xref: /linux/drivers/gpu/drm/xe/xe_ttm_stolen_mgr.c (revision 55a42f78ffd386e01a5404419f8c5ded7db70a21)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021-2023 Intel Corporation
4  * Copyright (C) 2021-2002 Red Hat
5  */
6 
7 #include <drm/drm_managed.h>
8 
9 #include <drm/ttm/ttm_device.h>
10 #include <drm/ttm/ttm_placement.h>
11 #include <drm/ttm/ttm_range_manager.h>
12 
13 #include <generated/xe_wa_oob.h>
14 
15 #include "regs/xe_bars.h"
16 #include "regs/xe_gt_regs.h"
17 #include "regs/xe_regs.h"
18 #include "xe_bo.h"
19 #include "xe_device.h"
20 #include "xe_gt.h"
21 #include "xe_gt_printk.h"
22 #include "xe_mmio.h"
23 #include "xe_res_cursor.h"
24 #include "xe_sriov.h"
25 #include "xe_ttm_stolen_mgr.h"
26 #include "xe_ttm_vram_mgr.h"
27 #include "xe_wa.h"
28 #include "xe_vram.h"
29 
30 struct xe_ttm_stolen_mgr {
31 	struct xe_ttm_vram_mgr base;
32 
33 	/* PCI base offset */
34 	resource_size_t io_base;
35 	/* GPU base offset */
36 	resource_size_t stolen_base;
37 
38 	void __iomem *mapping;
39 };
40 
41 static inline struct xe_ttm_stolen_mgr *
42 to_stolen_mgr(struct ttm_resource_manager *man)
43 {
44 	return container_of(man, struct xe_ttm_stolen_mgr, base.manager);
45 }
46 
47 /**
48  * xe_ttm_stolen_cpu_access_needs_ggtt() - If we can't directly CPU access
49  * stolen, can we then fallback to mapping through the GGTT.
50  * @xe: xe device
51  *
52  * Some older integrated platforms don't support reliable CPU access for stolen,
53  * however on such hardware we can always use the mappable part of the GGTT for
54  * CPU access. Check if that's the case for this device.
55  */
56 bool xe_ttm_stolen_cpu_access_needs_ggtt(struct xe_device *xe)
57 {
58 	return GRAPHICS_VERx100(xe) < 1270 && !IS_DGFX(xe);
59 }
60 
61 static u32 get_wopcm_size(struct xe_device *xe)
62 {
63 	u32 wopcm_size;
64 	u64 val;
65 
66 	val = xe_mmio_read64_2x32(xe_root_tile_mmio(xe), STOLEN_RESERVED);
67 	val = REG_FIELD_GET64(WOPCM_SIZE_MASK, val);
68 
69 	switch (val) {
70 	case 0x5 ... 0x6:
71 		val--;
72 		fallthrough;
73 	case 0x0 ... 0x3:
74 		wopcm_size = (1U << val) * SZ_1M;
75 		break;
76 	default:
77 		WARN(1, "Missing case wopcm_size=%llx\n", val);
78 		wopcm_size = 0;
79 	}
80 
81 	return wopcm_size;
82 }
83 
84 static s64 detect_bar2_dgfx(struct xe_device *xe, struct xe_ttm_stolen_mgr *mgr)
85 {
86 	struct xe_vram_region *tile_vram = xe_device_get_root_tile(xe)->mem.vram;
87 	resource_size_t tile_io_start = xe_vram_region_io_start(tile_vram);
88 	struct xe_mmio *mmio = xe_root_tile_mmio(xe);
89 	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
90 	u64 stolen_size, wopcm_size;
91 	u64 tile_offset;
92 	u64 tile_size;
93 
94 	tile_offset = tile_io_start - xe_vram_region_io_start(xe->mem.vram);
95 	tile_size = xe_vram_region_actual_physical_size(tile_vram);
96 
97 	/* Use DSM base address instead for stolen memory */
98 	mgr->stolen_base = (xe_mmio_read64_2x32(mmio, DSMBASE) & BDSM_MASK) - tile_offset;
99 	if (drm_WARN_ON(&xe->drm, tile_size < mgr->stolen_base))
100 		return 0;
101 
102 	/* Carve out the top of DSM as it contains the reserved WOPCM region */
103 	wopcm_size = get_wopcm_size(xe);
104 	if (drm_WARN_ON(&xe->drm, !wopcm_size))
105 		return 0;
106 
107 	stolen_size = tile_size - mgr->stolen_base;
108 	stolen_size -= wopcm_size;
109 
110 	/* Verify usage fits in the actual resource available */
111 	if (mgr->stolen_base + stolen_size <= pci_resource_len(pdev, LMEM_BAR))
112 		mgr->io_base = tile_io_start + mgr->stolen_base;
113 
114 	/*
115 	 * There may be few KB of platform dependent reserved memory at the end
116 	 * of vram which is not part of the DSM. Such reserved memory portion is
117 	 * always less then DSM granularity so align down the stolen_size to DSM
118 	 * granularity to accommodate such reserve vram portion.
