1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Dynamic DMA mapping support for AMD Hammer.
4 *
5 * Use the integrated AGP GART in the Hammer northbridge as an IOMMU for PCI.
6 * This allows to use PCI devices that only support 32bit addresses on systems
7 * with more than 4GB.
8 *
9 * See Documentation/core-api/dma-api-howto.rst for the interface specification.
10 *
11 * Copyright 2002 Andi Kleen, SuSE Labs.
12 */
13
14 #include <linux/types.h>
15 #include <linux/ctype.h>
16 #include <linux/agp_backend.h>
17 #include <linux/init.h>
18 #include <linux/mm.h>
19 #include <linux/sched.h>
20 #include <linux/sched/debug.h>
21 #include <linux/string.h>
22 #include <linux/spinlock.h>
23 #include <linux/pci.h>
24 #include <linux/topology.h>
25 #include <linux/interrupt.h>
26 #include <linux/bitmap.h>
27 #include <linux/kdebug.h>
28 #include <linux/scatterlist.h>
29 #include <linux/iommu-helper.h>
30 #include <linux/syscore_ops.h>
31 #include <linux/io.h>
32 #include <linux/gfp.h>
33 #include <linux/atomic.h>
34 #include <linux/dma-direct.h>
35 #include <linux/dma-map-ops.h>
36 #include <asm/mtrr.h>
37 #include <asm/proto.h>
38 #include <asm/iommu.h>
39 #include <asm/gart.h>
40 #include <asm/set_memory.h>
41 #include <asm/swiotlb.h>
42 #include <asm/dma.h>
43 #include <asm/amd_nb.h>
44 #include <asm/x86_init.h>
45 #include <asm/iommu_table.h>
46
47 static unsigned long iommu_bus_base; /* GART remapping area (physical) */
48 static unsigned long iommu_size; /* size of remapping area bytes */
49 static unsigned long iommu_pages; /* .. and in pages */
50
51 static u32 *iommu_gatt_base; /* Remapping table */
52
53 /*
54 * If this is disabled the IOMMU will use an optimized flushing strategy
55 * of only flushing when an mapping is reused. With it true the GART is
56 * flushed for every mapping. Problem is that doing the lazy flush seems
57 * to trigger bugs with some popular PCI cards, in particular 3ware (but
58 * has been also also seen with Qlogic at least).
59 */
60 static int iommu_fullflush = 1;
61
62 /* Allocation bitmap for the remapping area: */
63 static DEFINE_SPINLOCK(iommu_bitmap_lock);
64 /* Guarded by iommu_bitmap_lock: */
65 static unsigned long *iommu_gart_bitmap;
66
67 static u32 gart_unmapped_entry;
68
69 #define GPTE_VALID 1
70 #define GPTE_COHERENT 2
71 #define GPTE_ENCODE(x) \
72 (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
73 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
74
75 #ifdef CONFIG_AGP
76 #define AGPEXTERN extern
77 #else
78 #define AGPEXTERN
79 #endif
80
81 /* GART can only remap to physical addresses < 1TB */
82 #define GART_MAX_PHYS_ADDR (1ULL << 40)
83
84 /* backdoor interface to AGP driver */
85 AGPEXTERN int agp_memory_reserved;
86 AGPEXTERN __u32 *agp_gatt_table;
87
88 static unsigned long next_bit; /* protected by iommu_bitmap_lock */
89 static bool need_flush; /* global flush state. set for each gart wrap */
90
alloc_iommu(struct device * dev,int size,unsigned long align_mask)91 static unsigned long alloc_iommu(struct device *dev, int size,
92 unsigned long align_mask)
93 {
94 unsigned long offset, flags;
95 unsigned long boundary_size;
96 unsigned long base_index;
97
98 base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
99 PAGE_SIZE) >> PAGE_SHIFT;
100 boundary_size = dma_get_seg_boundary_nr_pages(dev, PAGE_SHIFT);
101
102 spin_lock_irqsave(&iommu_bitmap_lock, flags);
103 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
104 size, base_index, boundary_size, align_mask);
105 if (offset == -1) {
106 need_flush = true;
107 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
108 size, base_index, boundary_size,
109 align_mask);
110 }
111 if (offset != -1) {
112 next_bit = offset+size;
113 if (next_bit >= iommu_pages) {
114 next_bit = 0;
115 need_flush = true;
116 }
117 }
118 if (iommu_fullflush)
119 need_flush = true;
120 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
121
122 return offset;
123 }
124
free_iommu(unsigned long offset,int size)125 static void free_iommu(unsigned long offset, int size)
126 {
127 unsigned long flags;
128
129 spin_lock_irqsave(&iommu_bitmap_lock, flags);
130 bitmap_clear(iommu_gart_bitmap, offset, size);
131 if (offset >= next_bit)
132 next_bit = offset + size;
133 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
134 }
135
136 /*
137 * Use global flush state to avoid races with multiple flushers.
