| 1 |
/* |
| 2 |
* Copyright © 2022 Imagination Technologies Ltd. |
| 3 |
* |
| 4 |
* SPDX-License-Identifier: MIT |
| 5 |
*/ |
| 6 |
|
| 7 |
#include "pvr_free_list.h" |
| 8 |
|
| 9 |
#include "vk_log.h" |
| 10 |
|
| 11 |
#include "hwdef/pvr_hw_utils.h" |
| 12 |
|
| 13 |
#include "pvr_device.h" |
| 14 |
#include "pvr_physical_device.h" |
| 15 |
|
| 16 |
#define ROGUE_BIF_PM_FREELIST_BASE_ADDR_ALIGNSIZE 16U |
| 17 |
|
| 18 |
/* FIXME: Is there a hardware define we can use instead? */ |
| 19 |
/* 1 DWord per PM physical page stored in the free list */ |
| 20 |
#define ROGUE_FREE_LIST_ENTRY_SIZE ((uint32_t)sizeof(uint32_t)) |
| 21 |
|
| 22 |
VkResult pvr_free_list_create(struct pvr_device *device, |
| 23 |
uint32_t initial_size, |
| 24 |
uint32_t max_size, |
| 25 |
uint32_t grow_size, |
| 26 |
uint32_t grow_threshold, |
| 27 |
struct pvr_free_list *parent_free_list, |
| 28 |
struct pvr_free_list **const free_list_out) |
| 29 |
{ |
| 30 |
const struct pvr_device_runtime_info *runtime_info = |
| 31 |
&device->pdevice->dev_runtime_info; |
| 32 |
struct pvr_winsys_free_list *parent_ws_free_list = |
| 33 |
parent_free_list ? parent_free_list->ws_free_list : NULL; |
| 34 |
const uint64_t bo_flags = PVR_BO_ALLOC_FLAG_GPU_UNCACHED | |
| 35 |
PVR_BO_ALLOC_FLAG_PM_FW_PROTECT; |
| 36 |
struct pvr_free_list *free_list; |
| 37 |
uint32_t cache_line_size; |
| 38 |
uint32_t initial_num_pages; |
| 39 |
uint32_t grow_num_pages; |
| 40 |
uint32_t max_num_pages; |
| 41 |
uint64_t addr_alignment; |
| 42 |
uint64_t size_alignment; |
| 43 |
uint64_t size; |
| 44 |
VkResult result; |
| 45 |
|
| 46 |
assert((initial_size + grow_size) <= max_size); |
| 47 |
assert(max_size != 0); |
| 48 |
assert(grow_threshold <= 100); |
| 49 |
|
| 50 |
/* Make sure the free list is created with at least a single page. */ |
| 51 |
if (initial_size == 0) |
| 52 |
initial_size = ROGUE_BIF_PM_PHYSICAL_PAGE_SIZE; |
| 53 |
|
| 54 |
/* The freelists sizes must respect the PM freelist base address alignment |
| 55 |
* requirement. As the freelist entries are cached by the SLC, it's also |
| 56 |
* necessary to ensure the sizes respect the SLC cache line size to avoid |
| 57 |
* invalid entries appearing in the cache, which would be problematic after |
| 58 |
* a grow operation, as the SLC entries aren't invalidated. We do this by |
| 59 |
* making sure the freelist values are appropriately aligned. |
| 60 |
* |
| 61 |
* To calculate the alignment, we first take the largest of the freelist |
| 62 |
* base address alignment and the SLC cache line size. We then divide this |
| 63 |
* by the freelist entry size to determine the number of freelist entries |
| 64 |
* required by the PM. Finally, as each entry holds a single PM physical |
| 65 |
* page, we multiple the number of entries by the page size. |
| 66 |
* |
| 67 |
* As an example, if the base address alignment is 16 bytes, the SLC cache |
| 68 |
* line size is 64 bytes and the freelist entry size is 4 bytes then 16 |
| 69 |
* entries are required, as we take the SLC cacheline size (being the larger |
| 70 |
* of the two values) and divide this by 4. If the PM page size is 4096 |
| 71 |
* bytes then we end up with an alignment of 65536 bytes. |
| 72 |
*/ |
| 73 |
cache_line_size = pvr_get_slc_cache_line_size(&device->pdevice->dev_info); |
