Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Firmware replacement code.
4 *
5 * Work around broken BIOSes that don't set an aperture, only set the
6 * aperture in the AGP bridge, or set too small aperture.
7 *
8 * If all fails map the aperture over some low memory. This is cheaper than
9 * doing bounce buffering. The memory is lost. This is done at early boot
10 * because only the bootmem allocator can allocate 32+MB.
11 *
12 * Copyright 2002 Andi Kleen, SuSE Labs.
13 */
14#define pr_fmt(fmt) "AGP: " fmt
15
16#include <linux/kernel.h>
17#include <linux/kcore.h>
18#include <linux/types.h>
19#include <linux/init.h>
20#include <linux/memblock.h>
21#include <linux/mmzone.h>
22#include <linux/pci_ids.h>
23#include <linux/pci.h>
24#include <linux/bitops.h>
25#include <linux/suspend.h>
26#include <asm/e820/api.h>
27#include <asm/io.h>
28#include <asm/iommu.h>
29#include <asm/gart.h>
30#include <asm/pci-direct.h>
31#include <asm/dma.h>
32#include <asm/amd_nb.h>
33#include <asm/x86_init.h>
34#include <linux/crash_dump.h>
35
36/*
37 * Using 512M as goal, in case kexec will load kernel_big
38 * that will do the on-position decompress, and could overlap with
39 * the gart aperture that is used.
40 * Sequence:
41 * kernel_small
42 * ==> kexec (with kdump trigger path or gart still enabled)
43 * ==> kernel_small (gart area become e820_reserved)
44 * ==> kexec (with kdump trigger path or gart still enabled)
45 * ==> kerne_big (uncompressed size will be big than 64M or 128M)
46 * So don't use 512M below as gart iommu, leave the space for kernel
47 * code for safe.
48 */
49#define GART_MIN_ADDR (512ULL << 20)
50#define GART_MAX_ADDR (1ULL << 32)
51
52int gart_iommu_aperture;
53int gart_iommu_aperture_disabled __initdata;
54int gart_iommu_aperture_allowed __initdata;
55
56int fallback_aper_order __initdata = 1; /* 64MB */
57int fallback_aper_force __initdata;
58
59int fix_aperture __initdata = 1;
60
61#if defined(CONFIG_PROC_VMCORE) || defined(CONFIG_PROC_KCORE)
62/*
63 * If the first kernel maps the aperture over e820 RAM, the kdump kernel will
64 * use the same range because it will remain configured in the northbridge.
65 * Trying to dump this area via /proc/vmcore may crash the machine, so exclude
66 * it from vmcore.
67 */
68static unsigned long aperture_pfn_start, aperture_page_count;
69
70static int gart_mem_pfn_is_ram(unsigned long pfn)
71{
72 return likely((pfn < aperture_pfn_start) ||
73 (pfn >= aperture_pfn_start + aperture_page_count));
74}
75
76#ifdef CONFIG_PROC_VMCORE
77static bool gart_oldmem_pfn_is_ram(struct vmcore_cb *cb, unsigned long pfn)
78{
79 return !!gart_mem_pfn_is_ram(pfn);
80}
81
82static struct vmcore_cb gart_vmcore_cb = {
83 .pfn_is_ram = gart_oldmem_pfn_is_ram,
84};
85#endif
86
87static void __init exclude_from_core(u64 aper_base, u32 aper_order)
88{
89 aperture_pfn_start = aper_base >> PAGE_SHIFT;
90 aperture_page_count = (32 * 1024 * 1024) << aper_order >> PAGE_SHIFT;
91#ifdef CONFIG_PROC_VMCORE
92 register_vmcore_cb(&gart_vmcore_cb);
93#endif
94#ifdef CONFIG_PROC_KCORE
95 WARN_ON(register_mem_pfn_is_ram(&gart_mem_pfn_is_ram));
96#endif
97}
98#else
99static void exclude_from_core(u64 aper_base, u32 aper_order)
100{
101}
102#endif
103
104/* This code runs before the PCI subsystem is initialized, so just
105 access the northbridge directly. */
106
107static u32 __init allocate_aperture(void)
108{
109 u32 aper_size;
110 unsigned long addr;
111
112 /* aper_size should <= 1G */
113 if (fallback_aper_order > 5)
114 fallback_aper_order = 5;
115 aper_size = (32 * 1024 * 1024) << fallback_aper_order;
116
117 /*
118 * Aperture has to be naturally aligned. This means a 2GB aperture
119 * won't have much chance of finding a place in the lower 4GB of
120 * memory. Unfortunately we cannot move it up because that would
121 * make the IOMMU useless.
