Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright 2010
4 * by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
5 *
6 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
7 *
8 * PV guests under Xen are running in an non-contiguous memory architecture.
9 *
10 * When PCI pass-through is utilized, this necessitates an IOMMU for
11 * translating bus (DMA) to virtual and vice-versa and also providing a
12 * mechanism to have contiguous pages for device drivers operations (say DMA
13 * operations).
14 *
15 * Specifically, under Xen the Linux idea of pages is an illusion. It
16 * assumes that pages start at zero and go up to the available memory. To
17 * help with that, the Linux Xen MMU provides a lookup mechanism to
18 * translate the page frame numbers (PFN) to machine frame numbers (MFN)
19 * and vice-versa. The MFN are the "real" frame numbers. Furthermore
20 * memory is not contiguous. Xen hypervisor stitches memory for guests
21 * from different pools, which means there is no guarantee that PFN==MFN
22 * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
23 * allocated in descending order (high to low), meaning the guest might
24 * never get any MFN's under the 4GB mark.
25 */
26
27#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
28
29#include <linux/memblock.h>
30#include <linux/dma-direct.h>
31#include <linux/dma-noncoherent.h>
32#include <linux/export.h>
33#include <xen/swiotlb-xen.h>
34#include <xen/page.h>
35#include <xen/xen-ops.h>
36#include <xen/hvc-console.h>
37
38#include <asm/dma-mapping.h>
39#include <asm/xen/page-coherent.h>
40
41#include <trace/events/swiotlb.h>
42#define MAX_DMA_BITS 32
43/*
44 * Used to do a quick range check in swiotlb_tbl_unmap_single and
45 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
46 * API.
47 */
48
49static char *xen_io_tlb_start, *xen_io_tlb_end;
50static unsigned long xen_io_tlb_nslabs;
51/*
52 * Quick lookup value of the bus address of the IOTLB.
53 */
54
55static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr)
56{
57 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
58 phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT;
59
60 baddr |= paddr & ~XEN_PAGE_MASK;
61 return baddr;
62}
63
64static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr)
65{
66 return phys_to_dma(dev, xen_phys_to_bus(dev, paddr));
67}
68
69static inline phys_addr_t xen_bus_to_phys(struct device *dev,
70 phys_addr_t baddr)
71{
72 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
73 phys_addr_t paddr = (xen_pfn << XEN_PAGE_SHIFT) |
74 (baddr & ~XEN_PAGE_MASK);
75
76 return paddr;
77}
78
79static inline phys_addr_t xen_dma_to_phys(struct device *dev,
80 dma_addr_t dma_addr)
81{
82 return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr));
83}
84
85static inline dma_addr_t xen_virt_to_bus(struct device *dev, void *address)
86{
87 return xen_phys_to_dma(dev, virt_to_phys(address));
88}
89
90static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
91{
92 unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p);
93 unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size);
94
95 next_bfn = pfn_to_bfn(xen_pfn);
96
97 for (i = 1; i < nr_pages; i++)
98 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
99 return 1;
100
101 return 0;
102}
103
104static int is_xen_swiotlb_buffer(struct device *dev, dma_addr_t dma_addr)
105{
106 unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr));
107 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
108 phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT;
109
110 /* If the address is outside our domain, it CAN
111 * have the same virtual address as another address
112 * in our domain. Therefore _only_ check address within our domain.