| 74 |
|
| 75 |
addr_alignment = |
| 76 |
MAX2(ROGUE_BIF_PM_FREELIST_BASE_ADDR_ALIGNSIZE, cache_line_size); |
| 77 |
size_alignment = (addr_alignment / ROGUE_FREE_LIST_ENTRY_SIZE) * |
| 78 |
ROGUE_BIF_PM_PHYSICAL_PAGE_SIZE; |
| 79 |
|
| 80 |
assert(util_is_power_of_two_nonzero64(size_alignment)); |
| 81 |
|
| 82 |
initial_size = align64(initial_size, size_alignment); |
| 83 |
max_size = align64(max_size, size_alignment); |
| 84 |
grow_size = align64(grow_size, size_alignment); |
| 85 |
|
| 86 |
/* Make sure the 'max' size doesn't exceed what the firmware supports and |
| 87 |
* adjust the other sizes accordingly. |
| 88 |
*/ |
| 89 |
if (max_size > runtime_info->max_free_list_size) { |
| 90 |
max_size = runtime_info->max_free_list_size; |
| 91 |
assert(align64(max_size, size_alignment) == max_size); |
| 92 |
} |
| 93 |
|
| 94 |
if (initial_size > max_size) |
| 95 |
initial_size = max_size; |
| 96 |
|
| 97 |
if (initial_size == max_size) |
| 98 |
grow_size = 0; |
| 99 |
|
| 100 |
initial_num_pages = initial_size >> ROGUE_BIF_PM_PHYSICAL_PAGE_SHIFT; |
| 101 |
max_num_pages = max_size >> ROGUE_BIF_PM_PHYSICAL_PAGE_SHIFT; |
| 102 |
grow_num_pages = grow_size >> ROGUE_BIF_PM_PHYSICAL_PAGE_SHIFT; |
| 103 |
|
| 104 |
/* Calculate the size of the buffer needed to store the free list entries |
| 105 |
* based on the maximum number of pages we can have. |
| 106 |
*/ |
| 107 |
size = max_num_pages * ROGUE_FREE_LIST_ENTRY_SIZE; |
| 108 |
assert(align64(size, addr_alignment) == size); |
| 109 |
|
| 110 |
free_list = vk_alloc(&device->vk.alloc, |
| 111 |
sizeof(*free_list), |
| 112 |
8, |
| 113 |
VK_SYSTEM_ALLOCATION_SCOPE_DEVICE); |
| 114 |
if (!free_list) |
| 115 |
return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY); |
| 116 |
|
| 117 |
/* FIXME: The memory is mapped GPU uncached, but this seems to contradict |
| 118 |
* the comment above about aligning to the SLC cache line size. |
| 119 |
*/ |
| 120 |
result = pvr_bo_alloc(device, |
| 121 |
device->heaps.general_heap, |
| 122 |
size, |
| 123 |
addr_alignment, |
| 124 |
bo_flags, |
| 125 |
&free_list->bo); |
| 126 |
if (result != VK_SUCCESS) |
| 127 |
goto err_vk_free_free_list; |
| 128 |
|
| 129 |
result = device->ws->ops->free_list_create(device->ws, |
| 130 |
free_list->bo->vma, |
| 131 |
initial_num_pages, |
| 132 |
max_num_pages, |
| 133 |
grow_num_pages, |
| 134 |
grow_threshold, |
| 135 |
parent_ws_free_list, |
| 136 |
&free_list->ws_free_list); |
| 137 |
if (result != VK_SUCCESS) |
| 138 |
goto err_pvr_bo_free_bo; |
| 139 |
|
| 140 |
free_list->device = device; |
| 141 |
free_list->size = size; |
| 142 |
|
| 143 |
*free_list_out = free_list; |
| 144 |
|
| 145 |
return VK_SUCCESS; |
| 146 |
|
| 147 |
err_pvr_bo_free_bo: |
| 148 |
pvr_bo_free(device, free_list->bo); |
| 149 |
|
| 150 |
err_vk_free_free_list: |
| 151 |
vk_free(&device->vk.alloc, free_list); |
| 152 |
|
| 153 |
return result; |
| 154 |
} |
| 155 |
|
| 156 |
void pvr_free_list_destroy(struct pvr_free_list *free_list) |
| 157 |
{ |
| 158 |
struct pvr_device *device = free_list->device; |
| 159 |
|
| 160 |
device->ws->ops->free_list_destroy(free_list->ws_free_list); |
| 161 |
pvr_bo_free(device, free_list->bo); |
| 162 |
vk_free(&device->vk.alloc, free_list); |
| 163 |
} |