119 	 */
120 	return ALIGN_DOWN(stolen_size, SZ_1M);
121 }
122 
123 static u32 detect_bar2_integrated(struct xe_device *xe, struct xe_ttm_stolen_mgr *mgr)
124 {
125 	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
126 	struct xe_gt *media_gt = xe_device_get_root_tile(xe)->media_gt;
127 	u32 stolen_size, wopcm_size;
128 	u32 ggc, gms;
129 
130 	ggc = xe_mmio_read32(xe_root_tile_mmio(xe), GGC);
131 
132 	/*
133 	 * Check GGMS: it should be fixed 0x3 (8MB), which corresponds to the
134 	 * GTT size
135 	 */
136 	if (drm_WARN_ON(&xe->drm, (ggc & GGMS_MASK) != GGMS_MASK))
137 		return 0;
138 
139 	/*
140 	 * Graphics >= 1270 uses the offset to the GSMBASE as address in the
141 	 * PTEs, together with the DM flag being set. Previously there was no
142 	 * such flag so the address was the io_base.
143 	 *
144 	 * DSMBASE = GSMBASE + 8MB
145 	 */
146 	mgr->stolen_base = SZ_8M;
147 	mgr->io_base = pci_resource_start(pdev, 2) + mgr->stolen_base;
148 
149 	/* return valid GMS value, -EIO if invalid */
150 	gms = REG_FIELD_GET(GMS_MASK, ggc);
151 	switch (gms) {
152 	case 0x0 ... 0x04:
153 		stolen_size = gms * 32 * SZ_1M;
154 		break;
155 	case 0xf0 ... 0xfe:
156 		stolen_size = (gms - 0xf0 + 1) * 4 * SZ_1M;
157 		break;
158 	default:
159 		return 0;
160 	}
161 
162 	/* Carve out the top of DSM as it contains the reserved WOPCM region */
163 	wopcm_size = get_wopcm_size(xe);
164 	if (drm_WARN_ON(&xe->drm, !wopcm_size))
165 		return 0;
166 
167 	stolen_size -= wopcm_size;
168 
169 	if (media_gt && XE_GT_WA(media_gt, 14019821291)) {
170 		u64 gscpsmi_base = xe_mmio_read64_2x32(&media_gt->mmio, GSCPSMI_BASE)
171 			& ~GENMASK_ULL(5, 0);
172 
173 		/*
174 		 * This workaround is primarily implemented by the BIOS.  We
175 		 * just need to figure out whether the BIOS has applied the
176 		 * workaround (meaning the programmed address falls within
177 		 * the DSM) and, if so, reserve that part of the DSM to
178 		 * prevent accidental reuse.  The DSM location should be just
179 		 * below the WOPCM.
180 		 */
181 		if (gscpsmi_base >= mgr->io_base &&
182 		    gscpsmi_base < mgr->io_base + stolen_size) {
183 			xe_gt_dbg(media_gt,
184 				  "Reserving %llu bytes of DSM for Wa_14019821291\n",
185 				  mgr->io_base + stolen_size - gscpsmi_base);
186 			stolen_size = gscpsmi_base - mgr->io_base;
187 		}
188 	}
189 
190 	if (drm_WARN_ON(&xe->drm, stolen_size + SZ_8M > pci_resource_len(pdev, 2)))
191 		return 0;
192 
193 	return stolen_size;
194 }
195 
196 extern struct resource intel_graphics_stolen_res;
197 
198 static u64 detect_stolen(struct xe_device *xe, struct xe_ttm_stolen_mgr *mgr)
199 {
200 #ifdef CONFIG_X86
201 	/* Map into GGTT */
202 	mgr->io_base = pci_resource_start(to_pci_dev(xe->drm.dev), 2);
203 
204 	/* Stolen memory is x86 only */
205 	mgr->stolen_base = intel_graphics_stolen_res.start;
206 	return resource_size(&intel_graphics_stolen_res);
207 #else
208 	return 0;
209 #endif
210 }
211 
212 int xe_ttm_stolen_mgr_init(struct xe_device *xe)
213 {
214 	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
215 	struct xe_ttm_stolen_mgr *mgr;
216 	u64 stolen_size, io_size;
217 	int err;
218 
219 	mgr = drmm_kzalloc(&xe->drm, sizeof(*mgr), GFP_KERNEL);
220 	if (!mgr)
221 		return -ENOMEM;
222 
223 	if (IS_SRIOV_VF(xe))
224 		stolen_size = 0;
225 	else if (IS_DGFX(xe))
226 		stolen_size = detect_bar2_dgfx(xe, mgr);
227 	else if (GRAPHICS_VERx100(xe) >= 1270)
228 		stolen_size = detect_bar2_integrated(xe, mgr);
229 	else
230 		stolen_size = detect_stolen(xe, mgr);
231 
232 	if (!stolen_size) {
233 		drm_dbg_kms(&xe->drm, "No stolen memory support\n");
234 		return 0;
235 	}
236 
237 	/*
238 	 * We don't try to attempt partial visible support for stolen vram,
239 	 * since stolen is always at the end of vram, and the BAR size is pretty
240 	 * much always 256M, with small-bar.