113 */
114 if (pfn_valid(PFN_DOWN(paddr))) {
115 return paddr >= virt_to_phys(xen_io_tlb_start) &&
116 paddr < virt_to_phys(xen_io_tlb_end);
117 }
118 return 0;
119}
120
121static int
122xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
123{
124 int i, rc;
125 int dma_bits;
126 dma_addr_t dma_handle;
127 phys_addr_t p = virt_to_phys(buf);
128
129 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
130
131 i = 0;
132 do {
133 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
134
135 do {
136 rc = xen_create_contiguous_region(
137 p + (i << IO_TLB_SHIFT),
138 get_order(slabs << IO_TLB_SHIFT),
139 dma_bits, &dma_handle);
140 } while (rc && dma_bits++ < MAX_DMA_BITS);
141 if (rc)
142 return rc;
143
144 i += slabs;
145 } while (i < nslabs);
146 return 0;
147}
148static unsigned long xen_set_nslabs(unsigned long nr_tbl)
149{
150 if (!nr_tbl) {
151 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
152 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
153 } else
154 xen_io_tlb_nslabs = nr_tbl;
155
156 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
157}
158
159enum xen_swiotlb_err {
160 XEN_SWIOTLB_UNKNOWN = 0,
161 XEN_SWIOTLB_ENOMEM,
162 XEN_SWIOTLB_EFIXUP
163};
164
165static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
166{
167 switch (err) {
168 case XEN_SWIOTLB_ENOMEM:
169 return "Cannot allocate Xen-SWIOTLB buffer\n";
170 case XEN_SWIOTLB_EFIXUP:
171 return "Failed to get contiguous memory for DMA from Xen!\n"\
172 "You either: don't have the permissions, do not have"\
173 " enough free memory under 4GB, or the hypervisor memory"\
174 " is too fragmented!";
175 default:
176 break;
177 }
178 return "";
179}
180int __ref xen_swiotlb_init(int verbose, bool early)
181{
182 unsigned long bytes, order;
183 int rc = -ENOMEM;
184 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
185 unsigned int repeat = 3;
186
187 xen_io_tlb_nslabs = swiotlb_nr_tbl();
188retry:
189 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
190 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
191
192 /*
193 * IO TLB memory already allocated. Just use it.
194 */
195 if (io_tlb_start != 0) {
196 xen_io_tlb_start = phys_to_virt(io_tlb_start);
197 goto end;
198 }
199
200 /*
201 * Get IO TLB memory from any location.
202 */
203 if (early) {
204 xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes),
205 PAGE_SIZE);
206 if (!xen_io_tlb_start)
207 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
208 __func__, PAGE_ALIGN(bytes), PAGE_SIZE);
209 } else {
210#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
211#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
212 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
213 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
214 if (xen_io_tlb_start)
215 break;
216 order--;
217 }
218 if (order != get_order(bytes)) {
219 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
220 (PAGE_SIZE << order) >> 20);
221 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
222 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
223 }
224 }
225 if (!xen_io_tlb_start) {
226 m_ret = XEN_SWIOTLB_ENOMEM;
227 goto error;
228 }
229 /*
230 * And replace that memory with pages under 4GB.
231 */
232 rc = xen_swiotlb_fixup(xen_io_tlb_start,
233 bytes,
234 xen_io_tlb_nslabs);
235 if (rc) {
236 if (early)
237 memblock_free(__pa(xen_io_tlb_start),
238 PAGE_ALIGN(bytes));
239 else {
240 free_pages((unsigned long)xen_io_tlb_start, order);
241 xen_io_tlb_start = NULL;
242 }
243 m_ret = XEN_SWIOTLB_EFIXUP;
244 goto error;
245 }
246 if (early) {
247 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
248 verbose))
249 panic("Cannot allocate SWIOTLB buffer");
250 rc = 0;
251 } else
252 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
253
254end:
255 xen_io_tlb_end = xen_io_tlb_start + bytes;
256 if (!rc)
257 swiotlb_set_max_segment(PAGE_SIZE);
258
259 return rc;
260error:
261 if (repeat--) {
262 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
263 (xen_io_tlb_nslabs >> 1));
264 pr_info("Lowering to %luMB\n",
265 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
266 goto retry;
267 }
268 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
269 if (early)
270 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
271 else
272 free_pages((unsigned long)xen_io_tlb_start, order);
273 return rc;
274}
275
276static void *
277xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
278 dma_addr_t *dma_handle, gfp_t flags,
279 unsigned long attrs)
280{
281 void *ret;
282 int order = get_order(size);
283 u64 dma_mask = DMA_BIT_MASK(32);
284 phys_addr_t phys;
285 dma_addr_t dev_addr;
286
287 /*
288 * Ignore region specifiers - the kernel's ideas of
289 * pseudo-phys memory layout has nothing to do with the
290 * machine physical layout. We can't allocate highmem
291 * because we can't return a pointer to it.