138 */
flush_gart(void)139 static void flush_gart(void)
140 {
141 unsigned long flags;
142
143 spin_lock_irqsave(&iommu_bitmap_lock, flags);
144 if (need_flush) {
145 amd_flush_garts();
146 need_flush = false;
147 }
148 spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
149 }
150
151 #ifdef CONFIG_IOMMU_LEAK
152 /* Debugging aid for drivers that don't free their IOMMU tables */
dump_leak(void)153 static void dump_leak(void)
154 {
155 static int dump;
156
157 if (dump)
158 return;
159 dump = 1;
160
161 show_stack(NULL, NULL, KERN_ERR);
162 debug_dma_dump_mappings(NULL);
163 }
164 #endif
165
iommu_full(struct device * dev,size_t size,int dir)166 static void iommu_full(struct device *dev, size_t size, int dir)
167 {
168 /*
169 * Ran out of IOMMU space for this operation. This is very bad.
170 * Unfortunately the drivers cannot handle this operation properly.
171 * Return some non mapped prereserved space in the aperture and
172 * let the Northbridge deal with it. This will result in garbage
173 * in the IO operation. When the size exceeds the prereserved space
174 * memory corruption will occur or random memory will be DMAed
175 * out. Hopefully no network devices use single mappings that big.
176 */
177
178 dev_err(dev, "PCI-DMA: Out of IOMMU space for %lu bytes\n", size);
179 #ifdef CONFIG_IOMMU_LEAK
180 dump_leak();
181 #endif
182 }
183
184 static inline int
need_iommu(struct device * dev,unsigned long addr,size_t size)185 need_iommu(struct device *dev, unsigned long addr, size_t size)
186 {
187 return force_iommu || !dma_capable(dev, addr, size, true);
188 }
189
190 static inline int
nonforced_iommu(struct device * dev,unsigned long addr,size_t size)191 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
192 {
193 return !dma_capable(dev, addr, size, true);
194 }
195
196 /* Map a single continuous physical area into the IOMMU.
197 * Caller needs to check if the iommu is needed and flush.
198 */
dma_map_area(struct device * dev,dma_addr_t phys_mem,size_t size,int dir,unsigned long align_mask)199 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
200 size_t size, int dir, unsigned long align_mask)
201 {
202 unsigned long npages = iommu_num_pages(phys_mem, size, PAGE_SIZE);
203 unsigned long iommu_page;
204 int i;
205
206 if (unlikely(phys_mem + size > GART_MAX_PHYS_ADDR))
207 return DMA_MAPPING_ERROR;
208
209 iommu_page = alloc_iommu(dev, npages, align_mask);
210 if (iommu_page == -1) {
211 if (!nonforced_iommu(dev, phys_mem, size))
212 return phys_mem;
213 if (panic_on_overflow)
214 panic("dma_map_area overflow %lu bytes\n", size);
215 iommu_full(dev, size, dir);
216 return DMA_MAPPING_ERROR;
217 }
218
219 for (i = 0; i < npages; i++) {
220 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
221 phys_mem += PAGE_SIZE;
222 }
223 return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
224 }
225
226 /* Map a single area into the IOMMU */
gart_map_page(struct device * dev,struct page * page,unsigned long offset,size_t size,enum dma_data_direction dir,unsigned long attrs)227 static dma_addr_t gart_map_page(struct device *dev, struct page *page,
228 unsigned long offset, size_t size,
229 enum dma_data_direction dir,
230 unsigned long attrs)
231 {
232 unsigned long bus;
233 phys_addr_t paddr = page_to_phys(page) + offset;
234
235 if (!need_iommu(dev, paddr, size))
236 return paddr;
237
238 bus = dma_map_area(dev, paddr, size, dir, 0);
239 flush_gart();
240
241 return bus;
242 }
243
244 /*
245 * Free a DMA mapping.
