blob: c984768d98ca95d430b88d4af537183600ce4c53 [file] [log] [blame]
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -04001/*
2 * Copyright 2010
3 * by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
4 *
5 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License v2.0 as published by
9 * the Free Software Foundation
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * PV guests under Xen are running in an non-contiguous memory architecture.
17 *
18 * When PCI pass-through is utilized, this necessitates an IOMMU for
19 * translating bus (DMA) to virtual and vice-versa and also providing a
20 * mechanism to have contiguous pages for device drivers operations (say DMA
21 * operations).
22 *
23 * Specifically, under Xen the Linux idea of pages is an illusion. It
24 * assumes that pages start at zero and go up to the available memory. To
25 * help with that, the Linux Xen MMU provides a lookup mechanism to
26 * translate the page frame numbers (PFN) to machine frame numbers (MFN)
27 * and vice-versa. The MFN are the "real" frame numbers. Furthermore
28 * memory is not contiguous. Xen hypervisor stitches memory for guests
29 * from different pools, which means there is no guarantee that PFN==MFN
30 * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
31 * allocated in descending order (high to low), meaning the guest might
32 * never get any MFN's under the 4GB mark.
33 *
34 */
35
36#include <linux/bootmem.h>
37#include <linux/dma-mapping.h>
38#include <xen/swiotlb-xen.h>
39#include <xen/page.h>
40#include <xen/xen-ops.h>
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -040041#include <xen/hvc-console.h>
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040042/*
43 * Used to do a quick range check in swiotlb_tbl_unmap_single and
44 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
45 * API.
46 */
47
48static char *xen_io_tlb_start, *xen_io_tlb_end;
49static unsigned long xen_io_tlb_nslabs;
50/*
51 * Quick lookup value of the bus address of the IOTLB.
52 */
53
54u64 start_dma_addr;
55
56static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
57{
Justin P. Mattock6eab04a2011-04-08 19:49:08 -070058 return phys_to_machine(XPADDR(paddr)).maddr;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040059}
60
61static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
62{
63 return machine_to_phys(XMADDR(baddr)).paddr;
64}
65
66static dma_addr_t xen_virt_to_bus(void *address)
67{
68 return xen_phys_to_bus(virt_to_phys(address));
69}
70
71static int check_pages_physically_contiguous(unsigned long pfn,
72 unsigned int offset,
73 size_t length)
74{
75 unsigned long next_mfn;
76 int i;
77 int nr_pages;
78
79 next_mfn = pfn_to_mfn(pfn);
80 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
81
82 for (i = 1; i < nr_pages; i++) {
83 if (pfn_to_mfn(++pfn) != ++next_mfn)
84 return 0;
85 }
86 return 1;
87}
88
89static int range_straddles_page_boundary(phys_addr_t p, size_t size)
90{
91 unsigned long pfn = PFN_DOWN(p);
92 unsigned int offset = p & ~PAGE_MASK;
93
94 if (offset + size <= PAGE_SIZE)
95 return 0;
96 if (check_pages_physically_contiguous(pfn, offset, size))
97 return 0;
98 return 1;
99}
100
101static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
102{
103 unsigned long mfn = PFN_DOWN(dma_addr);
104 unsigned long pfn = mfn_to_local_pfn(mfn);
105 phys_addr_t paddr;
106
107 /* If the address is outside our domain, it CAN
108 * have the same virtual address as another address
109 * in our domain. Therefore _only_ check address within our domain.
