1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2021-2024 Intel Corporation
4 */
5
6 #include <linux/pci.h>
7
8 #include <drm/drm_managed.h>
9 #include <drm/drm_print.h>
10
11 #include "regs/xe_bars.h"
12 #include "regs/xe_gt_regs.h"
13 #include "regs/xe_regs.h"
14 #include "xe_assert.h"
15 #include "xe_device.h"
16 #include "xe_force_wake.h"
17 #include "xe_gt_mcr.h"
18 #include "xe_gt_sriov_vf.h"
19 #include "xe_mmio.h"
20 #include "xe_module.h"
21 #include "xe_sriov.h"
22 #include "xe_vram.h"
23
24 #define BAR_SIZE_SHIFT 20
25
26 static void
_resize_bar(struct xe_device * xe,int resno,resource_size_t size)27 _resize_bar(struct xe_device *xe, int resno, resource_size_t size)
28 {
29 struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
30 int bar_size = pci_rebar_bytes_to_size(size);
31 int ret;
32
33 if (pci_resource_len(pdev, resno))
34 pci_release_resource(pdev, resno);
35
36 ret = pci_resize_resource(pdev, resno, bar_size);
37 if (ret) {
38 drm_info(&xe->drm, "Failed to resize BAR%d to %dM (%pe). Consider enabling 'Resizable BAR' support in your BIOS\n",
39 resno, 1 << bar_size, ERR_PTR(ret));
40 return;
41 }
42
43 drm_info(&xe->drm, "BAR%d resized to %dM\n", resno, 1 << bar_size);
44 }
45
46 /*
47 * if force_vram_bar_size is set, attempt to set to the requested size
48 * else set to maximum possible size
49 */
resize_vram_bar(struct xe_device * xe)50 static void resize_vram_bar(struct xe_device *xe)
51 {
52 u64 force_vram_bar_size = xe_modparam.force_vram_bar_size;
53 struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
54 struct pci_bus *root = pdev->bus;
55 resource_size_t current_size;
56 resource_size_t rebar_size;
57 struct resource *root_res;
58 u32 bar_size_mask;
59 u32 pci_cmd;
60 int i;
61
62 /* gather some relevant info */
63 current_size = pci_resource_len(pdev, LMEM_BAR);
64 bar_size_mask = pci_rebar_get_possible_sizes(pdev, LMEM_BAR);
65
66 if (!bar_size_mask)
67 return;
68
69 /* set to a specific size? */
70 if (force_vram_bar_size) {
71 u32 bar_size_bit;
72
73 rebar_size = force_vram_bar_size * (resource_size_t)SZ_1M;
74
75 bar_size_bit = bar_size_mask & BIT(pci_rebar_bytes_to_size(rebar_size));
76
77 if (!bar_size_bit) {
78 drm_info(&xe->drm,
79 "Requested size: %lluMiB is not supported by rebar sizes: 0x%x. Leaving default: %lluMiB\n",
80 (u64)rebar_size >> 20, bar_size_mask, (u64)current_size >> 20);
81 return;
82 }
83
84 rebar_size = 1ULL << (__fls(bar_size_bit) + BAR_SIZE_SHIFT);
85
86 if (rebar_size == current_size)
87 return;
88 } else {
89 rebar_size = 1ULL << (__fls(bar_size_mask) + BAR_SIZE_SHIFT);
90
91 /* only resize if larger than current */
92 if (rebar_size <= current_size)
93 return;
94 }
95
96 drm_info(&xe->drm, "Attempting to resize bar from %lluMiB -> %lluMiB\n",
97 (u64)current_size >> 20, (u64)rebar_size >> 20);
98
99 while (root->parent)
100 root = root->parent;
101
102 pci_bus_for_each_resource(root, root_res, i) {
103 if (root_res && root_res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
104 (u64)root_res->start > 0x100000000ul)
105 break;
106 }
107
108 if (!root_res) {
109 drm_info(&xe->drm, "Can't resize VRAM BAR - platform support is missing. Consider enabling 'Resizable BAR' support in your BIOS\n");
110 return;
111 }
112
113 pci_read_config_dword(pdev, PCI_COMMAND, &pci_cmd);
114 pci_write_config_dword(pdev, PCI_COMMAND, pci_cmd & ~PCI_COMMAND_MEMORY);
115
116 _resize_bar(xe, LMEM_BAR, rebar_size);
117
118 pci_assign_unassigned_bus_resources(pdev->bus);
119 pci_write_config_dword(pdev, PCI_COMMAND, pci_cmd);
120 }
121
resource_is_valid(struct pci_dev * pdev,int bar)122 static bool resource_is_valid(struct pci_dev *pdev, int bar)
123 {
124 if (!pci_resource_flags(pdev, bar))
125 return false;
126
127 if (pci_resource_flags(pdev, bar) & IORESOURCE_UNSET)
128 return false;
129
130 if (!pci_resource_len(pdev, bar))
131 return false;
132
133 return true;
134 }
135
determine_lmem_bar_size(struct xe_device * xe)136 static int determine_lmem_bar_size(struct xe_device *xe)
137 {
138 struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
139
140 if (!resource_is_valid(pdev, LMEM_BAR)) {
141 drm_err(&xe->drm, "pci resource is not valid\n");
142 return -ENXIO;
143 }
144
145 resize_vram_bar(xe);
146
147 xe->mem.vram.io_start = pci_resource_start(pdev, LMEM_BAR);
148 xe->mem.vram.io_size = pci_resource_len(pdev, LMEM_BAR);
149 if (!xe->mem.vram.io_size)
150 return -EIO;
151
