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// 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-map-ops.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/*
45 * Quick lookup value of the bus address of the IOTLB.
46 */
47
48static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr)
49{
50 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
51 phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT;
52
53 baddr |= paddr & ~XEN_PAGE_MASK;
54 return baddr;
55}
56
57static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr)
58{
59 return phys_to_dma(dev, xen_phys_to_bus(dev, paddr));
60}
61
62static inline phys_addr_t xen_bus_to_phys(struct device *dev,
63 phys_addr_t baddr)
64{
65 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
66 phys_addr_t paddr = (xen_pfn << XEN_PAGE_SHIFT) |
67 (baddr & ~XEN_PAGE_MASK);
68
69 return paddr;
70}
71
72static inline phys_addr_t xen_dma_to_phys(struct device *dev,
73 dma_addr_t dma_addr)
74{
75 return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr));
76}
77
78static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
79{
80 unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p);
81 unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size);
82
83 next_bfn = pfn_to_bfn(xen_pfn);
84
85 for (i = 1; i < nr_pages; i++)
86 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
87 return 1;
88
89 return 0;
90}
91
92static int is_xen_swiotlb_buffer(struct device *dev, dma_addr_t dma_addr)
93{
94 unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr));
95 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
96 phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT;
97
98 /* If the address is outside our domain, it CAN
99 * have the same virtual address as another address
100 * in our domain. Therefore _only_ check address within our domain.
101 */
102 if (pfn_valid(PFN_DOWN(paddr)))
103 return is_swiotlb_buffer(paddr);
104 return 0;
105}
106
107static int xen_swiotlb_fixup(void *buf, unsigned long nslabs)
108{
109 int i, rc;
110 int dma_bits;
111 dma_addr_t dma_handle;
112 phys_addr_t p = virt_to_phys(buf);
113
114 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
115
116 i = 0;
117 do {
118 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
119
120 do {
121 rc = xen_create_contiguous_region(
122 p + (i << IO_TLB_SHIFT),
123 get_order(slabs << IO_TLB_SHIFT),
124 dma_bits, &dma_handle);
125 } while (rc && dma_bits++ < MAX_DMA_BITS);
126 if (rc)
127 return rc;
128
129 i += slabs;
130 } while (i < nslabs);
131 return 0;
132}
133
134enum xen_swiotlb_err {
135 XEN_SWIOTLB_UNKNOWN = 0,
136 XEN_SWIOTLB_ENOMEM,
137 XEN_SWIOTLB_EFIXUP
138};
139
140static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
141{
142 switch (err) {
143 case XEN_SWIOTLB_ENOMEM:
144 return "Cannot allocate Xen-SWIOTLB buffer\n";
145 case XEN_SWIOTLB_EFIXUP:
146 return "Failed to get contiguous memory for DMA from Xen!\n"\
147 "You either: don't have the permissions, do not have"\
148 " enough free memory under 4GB, or the hypervisor memory"\
149 " is too fragmented!";
150 default:
151 break;
152 }
153 return "";
154}
155
156#define DEFAULT_NSLABS ALIGN(SZ_64M >> IO_TLB_SHIFT, IO_TLB_SEGSIZE)
157
158int __ref xen_swiotlb_init(void)
159{
160 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
161 unsigned long bytes = swiotlb_size_or_default();
162 unsigned long nslabs = bytes >> IO_TLB_SHIFT;
163 unsigned int order, repeat = 3;
164 int rc = -ENOMEM;
165 char *start;
166
167 if (io_tlb_default_mem != NULL) {
168 pr_warn("swiotlb buffer already initialized\n");
169 return -EEXIST;
170 }
171
172retry:
173 m_ret = XEN_SWIOTLB_ENOMEM;
174 order = get_order(bytes);
175
176 /*
177 * Get IO TLB memory from any location.
