Loading...
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 <xen/swiotlb-xen.h>
39#include <xen/page.h>
40#include <xen/xen-ops.h>
41/*
42 * Used to do a quick range check in swiotlb_tbl_unmap_single and
43 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
44 * API.
45 */
46
47static char *xen_io_tlb_start, *xen_io_tlb_end;
48static unsigned long xen_io_tlb_nslabs;
49/*
50 * Quick lookup value of the bus address of the IOTLB.
51 */
52
53u64 start_dma_addr;
54
55static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
56{
57 return phys_to_machine(XPADDR(paddr)).maddr;
58}
59
60static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
61{
62 return machine_to_phys(XMADDR(baddr)).paddr;
63}
64
65static dma_addr_t xen_virt_to_bus(void *address)
66{
67 return xen_phys_to_bus(virt_to_phys(address));
68}
69
70static int check_pages_physically_contiguous(unsigned long pfn,
71 unsigned int offset,
72 size_t length)
73{
74 unsigned long next_mfn;
75 int i;
76 int nr_pages;
77
78 next_mfn = pfn_to_mfn(pfn);
79 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
80
81 for (i = 1; i < nr_pages; i++) {
82 if (pfn_to_mfn(++pfn) != ++next_mfn)
83 return 0;
84 }
85 return 1;
86}
87
88static int range_straddles_page_boundary(phys_addr_t p, size_t size)
89{
90 unsigned long pfn = PFN_DOWN(p);
91 unsigned int offset = p & ~PAGE_MASK;
92
93 if (offset + size <= PAGE_SIZE)
94 return 0;
95 if (check_pages_physically_contiguous(pfn, offset, size))
96 return 0;
97 return 1;
98}
99
100static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
101{
102 unsigned long mfn = PFN_DOWN(dma_addr);
103 unsigned long pfn = mfn_to_local_pfn(mfn);
104 phys_addr_t paddr;
105
106 /* If the address is outside our domain, it CAN
107 * have the same virtual address as another address
108 * in our domain. Therefore _only_ check address within our domain.
109 */
110 if (pfn_valid(pfn)) {
111 paddr = PFN_PHYS(pfn);
112 return paddr >= virt_to_phys(xen_io_tlb_start) &&
113 paddr < virt_to_phys(xen_io_tlb_end);
114 }
115 return 0;
116}
117
118static int max_dma_bits = 32;
119
120static int
121xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
122{
123 int i, rc;
124 int dma_bits;
125
126 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
127
128 i = 0;
129 do {
130 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
131
132 do {
133 rc = xen_create_contiguous_region(
134 (unsigned long)buf + (i << IO_TLB_SHIFT),
135 get_order(slabs << IO_TLB_SHIFT),
136 dma_bits);
137 } while (rc && dma_bits++ < max_dma_bits);
138 if (rc)
139 return rc;
140
141 i += slabs;
142 } while (i < nslabs);
143 return 0;
144}
145
146void __init xen_swiotlb_init(int verbose)
147{
148 unsigned long bytes;
149 int rc;
150 unsigned long nr_tbl;
151
152 nr_tbl = swioltb_nr_tbl();
153 if (nr_tbl)
154 xen_io_tlb_nslabs = nr_tbl;
155 else {
156 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
157 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
158 }
159
160 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
161
162 /*
163 * Get IO TLB memory from any location.
164 */
165 xen_io_tlb_start = alloc_bootmem(bytes);
166 if (!xen_io_tlb_start)
167 panic("Cannot allocate SWIOTLB buffer");
168
169 xen_io_tlb_end = xen_io_tlb_start + bytes;
170 /*
171 * And replace that memory with pages under 4GB.
172 */
173 rc = xen_swiotlb_fixup(xen_io_tlb_start,
174 bytes,
175 xen_io_tlb_nslabs);
176 if (rc)
177 goto error;
178
179 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
180 swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
181
182 return;
183error:
184 panic("DMA(%d): Failed to exchange pages allocated for DMA with Xen! "\
185 "We either don't have the permission or you do not have enough"\
186 "free memory under 4GB!\n", rc);
187}
188
189void *
190xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
191 dma_addr_t *dma_handle, gfp_t flags)
192{
193 void *ret;
194 int order = get_order(size);
195 u64 dma_mask = DMA_BIT_MASK(32);
196 unsigned long vstart;
197
198 /*
199 * Ignore region specifiers - the kernel's ideas of
200 * pseudo-phys memory layout has nothing to do with the
201 * machine physical layout. We can't allocate highmem
202 * because we can't return a pointer to it.