152 /* XXX: Need to change when xe link code is ready */
153 xe->mem.vram.dpa_base = 0;
154
155 /* set up a map to the total memory area. */
156 xe->mem.vram.mapping = ioremap_wc(xe->mem.vram.io_start, xe->mem.vram.io_size);
157
158 return 0;
159 }
160
get_flat_ccs_offset(struct xe_gt * gt,u64 tile_size)161 static inline u64 get_flat_ccs_offset(struct xe_gt *gt, u64 tile_size)
162 {
163 struct xe_device *xe = gt_to_xe(gt);
164 u64 offset;
165 u32 reg;
166
167 if (GRAPHICS_VER(xe) >= 20) {
168 u64 ccs_size = tile_size / 512;
169 u64 offset_hi, offset_lo;
170 u32 nodes, num_enabled;
171
172 reg = xe_mmio_read32(>->mmio, MIRROR_FUSE3);
173 nodes = REG_FIELD_GET(XE2_NODE_ENABLE_MASK, reg);
174 num_enabled = hweight32(nodes); /* Number of enabled l3 nodes */
175
176 reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_LOWER);
177 offset_lo = REG_FIELD_GET(XE2_FLAT_CCS_BASE_LOWER_ADDR_MASK, reg);
178
179 reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_UPPER);
180 offset_hi = REG_FIELD_GET(XE2_FLAT_CCS_BASE_UPPER_ADDR_MASK, reg);
181
182 offset = offset_hi << 32; /* HW view bits 39:32 */
183 offset |= offset_lo << 6; /* HW view bits 31:6 */
184 offset *= num_enabled; /* convert to SW view */
185 offset = round_up(offset, SZ_128K); /* SW must round up to nearest 128K */
186
187 /* We don't expect any holes */
188 xe_assert_msg(xe, offset == (xe_mmio_read64_2x32(>_to_tile(gt)->mmio, GSMBASE) -
189 ccs_size),
190 "Hole between CCS and GSM.\n");
191 } else {
192 reg = xe_gt_mcr_unicast_read_any(gt, XEHP_FLAT_CCS_BASE_ADDR);
193 offset = (u64)REG_FIELD_GET(XEHP_FLAT_CCS_PTR, reg) * SZ_64K;
194 }
195
196 return offset;
197 }
198
199 /*
200 * tile_vram_size() - Collect vram size and offset information
201 * @tile: tile to get info for
202 * @vram_size: available vram (size - device reserved portions)
203 * @tile_size: actual vram size
204 * @tile_offset: physical start point in the vram address space
205 *
206 * There are 4 places for size information:
207 * - io size (from pci_resource_len of LMEM bar) (only used for small bar and DG1)
208 * - TILEx size (actual vram size)
209 * - GSMBASE offset (TILEx - "stolen")
210 * - CSSBASE offset (TILEx - CSS space necessary)
211 *
212 * CSSBASE is always a lower/smaller offset then GSMBASE.
213 *
214 * The actual available size of memory is to the CCS or GSM base.
215 * NOTE: multi-tile bases will include the tile offset.
216 *
217 */
tile_vram_size(struct xe_tile * tile,u64 * vram_size,u64 * tile_size,u64 * tile_offset)218 static int tile_vram_size(struct xe_tile *tile, u64 *vram_size,
219 u64 *tile_size, u64 *tile_offset)
220 {
221 struct xe_device *xe = tile_to_xe(tile);
222 struct xe_gt *gt = tile->primary_gt;
223 unsigned int fw_ref;
224 u64 offset;
225 u32 reg;
226
227 if (IS_SRIOV_VF(xe)) {
228 struct xe_tile *t;
229 int id;
230
231 offset = 0;
232 for_each_tile(t, xe, id)
233 for_each_if(t->id < tile->id)
234 offset += xe_gt_sriov_vf_lmem(t->primary_gt);
235
236 *tile_size = xe_gt_sriov_vf_lmem(gt);
237 *vram_size = *tile_size;
238 *tile_offset = offset;
239
240 return 0;
241 }
242
243 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
244 if (!fw_ref)
245 return -ETIMEDOUT;
246
247 /* actual size */
248 if (unlikely(xe->info.platform == XE_DG1)) {
249 *tile_size = pci_resource_len(to_pci_dev(xe->drm.dev), LMEM_BAR);
250 *tile_offset = 0;
251 } else {
252 reg = xe_gt_mcr_unicast_read_any(gt, XEHP_TILE_ADDR_RANGE(gt->info.id));
253 *tile_size = (u64)REG_FIELD_GET(GENMASK(14, 8), reg) * SZ_1G;
254 *tile_offset = (u64)REG_FIELD_GET(GENMASK(7, 1), reg) * SZ_1G;
255 }
256
257 /* minus device usage */
258 if (xe->info.has_flat_ccs) {
259 offset = get_flat_ccs_offset(gt, *tile_size);
260 } else {
261 offset = xe_mmio_read64_2x32(&tile->mmio, GSMBASE);
262 }
263
264 /* remove the tile offset so we have just the available size */
265 *vram_size = offset - *tile_offset;
266
267 xe_force_wake_put(gt_to_fw(gt), fw_ref);
268
269 return 0;
270 }
271
vram_fini(void * arg)272 static void vram_fini(void *arg)
273 {
274 struct xe_device *xe = arg;
275 struct xe_tile *tile;
276 int id;
277
278 if (xe->mem.vram.mapping)
279 iounmap(xe->mem.vram.mapping);
280
281 xe->mem.vram.mapping = NULL;
282
283 for_each_tile(tile, xe, id)
284 tile->mem.vram.mapping = NULL;
285 }
286
287 /**
288 * xe_vram_probe() - Probe VRAM configuration
289 * @xe: the &xe_device
290 *
291 * Collect VRAM size and offset information for all tiles.