246 */
gart_unmap_page(struct device * dev,dma_addr_t dma_addr,size_t size,enum dma_data_direction dir,unsigned long attrs)247 static void gart_unmap_page(struct device *dev, dma_addr_t dma_addr,
248 size_t size, enum dma_data_direction dir,
249 unsigned long attrs)
250 {
251 unsigned long iommu_page;
252 int npages;
253 int i;
254
255 if (WARN_ON_ONCE(dma_addr == DMA_MAPPING_ERROR))
256 return;
257
258 /*
259 * This driver will not always use a GART mapping, but might have
260 * created a direct mapping instead. If that is the case there is
261 * nothing to unmap here.
262 */
263 if (dma_addr < iommu_bus_base ||
264 dma_addr >= iommu_bus_base + iommu_size)
265 return;
266
267 iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
268 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
269 for (i = 0; i < npages; i++) {
270 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
271 }
272 free_iommu(iommu_page, npages);
273 }
274
275 /*
276 * Wrapper for pci_unmap_single working with scatterlists.
277 */
gart_unmap_sg(struct device * dev,struct scatterlist * sg,int nents,enum dma_data_direction dir,unsigned long attrs)278 static void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
279 enum dma_data_direction dir, unsigned long attrs)
280 {
281 struct scatterlist *s;
282 int i;
283
284 for_each_sg(sg, s, nents, i) {
285 if (!s->dma_length || !s->length)
286 break;
287 gart_unmap_page(dev, s->dma_address, s->dma_length, dir, 0);
288 }
289 }
290
291 /* Fallback for dma_map_sg in case of overflow */
dma_map_sg_nonforce(struct device * dev,struct scatterlist * sg,int nents,int dir)292 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
293 int nents, int dir)
294 {
295 struct scatterlist *s;
296 int i;
297
298 #ifdef CONFIG_IOMMU_DEBUG
299 pr_debug("dma_map_sg overflow\n");
300 #endif
301
302 for_each_sg(sg, s, nents, i) {
303 unsigned long addr = sg_phys(s);
304
305 if (nonforced_iommu(dev, addr, s->length)) {
306 addr = dma_map_area(dev, addr, s->length, dir, 0);
307 if (addr == DMA_MAPPING_ERROR) {
308 if (i > 0)
309 gart_unmap_sg(dev, sg, i, dir, 0);
310 nents = 0;
311 sg[0].dma_length = 0;
312 break;
313 }
314 }
315 s->dma_address = addr;
316 s->dma_length = s->length;
317 }
318 flush_gart();
319
320 return nents;
321 }
322
323 /* Map multiple scatterlist entries continuous into the first. */
__dma_map_cont(struct device * dev,struct scatterlist * start,int nelems,struct scatterlist * sout,unsigned long pages)324 static int __dma_map_cont(struct device *dev, struct scatterlist *start,
325 int nelems, struct scatterlist *sout,
326 unsigned long pages)
327 {
328 unsigned long iommu_start = alloc_iommu(dev, pages, 0);
329 unsigned long iommu_page = iommu_start;
330 struct scatterlist *s;
331 int i;
332
333 if (iommu_start == -1)
334 return -1;
335
336 for_each_sg(start, s, nelems, i) {
337 unsigned long pages, addr;
338 unsigned long phys_addr = s->dma_address;
339
340 BUG_ON(s != start && s->offset);
341 if (s == start) {
342 sout->dma_address = iommu_bus_base;
343 sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
344 sout->dma_length = s->length;
345 } else {
346 sout->dma_length += s->length;
347 }
348
349 addr = phys_addr;
350 pages = iommu_num_pages(s->offset, s->length, PAGE_SIZE);
351 while (pages--) {
352 iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
353 addr += PAGE_SIZE;
354 iommu_page++;
355 }
356 }
357 BUG_ON(iommu_page - iommu_start != pages);
358
359 return 0;
360 }
361
362 static inline int
dma_map_cont(struct device * dev,struct scatterlist * start,int nelems,struct scatterlist * sout,unsigned long pages,int need)363 dma_map_cont(struct device *dev, struct scatterlist *start, int nelems,
364 struct scatterlist *sout, unsigned long pages, int need)
365 {
366 if (!need) {
367 BUG_ON(nelems != 1);
368 sout->dma_address = start->dma_address;
369 sout->dma_length = start->length;
370 return 0;
371 }
372 return __dma_map_cont(dev, start, nelems, sout, pages);
373 }
374
375 /*
376 * DMA map all entries in a scatterlist.
