+unsigned int curr; +unsigned int max; +}; + +/* Constructor for new iterator */ +static inline struct sgt_iter +__sgt_iter (struct scatterlist *sgl,bool dma) +{ +struct sgt_iter s = { .sgp = sgl }; + +if (s.sgp) {
Not acceptable just to GPF out if passed in a NULL? It's a BUG for all our use cases. And would save a conditional every chain step.
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.h +10
-0+ s.max = s.curr = s.sgp->offset; + s.max += s.sgp->length; +if (dma) + s.ix.dma =sg_dma_address (s.sgp); +else + s.ix.pfn =page_to_pfn (sg_page (s.sgp)); + } + +return s; +}
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object { · line 2297
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.h +10
-0+ +/** +* for_each_sgt_dma - iterate over the DMA addresses of the given sg_table +* @__dmap: DMA address (output) +* @__iter: 'struct sgt_iter' (iterator state, internal) +* @__sgt: sg_table to iterate over (input) +*/ +#define for_each_sgt_dma (__dmap, __iter, __sgt) \ + for ((__iter) = __sgt_iter((__sgt)->sgl, true); \ + ((__dmap) = (__iter).ix.dma + (__iter).curr); \
This prevents us using dma_addr_t 0, which is the error value iirc, so ok.
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object { · line 2310
Quoted 13 lines
drivers/gpu/drm/i915/i915_drv.h +13
-0+ (((__iter).curr += PAGE_SIZE) < (__iter).max) || \ + ((__iter) = __sgt_iter(sg_next((__iter).sgp), true), 0)) + +/** +* for_each_sgt_page - iterate over the pages of the given sg_table +* @__pp: page pointer (output) +* @__iter: 'struct sgt_iter' (iterator state, internal) +* @__sgt: sg_table to iterate over (input) +*/ +#define for_each_sgt_page (__pp, __iter, __sgt) \ + for ((__iter) = __sgt_iter((__sgt)->sgl, false); \ + ((__pp) = (__iter).ix.pfn == 0 ? NULL : \ + pfn_to_page((__iter).ix.pfn + ((__iter).curr >> PAGE_SHIFT)));\
If we shifted in the __sgt_iter() we could do simpler instructions inside the loop. The compiler won't know that pfn_to_page() is always true, so maybe a pfn_to_page(),1 here will help?
Quoted 2 lines
drivers/gpu/drm/i915/i915_drv.h +2
-0+ (((__iter).curr +=PAGE_SIZE ) < (__iter).max) || \ + ((__iter) = __sgt_iter(sg_next((__iter).sgp), false), 0))
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Chris Wilson
Reviewed-byRe: [PATCH v4 3/4] Introduce & use new lightweight SGL iterators
Quoted 66 lines · Chris Wilson
Quoted 31 lines · Dave Gordon
The existing
for_each_sg_page()iterator is somewhat heavyweight, and is limiting i915 driver performance in a few benchmarks. So here we introduce somewhat lighter weight iterators, primarily for use with GEM objects or other case where we need only deal with whole aligned pages.Unlike the old iterator, the new iterators use an internal state structure which is not intended to be accessed by the caller; instead each takes as a parameter an output variable which is set before each iteration. This makes them particularly simple to use :)
One of the new iterators provides the caller with the DMA address of each page in turn; the other provides the 'struct page' pointer required by many memory management operations.
Various uses of
for_each_sg_page()are then converted to the new macros.Signed-off-by Dave Gordon <[email protected]>Cc Chris Wilson <[email protected]>--- drivers/gpu/drm/i915/i915_drv.h | 62 +++++++++++++++++++++++++-- drivers/gpu/drm/i915/i915_gem.c | 20 ++++----- drivers/gpu/drm/i915/i915_gem_fence.c | 14 +++--- drivers/gpu/drm/i915/i915_gem_gtt.c | 76 ++++++++++++++++----------------- drivers/gpu/drm/i915/i915_gem_userptr.c | 7 ++- 5 files changed, 116 insertions(+), 63 deletions(-)drivers/gpu/drm/i915/i915_drv.h +0
-0drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object_ops { · line 2126
Quoted 7 lines
drivers/gpu/drm/i915/i915_drv.h +4
-0@@ -2120,6 +2120,10 @@ struct drm_i915_gem_object_ops {#define INTEL_FRONTBUFFER_ALL_MASK (pipe) \ (0xff << (INTEL_FRONTBUFFER_BITS_PER_PIPE * (pipe))) +void i915_gem_track_fb (struct drm_i915_gem_object *old, +struct drm_i915_gem_object *new, +unsigned frontbuffer_bits); +That's not a much better spot, since this is now before the object it acts upon. One day we'll get our types separated from their actors.