292 *
293 * Return: 0 on success, error code on failure
294 */
xe_vram_probe(struct xe_device * xe)295 int xe_vram_probe(struct xe_device *xe)
296 {
297 struct xe_tile *tile;
298 resource_size_t io_size;
299 u64 available_size = 0;
300 u64 total_size = 0;
301 u64 tile_offset;
302 u64 tile_size;
303 u64 vram_size;
304 int err;
305 u8 id;
306
307 if (!IS_DGFX(xe))
308 return 0;
309
310 /* Get the size of the root tile's vram for later accessibility comparison */
311 tile = xe_device_get_root_tile(xe);
312 err = tile_vram_size(tile, &vram_size, &tile_size, &tile_offset);
313 if (err)
314 return err;
315
316 err = determine_lmem_bar_size(xe);
317 if (err)
318 return err;
319
320 drm_info(&xe->drm, "VISIBLE VRAM: %pa, %pa\n", &xe->mem.vram.io_start,
321 &xe->mem.vram.io_size);
322
323 io_size = xe->mem.vram.io_size;
324
325 /* tile specific ranges */
326 for_each_tile(tile, xe, id) {
327 err = tile_vram_size(tile, &vram_size, &tile_size, &tile_offset);
328 if (err)
329 return err;
330
331 tile->mem.vram.actual_physical_size = tile_size;
332 tile->mem.vram.io_start = xe->mem.vram.io_start + tile_offset;
333 tile->mem.vram.io_size = min_t(u64, vram_size, io_size);
334
335 if (!tile->mem.vram.io_size) {
336 drm_err(&xe->drm, "Tile without any CPU visible VRAM. Aborting.\n");
337 return -ENODEV;
338 }
339
340 tile->mem.vram.dpa_base = xe->mem.vram.dpa_base + tile_offset;
341 tile->mem.vram.usable_size = vram_size;
342 tile->mem.vram.mapping = xe->mem.vram.mapping + tile_offset;
343
344 if (tile->mem.vram.io_size < tile->mem.vram.usable_size)
345 drm_info(&xe->drm, "Small BAR device\n");
346 drm_info(&xe->drm, "VRAM[%u, %u]: Actual physical size %pa, usable size exclude stolen %pa, CPU accessible size %pa\n", id,
347 tile->id, &tile->mem.vram.actual_physical_size, &tile->mem.vram.usable_size, &tile->mem.vram.io_size);
348 drm_info(&xe->drm, "VRAM[%u, %u]: DPA range: [%pa-%llx], io range: [%pa-%llx]\n", id, tile->id,
349 &tile->mem.vram.dpa_base, tile->mem.vram.dpa_base + (u64)tile->mem.vram.actual_physical_size,
350 &tile->mem.vram.io_start, tile->mem.vram.io_start + (u64)tile->mem.vram.io_size);
351
352 /* calculate total size using tile size to get the correct HW sizing */
353 total_size += tile_size;
354 available_size += vram_size;
355
356 if (total_size > xe->mem.vram.io_size) {
357 drm_info(&xe->drm, "VRAM: %pa is larger than resource %pa\n",
358 &total_size, &xe->mem.vram.io_size);
359 }
360
361 io_size -= min_t(u64, tile_size, io_size);
362 }
363
364 xe->mem.vram.actual_physical_size = total_size;
365
366 drm_info(&xe->drm, "Total VRAM: %pa, %pa\n", &xe->mem.vram.io_start,
367 &xe->mem.vram.actual_physical_size);
368 drm_info(&xe->drm, "Available VRAM: %pa, %pa\n", &xe->mem.vram.io_start,
369 &available_size);
370
371 return devm_add_action_or_reset(xe->drm.dev, vram_fini, xe);
372 }
373