377 * Merge chunks that have page aligned sizes into a continuous mapping.
378 */
gart_map_sg(struct device * dev,struct scatterlist * sg,int nents,enum dma_data_direction dir,unsigned long attrs)379 static int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents,
380 enum dma_data_direction dir, unsigned long attrs)
381 {
382 struct scatterlist *s, *ps, *start_sg, *sgmap;
383 int need = 0, nextneed, i, out, start;
384 unsigned long pages = 0;
385 unsigned int seg_size;
386 unsigned int max_seg_size;
387
388 if (nents == 0)
389 return 0;
390
391 out = 0;
392 start = 0;
393 start_sg = sg;
394 sgmap = sg;
395 seg_size = 0;
396 max_seg_size = dma_get_max_seg_size(dev);
397 ps = NULL; /* shut up gcc */
398
399 for_each_sg(sg, s, nents, i) {
400 dma_addr_t addr = sg_phys(s);
401
402 s->dma_address = addr;
403 BUG_ON(s->length == 0);
404
405 nextneed = need_iommu(dev, addr, s->length);
406
407 /* Handle the previous not yet processed entries */
408 if (i > start) {
409 /*
410 * Can only merge when the last chunk ends on a
411 * page boundary and the new one doesn't have an
412 * offset.
413 */
414 if (!iommu_merge || !nextneed || !need || s->offset ||
415 (s->length + seg_size > max_seg_size) ||
416 (ps->offset + ps->length) % PAGE_SIZE) {
417 if (dma_map_cont(dev, start_sg, i - start,
418 sgmap, pages, need) < 0)
419 goto error;
420 out++;
421
422 seg_size = 0;
423 sgmap = sg_next(sgmap);
424 pages = 0;
425 start = i;
426 start_sg = s;
427 }
428 }
429
430 seg_size += s->length;
431 need = nextneed;
432 pages += iommu_num_pages(s->offset, s->length, PAGE_SIZE);
433 ps = s;
434 }
435 if (dma_map_cont(dev, start_sg, i - start, sgmap, pages, need) < 0)
436 goto error;
437 out++;
438 flush_gart();
439 if (out < nents) {
440 sgmap = sg_next(sgmap);
441 sgmap->dma_length = 0;
442 }
443 return out;
444
445 error:
446 flush_gart();
447 gart_unmap_sg(dev, sg, out, dir, 0);
448
449 /* When it was forced or merged try again in a dumb way */
450 if (force_iommu || iommu_merge) {
451 out = dma_map_sg_nonforce(dev, sg, nents, dir);
452 if (out > 0)
453 return out;
454 }
455 if (panic_on_overflow)
456 panic("dma_map_sg: overflow on %lu pages\n", pages);
457
458 iommu_full(dev, pages << PAGE_SHIFT, dir);
459 for_each_sg(sg, s, nents, i)
460 s->dma_address = DMA_MAPPING_ERROR;
461 return 0;
462 }
463
464 /* allocate and map a coherent mapping */
465 static void *
gart_alloc_coherent(struct device * dev,size_t size,dma_addr_t * dma_addr,gfp_t flag,unsigned long attrs)466 gart_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_addr,
467 gfp_t flag, unsigned long attrs)
468 {
469 void *vaddr;
470
471 vaddr = dma_direct_alloc(dev, size, dma_addr, flag, attrs);
472 if (!vaddr ||
473 !force_iommu || dev->coherent_dma_mask <= DMA_BIT_MASK(24))
474 return vaddr;
475
476 *dma_addr = dma_map_area(dev, virt_to_phys(vaddr), size,
477 DMA_BIDIRECTIONAL, (1UL << get_order(size)) - 1);
478 flush_gart();
479 if (unlikely(*dma_addr == DMA_MAPPING_ERROR))
480 goto out_free;
481 return vaddr;
482 out_free:
483 dma_direct_free(dev, size, vaddr, *dma_addr, attrs);
484 return NULL;
485 }
486
487 /* free a coherent mapping */
488 static void
gart_free_coherent(struct device * dev,size_t size,void * vaddr,dma_addr_t dma_addr,unsigned long attrs)489 gart_free_coherent(struct device *dev, size_t size, void *vaddr,
490 dma_addr_t dma_addr, unsigned long attrs)
491 {
492 gart_unmap_page(dev, dma_addr, size, DMA_BIDIRECTIONAL, 0);
493 dma_direct_free(dev, size, vaddr, dma_addr, attrs);
494 }
495
496 static int no_agp;
497
check_iommu_size(unsigned long aper,u64 aper_size)498 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
499 {
500 unsigned long a;
501
502 if (!iommu_size) {
503 iommu_size = aper_size;
504 if (!no_agp)
505 iommu_size /= 2;
506 }
507
508 a = aper + iommu_size;
509 iommu_size -= round_up(a, PMD_PAGE_SIZE) - a;
510
511 if (iommu_size < 64*1024*1024) {
512 pr_warn("PCI-DMA: Warning: Small IOMMU %luMB."
