[Intel-gfx] [PATCH 07/11] drm/i915/gtt: Use NULL to encode scratch shadow entries
Mika Kuoppala
mika.kuoppala at linux.intel.com
Wed Jul 10 16:21:15 UTC 2019
Chris Wilson <chris at chris-wilson.co.uk> writes:
> We can simplify our gtt walking code by comparing against NULL for
> scratch entries as opposed to looking up the distinct per-level scratch
> pointer.
>
> The only caveat is to remember to protect external parties and map the
> NULL to the scratch top pd.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
> drivers/gpu/drm/i915/i915_gem_gtt.c | 124 +++++++++-------------------
> drivers/gpu/drm/i915/i915_gem_gtt.h | 2 +-
> 2 files changed, 41 insertions(+), 85 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
> index b7882f06214a..a99b89502a90 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
> @@ -596,18 +596,17 @@ static void cleanup_page_dma(struct i915_address_space *vm,
>
> #define kmap_atomic_px(px) kmap_atomic(px_base(px)->page)
>
> -#define fill_px(px, v) fill_page_dma(px_base(px), (v))
> -#define fill32_px(px, v) fill_page_dma_32(px_base(px), (v))
> -
> -static void fill_page_dma(struct i915_page_dma *p, const u64 val)
> +static void
> +fill_page_dma(struct i915_page_dma *p, const u64 val, unsigned int count)
> {
> - kunmap_atomic(memset64(kmap_atomic(p->page), val, I915_PDES));
> + kunmap_atomic(memset64(kmap_atomic(p->page), val, count));
> }
>
> -static void fill_page_dma_32(struct i915_page_dma *p, const u32 v)
> -{
> - fill_page_dma(p, (u64)v << 32 | v);
> -}
> +#define fill_px(px, v) fill_page_dma(px_base(px), (v), I915_PDES)
> +#define fill32_px(px, v) do { \
> + u64 vv = lower_32_bits(v); \
> + fill_px(px, vv << 32 | vv); \
> +} while (0)
>
> static int
> setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
> @@ -711,7 +710,6 @@ static struct i915_page_table *alloc_pt(struct i915_address_space *vm)
> }
>
> atomic_set(&pt->used, 0);
> -
> return pt;
> }
>
> @@ -719,13 +717,11 @@ static struct i915_page_directory *__alloc_pd(void)
> {
> struct i915_page_directory *pd;
>
> - pd = kmalloc(sizeof(*pd), I915_GFP_ALLOW_FAIL);
> + pd = kzalloc(sizeof(*pd), I915_GFP_ALLOW_FAIL);
> if (unlikely(!pd))
> return NULL;
>
> - atomic_set(px_used(pd), 0);
> spin_lock_init(&pd->lock);
> -
> return pd;
> }
>
> @@ -753,63 +749,56 @@ static void free_pd(struct i915_address_space *vm, struct i915_page_dma *pd)
>
> #define free_px(vm, px) free_pd(vm, px_base(px))
>
> -static void init_pd(struct i915_page_directory *pd,
> - struct i915_page_scratch *scratch)
> -{
> - fill_px(pd, scratch->encode);
> - memset_p(pd->entry, scratch, 512);
> -}
> -
> static inline void
> write_dma_entry(struct i915_page_dma * const pdma,
> - const unsigned short pde,
> + const unsigned short idx,
> const u64 encoded_entry)
> {
> u64 * const vaddr = kmap_atomic(pdma->page);
>
> - vaddr[pde] = encoded_entry;
> + vaddr[idx] = encoded_entry;
> kunmap_atomic(vaddr);
> }
>
> static inline void
> __set_pd_entry(struct i915_page_directory * const pd,
> - const unsigned short pde,
> + const unsigned short idx,
My excuse was that as it is a pd, the pde fits.
Considering that now it is at any level and pde
in bspec is reserved for the last level, idx
is better.