292 */
293 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
294
295 /* Convert the size to actually allocated. */
296 size = 1UL << (order + XEN_PAGE_SHIFT);
297
298 /* On ARM this function returns an ioremap'ped virtual address for
299 * which virt_to_phys doesn't return the corresponding physical
300 * address. In fact on ARM virt_to_phys only works for kernel direct
301 * mapped RAM memory. Also see comment below.
302 */
303 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
304
305 if (!ret)
306 return ret;
307
308 if (hwdev && hwdev->coherent_dma_mask)
309 dma_mask = hwdev->coherent_dma_mask;
310
311 /* At this point dma_handle is the dma address, next we are
312 * going to set it to the machine address.
313 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
314 * to *dma_handle. */
315 phys = dma_to_phys(hwdev, *dma_handle);
316 dev_addr = xen_phys_to_dma(hwdev, phys);
317 if (((dev_addr + size - 1 <= dma_mask)) &&
318 !range_straddles_page_boundary(phys, size))
319 *dma_handle = dev_addr;
320 else {
321 if (xen_create_contiguous_region(phys, order,
322 fls64(dma_mask), dma_handle) != 0) {
323 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
324 return NULL;
325 }
326 *dma_handle = phys_to_dma(hwdev, *dma_handle);
327 SetPageXenRemapped(virt_to_page(ret));
328 }
329 memset(ret, 0, size);
330 return ret;
331}
332
333static void
334xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
335 dma_addr_t dev_addr, unsigned long attrs)
336{
337 int order = get_order(size);
338 phys_addr_t phys;
339 u64 dma_mask = DMA_BIT_MASK(32);
340 struct page *page;
341
342 if (hwdev && hwdev->coherent_dma_mask)
343 dma_mask = hwdev->coherent_dma_mask;
344
345 /* do not use virt_to_phys because on ARM it doesn't return you the
346 * physical address */
347 phys = xen_dma_to_phys(hwdev, dev_addr);
348
349 /* Convert the size to actually allocated. */
350 size = 1UL << (order + XEN_PAGE_SHIFT);
351
352 if (is_vmalloc_addr(vaddr))
353 page = vmalloc_to_page(vaddr);
354 else
355 page = virt_to_page(vaddr);
356
357 if (!WARN_ON((dev_addr + size - 1 > dma_mask) ||
358 range_straddles_page_boundary(phys, size)) &&
359 TestClearPageXenRemapped(page))
360 xen_destroy_contiguous_region(phys, order);
361
362 xen_free_coherent_pages(hwdev, size, vaddr, phys_to_dma(hwdev, phys),
363 attrs);
364}
365
366/*
367 * Map a single buffer of the indicated size for DMA in streaming mode. The
368 * physical address to use is returned.
369 *
370 * Once the device is given the dma address, the device owns this memory until
371 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
372 */
373static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
374 unsigned long offset, size_t size,
375 enum dma_data_direction dir,
376 unsigned long attrs)
377{
378 phys_addr_t map, phys = page_to_phys(page) + offset;
379 dma_addr_t dev_addr = xen_phys_to_dma(dev, phys);
380
381 BUG_ON(dir == DMA_NONE);
382 /*
383 * If the address happens to be in the device's DMA window,
384 * we can safely return the device addr and not worry about bounce
385 * buffering it.