122 */
123 addr = memblock_phys_alloc_range(aper_size, aper_size,
124 GART_MIN_ADDR, GART_MAX_ADDR);
125 if (!addr) {
126 pr_err("Cannot allocate aperture memory hole [mem %#010lx-%#010lx] (%uKB)\n",
127 addr, addr + aper_size - 1, aper_size >> 10);
128 return 0;
129 }
130 pr_info("Mapping aperture over RAM [mem %#010lx-%#010lx] (%uKB)\n",
131 addr, addr + aper_size - 1, aper_size >> 10);
132 register_nosave_region(addr >> PAGE_SHIFT,
133 (addr+aper_size) >> PAGE_SHIFT);
134
135 return (u32)addr;
136}
137
138
139/* Find a PCI capability */
140static u32 __init find_cap(int bus, int slot, int func, int cap)
141{
142 int bytes;
143 u8 pos;
144
145 if (!(read_pci_config_16(bus, slot, func, PCI_STATUS) &
146 PCI_STATUS_CAP_LIST))
147 return 0;
148
149 pos = read_pci_config_byte(bus, slot, func, PCI_CAPABILITY_LIST);
150 for (bytes = 0; bytes < 48 && pos >= 0x40; bytes++) {
151 u8 id;
152
153 pos &= ~3;
154 id = read_pci_config_byte(bus, slot, func, pos+PCI_CAP_LIST_ID);
155 if (id == 0xff)
156 break;
157 if (id == cap)
158 return pos;
159 pos = read_pci_config_byte(bus, slot, func,
160 pos+PCI_CAP_LIST_NEXT);
161 }
162 return 0;
163}
164
165/* Read a standard AGPv3 bridge header */
166static u32 __init read_agp(int bus, int slot, int func, int cap, u32 *order)
167{
168 u32 apsize;
169 u32 apsizereg;
170 int nbits;
171 u32 aper_low, aper_hi;
172 u64 aper;
173 u32 old_order;
174
175 pr_info("pci 0000:%02x:%02x:%02x: AGP bridge\n", bus, slot, func);
176 apsizereg = read_pci_config_16(bus, slot, func, cap + 0x14);
177 if (apsizereg == 0xffffffff) {
178 pr_err("pci 0000:%02x:%02x.%d: APSIZE unreadable\n",
179 bus, slot, func);
180 return 0;
181 }
182
183 /* old_order could be the value from NB gart setting */
184 old_order = *order;
185
186 apsize = apsizereg & 0xfff;
187 /* Some BIOS use weird encodings not in the AGPv3 table. */
188 if (apsize & 0xff)
189 apsize |= 0xf00;
190 nbits = hweight16(apsize);
191 *order = 7 - nbits;
192 if ((int)*order < 0) /* < 32MB */
193 *order = 0;
194
195 aper_low = read_pci_config(bus, slot, func, 0x10);
196 aper_hi = read_pci_config(bus, slot, func, 0x14);
197 aper = (aper_low & ~((1<<22)-1)) | ((u64)aper_hi << 32);
198
199 /*
200 * On some sick chips, APSIZE is 0. It means it wants 4G
201 * so let double check that order, and lets trust AMD NB settings:
202 */
203 pr_info("pci 0000:%02x:%02x.%d: AGP aperture [bus addr %#010Lx-%#010Lx] (old size %uMB)\n",
204 bus, slot, func, aper, aper + (32ULL << (old_order + 20)) - 1,
205 32 << old_order);
206 if (aper + (32ULL<<(20 + *order)) > 0x100000000ULL) {
207 pr_info("pci 0000:%02x:%02x.%d: AGP aperture size %uMB (APSIZE %#x) is not right, using settings from NB\n",
208 bus, slot, func, 32 << *order, apsizereg);
209 *order = old_order;
210 }
211
212 pr_info("pci 0000:%02x:%02x.%d: AGP aperture [bus addr %#010Lx-%#010Lx] (%uMB, APSIZE %#x)\n",
213 bus, slot, func, aper, aper + (32ULL << (*order + 20)) - 1,
214 32 << *order, apsizereg);
215
216 if (!aperture_valid(aper, (32*1024*1024) << *order, 32<<20))
217 return 0;
218 return (u32)aper;
219}
220
221/*
222 * Look for an AGP bridge. Windows only expects the aperture in the
223 * AGP bridge and some BIOS forget to initialize the Northbridge too.
224 * Work around this here.
225 *
226 * Do an PCI bus scan by hand because we're running before the PCI
227 * subsystem.
228 *
229 * All AMD AGP bridges are AGPv3 compliant, so we can do this scan
230 * generically. It's probably overkill to always scan all slots because
231 * the AGP bridges should be always an own bus on the HT hierarchy,
232 * but do it here for future safety.