> struct i915_page_dma * const to,
> u64 (*encode)(const dma_addr_t, const enum i915_cache_level))
> {
> GEM_BUG_ON(atomic_read(px_used(pd)) > 512);
>
> atomic_inc(px_used(pd));
> - pd->entry[pde] = to;
> - write_dma_entry(px_base(pd), pde, encode(to->daddr, I915_CACHE_LLC));
> + pd->entry[idx] = to;
> + write_dma_entry(px_base(pd), idx, encode(to->daddr, I915_CACHE_LLC));
> }
>
> -#define set_pd_entry(pd, pde, to) \
> - __set_pd_entry((pd), (pde), px_base(to), gen8_pde_encode)
> +#define set_pd_entry(pd, idx, to) \
> + __set_pd_entry((pd), (idx), px_base(to), gen8_pde_encode)
>
> static inline void
> clear_pd_entry(struct i915_page_directory * const pd,
> - const unsigned short pde,
> - struct i915_page_scratch * const scratch)
> + const unsigned short idx,
> + const struct i915_page_scratch * const scratch)
> {
> GEM_BUG_ON(atomic_read(px_used(pd)) == 0);
>
> - write_dma_entry(px_base(pd), pde, scratch->encode);
> - pd->entry[pde] = scratch;
> + write_dma_entry(px_base(pd), idx, scratch->encode);
> + pd->entry[idx] = NULL;
> atomic_dec(px_used(pd));
> }
>
> static bool
> release_pd_entry(struct i915_page_directory * const pd,
> - const unsigned short pde,
> + const unsigned short idx,
> struct i915_page_table * const pt,
> - struct i915_page_scratch * const scratch)
> + const struct i915_page_scratch * const scratch)
> {
> bool free = false;
>
> spin_lock(&pd->lock);
> if (atomic_dec_and_test(&pt->used)) {
> - clear_pd_entry(pd, pde, scratch);
> + clear_pd_entry(pd, idx, scratch);
> free = true;
> }
> spin_unlock(&pd->lock);
> @@ -910,7 +899,7 @@ static void gen8_free_page_tables(struct i915_address_space *vm,
> int i;
>
> for (i = 0; i < I915_PDES; i++) {
> - if (pd->entry[i] != &vm->scratch[1])
> + if (pd->entry[i])
> free_pd(vm, pd->entry[i]);
> }
> }
> @@ -922,7 +911,7 @@ static void gen8_ppgtt_cleanup_3lvl(struct i915_address_space *vm,
> int i;
>
> for (i = 0; i < pdpes; i++) {
> - if (pdp->entry[i] == &vm->scratch[2])
> + if (!pdp->entry[i])
> continue;
>
> gen8_free_page_tables(vm, pdp->entry[i]);
> @@ -940,7 +929,7 @@ static void gen8_ppgtt_cleanup_4lvl(struct i915_ppgtt *ppgtt)
> for (i = 0; i < GEN8_PML4ES_PER_PML4; i++) {
> struct i915_page_directory *pdp = i915_pdp_entry(pml4, i);
>
> - if (px_base(pdp) == px_base(&ppgtt->vm.scratch[3]))
> + if (!pdp)
> continue;
>
> gen8_ppgtt_cleanup_3lvl(&ppgtt->vm, pdp);
> @@ -994,8 +983,6 @@ static void gen8_ppgtt_clear_pd(struct i915_address_space *vm,
> u32 pde;
>
> gen8_for_each_pde(pt, pd, start, length, pde) {
> - GEM_BUG_ON(px_base(pt) == px_base(&vm->scratch[1]));
> -
Null oops is atleast equally descriptive.
> atomic_inc(&pt->used);
> gen8_ppgtt_clear_pt(vm, pt, start, length);
> if (release_pd_entry(pd, pde, pt, &vm->scratch[1]))
> @@ -1014,8 +1001,6 @@ static void gen8_ppgtt_clear_pdp(struct i915_address_space *vm,
> unsigned int pdpe;
>
> gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
> - GEM_BUG_ON(px_base(pd) == px_base(&vm->scratch[2]));
> -
> atomic_inc(px_used(pd));
> gen8_ppgtt_clear_pd(vm, pd, start, length);
> if (release_pd_entry(pdp, pdpe, &pd->pt, &vm->scratch[2]))
> @@ -1044,8 +1029,6 @@ static void gen8_ppgtt_clear_4lvl(struct i915_address_space *vm,
> GEM_BUG_ON(!i915_vm_is_4lvl(vm));
>
> gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
> - GEM_BUG_ON(px_base(pdp) == px_base(&vm->scratch[3]));
> -
> atomic_inc(px_used(pdp));
> gen8_ppgtt_clear_pdp(vm, pdp, start, length);
> if (release_pd_entry(pml4, pml4e, &pdp->pt, &vm->scratch[3]))
> @@ -1066,7 +1049,7 @@ static int gen8_ppgtt_alloc_pd(struct i915_address_space *vm,
> gen8_for_each_pde(pt, pd, start, length, pde) {
> const int count = gen8_pte_count(start, length);
>
> - if (px_base(pt) == px_base(&vm->scratch[1])) {
> + if (!pt) {
> spin_unlock(&pd->lock);
>
> pt = fetch_and_zero(&alloc);
> @@ -1081,7 +1064,7 @@ static int gen8_ppgtt_alloc_pd(struct i915_address_space *vm,
> fill_px(pt, vm->scratch[0].encode);
>
> spin_lock(&pd->lock);
