xref: /linux/drivers/gpu/drm/xe/xe_vram.c (revision 83675851547e835c15252c601f41acf269c351d9)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021-2024 Intel Corporation
4  */
5 
6 #include <kunit/visibility.h>
7 #include <linux/pci.h>
8 
9 #include <drm/drm_managed.h>
10 #include <drm/drm_print.h>
11 
12 #include "regs/xe_bars.h"
13 #include "regs/xe_gt_regs.h"
14 #include "regs/xe_regs.h"
15 #include "xe_assert.h"
16 #include "xe_device.h"
17 #include "xe_force_wake.h"
18 #include "xe_gt_mcr.h"
19 #include "xe_mmio.h"
20 #include "xe_sriov.h"
21 #include "xe_tile_sriov_vf.h"
22 #include "xe_ttm_vram_mgr.h"
23 #include "xe_vram.h"
24 #include "xe_vram_types.h"
25 
26 static bool resource_is_valid(struct pci_dev *pdev, int bar)
27 {
28 	if (!pci_resource_flags(pdev, bar))
29 		return false;
30 
31 	if (pci_resource_flags(pdev, bar) & IORESOURCE_UNSET)
32 		return false;
33 
34 	if (!pci_resource_len(pdev, bar))
35 		return false;
36 
37 	return true;
38 }
39 
40 static int determine_lmem_bar_size(struct xe_device *xe, struct xe_vram_region *lmem_bar)
41 {
42 	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
43 
44 	if (!resource_is_valid(pdev, LMEM_BAR)) {
45 		drm_err(&xe->drm, "pci resource is not valid\n");
46 		return -ENXIO;
47 	}
48 
49 	lmem_bar->io_start = pci_resource_start(pdev, LMEM_BAR);
50 	lmem_bar->io_size = pci_resource_len(pdev, LMEM_BAR);
51 	if (!lmem_bar->io_size)
52 		return -EIO;
53 
54 	/* XXX: Need to change when xe link code is ready */
55 	lmem_bar->dpa_base = 0;
56 
57 	/* set up a map to the total memory area. */
58 	lmem_bar->mapping = devm_ioremap_wc(&pdev->dev, lmem_bar->io_start, lmem_bar->io_size);
59 
60 	return 0;
61 }
62 
63 static int get_flat_ccs_offset(struct xe_gt *gt, u64 tile_size, u64 *poffset)
64 {
65 	struct xe_device *xe = gt_to_xe(gt);
66 	u64 offset;
67 	u32 reg;
68 
69 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
70 	if (!fw_ref.domains)
71 		return -ETIMEDOUT;
72 
73 	if (GRAPHICS_VER(xe) >= 20) {
74 		u64 ccs_size = tile_size / 512;
75 		u64 offset_hi, offset_lo;
76 		u32 nodes, num_enabled;
77 
78 		reg = xe_mmio_read32(&gt->mmio, MIRROR_FUSE3);
79 		nodes = REG_FIELD_GET(XE2_NODE_ENABLE_MASK, reg);
80 		num_enabled = hweight32(nodes); /* Number of enabled l3 nodes */
81 
82 		reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_LOWER);
83 		offset_lo = REG_FIELD_GET(XE2_FLAT_CCS_BASE_LOWER_ADDR_MASK, reg);
84 
85 		reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_UPPER);
86 		offset_hi = REG_FIELD_GET(XE2_FLAT_CCS_BASE_UPPER_ADDR_MASK, reg);
87 
88 		offset = offset_hi << 32; /* HW view bits 39:32 */
89 		offset |= offset_lo << 6; /* HW view bits 31:6 */
90 		offset *= num_enabled; /* convert to SW view */
91 		offset = round_up(offset, SZ_128K); /* SW must round up to nearest 128K */
92 
93 		/* We don't expect any holes */
94 		xe_assert_msg(xe, offset == (xe_mmio_read64_2x32(&gt_to_tile(gt)->mmio, GSMBASE) -
95 					     ccs_size),
96 			      "Hole between CCS and GSM.\n");
97 	} else {
98 		reg = xe_gt_mcr_unicast_read_any(gt, XEHP_FLAT_CCS_BASE_ADDR);
99 		offset = (u64)REG_FIELD_GET(XEHP_FLAT_CCS_PTR, reg) * SZ_64K;
100 	}
101 
102 	*poffset = offset;
103 
104 	return 0;
105 }
106 
107 /*
108  * tile_vram_size() - Collect vram size and offset information
109  * @tile: tile to get info for
110  * @vram_size: available vram (size - device reserved portions)
111  * @tile_size: actual vram size
112  * @tile_offset: physical start point in the vram address space
113  *
114  * There are 4 places for size information:
115  * - io size (from pci_resource_len of LMEM bar) (only used for small bar and DG1)
116  * - TILEx size (actual vram size)
117  * - GSMBASE offset (TILEx - "stolen")
118  * - CSSBASE offset (TILEx - CSS space necessary)
119  *
120  * CSSBASE is always a lower/smaller offset then GSMBASE.