233 */
234static u32 __init search_agp_bridge(u32 *order, int *valid_agp)
235{
236 int bus, slot, func;
237
238 /* Poor man's PCI discovery */
239 for (bus = 0; bus < 256; bus++) {
240 for (slot = 0; slot < 32; slot++) {
241 for (func = 0; func < 8; func++) {
242 u32 class, cap;
243 u8 type;
244 class = read_pci_config(bus, slot, func,
245 PCI_CLASS_REVISION);
246 if (class == 0xffffffff)
247 break;
248
249 switch (class >> 16) {
250 case PCI_CLASS_BRIDGE_HOST:
251 case PCI_CLASS_BRIDGE_OTHER: /* needed? */
252 /* AGP bridge? */
253 cap = find_cap(bus, slot, func,
254 PCI_CAP_ID_AGP);
255 if (!cap)
256 break;
257 *valid_agp = 1;
258 return read_agp(bus, slot, func, cap,
259 order);
260 }
261
262 type = read_pci_config_byte(bus, slot, func,
263 PCI_HEADER_TYPE);
264 if (!(type & PCI_HEADER_TYPE_MFD))
265 break;
266 }
267 }
268 }
269 pr_info("No AGP bridge found\n");
270
271 return 0;
272}
273
274static bool gart_fix_e820 __initdata = true;
275
276static int __init parse_gart_mem(char *p)
277{
278 return kstrtobool(p, &gart_fix_e820);
279}
280early_param("gart_fix_e820", parse_gart_mem);
281
282/*
283 * With kexec/kdump, if the first kernel doesn't shut down the GART and the
284 * second kernel allocates a different GART region, there might be two
285 * overlapping GART regions present:
286 *
287 * - the first still used by the GART initialized in the first kernel.
288 * - (sub-)set of it used as normal RAM by the second kernel.
289 *
290 * which leads to memory corruptions and a kernel panic eventually.
291 *
292 * This can also happen if the BIOS has forgotten to mark the GART region
293 * as reserved.
294 *
295 * Try to update the e820 map to mark that new region as reserved.
296 */
297void __init early_gart_iommu_check(void)
298{
299 u32 agp_aper_order = 0;
300 int i, fix, slot, valid_agp = 0;
301 u32 ctl;
302 u32 aper_size = 0, aper_order = 0, last_aper_order = 0;
303 u64 aper_base = 0, last_aper_base = 0;
304 int aper_enabled = 0, last_aper_enabled = 0, last_valid = 0;
305
306 if (!amd_gart_present())
307 return;
308
309 if (!early_pci_allowed())
310 return;
311
312 /* This is mostly duplicate of iommu_hole_init */
313 search_agp_bridge(&agp_aper_order, &valid_agp);
314
315 fix = 0;
316 for (i = 0; amd_nb_bus_dev_ranges[i].dev_limit; i++) {
317 int bus;
318 int dev_base, dev_limit;
319
320 bus = amd_nb_bus_dev_ranges[i].bus;
321 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
322 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
323
324 for (slot = dev_base; slot < dev_limit; slot++) {
325 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
326 continue;
327
328 ctl = read_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL);
329 aper_enabled = ctl & GARTEN;
330 aper_order = (ctl >> 1) & 7;
331 aper_size = (32 * 1024 * 1024) << aper_order;
332 aper_base = read_pci_config(bus, slot, 3, AMD64_GARTAPERTUREBASE) & 0x7fff;
333 aper_base <<= 25;
334
335 if (last_valid) {
336 if ((aper_order != last_aper_order) ||
337 (aper_base != last_aper_base) ||
338 (aper_enabled != last_aper_enabled)) {
339 fix = 1;
340 break;
341 }
342 }
343
344 last_aper_order = aper_order;
345 last_aper_base = aper_base;
346 last_aper_enabled = aper_enabled;
347 last_valid = 1;
348 }
349 }
350
351 if (!fix && !aper_enabled)
352 return;
353
354 if (!aper_base || !aper_size || aper_base + aper_size > 0x100000000UL)
355 fix = 1;
356
357 if (gart_fix_e820 && !fix && aper_enabled) {
358 if (e820__mapped_any(aper_base, aper_base + aper_size,
359 E820_TYPE_RAM)) {
360 /* reserve it, so we can reuse it in second kernel */
361 pr_info("e820: reserve [mem %#010Lx-%#010Lx] for GART\n",
362 aper_base, aper_base + aper_size - 1);
363 e820__range_add(aper_base, aper_size, E820_TYPE_RESERVED);
364 e820__update_table_print();
365 }
366 }
367
368 if (valid_agp)
369 return;
370
371 /* disable them all at first */
372 for (i = 0; i < amd_nb_bus_dev_ranges[i].dev_limit; i++) {
373 int bus;
374 int dev_base, dev_limit;
375
376 bus = amd_nb_bus_dev_ranges[i].bus;
377 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
378 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
379
380 for (slot = dev_base; slot < dev_limit; slot++) {
