xref: /linux/drivers/gpu/drm/xe/xe_tile.c (revision 44343e8b250abb2f6bfd615493ca07a7f11f3cc2)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2023 Intel Corporation
4  */
5 
6 #include <linux/fault-inject.h>
7 
8 #include <drm/drm_managed.h>
9 
10 #include "xe_bo.h"
11 #include "xe_device.h"
12 #include "xe_ggtt.h"
13 #include "xe_gt.h"
14 #include "xe_memirq.h"
15 #include "xe_migrate.h"
16 #include "xe_pcode.h"
17 #include "xe_sa.h"
18 #include "xe_svm.h"
19 #include "xe_tile.h"
20 #include "xe_tile_sysfs.h"
21 #include "xe_ttm_vram_mgr.h"
22 #include "xe_wa.h"
23 #include "xe_vram.h"
24 #include "xe_vram_types.h"
25 
26 /**
27  * DOC: Multi-tile Design
28  *
29  * Different vendors use the term "tile" a bit differently, but in the Intel
30  * world, a 'tile' is pretty close to what most people would think of as being
31  * a complete GPU.  When multiple GPUs are placed behind a single PCI device,
32  * that's what is referred to as a "multi-tile device."  In such cases, pretty
33  * much all hardware is replicated per-tile, although certain responsibilities
34  * like PCI communication, reporting of interrupts to the OS, etc. are handled
35  * solely by the "root tile."  A multi-tile platform takes care of tying the
36  * tiles together in a way such that interrupt notifications from remote tiles
37  * are forwarded to the root tile, the per-tile vram is combined into a single
38  * address space, etc.
39  *
40  * In contrast, a "GT" (which officially stands for "Graphics Technology") is
41  * the subset of a GPU/tile that is responsible for implementing graphics
42  * and/or media operations.  The GT is where a lot of the driver implementation
43  * happens since it's where the hardware engines, the execution units, and the
44  * GuC all reside.
45  *
46  * Historically most Intel devices were single-tile devices that contained a
47  * single GT.  PVC is an example of an Intel platform built on a multi-tile
48  * design (i.e., multiple GPUs behind a single PCI device); each PVC tile only
49  * has a single GT.  In contrast, platforms like MTL that have separate chips
50  * for render and media IP are still only a single logical GPU, but the
51  * graphics and media IP blocks are each exposed as a separate GT within that
52  * single GPU.  This is important from a software perspective because multi-GT
53  * platforms like MTL only replicate a subset of the GPU hardware and behave
54  * differently than multi-tile platforms like PVC where nearly everything is
55  * replicated.
56  *
57  * Per-tile functionality (shared by all GTs within the tile):
58  *  - Complete 4MB MMIO space (containing SGunit/SoC registers, GT
59  *    registers, display registers, etc.)
60  *  - Global GTT
61  *  - VRAM (if discrete)
62  *  - Interrupt flows
63  *  - Migration context
64  *  - kernel batchbuffer pool
65  *  - Primary GT
66  *  - Media GT (if media version >= 13)
67  *
68  * Per-GT functionality:
69  *  - GuC
70  *  - Hardware engines
71  *  - Programmable hardware units (subslices, EUs)
72  *  - GSI subset of registers (multiple copies of these registers reside
73  *    within the complete MMIO space provided by the tile, but at different
74  *    offsets --- 0 for render, 0x380000 for media)
75  *  - Multicast register steering
76  *  - TLBs to cache page table translations
77  *  - Reset capability
78  *  - Low-level power management (e.g., C6)
79  *  - Clock frequency
80  *  - MOCS and PAT programming
81  */
82 
83 /**
84  * xe_tile_alloc - Perform per-tile memory allocation
85  * @tile: Tile to perform allocations for
86  *
87  * Allocates various per-tile data structures using DRM-managed allocations.
88  * Does not touch the hardware.
89  *
90  * Returns -ENOMEM if allocations fail, otherwise 0.