110 */
111 if (pfn_valid(pfn)) {
112 paddr = PFN_PHYS(pfn);
113 return paddr >= virt_to_phys(xen_io_tlb_start) &&
114 paddr < virt_to_phys(xen_io_tlb_end);
115 }
116 return 0;
117}
118
119static int max_dma_bits = 32;
120
121static int
122xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
123{
124 int i, rc;
125 int dma_bits;
126
127 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
128
129 i = 0;
130 do {
131 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
132
133 do {
134 rc = xen_create_contiguous_region(
135 (unsigned long)buf + (i << IO_TLB_SHIFT),
136 get_order(slabs << IO_TLB_SHIFT),
137 dma_bits);
138 } while (rc && dma_bits++ < max_dma_bits);
139 if (rc)
140 return rc;
141
142 i += slabs;
143 } while (i < nslabs);
144 return 0;
145}
146
147void __init xen_swiotlb_init(int verbose)
148{
149 unsigned long bytes;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400150 int rc = -ENOMEM;
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900151 unsigned long nr_tbl;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400152 char *m = NULL;
153 unsigned int repeat = 3;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400154
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900155 nr_tbl = swioltb_nr_tbl();
156 if (nr_tbl)
157 xen_io_tlb_nslabs = nr_tbl;
158 else {
159 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
160 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
161 }
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400162retry:
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400163 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
164
165 /*
166 * Get IO TLB memory from any location.
167 */
168 xen_io_tlb_start = alloc_bootmem(bytes);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400169 if (!xen_io_tlb_start) {
170 m = "Cannot allocate Xen-SWIOTLB buffer!\n";
171 goto error;
172 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400173 xen_io_tlb_end = xen_io_tlb_start + bytes;
174 /*
175 * And replace that memory with pages under 4GB.
176 */
177 rc = xen_swiotlb_fixup(xen_io_tlb_start,
178 bytes,
179 xen_io_tlb_nslabs);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400180 if (rc) {
181 free_bootmem(__pa(xen_io_tlb_start), bytes);
182 m = "Failed to get contiguous memory for DMA from Xen!\n"\
183 "You either: don't have the permissions, do not have"\
184 " enough free memory under 4GB, or the hypervisor memory"\
185 "is too fragmented!";
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400186 goto error;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400187 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400188 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
189 swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
190
191 return;
192error:
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400193 if (repeat--) {
194 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
195 (xen_io_tlb_nslabs >> 1));
196 printk(KERN_INFO "Xen-SWIOTLB: Lowering to %luMB\n",
197 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
198 goto retry;
199 }
Randy Dunlap61ca7982011-08-11 13:57:07 -0700200 xen_raw_printk("%s (rc:%d)", m, rc);
201 panic("%s (rc:%d)", m, rc);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400202}
203
204void *
205xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
206 dma_addr_t *dma_handle, gfp_t flags)
207{
208 void *ret;
209 int order = get_order(size);
210 u64 dma_mask = DMA_BIT_MASK(32);
211 unsigned long vstart;
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400212 phys_addr_t phys;
213 dma_addr_t dev_addr;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400214
215 /*
216 * Ignore region specifiers - the kernel's ideas of
217 * pseudo-phys memory layout has nothing to do with the
218 * machine physical layout. We can't allocate highmem
219 * because we can't return a pointer to it.
220 */
221 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
222
223 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
224 return ret;
225
226 vstart = __get_free_pages(flags, order);
227 ret = (void *)vstart;
228
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400229 if (!ret)
230 return ret;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400231
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400232 if (hwdev && hwdev->coherent_dma_mask)
233 dma_mask = hwdev->coherent_dma_mask;
234
235 phys = virt_to_phys(ret);
236 dev_addr = xen_phys_to_bus(phys);
237 if (((dev_addr + size - 1 <= dma_mask)) &&
238 !range_straddles_page_boundary(phys, size))
239 *dma_handle = dev_addr;
240 else {
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400241 if (xen_create_contiguous_region(vstart, order,
242 fls64(dma_mask)) != 0) {
243 free_pages(vstart, order);
244 return NULL;
245 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400246 *dma_handle = virt_to_machine(ret).maddr;
247 }
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400248 memset(ret, 0, size);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400249 return ret;
250}
251EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
252
253void
254xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
255 dma_addr_t dev_addr)
256{
257 int order = get_order(size);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400258 phys_addr_t phys;
259 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400260
261 if (dma_release_from_coherent(hwdev, order, vaddr))
262 return;
263
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400264 if (hwdev && hwdev->coherent_dma_mask)
265 dma_mask = hwdev->coherent_dma_mask;
266
267 phys = virt_to_phys(vaddr);
268
269 if (((dev_addr + size - 1 > dma_mask)) ||
270 range_straddles_page_boundary(phys, size))
271 xen_destroy_contiguous_region((unsigned long)vaddr, order);
272
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400273 free_pages((unsigned long)vaddr, order);
274}
275EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
276
277
278/*
279 * Map a single buffer of the indicated size for DMA in streaming mode. The
280 * physical address to use is returned.