381 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
382 continue;
383
384 ctl = read_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL);
385 ctl &= ~GARTEN;
386 write_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL, ctl);
387 }
388 }
389
390}
391
392static int __initdata printed_gart_size_msg;
393
394void __init gart_iommu_hole_init(void)
395{
396 u32 agp_aper_base = 0, agp_aper_order = 0;
397 u32 aper_size, aper_alloc = 0, aper_order = 0, last_aper_order = 0;
398 u64 aper_base, last_aper_base = 0;
399 int fix, slot, valid_agp = 0;
400 int i, node;
401
402 if (!amd_gart_present())
403 return;
404
405 if (gart_iommu_aperture_disabled || !fix_aperture ||
406 !early_pci_allowed())
407 return;
408
409 pr_info("Checking aperture...\n");
410
411 if (!fallback_aper_force)
412 agp_aper_base = search_agp_bridge(&agp_aper_order, &valid_agp);
413
414 fix = 0;
415 node = 0;
416 for (i = 0; i < amd_nb_bus_dev_ranges[i].dev_limit; i++) {
417 int bus;
418 int dev_base, dev_limit;
419 u32 ctl;
420
421 bus = amd_nb_bus_dev_ranges[i].bus;
422 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
423 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
424
425 for (slot = dev_base; slot < dev_limit; slot++) {
426 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
427 continue;
428
429 iommu_detected = 1;
430 gart_iommu_aperture = 1;
431 x86_init.iommu.iommu_init = gart_iommu_init;
432
433 ctl = read_pci_config(bus, slot, 3,
434 AMD64_GARTAPERTURECTL);
435
436 /*
437 * Before we do anything else disable the GART. It may
438 * still be enabled if we boot into a crash-kernel here.
439 * Reconfiguring the GART while it is enabled could have
440 * unknown side-effects.
441 */
442 ctl &= ~GARTEN;
443 write_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL, ctl);
444
445 aper_order = (ctl >> 1) & 7;
446 aper_size = (32 * 1024 * 1024) << aper_order;
447 aper_base = read_pci_config(bus, slot, 3, AMD64_GARTAPERTUREBASE) & 0x7fff;
448 aper_base <<= 25;
449
450 pr_info("Node %d: aperture [bus addr %#010Lx-%#010Lx] (%uMB)\n",
451 node, aper_base, aper_base + aper_size - 1,
452 aper_size >> 20);
453 node++;
454
455 if (!aperture_valid(aper_base, aper_size, 64<<20)) {
456 if (valid_agp && agp_aper_base &&
457 agp_aper_base == aper_base &&
458 agp_aper_order == aper_order) {
459 /* the same between two setting from NB and agp */
460 if (!no_iommu &&
461 max_pfn > MAX_DMA32_PFN &&
462 !printed_gart_size_msg) {
463 pr_err("you are using iommu with agp, but GART size is less than 64MB\n");
464 pr_err("please increase GART size in your BIOS setup\n");
465 pr_err("if BIOS doesn't have that option, contact your HW vendor!\n");
466 printed_gart_size_msg = 1;
467 }
468 } else {
469 fix = 1;
470 goto out;
471 }
472 }
473
474 if ((last_aper_order && aper_order != last_aper_order) ||
475 (last_aper_base && aper_base != last_aper_base)) {
476 fix = 1;
477 goto out;
478 }
479 last_aper_order = aper_order;
480 last_aper_base = aper_base;
481 }
482 }
483
484out:
485 if (!fix && !fallback_aper_force) {
486 if (last_aper_base) {
487 /*
488 * If this is the kdump kernel, the first kernel
489 * may have allocated the range over its e820 RAM
490 * and fixed up the northbridge
491 */
492 exclude_from_core(last_aper_base, last_aper_order);
493 }
494 return;
495 }
496
497 if (!fallback_aper_force) {
498 aper_alloc = agp_aper_base;
499 aper_order = agp_aper_order;
500 }
501
502 if (aper_alloc) {
503 /* Got the aperture from the AGP bridge */
504 } else if ((!no_iommu && max_pfn > MAX_DMA32_PFN) ||
505 force_iommu ||
506 valid_agp ||
507 fallback_aper_force) {
508 pr_info("Your BIOS doesn't leave an aperture memory hole\n");
509 pr_info("Please enable the IOMMU option in the BIOS setup\n");
510 pr_info("This costs you %dMB of RAM\n",
511 32 << fallback_aper_order);
512
513 aper_order = fallback_aper_order;
514 aper_alloc = allocate_aperture();
515 if (!aper_alloc) {
516 /*
517 * Could disable AGP and IOMMU here, but it's
518 * probably not worth it. But the later users
519 * cannot deal with bad apertures and turning
520 * on the aperture over memory causes very
521 * strange problems, so it's better to panic
522 * early.