121  *
122  * The actual available size of memory is to the CCS or GSM base.
123  * NOTE: multi-tile bases will include the tile offset.
124  *
125  */
126 static int tile_vram_size(struct xe_tile *tile, u64 *vram_size,
127 			  u64 *tile_size, u64 *tile_offset)
128 {
129 	struct xe_device *xe = tile_to_xe(tile);
130 	struct xe_gt *gt = tile->primary_gt;
131 	u64 offset;
132 	u32 reg;
133 
134 	if (IS_SRIOV_VF(xe)) {
135 		struct xe_tile *t;
136 		int id;
137 
138 		offset = 0;
139 		for_each_tile(t, xe, id)
140 			for_each_if(t->id < tile->id)
141 				offset += xe_tile_sriov_vf_lmem(t);
142 
143 		*tile_size = xe_tile_sriov_vf_lmem(tile);
144 		*vram_size = *tile_size;
145 		*tile_offset = offset;
146 
147 		return 0;
148 	}
149 
150 	/* actual size */
151 	if (unlikely(xe->info.platform == XE_DG1)) {
152 		*tile_size = pci_resource_len(to_pci_dev(xe->drm.dev), LMEM_BAR);
153 		*tile_offset = 0;
154 	} else {
155 		reg = xe_mmio_read32(&tile->mmio, SG_TILE_ADDR_RANGE(tile->id));
156 		*tile_size = (u64)REG_FIELD_GET(GENMASK(17, 8), reg) * SZ_1G;
157 		*tile_offset = (u64)REG_FIELD_GET(GENMASK(7, 1), reg) * SZ_1G;
158 	}
159 
160 	/* minus device usage */
161 	if (xe->info.has_flat_ccs) {
162 		int ret = get_flat_ccs_offset(gt, *tile_size, &offset);
163 
164 		if (ret)
165 			return ret;
166 	} else {
167 		offset = xe_mmio_read64_2x32(&tile->mmio, GSMBASE);
168 	}
169 
170 	/* remove the tile offset so we have just the available size */
171 	*vram_size = offset - *tile_offset;
172 
173 	return 0;
174 }
175 
176 static void vram_fini(void *arg)
177 {
178 	struct xe_device *xe = arg;
179 	struct xe_tile *tile;
180 	int id;
181 
182 	xe->mem.vram->mapping = NULL;
183 
184 	for_each_tile(tile, xe, id) {
185 		tile->mem.vram->mapping = NULL;
186 		if (tile->mem.kernel_vram)
187 			tile->mem.kernel_vram->mapping = NULL;
188 	}
189 }
190 
191 struct xe_vram_region *xe_vram_region_alloc(struct xe_device *xe, u8 id, u32 placement)
192 {
193 	struct xe_vram_region *vram;
194 	struct drm_device *drm = &xe->drm;
195 
196 	xe_assert(xe, id < xe->info.tile_count);
197 
198 	vram = drmm_kzalloc(drm, sizeof(*vram), GFP_KERNEL);
199 	if (!vram)
200 		return NULL;
201 
202 	vram->xe = xe;
203 	vram->id = id;
204 	vram->placement = placement;
205 #if defined(CONFIG_DRM_XE_PAGEMAP)
206 	vram->migrate = xe->tiles[id].migrate;
207 #endif
208 	return vram;
209 }
210 
211 static void print_vram_region_info(struct xe_device *xe, struct xe_vram_region *vram)
212 {
213 	struct drm_device *drm = &xe->drm;
214 
215 	if (vram->io_size < vram->usable_size)
216 		drm_info(drm, "Small BAR device\n");
217 
218 	drm_info(drm,
219 		 "VRAM[%u]: Actual physical size %pa, usable size exclude stolen %pa, CPU accessible size %pa\n",
220 		 vram->id, &vram->actual_physical_size, &vram->usable_size, &vram->io_size);
221 	drm_info(drm, "VRAM[%u]: DPA range: [%pa-%llx], io range: [%pa-%llx]\n",
222 		 vram->id, &vram->dpa_base, vram->dpa_base + (u64)vram->actual_physical_size,
223 		 &vram->io_start, vram->io_start + (u64)vram->io_size);
224 }
225 
226 static int vram_region_init(struct xe_device *xe, struct xe_vram_region *vram,
227 			    struct xe_vram_region *lmem_bar, u64 offset, u64 usable_size,
228 			    u64 region_size, resource_size_t remain_io_size)
229 {
230 	/* Check if VRAM region is already initialized */
231 	if (vram->mapping)
232 		return 0;
233 
234 	vram->actual_physical_size = region_size;
235 	vram->io_start = lmem_bar->io_start + offset;
236 	vram->io_size = min_t(u64, usable_size, remain_io_size);
237 
238 	if (!vram->io_size) {
239 		drm_err(&xe->drm, "Tile without any CPU visible VRAM. Aborting.\n");
240 		return -ENODEV;
241 	}
242 
243 	vram->dpa_base = lmem_bar->dpa_base + offset;
244 	vram->mapping = lmem_bar->mapping + offset;
245 	vram->usable_size = usable_size;
246 
247 	print_vram_region_info(xe, vram);
248 
249 	return 0;
250 }
251 
252 /**
253  * xe_vram_probe() - Probe VRAM configuration
254  * @xe: the &xe_device
255  *
256  * Collect VRAM size and offset information for all tiles.