386 */
387 if (dma_capable(dev, dev_addr, size, true) &&
388 !range_straddles_page_boundary(phys, size) &&
389 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
390 swiotlb_force != SWIOTLB_FORCE)
391 goto done;
392
393 /*
394 * Oh well, have to allocate and map a bounce buffer.
395 */
396 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
397
398 map = swiotlb_tbl_map_single(dev, virt_to_phys(xen_io_tlb_start),
399 phys, size, size, dir, attrs);
400 if (map == (phys_addr_t)DMA_MAPPING_ERROR)
401 return DMA_MAPPING_ERROR;
402
403 phys = map;
404 dev_addr = xen_phys_to_dma(dev, map);
405
406 /*
407 * Ensure that the address returned is DMA'ble
408 */
409 if (unlikely(!dma_capable(dev, dev_addr, size, true))) {
410 swiotlb_tbl_unmap_single(dev, map, size, size, dir,
411 attrs | DMA_ATTR_SKIP_CPU_SYNC);
412 return DMA_MAPPING_ERROR;
413 }
414
415done:
416 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
417 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr))))
418 arch_sync_dma_for_device(phys, size, dir);
419 else
420 xen_dma_sync_for_device(dev, dev_addr, size, dir);
421 }
422 return dev_addr;
423}
424
425/*
426 * Unmap a single streaming mode DMA translation. The dma_addr and size must
427 * match what was provided for in a previous xen_swiotlb_map_page call. All
428 * other usages are undefined.
429 *
430 * After this call, reads by the cpu to the buffer are guaranteed to see
431 * whatever the device wrote there.
432 */
433static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
434 size_t size, enum dma_data_direction dir, unsigned long attrs)
435{
436 phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr);
437
438 BUG_ON(dir == DMA_NONE);
439
440 if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
441 if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr))))
442 arch_sync_dma_for_cpu(paddr, size, dir);
443 else
444 xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir);
445 }
446
447 /* NOTE: We use dev_addr here, not paddr! */
448 if (is_xen_swiotlb_buffer(hwdev, dev_addr))
449 swiotlb_tbl_unmap_single(hwdev, paddr, size, size, dir, attrs);
450}
451
452static void
453xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
454 size_t size, enum dma_data_direction dir)
455{
456 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
457
458 if (!dev_is_dma_coherent(dev)) {
459 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
460 arch_sync_dma_for_cpu(paddr, size, dir);
461 else
462 xen_dma_sync_for_cpu(dev, dma_addr, size, dir);
463 }
464
465 if (is_xen_swiotlb_buffer(dev, dma_addr))
466 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_CPU);
467}
468
469static void
470xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
471 size_t size, enum dma_data_direction dir)
472{
473 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
474
475 if (is_xen_swiotlb_buffer(dev, dma_addr))
476 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
477
478 if (!dev_is_dma_coherent(dev)) {
479 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
480 arch_sync_dma_for_device(paddr, size, dir);
481 else
482 xen_dma_sync_for_device(dev, dma_addr, size, dir);
483 }
484}
485
486/*
487 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
488 * concerning calls here are the same as for swiotlb_unmap_page() above.
489 */
490static void
491xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
492 enum dma_data_direction dir, unsigned long attrs)
493{
494 struct scatterlist *sg;
495 int i;
496
497 BUG_ON(dir == DMA_NONE);
498
499 for_each_sg(sgl, sg, nelems, i)
500 xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg),
501 dir, attrs);
502
503}
504
505static int
506xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
507 enum dma_data_direction dir, unsigned long attrs)
508{
509 struct scatterlist *sg;
510 int i;
511
512 BUG_ON(dir == DMA_NONE);
513
514 for_each_sg(sgl, sg, nelems, i) {
515 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
516 sg->offset, sg->length, dir, attrs);
517 if (sg->dma_address == DMA_MAPPING_ERROR)
518 goto out_unmap;
519 sg_dma_len(sg) = sg->length;
520 }
521
522 return nelems;
523out_unmap:
524 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
525 sg_dma_len(sgl) = 0;
526 return 0;
527}
528
529static void
530xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
531 int nelems, enum dma_data_direction dir)
532{
533 struct scatterlist *sg;
534 int i;
535
536 for_each_sg(sgl, sg, nelems, i) {
537 xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
538 sg->length, dir);
539 }
540}
541
542static void
543xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
544 int nelems, enum dma_data_direction dir)
545{
546 struct scatterlist *sg;
547 int i;
548
549 for_each_sg(sgl, sg, nelems, i) {
550 xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
551 sg->length, dir);
552 }
553}
554
555/*
556 * Return whether the given device DMA address mask can be supported
557 * properly. For example, if your device can only drive the low 24-bits
558 * during bus mastering, then you would pass 0x00ffffff as the mask to
559 * this function.