523 */
524 panic("Not enough memory for aperture");
525 }
526 } else {
527 return;
528 }
529
530 /*
531 * If this is the kdump kernel _and_ the first kernel did not
532 * configure the aperture in the northbridge, this range may
533 * overlap with the first kernel's memory. We can't access the
534 * range through vmcore even though it should be part of the dump.
535 */
536 exclude_from_core(aper_alloc, aper_order);
537
538 /* Fix up the north bridges */
539 for (i = 0; i < amd_nb_bus_dev_ranges[i].dev_limit; i++) {
540 int bus, dev_base, dev_limit;
541
542 /*
543 * Don't enable translation yet but enable GART IO and CPU
544 * accesses and set DISTLBWALKPRB since GART table memory is UC.
545 */
546 u32 ctl = aper_order << 1;
547
548 bus = amd_nb_bus_dev_ranges[i].bus;
549 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
550 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
551 for (slot = dev_base; slot < dev_limit; slot++) {
552 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
553 continue;
554
555 write_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL, ctl);
556 write_pci_config(bus, slot, 3, AMD64_GARTAPERTUREBASE, aper_alloc >> 25);
557 }
558 }
559
560 set_up_gart_resume(aper_order, aper_alloc);
561}
1/*
2 * Firmware replacement code.
3 *
4 * Work around broken BIOSes that don't set an aperture, only set the
5 * aperture in the AGP bridge, or set too small aperture.
6 *
7 * If all fails map the aperture over some low memory. This is cheaper than
8 * doing bounce buffering. The memory is lost. This is done at early boot
9 * because only the bootmem allocator can allocate 32+MB.
10 *
11 * Copyright 2002 Andi Kleen, SuSE Labs.
12 */
13#include <linux/kernel.h>
14#include <linux/types.h>
15#include <linux/init.h>
16#include <linux/memblock.h>
17#include <linux/mmzone.h>
18#include <linux/pci_ids.h>
19#include <linux/pci.h>
20#include <linux/bitops.h>
21#include <linux/suspend.h>
22#include <asm/e820.h>
23#include <asm/io.h>
24#include <asm/iommu.h>
25#include <asm/gart.h>
26#include <asm/pci-direct.h>
27#include <asm/dma.h>
28#include <asm/amd_nb.h>
29#include <asm/x86_init.h>
30
31/*
32 * Using 512M as goal, in case kexec will load kernel_big
33 * that will do the on-position decompress, and could overlap with
34 * with the gart aperture that is used.
35 * Sequence:
36 * kernel_small
37 * ==> kexec (with kdump trigger path or gart still enabled)
38 * ==> kernel_small (gart area become e820_reserved)
39 * ==> kexec (with kdump trigger path or gart still enabled)
40 * ==> kerne_big (uncompressed size will be big than 64M or 128M)
41 * So don't use 512M below as gart iommu, leave the space for kernel
42 * code for safe.
43 */
44#define GART_MIN_ADDR (512ULL << 20)
45#define GART_MAX_ADDR (1ULL << 32)
46
47int gart_iommu_aperture;
48int gart_iommu_aperture_disabled __initdata;
49int gart_iommu_aperture_allowed __initdata;
50
51int fallback_aper_order __initdata = 1; /* 64MB */
52int fallback_aper_force __initdata;
53
54int fix_aperture __initdata = 1;
55
56/* This code runs before the PCI subsystem is initialized, so just
57 access the northbridge directly. */
58
59static u32 __init allocate_aperture(void)
60{
61 u32 aper_size;
62 unsigned long addr;
63
64 /* aper_size should <= 1G */
65 if (fallback_aper_order > 5)
66 fallback_aper_order = 5;
67 aper_size = (32 * 1024 * 1024) << fallback_aper_order;
68
69 /*
70 * Aperture has to be naturally aligned. This means a 2GB aperture
71 * won't have much chance of finding a place in the lower 4GB of
72 * memory. Unfortunately we cannot move it up because that would
73 * make the IOMMU useless.