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.h +0
-2struct drm_i915_gem_object {struct drm_gem_object base;@@ -2251,9 +2255,61 @@ struct drm_i915_gem_object { };#define to_intel_bo (x) container_of(x, struct drm_i915_gem_object, base) -void i915_gem_track_fb(struct drm_i915_gem_object *old, - struct drm_i915_gem_object *new,drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object { · line 2266
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.h +9
-1- unsigned frontbuffer_bits); +/* +* New optimised SGL iterator for GEM objects +*/ +struct sgt_iter { +struct scatterlist *sgp; +union { +unsigned long pfn; +unsigned long dma; + } ix;An anonymous union here would be slightly more pleasant. And we should be using the correct types.
Quoted 14 lines
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_gem_object { · line 2277
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.h +11
-0+unsigned int curr; +unsigned int max; +}; + +/* Constructor for new iterator */ +static inline struct sgt_iter +__sgt_iter (struct scatterlist *sgl,bool dma) +{ +struct sgt_iter s = { .sgp = sgl }; + +if (s.sgp) {Not acceptable just to GPF out if passed in a NULL? It's a BUG for all our use cases. And would save a conditional every chain step.
Oh, it's how we stop after __sg_next().
Ok, played a bit and the only thing that sticks is
/*
- * New optimised SGL iterator for GEM objects
+ * Optimised SGL iterator for GEM objects
*/
struct sgt_iter {
struct scatterlist *sgp;
union {
unsigned long pfn;
- unsigned long dma;
- } ix;
+ dma_addr_t dma;
+ };
unsigned int curr;
unsigned int max;};
-/* Constructor for new iterator */
-static inline struct sgt_iter
+static __always_inline struct sgt_iter
Nothing is ever new past the commit that introduces it!
Final diff (a few % in micro, consistently above the noise):
drivers/gpu/drm/i915/i915_drv.h +12 -17
@@ -2256,31 +2256,26 @@ struct drm_i915_gem_object {#define to_intel_bo (x) container_of(x, struct drm_i915_gem_object, base)/* -* New optimised SGL iterator for GEM objects +* Optimised SGL iterator for GEM objects */ -struct sgt_iter { +static __always_inline struct sgt_iter {struct scatterlist *sgp;union {unsigned long pfn; -unsigned long dma; - } ix; +dma_addr_t dma; + };unsigned int curr;unsigned int max; -}; - -/* Constructor for new iterator */ -static inline struct sgt_iter -__sgt_iter (struct scatterlist *sgl,bool dma) -{ +} __sgt_iter(struct scatterlist *sgl, bool dma) {struct sgt_iter s = { .sgp = sgl };if (s.sgp) { s.max = s.curr = s.sgp->offset; s.max += s.sgp->length;if (dma) - s.ix.dma =sg_dma_address (s.sgp); + s.dma =sg_dma_address (s.sgp);else - s.ix.pfn =page_to_pfn (sg_page (s.sgp)); + s.pfn =page_to_pfn (sg_page (s.sgp)); }return s;
@@ -2313,9 +2308,9 @@ static inline struct scatterlist *__sg_next(struct scatterlist *sg)*/ #define for_each_sgt_dma (__dmap, __iter, __sgt) \ for ((__iter) = __sgt_iter((__sgt)->sgl, true); \ - ((__dmap) = (__iter).ix.dma + (__iter).curr); \ + ((__dmap) = (__iter).dma + (__iter).curr); \ (((__iter).curr += PAGE_SIZE) < (__iter).max) || \ - ((__iter) = __sgt_iter(__sg_next((__iter).sgp), true), 0)) + ((__iter) = __sgt_iter(__sg_next((__iter).sgp), true), 0))/** * for_each_sgt_page - iterate over the pages of the given sg_table