560 */
561static int
562xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
563{
564 return xen_virt_to_bus(hwdev, xen_io_tlb_end - 1) <= mask;
565}
566
567const struct dma_map_ops xen_swiotlb_dma_ops = {
568 .alloc = xen_swiotlb_alloc_coherent,
569 .free = xen_swiotlb_free_coherent,
570 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
571 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
572 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
573 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
574 .map_sg = xen_swiotlb_map_sg,
575 .unmap_sg = xen_swiotlb_unmap_sg,
576 .map_page = xen_swiotlb_map_page,
577 .unmap_page = xen_swiotlb_unmap_page,
578 .dma_supported = xen_swiotlb_dma_supported,
579 .mmap = dma_common_mmap,
580 .get_sgtable = dma_common_get_sgtable,
581};
1/*
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 <linux/export.h>
39#include <xen/swiotlb-xen.h>
40#include <xen/page.h>
41#include <xen/xen-ops.h>
42#include <xen/hvc-console.h>
43/*
44 * Used to do a quick range check in swiotlb_tbl_unmap_single and
45 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
46 * API.
47 */
48
49static char *xen_io_tlb_start, *xen_io_tlb_end;
50static unsigned long xen_io_tlb_nslabs;
51/*
52 * Quick lookup value of the bus address of the IOTLB.
53 */
54
55u64 start_dma_addr;
56
57static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
58{
59 return phys_to_machine(XPADDR(paddr)).maddr;
60}
61
62static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
63{
64 return machine_to_phys(XMADDR(baddr)).paddr;
65}
66
67static dma_addr_t xen_virt_to_bus(void *address)
68{
69 return xen_phys_to_bus(virt_to_phys(address));
70}
71
72static int check_pages_physically_contiguous(unsigned long pfn,
73 unsigned int offset,
74 size_t length)
75{
76 unsigned long next_mfn;
77 int i;
78 int nr_pages;
79
80 next_mfn = pfn_to_mfn(pfn);
81 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
82
83 for (i = 1; i < nr_pages; i++) {
84 if (pfn_to_mfn(++pfn) != ++next_mfn)
85 return 0;
86 }
87 return 1;
88}
89
90static int range_straddles_page_boundary(phys_addr_t p, size_t size)
91{
92 unsigned long pfn = PFN_DOWN(p);
93 unsigned int offset = p & ~PAGE_MASK;
94
95 if (offset + size <= PAGE_SIZE)
96 return 0;
97 if (check_pages_physically_contiguous(pfn, offset, size))
98 return 0;
99 return 1;
100}
101
102static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
103{
104 unsigned long mfn = PFN_DOWN(dma_addr);
105 unsigned long pfn = mfn_to_local_pfn(mfn);
106 phys_addr_t paddr;
107
108 /* If the address is outside our domain, it CAN
109 * have the same virtual address as another address
110 * in our domain. Therefore _only_ check address within our domain.