203 */
204 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
205
206 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
207 return ret;
208
209 vstart = __get_free_pages(flags, order);
210 ret = (void *)vstart;
211
212 if (hwdev && hwdev->coherent_dma_mask)
213 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
214
215 if (ret) {
216 if (xen_create_contiguous_region(vstart, order,
217 fls64(dma_mask)) != 0) {
218 free_pages(vstart, order);
219 return NULL;
220 }
221 memset(ret, 0, size);
222 *dma_handle = virt_to_machine(ret).maddr;
223 }
224 return ret;
225}
226EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
227
228void
229xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
230 dma_addr_t dev_addr)
231{
232 int order = get_order(size);
233
234 if (dma_release_from_coherent(hwdev, order, vaddr))
235 return;
236
237 xen_destroy_contiguous_region((unsigned long)vaddr, order);
238 free_pages((unsigned long)vaddr, order);
239}
240EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
241
242
243/*
244 * Map a single buffer of the indicated size for DMA in streaming mode. The
245 * physical address to use is returned.
246 *
247 * Once the device is given the dma address, the device owns this memory until
248 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
249 */
250dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
251 unsigned long offset, size_t size,
252 enum dma_data_direction dir,
253 struct dma_attrs *attrs)
254{
255 phys_addr_t phys = page_to_phys(page) + offset;
256 dma_addr_t dev_addr = xen_phys_to_bus(phys);
257 void *map;
258
259 BUG_ON(dir == DMA_NONE);
260 /*
261 * If the address happens to be in the device's DMA window,
262 * we can safely return the device addr and not worry about bounce
263 * buffering it.
264 */
265 if (dma_capable(dev, dev_addr, size) &&
266 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
267 return dev_addr;
268
269 /*
270 * Oh well, have to allocate and map a bounce buffer.
271 */
272 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
273 if (!map)
274 return DMA_ERROR_CODE;
275
276 dev_addr = xen_virt_to_bus(map);
277
278 /*
279 * Ensure that the address returned is DMA'ble
280 */
281 if (!dma_capable(dev, dev_addr, size))
282 panic("map_single: bounce buffer is not DMA'ble");
283
284 return dev_addr;
285}
286EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
287
288/*
289 * Unmap a single streaming mode DMA translation. The dma_addr and size must
290 * match what was provided for in a previous xen_swiotlb_map_page call. All
291 * other usages are undefined.
292 *
293 * After this call, reads by the cpu to the buffer are guaranteed to see
294 * whatever the device wrote there.
295 */
296static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
297 size_t size, enum dma_data_direction dir)
298{
299 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
300
301 BUG_ON(dir == DMA_NONE);
302
303 /* NOTE: We use dev_addr here, not paddr! */
304 if (is_xen_swiotlb_buffer(dev_addr)) {
305 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
306 return;
307 }
308
309 if (dir != DMA_FROM_DEVICE)
310 return;
311
312 /*
313 * phys_to_virt doesn't work with hihgmem page but we could
314 * call dma_mark_clean() with hihgmem page here. However, we
315 * are fine since dma_mark_clean() is null on POWERPC. We can
316 * make dma_mark_clean() take a physical address if necessary.
317 */
318 dma_mark_clean(phys_to_virt(paddr), size);
319}
320
321void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
322 size_t size, enum dma_data_direction dir,
323 struct dma_attrs *attrs)
324{
325 xen_unmap_single(hwdev, dev_addr, size, dir);
326}
327EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
328
329/*
330 * Make physical memory consistent for a single streaming mode DMA translation
331 * after a transfer.