74 */
75 addr = memblock_find_in_range(GART_MIN_ADDR, GART_MAX_ADDR,
76 aper_size, aper_size);
77 if (!addr) {
78 printk(KERN_ERR
79 "Cannot allocate aperture memory hole (%lx,%uK)\n",
80 addr, aper_size>>10);
81 return 0;
82 }
83 memblock_reserve(addr, aper_size);
84 printk(KERN_INFO "Mapping aperture over %d KB of RAM @ %lx\n",
85 aper_size >> 10, addr);
86 register_nosave_region(addr >> PAGE_SHIFT,
87 (addr+aper_size) >> PAGE_SHIFT);
88
89 return (u32)addr;
90}
91
92
93/* Find a PCI capability */
94static u32 __init find_cap(int bus, int slot, int func, int cap)
95{
96 int bytes;
97 u8 pos;
98
99 if (!(read_pci_config_16(bus, slot, func, PCI_STATUS) &
100 PCI_STATUS_CAP_LIST))
101 return 0;
102
103 pos = read_pci_config_byte(bus, slot, func, PCI_CAPABILITY_LIST);
104 for (bytes = 0; bytes < 48 && pos >= 0x40; bytes++) {
105 u8 id;
106
107 pos &= ~3;
108 id = read_pci_config_byte(bus, slot, func, pos+PCI_CAP_LIST_ID);
109 if (id == 0xff)
110 break;
111 if (id == cap)
112 return pos;
113 pos = read_pci_config_byte(bus, slot, func,
114 pos+PCI_CAP_LIST_NEXT);
115 }
116 return 0;
117}
118
119/* Read a standard AGPv3 bridge header */
120static u32 __init read_agp(int bus, int slot, int func, int cap, u32 *order)
121{
122 u32 apsize;
123 u32 apsizereg;
124 int nbits;
125 u32 aper_low, aper_hi;
126 u64 aper;
127 u32 old_order;
128
129 printk(KERN_INFO "AGP bridge at %02x:%02x:%02x\n", bus, slot, func);
130 apsizereg = read_pci_config_16(bus, slot, func, cap + 0x14);
131 if (apsizereg == 0xffffffff) {
132 printk(KERN_ERR "APSIZE in AGP bridge unreadable\n");
133 return 0;
134 }
135
136 /* old_order could be the value from NB gart setting */
137 old_order = *order;
138
139 apsize = apsizereg & 0xfff;
140 /* Some BIOS use weird encodings not in the AGPv3 table. */
141 if (apsize & 0xff)
142 apsize |= 0xf00;
143 nbits = hweight16(apsize);
144 *order = 7 - nbits;
145 if ((int)*order < 0) /* < 32MB */
146 *order = 0;
147
148 aper_low = read_pci_config(bus, slot, func, 0x10);
149 aper_hi = read_pci_config(bus, slot, func, 0x14);
150 aper = (aper_low & ~((1<<22)-1)) | ((u64)aper_hi << 32);
151
152 /*
153 * On some sick chips, APSIZE is 0. It means it wants 4G
154 * so let double check that order, and lets trust AMD NB settings:
155 */
156 printk(KERN_INFO "Aperture from AGP @ %Lx old size %u MB\n",
157 aper, 32 << old_order);
158 if (aper + (32ULL<<(20 + *order)) > 0x100000000ULL) {
159 printk(KERN_INFO "Aperture size %u MB (APSIZE %x) is not right, using settings from NB\n",
160 32 << *order, apsizereg);
161 *order = old_order;
162 }
163
164 printk(KERN_INFO "Aperture from AGP @ %Lx size %u MB (APSIZE %x)\n",
165 aper, 32 << *order, apsizereg);
166
167 if (!aperture_valid(aper, (32*1024*1024) << *order, 32<<20))
168 return 0;
169 return (u32)aper;
170}
171
172/*
173 * Look for an AGP bridge. Windows only expects the aperture in the
174 * AGP bridge and some BIOS forget to initialize the Northbridge too.
175 * Work around this here.
176 *
177 * Do an PCI bus scan by hand because we're running before the PCI
178 * subsystem.
179 *
180 * All AMD AGP bridges are AGPv3 compliant, so we can do this scan
181 * generically. It's probably overkill to always scan all slots because
182 * the AGP bridges should be always an own bus on the HT hierarchy,
183 * but do it here for future safety.
184 */
185static u32 __init search_agp_bridge(u32 *order, int *valid_agp)
186{
187 int bus, slot, func;
188
189 /* Poor man's PCI discovery */
190 for (bus = 0; bus < 256; bus++) {
191 for (slot = 0; slot < 32; slot++) {
192 for (func = 0; func < 8; func++) {
193 u32 class, cap;
194 u8 type;
195 class = read_pci_config(bus, slot, func,
196 PCI_CLASS_REVISION);
197 if (class == 0xffffffff)
198 break;
199
200 switch (class >> 16) {
201 case PCI_CLASS_BRIDGE_HOST:
202 case PCI_CLASS_BRIDGE_OTHER: /* needed? */
203 /* AGP bridge? */
204 cap = find_cap(bus, slot, func,
205 PCI_CAP_ID_AGP);
206 if (!cap)
207 break;
208 *valid_agp = 1;
209 return read_agp(bus, slot, func, cap,
210 order);
211 }
212
213 /* No multi-function device? */
214 type = read_pci_config_byte(bus, slot, func,
215 PCI_HEADER_TYPE);
216 if (!(type & 0x80))
217 break;
218 }
219 }
220 }
221 printk(KERN_INFO "No AGP bridge found\n");
222
223 return 0;
224}
225
226static int gart_fix_e820 __initdata = 1;
227
228static int __init parse_gart_mem(char *p)
229{
230 if (!p)
231 return -EINVAL;
232
233 if (!strncmp(p, "off", 3))
234 gart_fix_e820 = 0;
235 else if (!strncmp(p, "on", 2))
236 gart_fix_e820 = 1;
237
238 return 0;
239}
240early_param("gart_fix_e820", parse_gart_mem);
241
242void __init early_gart_iommu_check(void)
243{
244 /*
245 * in case it is enabled before, esp for kexec/kdump,
246 * previous kernel already enable that. memset called
247 * by allocate_aperture/__alloc_bootmem_nopanic cause restart.
248 * or second kernel have different position for GART hole. and new
249 * kernel could use hole as RAM that is still used by GART set by
250 * first kernel
251 * or BIOS forget to put that in reserved.
252 * try to update e820 to make that region as reserved.
253 */
254 u32 agp_aper_order = 0;
255 int i, fix, slot, valid_agp = 0;
256 u32 ctl;
257 u32 aper_size = 0, aper_order = 0, last_aper_order = 0;
258 u64 aper_base = 0, last_aper_base = 0;
259 int aper_enabled = 0, last_aper_enabled = 0, last_valid = 0;
260
261 if (!early_pci_allowed())
262 return;
263
264 /* This is mostly duplicate of iommu_hole_init */
265 search_agp_bridge(&agp_aper_order, &valid_agp);
266
267 fix = 0;
268 for (i = 0; amd_nb_bus_dev_ranges[i].dev_limit; i++) {
269 int bus;
270 int dev_base, dev_limit;
271
272 bus = amd_nb_bus_dev_ranges[i].bus;
273 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
274 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
275
276 for (slot = dev_base; slot < dev_limit; slot++) {
277 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
278 continue;
279
280 ctl = read_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL);
281 aper_enabled = ctl & GARTEN;
282 aper_order = (ctl >> 1) & 7;
283 aper_size = (32 * 1024 * 1024) << aper_order;
284 aper_base = read_pci_config(bus, slot, 3, AMD64_GARTAPERTUREBASE) & 0x7fff;
285 aper_base <<= 25;
286
287 if (last_valid) {
288 if ((aper_order != last_aper_order) ||
289 (aper_base != last_aper_base) ||
290 (aper_enabled != last_aper_enabled)) {
291 fix = 1;
292 break;
293 }
294 }
295
296 last_aper_order = aper_order;
297 last_aper_base = aper_base;
298 last_aper_enabled = aper_enabled;
299 last_valid = 1;
300 }
301 }
302
303 if (!fix && !aper_enabled)
304 return;
305
306 if (!aper_base || !aper_size || aper_base + aper_size > 0x100000000UL)
307 fix = 1;
308
309 if (gart_fix_e820 && !fix && aper_enabled) {
310 if (e820_any_mapped(aper_base, aper_base + aper_size,
311 E820_RAM)) {
312 /* reserve it, so we can reuse it in second kernel */
313 printk(KERN_INFO "update e820 for GART\n");
314 e820_add_region(aper_base, aper_size, E820_RESERVED);
315 update_e820();
316 }
317 }
318
319 if (valid_agp)
320 return;
321
322 /* disable them all at first */
323 for (i = 0; i < amd_nb_bus_dev_ranges[i].dev_limit; i++) {
324 int bus;
325 int dev_base, dev_limit;
326
327 bus = amd_nb_bus_dev_ranges[i].bus;
328 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
329 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
330
331 for (slot = dev_base; slot < dev_limit; slot++) {
332 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
333 continue;
334
335 ctl = read_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL);
336 ctl &= ~GARTEN;
337 write_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL, ctl);
338 }
339 }
340
341}
342
343static int __initdata printed_gart_size_msg;
344
345int __init gart_iommu_hole_init(void)
346{
347 u32 agp_aper_base = 0, agp_aper_order = 0;
348 u32 aper_size, aper_alloc = 0, aper_order = 0, last_aper_order = 0;
349 u64 aper_base, last_aper_base = 0;
350 int fix, slot, valid_agp = 0;
351 int i, node;
352
353 if (gart_iommu_aperture_disabled || !fix_aperture ||
354 !early_pci_allowed())
355 return -ENODEV;
356
357 printk(KERN_INFO "Checking aperture...\n");
358
359 if (!fallback_aper_force)
360 agp_aper_base = search_agp_bridge(&agp_aper_order, &valid_agp);
361
362 fix = 0;
363 node = 0;
364 for (i = 0; i < amd_nb_bus_dev_ranges[i].dev_limit; i++) {
365 int bus;
366 int dev_base, dev_limit;
367 u32 ctl;
368
369 bus = amd_nb_bus_dev_ranges[i].bus;
370 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
371 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
372
373 for (slot = dev_base; slot < dev_limit; slot++) {
374 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
375 continue;
376
377 iommu_detected = 1;
378 gart_iommu_aperture = 1;
379 x86_init.iommu.iommu_init = gart_iommu_init;
380
381 ctl = read_pci_config(bus, slot, 3,
382 AMD64_GARTAPERTURECTL);
383
384 /*
385 * Before we do anything else disable the GART. It may
386 * still be enabled if we boot into a crash-kernel here.