111 */
112 if (pfn_valid(pfn)) {
113 paddr = PFN_PHYS(pfn);
114 return paddr >= virt_to_phys(xen_io_tlb_start) &&
115 paddr < virt_to_phys(xen_io_tlb_end);
116 }
117 return 0;
118}
119
120static int max_dma_bits = 32;
121
122static int
123xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
124{
125 int i, rc;
126 int dma_bits;
127
128 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
129
130 i = 0;
131 do {
132 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
133
134 do {
135 rc = xen_create_contiguous_region(
136 (unsigned long)buf + (i << IO_TLB_SHIFT),
137 get_order(slabs << IO_TLB_SHIFT),
138 dma_bits);
139 } while (rc && dma_bits++ < max_dma_bits);
140 if (rc)
141 return rc;
142
143 i += slabs;
144 } while (i < nslabs);
145 return 0;
146}
147
148void __init xen_swiotlb_init(int verbose)
149{
150 unsigned long bytes;
151 int rc = -ENOMEM;
152 unsigned long nr_tbl;
153 char *m = NULL;
154 unsigned int repeat = 3;
155
156 nr_tbl = swiotlb_nr_tbl();
157 if (nr_tbl)
158 xen_io_tlb_nslabs = nr_tbl;
159 else {
160 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
161 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
162 }
163retry:
164 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
165
166 /*
167 * Get IO TLB memory from any location.
168 */
169 xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
170 if (!xen_io_tlb_start) {
171 m = "Cannot allocate Xen-SWIOTLB buffer!\n";
172 goto error;
173 }
174 xen_io_tlb_end = xen_io_tlb_start + bytes;
175 /*
176 * And replace that memory with pages under 4GB.
177 */
178 rc = xen_swiotlb_fixup(xen_io_tlb_start,
179 bytes,
180 xen_io_tlb_nslabs);
181 if (rc) {
182 free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
183 m = "Failed to get contiguous memory for DMA from Xen!\n"\
184 "You either: don't have the permissions, do not have"\
185 " enough free memory under 4GB, or the hypervisor memory"\
186 "is too fragmented!";
187 goto error;
188 }
189 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
190 swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
191
192 return;
193error:
194 if (repeat--) {
195 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
196 (xen_io_tlb_nslabs >> 1));
197 printk(KERN_INFO "Xen-SWIOTLB: Lowering to %luMB\n",
198 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
199 goto retry;
200 }
201 xen_raw_printk("%s (rc:%d)", m, rc);
202 panic("%s (rc:%d)", m, rc);
203}
204
205void *
206xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
207 dma_addr_t *dma_handle, gfp_t flags,
208 struct dma_attrs *attrs)
209{
210 void *ret;
211 int order = get_order(size);
212 u64 dma_mask = DMA_BIT_MASK(32);
213 unsigned long vstart;
214 phys_addr_t phys;
215 dma_addr_t dev_addr;
216
217 /*
218 * Ignore region specifiers - the kernel's ideas of
219 * pseudo-phys memory layout has nothing to do with the
220 * machine physical layout. We can't allocate highmem
221 * because we can't return a pointer to it.
222 */
223 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
224
225 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
226 return ret;
227
228 vstart = __get_free_pages(flags, order);
229 ret = (void *)vstart;
230
231 if (!ret)
232 return ret;
233
234 if (hwdev && hwdev->coherent_dma_mask)
235 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
236
237 phys = virt_to_phys(ret);
238 dev_addr = xen_phys_to_bus(phys);
239 if (((dev_addr + size - 1 <= dma_mask)) &&
240 !range_straddles_page_boundary(phys, size))
241 *dma_handle = dev_addr;
242 else {
243 if (xen_create_contiguous_region(vstart, order,
244 fls64(dma_mask)) != 0) {
245 free_pages(vstart, order);
246 return NULL;
247 }
248 *dma_handle = virt_to_machine(ret).maddr;
249 }
250 memset(ret, 0, size);
251 return ret;
252}
253EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
254
255void
256xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
257 dma_addr_t dev_addr, struct dma_attrs *attrs)
258{
259 int order = get_order(size);
260 phys_addr_t phys;
261 u64 dma_mask = DMA_BIT_MASK(32);
262
263 if (dma_release_from_coherent(hwdev, order, vaddr))
264 return;
265
266 if (hwdev && hwdev->coherent_dma_mask)
267 dma_mask = hwdev->coherent_dma_mask;
268
269 phys = virt_to_phys(vaddr);
270
271 if (((dev_addr + size - 1 > dma_mask)) ||
272 range_straddles_page_boundary(phys, size))
273 xen_destroy_contiguous_region((unsigned long)vaddr, order);
274
275 free_pages((unsigned long)vaddr, order);
276}
277EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
278
279
280/*
281 * Map a single buffer of the indicated size for DMA in streaming mode. The
282 * physical address to use is returned.
