summaryrefslogtreecommitdiffstats
path: root/drivers/gpu/nvgpu/gk20a/page_allocator_priv.h
diff options
context:
space:
mode:
authorAlex Waterman <alexw@nvidia.com>2016-09-22 13:13:31 -0400
committermobile promotions <svcmobile_promotions@nvidia.com>2016-10-18 15:24:33 -0400
commitfc4f0ddddb50da826b527b7791028a67539f34c8 (patch)
tree0aa570dda2b3714e4ddc103c48821eb2f9903d10 /drivers/gpu/nvgpu/gk20a/page_allocator_priv.h
parent641444188f18dbf56dda980e31f1b404dbb6f166 (diff)
gpu: nvgpu: SLAB allocation for page allocator
Add the ability to do "SLAB" allocation in the page allocator. This is generally useful since the allocator manages 64K pages but often we only need 4k chunks (for example when allocating memory for page table entries). Bug 1799159 JIRA DNVGPU-100 Signed-off-by: Alex Waterman <alexw@nvidia.com> Reviewed-on: http://git-master/r/1225322 (cherry picked from commit 299a5639243e44be504391d9155b4ae17d914aa2) Change-Id: Ib3a8558d40ba16bd3a413f4fd38b146beaa3c66b Reviewed-on: http://git-master/r/1227924 Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com> Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Diffstat (limited to 'drivers/gpu/nvgpu/gk20a/page_allocator_priv.h')
-rw-r--r--drivers/gpu/nvgpu/gk20a/page_allocator_priv.h69
1 files changed, 66 insertions, 3 deletions
diff --git a/drivers/gpu/nvgpu/gk20a/page_allocator_priv.h b/drivers/gpu/nvgpu/gk20a/page_allocator_priv.h
index 3d4e3c43..7d7f43c2 100644
--- a/drivers/gpu/nvgpu/gk20a/page_allocator_priv.h
+++ b/drivers/gpu/nvgpu/gk20a/page_allocator_priv.h
@@ -19,12 +19,63 @@
19 19
20#include <linux/list.h> 20#include <linux/list.h>
21#include <linux/rbtree.h> 21#include <linux/rbtree.h>
22#include <gk20a/gk20a_allocator.h>
23 22
24#include "gk20a_allocator.h" 23#include "gk20a_allocator.h"
25 24
26struct gk20a_allocator; 25struct gk20a_allocator;
27 26
27/*
28 * This allocator implements the ability to do SLAB style allocation since the
29 * GPU has two page sizes available - 4k and 64k/128k. When the default
30 * granularity is the large page size (64k/128k) small allocations become very
31 * space inefficient. This is most notable in PDE and PTE blocks which are 4k
32 * in size.
33 *
34 * Thus we need the ability to suballocate in 64k pages. The way we do this for
35 * the GPU is as follows. We have several buckets for sub-64K allocations:
36 *
37 * B0 - 4k
38 * B1 - 8k
39 * B3 - 16k
40 * B4 - 32k
41 * B5 - 64k (for when large pages are 128k)
42 *
43 * When an allocation comes in for less than the large page size (from now on
44 * assumed to be 64k) the allocation is satisfied by one of the buckets.
45 */
46struct page_alloc_slab {
47 struct list_head empty;
48 struct list_head partial;
49 struct list_head full;
50
51 int nr_empty;
52 int nr_partial;
53 int nr_full;
54
55 u32 slab_size;
56};
57
58enum slab_page_state {
59 SP_EMPTY,
60 SP_PARTIAL,
61 SP_FULL,
62 SP_NONE
63};
64
65struct page_alloc_slab_page {
66 unsigned long bitmap;
67 u64 page_addr;
68 u32 slab_size;
69
70 u32 nr_objects;
71 u32 nr_objects_alloced;
72
73 enum slab_page_state state;
74
75 struct page_alloc_slab *owner;
76 struct list_head list_entry;
77};
78
28struct page_alloc_chunk { 79struct page_alloc_chunk {
29 struct list_head list_entry; 80 struct list_head list_entry;
30 81
@@ -34,7 +85,7 @@ struct page_alloc_chunk {
34 85
35/* 86/*
36 * Struct to handle internal management of page allocation. It holds a list 87 * Struct to handle internal management of page allocation. It holds a list
37 * of the chunks of page that make up the overall allocation - much like a 88 * of the chunks of pages that make up the overall allocation - much like a
38 * scatter gather table. 89 * scatter gather table.
39 */ 90 */
40struct gk20a_page_alloc { 91struct gk20a_page_alloc {
@@ -44,13 +95,20 @@ struct gk20a_page_alloc {
44 u64 length; 95 u64 length;
45 96
46 /* 97 /*
47 * Only useful for the RB tree - since the alloc will have discontiguous 98 * Only useful for the RB tree - since the alloc may have discontiguous
48 * pages the base is essentially irrelevant except for the fact that it 99 * pages the base is essentially irrelevant except for the fact that it
49 * is guarenteed to be unique. 100 * is guarenteed to be unique.
50 */ 101 */
51 u64 base; 102 u64 base;
52 103
53 struct rb_node tree_entry; 104 struct rb_node tree_entry;
105
106 /*
107 * Set if this is a slab alloc. Points back to the slab page that owns
108 * this particular allocation. nr_chunks will always be 1 if this is
109 * set.
110 */
111 struct page_alloc_slab_page *slab_page;
54}; 112};
55 113
56struct gk20a_page_allocator { 114struct gk20a_page_allocator {
@@ -73,6 +131,9 @@ struct gk20a_page_allocator {
73 131
74 struct rb_root allocs; /* Outstanding allocations. */ 132 struct rb_root allocs; /* Outstanding allocations. */
75 133
134 struct page_alloc_slab *slabs;
135 int nr_slabs;
136
76 u64 flags; 137 u64 flags;
77 138
78 /* 139 /*
@@ -82,6 +143,8 @@ struct gk20a_page_allocator {
82 u64 nr_frees; 143 u64 nr_frees;
83 u64 nr_fixed_allocs; 144 u64 nr_fixed_allocs;
84 u64 nr_fixed_frees; 145 u64 nr_fixed_frees;
146 u64 nr_slab_allocs;
147 u64 nr_slab_frees;
85 u64 pages_alloced; 148 u64 pages_alloced;
86 u64 pages_freed; 149 u64 pages_freed;
87}; 150};