332 *
333 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
334 * using the cpu, yet do not wish to teardown the dma mapping, you must
335 * call this function before doing so. At the next point you give the dma
336 * address back to the card, you must first perform a
337 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
338 */
339static void
340xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
341 size_t size, enum dma_data_direction dir,
342 enum dma_sync_target target)
343{
344 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
345
346 BUG_ON(dir == DMA_NONE);
347
348 /* NOTE: We use dev_addr here, not paddr! */
349 if (is_xen_swiotlb_buffer(dev_addr)) {
350 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
351 target);
352 return;
353 }
354
355 if (dir != DMA_FROM_DEVICE)
356 return;
357
358 dma_mark_clean(phys_to_virt(paddr), size);
359}
360
361void
362xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
363 size_t size, enum dma_data_direction dir)
364{
365 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
366}
367EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
368
369void
370xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
371 size_t size, enum dma_data_direction dir)
372{
373 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
374}
375EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
376
377/*
378 * Map a set of buffers described by scatterlist in streaming mode for DMA.
379 * This is the scatter-gather version of the above xen_swiotlb_map_page
380 * interface. Here the scatter gather list elements are each tagged with the
381 * appropriate dma address and length. They are obtained via
382 * sg_dma_{address,length}(SG).
383 *
384 * NOTE: An implementation may be able to use a smaller number of
385 * DMA address/length pairs than there are SG table elements.
386 * (for example via virtual mapping capabilities)
387 * The routine returns the number of addr/length pairs actually
388 * used, at most nents.
389 *
390 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
391 * same here.
392 */
393int
394xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
395 int nelems, enum dma_data_direction dir,
396 struct dma_attrs *attrs)
397{
398 struct scatterlist *sg;
399 int i;
400
401 BUG_ON(dir == DMA_NONE);
402
403 for_each_sg(sgl, sg, nelems, i) {
404 phys_addr_t paddr = sg_phys(sg);
405 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
406
407 if (swiotlb_force ||
408 !dma_capable(hwdev, dev_addr, sg->length) ||
409 range_straddles_page_boundary(paddr, sg->length)) {
410 void *map = swiotlb_tbl_map_single(hwdev,
411 start_dma_addr,
412 sg_phys(sg),
413 sg->length, dir);
414 if (!map) {
415 /* Don't panic here, we expect map_sg users
416 to do proper error handling. */
417 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
418 attrs);
419 sgl[0].dma_length = 0;
420 return DMA_ERROR_CODE;
421 }
422 sg->dma_address = xen_virt_to_bus(map);
423 } else
424 sg->dma_address = dev_addr;
425 sg->dma_length = sg->length;
426 }
427 return nelems;
428}
429EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
430
431int
432xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
433 enum dma_data_direction dir)
434{
435 return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
436}
437EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
438
439/*
440 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
441 * concerning calls here are the same as for swiotlb_unmap_page() above.
442 */
443void
444xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
445 int nelems, enum dma_data_direction dir,
446 struct dma_attrs *attrs)
447{
448 struct scatterlist *sg;
449 int i;
450
451 BUG_ON(dir == DMA_NONE);
452
453 for_each_sg(sgl, sg, nelems, i)
454 xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
455
456}
457EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
458
459void
460xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
461 enum dma_data_direction dir)
462{
463 return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
464}
465EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
466
467/*
468 * Make physical memory consistent for a set of streaming mode DMA translations
469 * after a transfer.
470 *
471 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
472 * and usage.
473 */
474static void
475xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
476 int nelems, enum dma_data_direction dir,
477 enum dma_sync_target target)
478{
479 struct scatterlist *sg;
480 int i;
481
482 for_each_sg(sgl, sg, nelems, i)
483 xen_swiotlb_sync_single(hwdev, sg->dma_address,
484 sg->dma_length, dir, target);
485}
486
487void
488xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
489 int nelems, enum dma_data_direction dir)
490{
491 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
492}
493EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
494
495void
496xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
497 int nelems, enum dma_data_direction dir)
498{
499 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
500}
501EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
502
503int
504xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
505{
506 return !dma_addr;
507}
508EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
509
510/*
511 * Return whether the given device DMA address mask can be supported
512 * properly. For example, if your device can only drive the low 24-bits
513 * during bus mastering, then you would pass 0x00ffffff as the mask to
514 * this function.
515 */
516int
517xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
518{
519 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
520}
521EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
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};