283 *
284 * Once the device is given the dma address, the device owns this memory until
285 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
286 */
287dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
288 unsigned long offset, size_t size,
289 enum dma_data_direction dir,
290 struct dma_attrs *attrs)
291{
292 phys_addr_t phys = page_to_phys(page) + offset;
293 dma_addr_t dev_addr = xen_phys_to_bus(phys);
294 void *map;
295
296 BUG_ON(dir == DMA_NONE);
297 /*
298 * If the address happens to be in the device's DMA window,
299 * we can safely return the device addr and not worry about bounce
300 * buffering it.
301 */
302 if (dma_capable(dev, dev_addr, size) &&
303 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
304 return dev_addr;
305
306 /*
307 * Oh well, have to allocate and map a bounce buffer.
308 */
309 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
310 if (!map)
311 return DMA_ERROR_CODE;
312
313 dev_addr = xen_virt_to_bus(map);
314
315 /*
316 * Ensure that the address returned is DMA'ble
317 */
318 if (!dma_capable(dev, dev_addr, size)) {
319 swiotlb_tbl_unmap_single(dev, map, size, dir);
320 dev_addr = 0;
321 }
322 return dev_addr;
323}
324EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
325
326/*
327 * Unmap a single streaming mode DMA translation. The dma_addr and size must
328 * match what was provided for in a previous xen_swiotlb_map_page call. All
329 * other usages are undefined.
330 *
331 * After this call, reads by the cpu to the buffer are guaranteed to see
332 * whatever the device wrote there.
333 */
334static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
335 size_t size, enum dma_data_direction dir)
336{
337 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
338
339 BUG_ON(dir == DMA_NONE);
340
341 /* NOTE: We use dev_addr here, not paddr! */
342 if (is_xen_swiotlb_buffer(dev_addr)) {
343 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
344 return;
345 }
346
347 if (dir != DMA_FROM_DEVICE)
348 return;
349
350 /*
351 * phys_to_virt doesn't work with hihgmem page but we could
352 * call dma_mark_clean() with hihgmem page here. However, we
353 * are fine since dma_mark_clean() is null on POWERPC. We can
354 * make dma_mark_clean() take a physical address if necessary.
355 */
356 dma_mark_clean(phys_to_virt(paddr), size);
357}
358
359void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
360 size_t size, enum dma_data_direction dir,
361 struct dma_attrs *attrs)
362{
363 xen_unmap_single(hwdev, dev_addr, size, dir);
364}
365EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
366
367/*
368 * Make physical memory consistent for a single streaming mode DMA translation
369 * after a transfer.