241 	 */
242 	io_size = 0;
243 	if (mgr->io_base && !xe_ttm_stolen_cpu_access_needs_ggtt(xe))
244 		io_size = stolen_size;
245 
246 	err = __xe_ttm_vram_mgr_init(xe, &mgr->base, XE_PL_STOLEN, stolen_size,
247 				     io_size, PAGE_SIZE);
248 	if (err) {
249 		drm_dbg_kms(&xe->drm, "Stolen mgr init failed: %i\n", err);
250 		return err;
251 	}
252 
253 	drm_dbg_kms(&xe->drm, "Initialized stolen memory support with %llu bytes\n",
254 		    stolen_size);
255 
256 	if (io_size)
257 		mgr->mapping = devm_ioremap_wc(&pdev->dev, mgr->io_base, io_size);
258 
259 	return 0;
260 }
261 
262 u64 xe_ttm_stolen_io_offset(struct xe_bo *bo, u32 offset)
263 {
264 	struct xe_device *xe = xe_bo_device(bo);
265 	struct ttm_resource_manager *ttm_mgr = ttm_manager_type(&xe->ttm, XE_PL_STOLEN);
266 	struct xe_ttm_stolen_mgr *mgr = to_stolen_mgr(ttm_mgr);
267 	struct xe_res_cursor cur;
268 
269 	XE_WARN_ON(!mgr->io_base);
270 
271 	if (xe_ttm_stolen_cpu_access_needs_ggtt(xe))
272 		return mgr->io_base + xe_bo_ggtt_addr(bo) + offset;
273 
274 	xe_res_first(bo->ttm.resource, offset, 4096, &cur);
275 	return mgr->io_base + cur.start;
276 }
277 
278 static int __xe_ttm_stolen_io_mem_reserve_bar2(struct xe_device *xe,
279 					       struct xe_ttm_stolen_mgr *mgr,
280 					       struct ttm_resource *mem)
281 {
282 	struct xe_res_cursor cur;
283 
284 	if (!mgr->io_base)
285 		return -EIO;
286 
287 	xe_res_first(mem, 0, 4096, &cur);
288 	mem->bus.offset = cur.start;
289 
290 	drm_WARN_ON(&xe->drm, !(mem->placement & TTM_PL_FLAG_CONTIGUOUS));
291 
292 	if (mem->placement & TTM_PL_FLAG_CONTIGUOUS && mgr->mapping)
293 		mem->bus.addr = (u8 __force *)mgr->mapping + mem->bus.offset;
294 
295 	mem->bus.offset += mgr->io_base;
296 	mem->bus.is_iomem = true;
297 	mem->bus.caching = ttm_write_combined;
298 
299 	return 0;
300 }
301 
302 static int __xe_ttm_stolen_io_mem_reserve_stolen(struct xe_device *xe,
303 						 struct xe_ttm_stolen_mgr *mgr,
304 						 struct ttm_resource *mem)
305 {
306 #ifdef CONFIG_X86
307 	struct xe_bo *bo = ttm_to_xe_bo(mem->bo);
308 
309 	XE_WARN_ON(IS_DGFX(xe));
310 
311 	/* XXX: Require BO to be mapped to GGTT? */
312 	if (drm_WARN_ON(&xe->drm, !(bo->flags & XE_BO_FLAG_GGTT)))
313 		return -EIO;
314 
315 	/* GGTT is always contiguously mapped */
316 	mem->bus.offset = xe_bo_ggtt_addr(bo) + mgr->io_base;
317 
318 	mem->bus.is_iomem = true;
319 	mem->bus.caching = ttm_write_combined;
320 
321 	return 0;
322 #else
323 	/* How is it even possible to get here without gen12 stolen? */
324 	drm_WARN_ON(&xe->drm, 1);
325 	return -EIO;
326 #endif
327 }
328 
329 int xe_ttm_stolen_io_mem_reserve(struct xe_device *xe, struct ttm_resource *mem)
330 {
331 	struct ttm_resource_manager *ttm_mgr = ttm_manager_type(&xe->ttm, XE_PL_STOLEN);
332 	struct xe_ttm_stolen_mgr *mgr = ttm_mgr ? to_stolen_mgr(ttm_mgr) : NULL;
333 
334 	if (!mgr || !mgr->io_base)
335 		return -EIO;
336 
337 	if (xe_ttm_stolen_cpu_access_needs_ggtt(xe))
338 		return __xe_ttm_stolen_io_mem_reserve_stolen(xe, mgr, mem);
339 	else
340 		return __xe_ttm_stolen_io_mem_reserve_bar2(xe, mgr, mem);
341 }
342 
343 u64 xe_ttm_stolen_gpu_offset(struct xe_device *xe)
344 {
345 	struct xe_ttm_stolen_mgr *mgr =
346 		to_stolen_mgr(ttm_manager_type(&xe->ttm, XE_PL_STOLEN));
347 
348 	return mgr->stolen_base;
349 }
350