91  */
92 static int xe_tile_alloc(struct xe_tile *tile)
93 {
94 	tile->mem.ggtt = xe_ggtt_alloc(tile);
95 	if (!tile->mem.ggtt)
96 		return -ENOMEM;
97 
98 	tile->migrate = xe_migrate_alloc(tile);
99 	if (!tile->migrate)
100 		return -ENOMEM;
101 
102 	return 0;
103 }
104 
105 /**
106  * xe_tile_alloc_vram - Perform per-tile VRAM structs allocation
107  * @tile: Tile to perform allocations for
108  *
109  * Allocates VRAM per-tile data structures using DRM-managed allocations.
110  * Does not touch the hardware.
111  *
112  * Returns -ENOMEM if allocations fail, otherwise 0.
113  */
114 int xe_tile_alloc_vram(struct xe_tile *tile)
115 {
116 	struct xe_device *xe = tile_to_xe(tile);
117 	struct xe_vram_region *vram;
118 
119 	if (!IS_DGFX(xe))
120 		return 0;
121 
122 	vram = xe_vram_region_alloc(xe, tile->id, XE_PL_VRAM0 + tile->id);
123 	if (!vram)
124 		return -ENOMEM;
125 	tile->mem.vram = vram;
126 
127 	return 0;
128 }
129 
130 /**
131  * xe_tile_init_early - Initialize the tile and primary GT
132  * @tile: Tile to initialize
133  * @xe: Parent Xe device
134  * @id: Tile ID
135  *
136  * Initializes per-tile resources that don't require any interactions with the
137  * hardware or any knowledge about the Graphics/Media IP version.
138  *
139  * Returns: 0 on success, negative error code on error.
140  */
141 int xe_tile_init_early(struct xe_tile *tile, struct xe_device *xe, u8 id)
142 {
143 	int err;
144 
145 	tile->xe = xe;
146 	tile->id = id;
147 
148 	err = xe_tile_alloc(tile);
149 	if (err)
150 		return err;
151 
152 	tile->primary_gt = xe_gt_alloc(tile);
153 	if (IS_ERR(tile->primary_gt))
154 		return PTR_ERR(tile->primary_gt);
155 
156 	xe_pcode_init(tile);
157 
158 	return 0;
159 }
160 ALLOW_ERROR_INJECTION(xe_tile_init_early, ERRNO); /* See xe_pci_probe() */
161 
162 /**
163  * xe_tile_init_noalloc - Init tile up to the point where allocations can happen.
164  * @tile: The tile to initialize.
165  *
166  * This function prepares the tile to allow memory allocations to VRAM, but is
167  * not allowed to allocate memory itself. This state is useful for display
168  * readout, because the inherited display framebuffer will otherwise be
169  * overwritten as it is usually put at the start of VRAM.
170  *
171  * Note that since this is tile initialization, it should not perform any
172  * GT-specific operations, and thus does not need to hold GT forcewake.
173  *
174  * Returns: 0 on success, negative error code on error.
175  */
176 int xe_tile_init_noalloc(struct xe_tile *tile)
177 {
178 	struct xe_device *xe = tile_to_xe(tile);
179 
180 	xe_wa_apply_tile_workarounds(tile);
181 
182 	if (xe->info.has_usm && IS_DGFX(xe))
183 		xe_devm_add(tile, tile->mem.vram);
184 
185 	if (IS_DGFX(xe) && !ttm_resource_manager_used(&tile->mem.vram->ttm.manager)) {
186 		int err = xe_ttm_vram_mgr_init(xe, tile->mem.vram);
187 
188 		if (err)
189 			return err;
190 		xe->info.mem_region_mask |= BIT(tile->mem.vram->id) << 1;
191 	}
192 
193 	return xe_tile_sysfs_init(tile);
194 }
195 
196 int xe_tile_init(struct xe_tile *tile)
197 {
198 	int err;
199 
200 	err = xe_memirq_init(&tile->memirq);
201 	if (err)
202 		return err;
203 
204 	tile->mem.kernel_bb_pool = xe_sa_bo_manager_init(tile, SZ_1M, 16);
205 	if (IS_ERR(tile->mem.kernel_bb_pool))
206 		return PTR_ERR(tile->mem.kernel_bb_pool);
207 
208 	return 0;
209 }
210 void xe_tile_migrate_wait(struct xe_tile *tile)
211 {
212 	xe_migrate_wait(tile->migrate);
213 }
214