281 *
282 * Once the device is given the dma address, the device owns this memory until
283 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
284 */
285dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
286 unsigned long offset, size_t size,
287 enum dma_data_direction dir,
288 struct dma_attrs *attrs)
289{
290 phys_addr_t phys = page_to_phys(page) + offset;
291 dma_addr_t dev_addr = xen_phys_to_bus(phys);
292 void *map;
293
294 BUG_ON(dir == DMA_NONE);
295 /*
296 * If the address happens to be in the device's DMA window,
297 * we can safely return the device addr and not worry about bounce
298 * buffering it.
299 */
300 if (dma_capable(dev, dev_addr, size) &&
301 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
302 return dev_addr;
303
304 /*
305 * Oh well, have to allocate and map a bounce buffer.
306 */
307 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
308 if (!map)
309 return DMA_ERROR_CODE;
310
311 dev_addr = xen_virt_to_bus(map);
312
313 /*
314 * Ensure that the address returned is DMA'ble
315 */
Konrad Rzeszutek Wilkab2a47b2011-07-22 12:51:48 -0400316 if (!dma_capable(dev, dev_addr, size)) {
317 swiotlb_tbl_unmap_single(dev, map, size, dir);
318 dev_addr = 0;
319 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400320 return dev_addr;
321}
322EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
323
324/*
325 * Unmap a single streaming mode DMA translation. The dma_addr and size must
326 * match what was provided for in a previous xen_swiotlb_map_page call. All
327 * other usages are undefined.
328 *
329 * After this call, reads by the cpu to the buffer are guaranteed to see
330 * whatever the device wrote there.
331 */
332static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
333 size_t size, enum dma_data_direction dir)
334{
335 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
336
337 BUG_ON(dir == DMA_NONE);
338
339 /* NOTE: We use dev_addr here, not paddr! */
340 if (is_xen_swiotlb_buffer(dev_addr)) {
341 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
342 return;
343 }
344
345 if (dir != DMA_FROM_DEVICE)
346 return;
347
348 /*
349 * phys_to_virt doesn't work with hihgmem page but we could
350 * call dma_mark_clean() with hihgmem page here. However, we
351 * are fine since dma_mark_clean() is null on POWERPC. We can
352 * make dma_mark_clean() take a physical address if necessary.
353 */
354 dma_mark_clean(phys_to_virt(paddr), size);
355}
356
357void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
358 size_t size, enum dma_data_direction dir,
359 struct dma_attrs *attrs)
360{
361 xen_unmap_single(hwdev, dev_addr, size, dir);
362}
363EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
364
365/*
366 * Make physical memory consistent for a single streaming mode DMA translation
367 * after a transfer.