513 " Consider increasing the AGP aperture in BIOS\n",
514 iommu_size >> 20);
515 }
516
517 return iommu_size;
518 }
519
read_aperture(struct pci_dev * dev,u32 * size)520 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
521 {
522 unsigned aper_size = 0, aper_base_32, aper_order;
523 u64 aper_base;
524
525 pci_read_config_dword(dev, AMD64_GARTAPERTUREBASE, &aper_base_32);
526 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &aper_order);
527 aper_order = (aper_order >> 1) & 7;
528
529 aper_base = aper_base_32 & 0x7fff;
530 aper_base <<= 25;
531
532 aper_size = (32 * 1024 * 1024) << aper_order;
533 if (aper_base + aper_size > 0x100000000UL || !aper_size)
534 aper_base = 0;
535
536 *size = aper_size;
537 return aper_base;
538 }
539
enable_gart_translations(void)540 static void enable_gart_translations(void)
541 {
542 int i;
543
544 if (!amd_nb_has_feature(AMD_NB_GART))
545 return;
546
547 for (i = 0; i < amd_nb_num(); i++) {
548 struct pci_dev *dev = node_to_amd_nb(i)->misc;
549
550 enable_gart_translation(dev, __pa(agp_gatt_table));
551 }
552
553 /* Flush the GART-TLB to remove stale entries */
554 amd_flush_garts();
555 }
556
557 /*
558 * If fix_up_north_bridges is set, the north bridges have to be fixed up on
559 * resume in the same way as they are handled in gart_iommu_hole_init().
560 */
561 static bool fix_up_north_bridges;
562 static u32 aperture_order;
563 static u32 aperture_alloc;
564
set_up_gart_resume(u32 aper_order,u32 aper_alloc)565 void set_up_gart_resume(u32 aper_order, u32 aper_alloc)
566 {
567 fix_up_north_bridges = true;
568 aperture_order = aper_order;
569 aperture_alloc = aper_alloc;
570 }
571
gart_fixup_northbridges(void)572 static void gart_fixup_northbridges(void)
573 {
574 int i;
575
576 if (!fix_up_north_bridges)
577 return;
578
579 if (!amd_nb_has_feature(AMD_NB_GART))
580 return;
581
582 pr_info("PCI-DMA: Restoring GART aperture settings\n");
583
584 for (i = 0; i < amd_nb_num(); i++) {
585 struct pci_dev *dev = node_to_amd_nb(i)->misc;
586
587 /*
588 * Don't enable translations just yet. That is the next
589 * step. Restore the pre-suspend aperture settings.
590 */
591 gart_set_size_and_enable(dev, aperture_order);
592 pci_write_config_dword(dev, AMD64_GARTAPERTUREBASE, aperture_alloc >> 25);
593 }
594 }
595
gart_resume(void)596 static void gart_resume(void)
597 {
598 pr_info("PCI-DMA: Resuming GART IOMMU\n");
599
600 gart_fixup_northbridges();
601
602 enable_gart_translations();
603 }
604
605 static struct syscore_ops gart_syscore_ops = {
606 .resume = gart_resume,
607
608 };
609
610 /*
611 * Private Northbridge GATT initialization in case we cannot use the
612 * AGP driver for some reason.