370 *
371 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
372 * using the cpu, yet do not wish to teardown the dma mapping, you must
373 * call this function before doing so. At the next point you give the dma
374 * address back to the card, you must first perform a
375 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
376 */
377static void
378xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
379 size_t size, enum dma_data_direction dir,
380 enum dma_sync_target target)
381{
382 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
383
384 BUG_ON(dir == DMA_NONE);
385
386 /* NOTE: We use dev_addr here, not paddr! */
387 if (is_xen_swiotlb_buffer(dev_addr)) {
388 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
389 target);
390 return;
391 }
392
393 if (dir != DMA_FROM_DEVICE)
394 return;
395
396 dma_mark_clean(phys_to_virt(paddr), size);
397}
398
399void
400xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
401 size_t size, enum dma_data_direction dir)
402{
403 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
404}
405EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
406
407void
408xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
409 size_t size, enum dma_data_direction dir)
410{
411 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
412}
413EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
414
415/*
416 * Map a set of buffers described by scatterlist in streaming mode for DMA.
417 * This is the scatter-gather version of the above xen_swiotlb_map_page
418 * interface. Here the scatter gather list elements are each tagged with the
419 * appropriate dma address and length. They are obtained via
420 * sg_dma_{address,length}(SG).
421 *
422 * NOTE: An implementation may be able to use a smaller number of
423 * DMA address/length pairs than there are SG table elements.
424 * (for example via virtual mapping capabilities)
425 * The routine returns the number of addr/length pairs actually
426 * used, at most nents.
427 *
428 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
429 * same here.
430 */
431int
432xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
433 int nelems, enum dma_data_direction dir,
434 struct dma_attrs *attrs)
435{
436 struct scatterlist *sg;
437 int i;
438
439 BUG_ON(dir == DMA_NONE);
440
441 for_each_sg(sgl, sg, nelems, i) {
442 phys_addr_t paddr = sg_phys(sg);
443 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
444
445 if (swiotlb_force ||
446 !dma_capable(hwdev, dev_addr, sg->length) ||
447 range_straddles_page_boundary(paddr, sg->length)) {
448 void *map = swiotlb_tbl_map_single(hwdev,
449 start_dma_addr,
450 sg_phys(sg),
451 sg->length, dir);
452 if (!map) {
453 /* Don't panic here, we expect map_sg users
454 to do proper error handling. */
455 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
456 attrs);
457 sgl[0].dma_length = 0;
458 return DMA_ERROR_CODE;
459 }
460 sg->dma_address = xen_virt_to_bus(map);
461 } else
462 sg->dma_address = dev_addr;
463 sg->dma_length = sg->length;
464 }
465 return nelems;
466}
467EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
468
469int
470xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
471 enum dma_data_direction dir)
472{
473 return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
474}
475EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
476
477/*
478 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
479 * concerning calls here are the same as for swiotlb_unmap_page() above.
480 */
481void
482xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
483 int nelems, enum dma_data_direction dir,
484 struct dma_attrs *attrs)
485{
486 struct scatterlist *sg;
487 int i;
488
489 BUG_ON(dir == DMA_NONE);
490
491 for_each_sg(sgl, sg, nelems, i)
492 xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
493
494}
495EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
496
497void
498xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
499 enum dma_data_direction dir)
500{
501 return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
502}
503EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
504
505/*
506 * Make physical memory consistent for a set of streaming mode DMA translations
507 * after a transfer.
508 *
509 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
510 * and usage.
511 */
512static void
513xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
514 int nelems, enum dma_data_direction dir,
515 enum dma_sync_target target)
516{
517 struct scatterlist *sg;
518 int i;
519
520 for_each_sg(sgl, sg, nelems, i)
521 xen_swiotlb_sync_single(hwdev, sg->dma_address,
522 sg->dma_length, dir, target);
523}
524
525void
526xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
527 int nelems, enum dma_data_direction dir)
528{
529 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
530}
531EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
532
533void
534xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
535 int nelems, enum dma_data_direction dir)
536{
537 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
538}
539EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
540
541int
542xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
543{
544 return !dma_addr;
545}
546EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
547
548/*
549 * Return whether the given device DMA address mask can be supported
550 * properly. For example, if your device can only drive the low 24-bits
551 * during bus mastering, then you would pass 0x00ffffff as the mask to
552 * this function.
553 */
554int
555xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
556{
557 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
558}
559EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);