257  *
258  * Return: 0 on success, error code on failure
259  */
260 int xe_vram_probe(struct xe_device *xe)
261 {
262 	struct xe_tile *tile;
263 	struct xe_vram_region lmem_bar;
264 	resource_size_t remain_io_size;
265 	u64 available_size = 0;
266 	u64 total_size = 0;
267 	int err;
268 	u8 id;
269 
270 	if (!IS_DGFX(xe))
271 		return 0;
272 
273 	err = determine_lmem_bar_size(xe, &lmem_bar);
274 	if (err)
275 		return err;
276 	drm_info(&xe->drm, "VISIBLE VRAM: %pa, %pa\n", &lmem_bar.io_start, &lmem_bar.io_size);
277 
278 	remain_io_size = lmem_bar.io_size;
279 
280 	for_each_tile(tile, xe, id) {
281 		u64 region_size;
282 		u64 usable_size;
283 		u64 tile_offset;
284 
285 		err = tile_vram_size(tile, &usable_size, &region_size, &tile_offset);
286 		if (err)
287 			return err;
288 
289 		total_size += region_size;
290 		available_size += usable_size;
291 
292 		err = vram_region_init(xe, tile->mem.vram, &lmem_bar, tile_offset, usable_size,
293 				       region_size, remain_io_size);
294 		if (err)
295 			return err;
296 
297 		if (total_size > lmem_bar.io_size) {
298 			drm_info(&xe->drm, "VRAM: %pa is larger than resource %pa\n",
299 				 &total_size, &lmem_bar.io_size);
300 		}
301 
302 		remain_io_size -= min_t(u64, tile->mem.vram->actual_physical_size, remain_io_size);
303 	}
304 
305 	err = vram_region_init(xe, xe->mem.vram, &lmem_bar, 0, available_size, total_size,
306 			       lmem_bar.io_size);
307 	if (err)
308 		return err;
309 
310 	return devm_add_action_or_reset(xe->drm.dev, vram_fini, xe);
311 }
312 
313 /**
314  * xe_vram_region_io_start - Get the IO start of a VRAM region
315  * @vram: the VRAM region
316  *
317  * Return: the IO start of the VRAM region, or 0 if not valid
318  */
319 resource_size_t xe_vram_region_io_start(const struct xe_vram_region *vram)
320 {
321 	return vram ? vram->io_start : 0;
322 }
323 
324 /**
325  * xe_vram_region_io_size - Get the IO size of a VRAM region
326  * @vram: the VRAM region
327  *
328  * Return: the IO size of the VRAM region, or 0 if not valid
329  */
330 resource_size_t xe_vram_region_io_size(const struct xe_vram_region *vram)
331 {
332 	return vram ? vram->io_size : 0;
333 }
334 
335 /**
336  * xe_vram_region_dpa_base - Get the DPA base of a VRAM region
337  * @vram: the VRAM region
338  *
339  * Return: the DPA base of the VRAM region, or 0 if not valid
340  */
341 resource_size_t xe_vram_region_dpa_base(const struct xe_vram_region *vram)
342 {
343 	return vram ? vram->dpa_base : 0;
344 }
345 
346 /**
347  * xe_vram_region_usable_size - Get the usable size of a VRAM region
348  * @vram: the VRAM region
349  *
350  * Return: the usable size of the VRAM region, or 0 if not valid
351  */
352 resource_size_t xe_vram_region_usable_size(const struct xe_vram_region *vram)
353 {
354 	return vram ? vram->usable_size : 0;
355 }
356 
357 /**
358  * xe_vram_region_actual_physical_size - Get the actual physical size of a VRAM region
359  * @vram: the VRAM region
360  *
361  * Return: the actual physical size of the VRAM region, or 0 if not valid
362  */
363 resource_size_t xe_vram_region_actual_physical_size(const struct xe_vram_region *vram)
364 {
365 	return vram ? vram->actual_physical_size : 0;
366 }
367 EXPORT_SYMBOL_IF_KUNIT(xe_vram_region_actual_physical_size);
368