613 */
init_amd_gatt(struct agp_kern_info * info)614 static __init int init_amd_gatt(struct agp_kern_info *info)
615 {
616 unsigned aper_size, gatt_size, new_aper_size;
617 unsigned aper_base, new_aper_base;
618 struct pci_dev *dev;
619 void *gatt;
620 int i;
621
622 pr_info("PCI-DMA: Disabling AGP.\n");
623
624 aper_size = aper_base = info->aper_size = 0;
625 dev = NULL;
626 for (i = 0; i < amd_nb_num(); i++) {
627 dev = node_to_amd_nb(i)->misc;
628 new_aper_base = read_aperture(dev, &new_aper_size);
629 if (!new_aper_base)
630 goto nommu;
631
632 if (!aper_base) {
633 aper_size = new_aper_size;
634 aper_base = new_aper_base;
635 }
636 if (aper_size != new_aper_size || aper_base != new_aper_base)
637 goto nommu;
638 }
639 if (!aper_base)
640 goto nommu;
641
642 info->aper_base = aper_base;
643 info->aper_size = aper_size >> 20;
644
645 gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
646 gatt = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
647 get_order(gatt_size));
648 if (!gatt)
649 panic("Cannot allocate GATT table");
650 if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
651 panic("Could not set GART PTEs to uncacheable pages");
652
653 agp_gatt_table = gatt;
654
655 register_syscore_ops(&gart_syscore_ops);
656
657 flush_gart();
658
659 pr_info("PCI-DMA: aperture base @ %x size %u KB\n",
660 aper_base, aper_size>>10);
661
662 return 0;
663
664 nommu:
665 /* Should not happen anymore */
666 pr_warn("PCI-DMA: More than 4GB of RAM and no IOMMU - falling back to iommu=soft.\n");
667 return -1;
668 }
669
670 static const struct dma_map_ops gart_dma_ops = {
671 .map_sg = gart_map_sg,
672 .unmap_sg = gart_unmap_sg,
673 .map_page = gart_map_page,
674 .unmap_page = gart_unmap_page,
675 .alloc = gart_alloc_coherent,
676 .free = gart_free_coherent,
677 .mmap = dma_common_mmap,
678 .get_sgtable = dma_common_get_sgtable,
679 .dma_supported = dma_direct_supported,
680 .get_required_mask = dma_direct_get_required_mask,
681 .alloc_pages = dma_direct_alloc_pages,
682 .free_pages = dma_direct_free_pages,
683 };
684
gart_iommu_shutdown(void)685 static void gart_iommu_shutdown(void)
686 {
687 struct pci_dev *dev;
688 int i;
689
690 /* don't shutdown it if there is AGP installed */
691 if (!no_agp)
692 return;
693
694 if (!amd_nb_has_feature(AMD_NB_GART))
695 return;
696
697 for (i = 0; i < amd_nb_num(); i++) {
698 u32 ctl;
699
700 dev = node_to_amd_nb(i)->misc;
701 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &ctl);
702
703 ctl &= ~GARTEN;
704
705 pci_write_config_dword(dev, AMD64_GARTAPERTURECTL, ctl);
706 }
707 }
708
gart_iommu_init(void)709 int __init gart_iommu_init(void)
710 {
711 struct agp_kern_info info;
712 unsigned long iommu_start;
713 unsigned long aper_base, aper_size;
714 unsigned long start_pfn, end_pfn;
715 unsigned long scratch;
716
717 if (!amd_nb_has_feature(AMD_NB_GART))
718 return 0;
719
720 #ifndef CONFIG_AGP_AMD64
721 no_agp = 1;
722 #else
723 /* Makefile puts PCI initialization via subsys_initcall first. */
724 /* Add other AMD AGP bridge drivers here */
725 no_agp = no_agp ||
726 (agp_amd64_init() < 0) ||
727 (agp_copy_info(agp_bridge, &info) < 0);
728 #endif
729
730 if (no_iommu ||
731 (!force_iommu && max_pfn <= MAX_DMA32_PFN) ||
732 !gart_iommu_aperture ||
733 (no_agp && init_amd_gatt(&info) < 0)) {
734 if (max_pfn > MAX_DMA32_PFN) {
735 pr_warn("More than 4GB of memory but GART IOMMU not available.\n");
736 pr_warn("falling back to iommu=soft.\n");
737 }
738 return 0;
739 }
740