@@ -2325,10 +2320,10 @@ static inline struct scatterlist *__sg_next(struct scatterlist *sg)*/ #define for_each_sgt_page (__pp, __iter, __sgt) \ for ((__iter) = __sgt_iter((__sgt)->sgl, false); \ - ((__pp) = (__iter).ix.pfn == 0 ? NULL : \ - pfn_to_page((__iter).ix.pfn + ((__iter).curr >> PAGE_SHIFT)));\ + ((__pp) = (__iter).pfn == 0 ? NULL : \ + pfn_to_page((__iter).pfn + ((__iter).curr >> PAGE_SHIFT))); \ (((__iter).curr += PAGE_SIZE) < (__iter).max) || \ - ((__iter) = __sgt_iter(__sg_next((__iter).sgp), false), 0)) + ((__iter) = __sgt_iter(__sg_next((__iter).sgp), false), 0))
And with that,
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 4/4 Try inlining sg_next()
In reply to Dave Gordon--- drivers/gpu/drm/i915/i915_drv.h | 23 +++++++++++++++++++++-- 1 file changed, 21 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +21 -2
@@ -2287,6 +2287,25 @@ __sgt_iter(struct scatterlist *sgl, bool dma) }/** +* __sg_next - return the next scatterlist entry in a list +* @sg: The current sg entry +* +* Description: +* If the entry is the last, return NULL; otherwise, step to the next +* element in the array (@sg@+1). If that's a chain pointer, follow it; +* otherwise just return the pointer to the current element. +**/ +static inline struct scatterlist *__sg_next (struct scatterlist *sg) +{ +#ifdef CONFIG_DEBUG_SG +BUG_ON (sg->sg_magic !=SG_MAGIC ); +#endif +return sg_is_last (sg) ?NULL : +likely (!sg_is_chain (++sg)) ? sg : +sg_chain_ptr (sg); +} + +/** * for_each_sgt_dma - iterate over the DMA addresses of the given sg_table * @__dmap: DMA address (output) * @__iter: 'struct sgt_iter' (iterator state, internal)
@@ -2296,7 +2315,7 @@ __sgt_iter(struct scatterlist *sgl, bool dma)for ((__iter) =__sgt_iter ((__sgt)->sgl, true); \ ((__dmap) = (__iter).ix.dma + (__iter).curr); \ (((__iter).curr +=PAGE_SIZE ) < (__iter).max) || \ - ((__iter) =__sgt_iter (sg_next ((__iter).sgp), true),0 )) + ((__iter) =__sgt_iter (__sg_next ((__iter).sgp), true),0 ))/** * for_each_sgt_page - iterate over the pages of the given sg_table
@@ -2309,7 +2328,7 @@ __sgt_iter(struct scatterlist *sgl, bool dma) ((__pp) = (__iter).ix.pfn ==0 ?NULL : \pfn_to_page ((__iter).ix.pfn + ((__iter).curr >>PAGE_SHIFT )));\ (((__iter).curr +=PAGE_SIZE ) < (__iter).max) || \ - ((__iter) = __sgt_iter(sg_next((__iter).sgp), false), 0)) + ((__iter) = __sgt_iter(__sg_next((__iter).sgp), false), 0))/** * Request queue structure.
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4 4/4] Try inlining sg_next()
Quoted 2 lines · Dave Gordon
Signed-off-by Dave Gordon <[email protected]>Cc Chris Wilson <[email protected]>
Much better. The effect of the inline
gem:exec:fault:1MiB: +4.90%
gem:exec:fault:1MiB:forked: +7.99%
gem:exec:fault:16MiB: +22.94%
gem:exec:fault:16MiB:forked: +19.96%
gem:exec:fault:256MiB: +27.45%
gem:exec:fault:256MiB:forked: +36.89%And it brings this series into parity with mine.