368 *
369 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
370 * using the cpu, yet do not wish to teardown the dma mapping, you must
371 * call this function before doing so. At the next point you give the dma
372 * address back to the card, you must first perform a
373 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
374 */
375static void
376xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
377 size_t size, enum dma_data_direction dir,
378 enum dma_sync_target target)
379{
380 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
381
382 BUG_ON(dir == DMA_NONE);
383
384 /* NOTE: We use dev_addr here, not paddr! */
385 if (is_xen_swiotlb_buffer(dev_addr)) {
386 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
387 target);
388 return;
389 }
390
391 if (dir != DMA_FROM_DEVICE)
392 return;
393
394 dma_mark_clean(phys_to_virt(paddr), size);
395}
396
397void
398xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
399 size_t size, enum dma_data_direction dir)
400{
401 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
402}
403EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
404
405void
406xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
407 size_t size, enum dma_data_direction dir)
408{
409 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
410}
411EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
412
413/*
414 * Map a set of buffers described by scatterlist in streaming mode for DMA.
415 * This is the scatter-gather version of the above xen_swiotlb_map_page
416 * interface. Here the scatter gather list elements are each tagged with the
417 * appropriate dma address and length. They are obtained via
418 * sg_dma_{address,length}(SG).
419 *
420 * NOTE: An implementation may be able to use a smaller number of
421 * DMA address/length pairs than there are SG table elements.
422 * (for example via virtual mapping capabilities)
423 * The routine returns the number of addr/length pairs actually
424 * used, at most nents.
425 *
426 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
427 * same here.
428 */
429int
430xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
431 int nelems, enum dma_data_direction dir,
432 struct dma_attrs *attrs)
433{
434 struct scatterlist *sg;
435 int i;
436
437 BUG_ON(dir == DMA_NONE);
438
439 for_each_sg(sgl, sg, nelems, i) {
440 phys_addr_t paddr = sg_phys(sg);
441 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
442
443 if (swiotlb_force ||
444 !dma_capable(hwdev, dev_addr, sg->length) ||
445 range_straddles_page_boundary(paddr, sg->length)) {
446 void *map = swiotlb_tbl_map_single(hwdev,
447 start_dma_addr,
448 sg_phys(sg),
449 sg->length, dir);
450 if (!map) {
451 /* Don't panic here, we expect map_sg users
452 to do proper error handling. */
453 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
454 attrs);
455 sgl[0].dma_length = 0;
456 return DMA_ERROR_CODE;
457 }
458 sg->dma_address = xen_virt_to_bus(map);
459 } else
460 sg->dma_address = dev_addr;
461 sg->dma_length = sg->length;
462 }
463 return nelems;
464}
465EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
466
467int
468xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
469 enum dma_data_direction dir)
470{
471 return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
472}
473EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
474
475/*
476 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
477 * concerning calls here are the same as for swiotlb_unmap_page() above.
478 */
479void
480xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
481 int nelems, enum dma_data_direction dir,
482 struct dma_attrs *attrs)
483{
484 struct scatterlist *sg;
485 int i;
486
487 BUG_ON(dir == DMA_NONE);
488
489 for_each_sg(sgl, sg, nelems, i)
490 xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
491
492}
493EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
494
495void
496xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
497 enum dma_data_direction dir)
498{
499 return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
500}
501EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
502
503/*
504 * Make physical memory consistent for a set of streaming mode DMA translations
505 * after a transfer.
506 *
507 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
508 * and usage.
509 */
510static void
511xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
512 int nelems, enum dma_data_direction dir,
513 enum dma_sync_target target)
514{
515 struct scatterlist *sg;
516 int i;
517
518 for_each_sg(sgl, sg, nelems, i)
519 xen_swiotlb_sync_single(hwdev, sg->dma_address,
520 sg->dma_length, dir, target);
521}
522
523void
524xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
525 int nelems, enum dma_data_direction dir)
526{
527 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
528}
529EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
530
531void
532xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
533 int nelems, enum dma_data_direction dir)
534{
535 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
536}
537EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
538
539int
540xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
541{
542 return !dma_addr;
543}
544EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
545
546/*
547 * Return whether the given device DMA address mask can be supported
548 * properly. For example, if your device can only drive the low 24-bits
549 * during bus mastering, then you would pass 0x00ffffff as the mask to
550 * this function.
551 */
552int
553xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
554{
555 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
556}
557EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);