741 /* need to map that range */
742 aper_size = info.aper_size << 20;
743 aper_base = info.aper_base;
744 end_pfn = (aper_base>>PAGE_SHIFT) + (aper_size>>PAGE_SHIFT);
745
746 start_pfn = PFN_DOWN(aper_base);
747 if (!pfn_range_is_mapped(start_pfn, end_pfn))
748 init_memory_mapping(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT,
749 PAGE_KERNEL);
750
751 pr_info("PCI-DMA: using GART IOMMU.\n");
752 iommu_size = check_iommu_size(info.aper_base, aper_size);
753 iommu_pages = iommu_size >> PAGE_SHIFT;
754
755 iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO,
756 get_order(iommu_pages/8));
757 if (!iommu_gart_bitmap)
758 panic("Cannot allocate iommu bitmap\n");
759
760 pr_info("PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
761 iommu_size >> 20);
762
763 agp_memory_reserved = iommu_size;
764 iommu_start = aper_size - iommu_size;
765 iommu_bus_base = info.aper_base + iommu_start;
766 iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
767
768 /*
769 * Unmap the IOMMU part of the GART. The alias of the page is
770 * always mapped with cache enabled and there is no full cache
771 * coherency across the GART remapping. The unmapping avoids
772 * automatic prefetches from the CPU allocating cache lines in
773 * there. All CPU accesses are done via the direct mapping to
774 * the backing memory. The GART address is only used by PCI
775 * devices.
776 */
777 set_memory_np((unsigned long)__va(iommu_bus_base),
778 iommu_size >> PAGE_SHIFT);
779 /*
780 * Tricky. The GART table remaps the physical memory range,
781 * so the CPU wont notice potential aliases and if the memory
782 * is remapped to UC later on, we might surprise the PCI devices
783 * with a stray writeout of a cacheline. So play it sure and
784 * do an explicit, full-scale wbinvd() _after_ having marked all
785 * the pages as Not-Present:
786 */
787 wbinvd();
788
789 /*
790 * Now all caches are flushed and we can safely enable
791 * GART hardware. Doing it early leaves the possibility
792 * of stale cache entries that can lead to GART PTE
793 * errors.
794 */
795 enable_gart_translations();
796
797 /*
798 * Try to workaround a bug (thanks to BenH):
799 * Set unmapped entries to a scratch page instead of 0.
800 * Any prefetches that hit unmapped entries won't get an bus abort
801 * then. (P2P bridge may be prefetching on DMA reads).
802 */
803 scratch = get_zeroed_page(GFP_KERNEL);
804 if (!scratch)
805 panic("Cannot allocate iommu scratch page");
806 gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
807
808 flush_gart();
809 dma_ops = &gart_dma_ops;
810 x86_platform.iommu_shutdown = gart_iommu_shutdown;
811 swiotlb = 0;
812
813 return 0;
814 }
815
gart_parse_options(char * p)816 void __init gart_parse_options(char *p)
817 {
818 int arg;
819
820 if (isdigit(*p) && get_option(&p, &arg))
821 iommu_size = arg;
822 if (!strncmp(p, "fullflush", 9))
823 iommu_fullflush = 1;
824 if (!strncmp(p, "nofullflush", 11))
825 iommu_fullflush = 0;
826 if (!strncmp(p, "noagp", 5))
827 no_agp = 1;
828 if (!strncmp(p, "noaperture", 10))
829 fix_aperture = 0;
830 /* duplicated from pci-dma.c */
831 if (!strncmp(p, "force", 5))
832 gart_iommu_aperture_allowed = 1;
833 if (!strncmp(p, "allowed", 7))
834 gart_iommu_aperture_allowed = 1;
835 if (!strncmp(p, "memaper", 7)) {
836 fallback_aper_force = 1;
837 p += 7;
838 if (*p == '=') {
839 ++p;
840 if (get_option(&p, &arg))
841 fallback_aper_order = arg;
842 }
843 }
844 }
845 IOMMU_INIT_POST(gart_iommu_hole_init);
846