178 */
179#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
180#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
181 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
182 start = (void *)xen_get_swiotlb_free_pages(order);
183 if (start)
184 break;
185 order--;
186 }
187 if (!start)
188 goto error;
189 if (order != get_order(bytes)) {
190 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
191 (PAGE_SIZE << order) >> 20);
192 nslabs = SLABS_PER_PAGE << order;
193 bytes = nslabs << IO_TLB_SHIFT;
194 }
195
196 /*
197 * And replace that memory with pages under 4GB.
198 */
199 rc = xen_swiotlb_fixup(start, nslabs);
200 if (rc) {
201 free_pages((unsigned long)start, order);
202 m_ret = XEN_SWIOTLB_EFIXUP;
203 goto error;
204 }
205 rc = swiotlb_late_init_with_tbl(start, nslabs);
206 if (rc)
207 return rc;
208 swiotlb_set_max_segment(PAGE_SIZE);
209 return 0;
210error:
211 if (repeat--) {
212 /* Min is 2MB */
213 nslabs = max(1024UL, (nslabs >> 1));
214 bytes = nslabs << IO_TLB_SHIFT;
215 pr_info("Lowering to %luMB\n", bytes >> 20);
216 goto retry;
217 }
218 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
219 return rc;
220}
221
222#ifdef CONFIG_X86
223void __init xen_swiotlb_init_early(void)
224{
225 unsigned long bytes = swiotlb_size_or_default();
226 unsigned long nslabs = bytes >> IO_TLB_SHIFT;
227 unsigned int repeat = 3;
228 char *start;
229 int rc;
230
231retry:
232 /*
233 * Get IO TLB memory from any location.
234 */
235 start = memblock_alloc(PAGE_ALIGN(bytes),
236 IO_TLB_SEGSIZE << IO_TLB_SHIFT);
237 if (!start)
238 panic("%s: Failed to allocate %lu bytes\n",
239 __func__, PAGE_ALIGN(bytes));
240
241 /*
242 * And replace that memory with pages under 4GB.
243 */
244 rc = xen_swiotlb_fixup(start, nslabs);
245 if (rc) {
246 memblock_free(__pa(start), PAGE_ALIGN(bytes));
247 if (repeat--) {
248 /* Min is 2MB */
249 nslabs = max(1024UL, (nslabs >> 1));
250 bytes = nslabs << IO_TLB_SHIFT;
251 pr_info("Lowering to %luMB\n", bytes >> 20);
252 goto retry;
253 }
254 panic("%s (rc:%d)", xen_swiotlb_error(XEN_SWIOTLB_EFIXUP), rc);
255 }
256
257 if (swiotlb_init_with_tbl(start, nslabs, false))
258 panic("Cannot allocate SWIOTLB buffer");
259 swiotlb_set_max_segment(PAGE_SIZE);
260}
261#endif /* CONFIG_X86 */
262
263static void *
264xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
265 dma_addr_t *dma_handle, gfp_t flags,
266 unsigned long attrs)
267{
268 void *ret;
269 int order = get_order(size);
270 u64 dma_mask = DMA_BIT_MASK(32);
271 phys_addr_t phys;
272 dma_addr_t dev_addr;
273
274 /*
275 * Ignore region specifiers - the kernel's ideas of
276 * pseudo-phys memory layout has nothing to do with the
277 * machine physical layout. We can't allocate highmem
278 * because we can't return a pointer to it.
279 */
280 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
281
282 /* Convert the size to actually allocated. */
283 size = 1UL << (order + XEN_PAGE_SHIFT);
284
285 /* On ARM this function returns an ioremap'ped virtual address for
286 * which virt_to_phys doesn't return the corresponding physical
287 * address. In fact on ARM virt_to_phys only works for kernel direct
288 * mapped RAM memory. Also see comment below.
289 */
290 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
291
292 if (!ret)
293 return ret;
294
295 if (hwdev && hwdev->coherent_dma_mask)
296 dma_mask = hwdev->coherent_dma_mask;
297
298 /* At this point dma_handle is the dma address, next we are
299 * going to set it to the machine address.
