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/*
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#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
37
38#include <linux/bootmem.h>
39#include <linux/dma-direct.h>
40#include <linux/export.h>
41#include <xen/swiotlb-xen.h>
42#include <xen/page.h>
43#include <xen/xen-ops.h>
44#include <xen/hvc-console.h>
45
46#include <asm/dma-mapping.h>
47#include <asm/xen/page-coherent.h>
48
49#include <trace/events/swiotlb.h>
50/*
51 * Used to do a quick range check in swiotlb_tbl_unmap_single and
52 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
53 * API.
54 */
55
56#define XEN_SWIOTLB_ERROR_CODE (~(dma_addr_t)0x0)
57
58static char *xen_io_tlb_start, *xen_io_tlb_end;
59static unsigned long xen_io_tlb_nslabs;
60/*
61 * Quick lookup value of the bus address of the IOTLB.
62 */
63
64static u64 start_dma_addr;
65
66/*
67 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
68 * can be 32bit when dma_addr_t is 64bit leading to a loss in
69 * information if the shift is done before casting to 64bit.
70 */
71static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
72{
73 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
74 dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
75
76 dma |= paddr & ~XEN_PAGE_MASK;
77
78 return dma;
79}
80
81static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
82{
83 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
84 dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
85 phys_addr_t paddr = dma;
86
87 paddr |= baddr & ~XEN_PAGE_MASK;
88
89 return paddr;
90}
91
92static inline dma_addr_t xen_virt_to_bus(void *address)
93{
94 return xen_phys_to_bus(virt_to_phys(address));
95}
96
97static int check_pages_physically_contiguous(unsigned long xen_pfn,
98 unsigned int offset,
99 size_t length)
100{
101 unsigned long next_bfn;
102 int i;
103 int nr_pages;
104
105 next_bfn = pfn_to_bfn(xen_pfn);
106 nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
107
108 for (i = 1; i < nr_pages; i++) {
109 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
110 return 0;
111 }
112 return 1;
113}
114
115static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
116{
117 unsigned long xen_pfn = XEN_PFN_DOWN(p);
118 unsigned int offset = p & ~XEN_PAGE_MASK;
119
120 if (offset + size <= XEN_PAGE_SIZE)
121 return 0;
122 if (check_pages_physically_contiguous(xen_pfn, offset, size))
123 return 0;
124 return 1;
125}
126
127static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
128{
129 unsigned long bfn = XEN_PFN_DOWN(dma_addr);
130 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
131 phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
132
133 /* If the address is outside our domain, it CAN
134 * have the same virtual address as another address
135 * in our domain. Therefore _only_ check address within our domain.
136 */
137 if (pfn_valid(PFN_DOWN(paddr))) {
138 return paddr >= virt_to_phys(xen_io_tlb_start) &&
139 paddr < virt_to_phys(xen_io_tlb_end);
140 }
141 return 0;
142}
143
144static int max_dma_bits = 32;
145
146static int
147xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
148{
149 int i, rc;
150 int dma_bits;
151 dma_addr_t dma_handle;
152 phys_addr_t p = virt_to_phys(buf);
153
154 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
155
156 i = 0;
157 do {
158 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
159
160 do {
161 rc = xen_create_contiguous_region(
162 p + (i << IO_TLB_SHIFT),
163 get_order(slabs << IO_TLB_SHIFT),
164 dma_bits, &dma_handle);
165 } while (rc && dma_bits++ < max_dma_bits);
166 if (rc)
167 return rc;
168
169 i += slabs;
170 } while (i < nslabs);
171 return 0;
172}
173static unsigned long xen_set_nslabs(unsigned long nr_tbl)
174{
175 if (!nr_tbl) {
176 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
177 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
178 } else
179 xen_io_tlb_nslabs = nr_tbl;
180
181 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
182}
183
184enum xen_swiotlb_err {
185 XEN_SWIOTLB_UNKNOWN = 0,
186 XEN_SWIOTLB_ENOMEM,
187 XEN_SWIOTLB_EFIXUP
188};
189
190static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
191{
192 switch (err) {
193 case XEN_SWIOTLB_ENOMEM:
194 return "Cannot allocate Xen-SWIOTLB buffer\n";
195 case XEN_SWIOTLB_EFIXUP:
196 return "Failed to get contiguous memory for DMA from Xen!\n"\
197 "You either: don't have the permissions, do not have"\
198 " enough free memory under 4GB, or the hypervisor memory"\
199 " is too fragmented!";
200 default:
201 break;
202 }
203 return "";
204}
205int __ref xen_swiotlb_init(int verbose, bool early)
206{
207 unsigned long bytes, order;
208 int rc = -ENOMEM;
209 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
210 unsigned int repeat = 3;
211
212 xen_io_tlb_nslabs = swiotlb_nr_tbl();
213retry:
214 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
215 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
216 /*
217 * Get IO TLB memory from any location.
218 */
219 if (early)
220 xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
221 else {
222#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
223#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
224 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
225 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
226 if (xen_io_tlb_start)
227 break;
228 order--;
229 }
230 if (order != get_order(bytes)) {
231 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
232 (PAGE_SIZE << order) >> 20);
233 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
234 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
235 }
236 }
237 if (!xen_io_tlb_start) {
238 m_ret = XEN_SWIOTLB_ENOMEM;
239 goto error;
240 }
241 xen_io_tlb_end = xen_io_tlb_start + bytes;
242 /*
243 * And replace that memory with pages under 4GB.