----
Avoiding the out-of-line call to sg_next() reduces the kernel execution overhead by 10% in some workloads (for example the Unreal Engine 4 demo Atlantis on 2GiB GTTs) which are dominated by the cost of inserting PTEs due to texture thrashing. We can demonstrate this in a microbenchmark that forces us to rebind the object on every execbuf, where we can measure a 25% improvement, in the time required to execute an execbuf requiring a texture to be rebound, for inlining the sg_next() for large texture sizes.
Benchmark: igt/benchmarks/gem_exec_fault
Benchmark: igt/benchmarks/gem_exec_trace/Atlantis
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Patchwork
CI✗ Ro.CI.BAT: failure for series starting with [v4,1/4] drm/i915: refactor i915_gem_object_pin_map()
In reply to Dave Gordon== Series Details ==
== Series Details ==
Series: series starting with [v4,1/4] drm/i915: refactor i915_gem_object_pin_map()
URL : https://patchwork.freedesktop.org/series/7432/
State : failure
== Summary ==
Series 7432v1 Series without cover letter
http://patchwork.freedesktop.org/api/1.0/series/7432/revisions/1/mbox
Test drv_module_reload_basic:
dmesg-warn -> PASS (ro-bdw-i7-5600u)
Test kms_flip:
Subgroup basic-flip-vs-wf_vblank:
pass -> FAIL (ro-bdw-i7-5600u)
Test kms_pipe_crc_basic:
Subgroup suspend-read-crc-pipe-b:
dmesg-warn -> PASS (ro-ivb-i7-3770)
Test kms_sink_crc_basic:
pass -> SKIP (ro-skl-i7-6700hq)fi-bdw-i7-5557u total:217 pass:204 dwarn:0 dfail:0 fail:0 skip:13 fi-bsw-n3050 total:216 pass:172 dwarn:0 dfail:0 fail:2 skip:42 fi-hsw-i7-4770k total:217 pass:195 dwarn:0 dfail:0 fail:0 skip:22 fi-hsw-i7-4770r total:217 pass:191 dwarn:0 dfail:0 fail:0 skip:26 fi-skl-i7-6700k total:217 pass:189 dwarn:0 dfail:0 fail:0 skip:28 ro-bdw-i5-5250u total:217 pass:179 dwarn:0 dfail:0 fail:0 skip:38 ro-bdw-i7-5557U total:217 pass:204 dwarn:0 dfail:0 fail:0 skip:13 ro-bdw-i7-5600u total:217 pass:184 dwarn:0 dfail:0 fail:2 skip:31 ro-bsw-n3050 total:217 pass:173 dwarn:0 dfail:0 fail:2 skip:42 ro-byt-n2820 total:216 pass:172 dwarn:0 dfail:0 fail:3 skip:41 ro-hsw-i3-4010u total:216 pass:191 dwarn:0 dfail:0 fail:0 skip:25 ro-hsw-i7-4770r total:217 pass:191 dwarn:0 dfail:0 fail:0 skip:26 ro-ilk-i7-620lm total:217 pass:149 dwarn:0 dfail:0 fail:1 skip:67 ro-ilk1-i5-650 total:212 pass:150 dwarn:0 dfail:0 fail:1 skip:61 ro-ivb-i7-3770 total:217 pass:181 dwarn:0 dfail:0 fail:0 skip:36 ro-ivb2-i7-3770 total:217 pass:185 dwarn:0 dfail:0 fail:0 skip:32 ro-skl-i7-6700hq total:212 pass:187 dwarn:0 dfail:0 fail:0 skip:25 ro-snb-i7-2620M total:217 pass:175 dwarn:0 dfail:0 fail:1 skip:41
fi-byt-n2820 failed to connect after reboot
Results at /archive/results/CI_IGT_test/RO_Patchwork_945/
9d15199 drm-intel-nightly: 2016y-05m-20d-07h-54m-59s UTC integration manifest 7854a0c Try inlining sg_next()
cd4e1f3 Introduce & use new lightweight SGL iterators
bf7123c drm/i915: optimise i915_gem_object_map() for small objects afffbfb drm/i915: refactor i915_gem_object_pin_map()
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx