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