300 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
301 * to *dma_handle. */
302 phys = dma_to_phys(hwdev, *dma_handle);
303 dev_addr = xen_phys_to_dma(hwdev, phys);
304 if (((dev_addr + size - 1 <= dma_mask)) &&
305 !range_straddles_page_boundary(phys, size))
306 *dma_handle = dev_addr;
307 else {
308 if (xen_create_contiguous_region(phys, order,
309 fls64(dma_mask), dma_handle) != 0) {
310 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
311 return NULL;
312 }
313 *dma_handle = phys_to_dma(hwdev, *dma_handle);
314 SetPageXenRemapped(virt_to_page(ret));
315 }
316 memset(ret, 0, size);
317 return ret;
318}
319
320static void
321xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
322 dma_addr_t dev_addr, unsigned long attrs)
323{
324 int order = get_order(size);
325 phys_addr_t phys;
326 u64 dma_mask = DMA_BIT_MASK(32);
327 struct page *page;
328
329 if (hwdev && hwdev->coherent_dma_mask)
330 dma_mask = hwdev->coherent_dma_mask;
331
332 /* do not use virt_to_phys because on ARM it doesn't return you the
333 * physical address */
334 phys = xen_dma_to_phys(hwdev, dev_addr);
335
336 /* Convert the size to actually allocated. */
337 size = 1UL << (order + XEN_PAGE_SHIFT);
338
339 if (is_vmalloc_addr(vaddr))
340 page = vmalloc_to_page(vaddr);
341 else
342 page = virt_to_page(vaddr);
343
344 if (!WARN_ON((dev_addr + size - 1 > dma_mask) ||
345 range_straddles_page_boundary(phys, size)) &&
346 TestClearPageXenRemapped(page))
347 xen_destroy_contiguous_region(phys, order);
348
349 xen_free_coherent_pages(hwdev, size, vaddr, phys_to_dma(hwdev, phys),
350 attrs);
351}
352
353/*
354 * Map a single buffer of the indicated size for DMA in streaming mode. The
355 * physical address to use is returned.
356 *
357 * Once the device is given the dma address, the device owns this memory until
358 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
359 */
360static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
361 unsigned long offset, size_t size,
362 enum dma_data_direction dir,
363 unsigned long attrs)
364{
365 phys_addr_t map, phys = page_to_phys(page) + offset;
366 dma_addr_t dev_addr = xen_phys_to_dma(dev, phys);
367
368 BUG_ON(dir == DMA_NONE);
369 /*
370 * If the address happens to be in the device's DMA window,
371 * we can safely return the device addr and not worry about bounce
372 * buffering it.
373 */
374 if (dma_capable(dev, dev_addr, size, true) &&
375 !range_straddles_page_boundary(phys, size) &&
376 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
377 swiotlb_force != SWIOTLB_FORCE)
378 goto done;
379
380 /*
381 * Oh well, have to allocate and map a bounce buffer.
382 */
383 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
384
385 map = swiotlb_tbl_map_single(dev, phys, size, size, dir, attrs);
386 if (map == (phys_addr_t)DMA_MAPPING_ERROR)
387 return DMA_MAPPING_ERROR;
388
389 phys = map;
390 dev_addr = xen_phys_to_dma(dev, map);
391
392 /*
393 * Ensure that the address returned is DMA'ble
394 */
395 if (unlikely(!dma_capable(dev, dev_addr, size, true))) {
396 swiotlb_tbl_unmap_single(dev, map, size, dir,
397 attrs | DMA_ATTR_SKIP_CPU_SYNC);
398 return DMA_MAPPING_ERROR;
399 }
400
401done:
402 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
403 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr))))
404 arch_sync_dma_for_device(phys, size, dir);
405 else
406 xen_dma_sync_for_device(dev, dev_addr, size, dir);
407 }
408 return dev_addr;
409}
410
411/*
412 * Unmap a single streaming mode DMA translation. The dma_addr and size must
413 * match what was provided for in a previous xen_swiotlb_map_page call. All
414 * other usages are undefined.
415 *
416 * After this call, reads by the cpu to the buffer are guaranteed to see
417 * whatever the device wrote there.
418 */
419static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
420 size_t size, enum dma_data_direction dir, unsigned long attrs)
421{
422 phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr);
423
424 BUG_ON(dir == DMA_NONE);
425
426 if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
427 if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr))))
428 arch_sync_dma_for_cpu(paddr, size, dir);
429 else
430 xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir);
431 }
432
433 /* NOTE: We use dev_addr here, not paddr! */
434 if (is_xen_swiotlb_buffer(hwdev, dev_addr))
435 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
436}
437
438static void
439xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
440 size_t size, enum dma_data_direction dir)
441{
442 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
443
444 if (!dev_is_dma_coherent(dev)) {
445 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
446 arch_sync_dma_for_cpu(paddr, size, dir);
447 else
448 xen_dma_sync_for_cpu(dev, dma_addr, size, dir);
449 }
450
451 if (is_xen_swiotlb_buffer(dev, dma_addr))
452 swiotlb_sync_single_for_cpu(dev, paddr, size, dir);
453}
454
455static void
456xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
457 size_t size, enum dma_data_direction dir)
458{
459 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
460
461 if (is_xen_swiotlb_buffer(dev, dma_addr))
462 swiotlb_sync_single_for_device(dev, paddr, size, dir);
463
464 if (!dev_is_dma_coherent(dev)) {
465 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
466 arch_sync_dma_for_device(paddr, size, dir);
467 else
468 xen_dma_sync_for_device(dev, dma_addr, size, dir);
469 }
470}
471
472/*
473 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
474 * concerning calls here are the same as for swiotlb_unmap_page() above.
475 */
476static void
477xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
478 enum dma_data_direction dir, unsigned long attrs)
479{
480 struct scatterlist *sg;
481 int i;
482
483 BUG_ON(dir == DMA_NONE);
484
485 for_each_sg(sgl, sg, nelems, i)
486 xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg),
487 dir, attrs);
488
489}
490
491static int
492xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
493 enum dma_data_direction dir, unsigned long attrs)
494{
495 struct scatterlist *sg;
496 int i;
497
498 BUG_ON(dir == DMA_NONE);
499
500 for_each_sg(sgl, sg, nelems, i) {
501 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
502 sg->offset, sg->length, dir, attrs);
503 if (sg->dma_address == DMA_MAPPING_ERROR)
504 goto out_unmap;
505 sg_dma_len(sg) = sg->length;
506 }
507
508 return nelems;
509out_unmap:
510 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
511 sg_dma_len(sgl) = 0;
512 return 0;
513}
514
515static void
516xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
517 int nelems, enum dma_data_direction dir)
518{
519 struct scatterlist *sg;
520 int i;
521
522 for_each_sg(sgl, sg, nelems, i) {
523 xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
524 sg->length, dir);
525 }
526}
527
528static void
529xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
530 int nelems, enum dma_data_direction dir)
531{
532 struct scatterlist *sg;
533 int i;
534
535 for_each_sg(sgl, sg, nelems, i) {
536 xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
537 sg->length, dir);
538 }
539}
540
541/*
542 * Return whether the given device DMA address mask can be supported
543 * properly. For example, if your device can only drive the low 24-bits
544 * during bus mastering, then you would pass 0x00ffffff as the mask to
545 * this function.
546 */
547static int
548xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
549{
550 return xen_phys_to_dma(hwdev, io_tlb_default_mem->end - 1) <= mask;
551}
552
553const struct dma_map_ops xen_swiotlb_dma_ops = {
554 .alloc = xen_swiotlb_alloc_coherent,
555 .free = xen_swiotlb_free_coherent,
556 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
557 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
558 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
559 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
560 .map_sg = xen_swiotlb_map_sg,
561 .unmap_sg = xen_swiotlb_unmap_sg,
562 .map_page = xen_swiotlb_map_page,
563 .unmap_page = xen_swiotlb_unmap_page,
564 .dma_supported = xen_swiotlb_dma_supported,
565 .mmap = dma_common_mmap,
566 .get_sgtable = dma_common_get_sgtable,
567 .alloc_pages = dma_common_alloc_pages,
568 .free_pages = dma_common_free_pages,
569};