> - if (pd->entry[pde] == &vm->scratch[1]) {
> + if (!pd->entry[pde]) {
> set_pd_entry(pd, pde, pt);
> } else {
> alloc = pt;
> @@ -1113,7 +1096,7 @@ static int gen8_ppgtt_alloc_pdp(struct i915_address_space *vm,
>
> spin_lock(&pdp->lock);
> gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
> - if (px_base(pd) == px_base(&vm->scratch[2])) {
> + if (!pd) {
> spin_unlock(&pdp->lock);
>
> pd = fetch_and_zero(&alloc);
> @@ -1124,10 +1107,10 @@ static int gen8_ppgtt_alloc_pdp(struct i915_address_space *vm,
> goto unwind;
> }
>
> - init_pd(pd, &vm->scratch[1]);
> + fill_px(pd, vm->scratch[1].encode);
>
> spin_lock(&pdp->lock);
> - if (pdp->entry[pdpe] == &vm->scratch[2]) {
> + if (!pdp->entry[pdpe]) {
> set_pd_entry(pdp, pdpe, pd);
> } else {
> alloc = pd;
> @@ -1177,7 +1160,7 @@ static int gen8_ppgtt_alloc_4lvl(struct i915_address_space *vm,
>
> spin_lock(&pml4->lock);
> gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
> - if (px_base(pdp) == px_base(&vm->scratch[3])) {
> + if (!pdp) {
> spin_unlock(&pml4->lock);
>
> pdp = fetch_and_zero(&alloc);
> @@ -1188,10 +1171,10 @@ static int gen8_ppgtt_alloc_4lvl(struct i915_address_space *vm,
> goto unwind;
> }
>
> - init_pd(pdp, &vm->scratch[2]);
> + fill_px(pdp, vm->scratch[2].encode);
>
> spin_lock(&pml4->lock);
> - if (pml4->entry[pml4e] == &vm->scratch[3]) {
> + if (!pml4->entry[pml4e]) {
> set_pd_entry(pml4, pml4e, pdp);
> } else {
> alloc = pdp;
> @@ -1527,7 +1510,7 @@ static int gen8_preallocate_top_level_pdp(struct i915_ppgtt *ppgtt)
> if (IS_ERR(pd))
> goto unwind;
>
> - init_pd(pd, &vm->scratch[1]);
> + fill_px(pd, vm->scratch[1].encode);
> set_pd_entry(pdp, pdpe, pd);
> }
>
> @@ -1558,46 +1541,19 @@ static void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt)
> ppgtt->vm.top = i915_vm_is_4lvl(&ppgtt->vm) ? 3 : 2;
> }
>
> -static void init_pd_n(struct i915_address_space *vm,
> - struct i915_page_directory *pd,
> - struct i915_page_scratch *scratch,
> - const unsigned int entries)
> -{
> - u64 * const vaddr = kmap_atomic_px(pd);
> -
> - memset64(vaddr, scratch->encode, entries);
> - kunmap_atomic(vaddr);
> -
> - memset_p(pd->entry, scratch, entries);
> -}
> -
> static struct i915_page_directory *
> gen8_alloc_top_pd(struct i915_address_space *vm)
> {
> + const unsigned int count = vm->total >> __gen8_pte_shift(vm->top);
Nice.
> struct i915_page_directory *pd;
>
> - if (i915_vm_is_4lvl(vm)) {
> - pd = alloc_pd(vm);
> - if (!IS_ERR(pd))
> - init_pd(pd, &vm->scratch[3]);
> + GEM_BUG_ON(count > ARRAY_SIZE(pd->entry));
I don't have to think. I like that.
>
> + pd = alloc_pd(vm);
> + if (IS_ERR(pd))
> return pd;
> - }
> -
> - /* 3lvl */
> - pd = __alloc_pd();
> - if (!pd)
> - return ERR_PTR(-ENOMEM);
> -
> - pd->entry[GEN8_3LVL_PDPES] = NULL;
Ok you dont like the sentry. Perhaps you could write
a few soothing words how noisily we crash if we
run long on this runway. If the tower sees and
sends firetrucks, it is fine.
> -
> - if (unlikely(setup_page_dma(vm, px_base(pd)))) {
> - kfree(pd);
> - return ERR_PTR(-ENOMEM);
> - }
> -
> - init_pd_n(vm, pd, &vm->scratch[2], GEN8_3LVL_PDPES);
>
> + fill_page_dma(px_base(pd), vm->scratch[vm->top].encode, count);
> return pd;
> }
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h
> index 669b204d4c13..2341944b9b17 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.h
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h
> @@ -610,7 +610,7 @@ i915_page_dir_dma_addr(const struct i915_ppgtt *ppgtt, const unsigned int n)
> {
> struct i915_page_dma *pt = ppgtt->pd->entry[n];
>
> - return px_dma(pt);
> + return px_dma(pt ?: px_base(&ppgtt->vm.scratch[ppgtt->vm.top]));
No other external users.
Earlier in series I yearned a little of is_entry_available|free|empty.
But the comparing against null beats any.
Reviewed-by: Mika Kuoppala <mika.kuoppala at linux.intel.com>
> }
>
> static inline struct i915_ggtt *
> --
> 2.20.1
More information about the Intel-gfx
mailing list