387 * Reconfiguring the GART while it is enabled could have
388 * unknown side-effects.
389 */
390 ctl &= ~GARTEN;
391 write_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL, ctl);
392
393 aper_order = (ctl >> 1) & 7;
394 aper_size = (32 * 1024 * 1024) << aper_order;
395 aper_base = read_pci_config(bus, slot, 3, AMD64_GARTAPERTUREBASE) & 0x7fff;
396 aper_base <<= 25;
397
398 printk(KERN_INFO "Node %d: aperture @ %Lx size %u MB\n",
399 node, aper_base, aper_size >> 20);
400 node++;
401
402 if (!aperture_valid(aper_base, aper_size, 64<<20)) {
403 if (valid_agp && agp_aper_base &&
404 agp_aper_base == aper_base &&
405 agp_aper_order == aper_order) {
406 /* the same between two setting from NB and agp */
407 if (!no_iommu &&
408 max_pfn > MAX_DMA32_PFN &&
409 !printed_gart_size_msg) {
410 printk(KERN_ERR "you are using iommu with agp, but GART size is less than 64M\n");
411 printk(KERN_ERR "please increase GART size in your BIOS setup\n");
412 printk(KERN_ERR "if BIOS doesn't have that option, contact your HW vendor!\n");
413 printed_gart_size_msg = 1;
414 }
415 } else {
416 fix = 1;
417 goto out;
418 }
419 }
420
421 if ((last_aper_order && aper_order != last_aper_order) ||
422 (last_aper_base && aper_base != last_aper_base)) {
423 fix = 1;
424 goto out;
425 }
426 last_aper_order = aper_order;
427 last_aper_base = aper_base;
428 }
429 }
430
431out:
432 if (!fix && !fallback_aper_force) {
433 if (last_aper_base)
434 return 1;
435 return 0;
436 }
437
438 if (!fallback_aper_force) {
439 aper_alloc = agp_aper_base;
440 aper_order = agp_aper_order;
441 }
442
443 if (aper_alloc) {
444 /* Got the aperture from the AGP bridge */
445 } else if ((!no_iommu && max_pfn > MAX_DMA32_PFN) ||
446 force_iommu ||
447 valid_agp ||
448 fallback_aper_force) {
449 printk(KERN_INFO
450 "Your BIOS doesn't leave a aperture memory hole\n");
451 printk(KERN_INFO
452 "Please enable the IOMMU option in the BIOS setup\n");
453 printk(KERN_INFO
454 "This costs you %d MB of RAM\n",
455 32 << fallback_aper_order);
456
457 aper_order = fallback_aper_order;
458 aper_alloc = allocate_aperture();
459 if (!aper_alloc) {
460 /*
461 * Could disable AGP and IOMMU here, but it's
462 * probably not worth it. But the later users
463 * cannot deal with bad apertures and turning
464 * on the aperture over memory causes very
465 * strange problems, so it's better to panic
466 * early.
467 */
468 panic("Not enough memory for aperture");
469 }
470 } else {
471 return 0;
472 }
473
474 /* Fix up the north bridges */
475 for (i = 0; i < amd_nb_bus_dev_ranges[i].dev_limit; i++) {
476 int bus, dev_base, dev_limit;
477
478 /*
479 * Don't enable translation yet but enable GART IO and CPU
480 * accesses and set DISTLBWALKPRB since GART table memory is UC.
481 */
482 u32 ctl = aper_order << 1;
483
484 bus = amd_nb_bus_dev_ranges[i].bus;
485 dev_base = amd_nb_bus_dev_ranges[i].dev_base;
486 dev_limit = amd_nb_bus_dev_ranges[i].dev_limit;
487 for (slot = dev_base; slot < dev_limit; slot++) {
488 if (!early_is_amd_nb(read_pci_config(bus, slot, 3, 0x00)))
489 continue;
490
491 write_pci_config(bus, slot, 3, AMD64_GARTAPERTURECTL, ctl);
492 write_pci_config(bus, slot, 3, AMD64_GARTAPERTUREBASE, aper_alloc >> 25);
493 }
494 }
495
496 set_up_gart_resume(aper_order, aper_alloc);
497
498 return 1;
499}