244 */
245 rc = xen_swiotlb_fixup(xen_io_tlb_start,
246 bytes,
247 xen_io_tlb_nslabs);
248 if (rc) {
249 if (early)
250 free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
251 else {
252 free_pages((unsigned long)xen_io_tlb_start, order);
253 xen_io_tlb_start = NULL;
254 }
255 m_ret = XEN_SWIOTLB_EFIXUP;
256 goto error;
257 }
258 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
259 if (early) {
260 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
261 verbose))
262 panic("Cannot allocate SWIOTLB buffer");
263 rc = 0;
264 } else
265 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
266
267 if (!rc)
268 swiotlb_set_max_segment(PAGE_SIZE);
269
270 return rc;
271error:
272 if (repeat--) {
273 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
274 (xen_io_tlb_nslabs >> 1));
275 pr_info("Lowering to %luMB\n",
276 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
277 goto retry;
278 }
279 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
280 if (early)
281 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
282 else
283 free_pages((unsigned long)xen_io_tlb_start, order);
284 return rc;
285}
286
287static void *
288xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
289 dma_addr_t *dma_handle, gfp_t flags,
290 unsigned long attrs)
291{
292 void *ret;
293 int order = get_order(size);
294 u64 dma_mask = DMA_BIT_MASK(32);
295 phys_addr_t phys;
296 dma_addr_t dev_addr;
297
298 /*
299 * Ignore region specifiers - the kernel's ideas of
300 * pseudo-phys memory layout has nothing to do with the
301 * machine physical layout. We can't allocate highmem
302 * because we can't return a pointer to it.
303 */
304 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
305
306 /* On ARM this function returns an ioremap'ped virtual address for
307 * which virt_to_phys doesn't return the corresponding physical
308 * address. In fact on ARM virt_to_phys only works for kernel direct
309 * mapped RAM memory. Also see comment below.
310 */
311 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
312
313 if (!ret)
314 return ret;
315
316 if (hwdev && hwdev->coherent_dma_mask)
317 dma_mask = hwdev->coherent_dma_mask;
318
319 /* At this point dma_handle is the physical address, next we are
320 * going to set it to the machine address.
321 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
322 * to *dma_handle. */
323 phys = *dma_handle;
324 dev_addr = xen_phys_to_bus(phys);
325 if (((dev_addr + size - 1 <= dma_mask)) &&
326 !range_straddles_page_boundary(phys, size))
327 *dma_handle = dev_addr;
328 else {
329 if (xen_create_contiguous_region(phys, order,
330 fls64(dma_mask), dma_handle) != 0) {
331 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
332 return NULL;
333 }
334 }
335 memset(ret, 0, size);
336 return ret;
337}
338
339static void
340xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
341 dma_addr_t dev_addr, unsigned long attrs)
342{
343 int order = get_order(size);
344 phys_addr_t phys;
345 u64 dma_mask = DMA_BIT_MASK(32);
346
347 if (hwdev && hwdev->coherent_dma_mask)
348 dma_mask = hwdev->coherent_dma_mask;
349
350 /* do not use virt_to_phys because on ARM it doesn't return you the
351 * physical address */
352 phys = xen_bus_to_phys(dev_addr);
353
354 if (((dev_addr + size - 1 <= dma_mask)) ||
355 range_straddles_page_boundary(phys, size))
356 xen_destroy_contiguous_region(phys, order);
357
358 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
359}
360
361/*
362 * Map a single buffer of the indicated size for DMA in streaming mode. The
363 * physical address to use is returned.
364 *
365 * Once the device is given the dma address, the device owns this memory until
366 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
367 */
368static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
369 unsigned long offset, size_t size,
370 enum dma_data_direction dir,
371 unsigned long attrs)
372{
373 phys_addr_t map, phys = page_to_phys(page) + offset;
374 dma_addr_t dev_addr = xen_phys_to_bus(phys);
375
376 BUG_ON(dir == DMA_NONE);
377 /*
378 * If the address happens to be in the device's DMA window,
379 * we can safely return the device addr and not worry about bounce
380 * buffering it.
381 */
382 if (dma_capable(dev, dev_addr, size) &&
383 !range_straddles_page_boundary(phys, size) &&
384 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
385 (swiotlb_force != SWIOTLB_FORCE)) {
386 /* we are not interested in the dma_addr returned by
387 * xen_dma_map_page, only in the potential cache flushes executed
388 * by the function. */
389 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
390 return dev_addr;
391 }
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, start_dma_addr, phys, size, dir,
399 attrs);
400 if (map == SWIOTLB_MAP_ERROR)
401 return XEN_SWIOTLB_ERROR_CODE;
402
403 dev_addr = xen_phys_to_bus(map);
404 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
405 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
406
407 /*
408 * Ensure that the address returned is DMA'ble
409 */
410 if (dma_capable(dev, dev_addr, size))
411 return dev_addr;
412
413 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
414 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
415
416 return XEN_SWIOTLB_ERROR_CODE;
417}
418
419/*
420 * Unmap a single streaming mode DMA translation. The dma_addr and size must
421 * match what was provided for in a previous xen_swiotlb_map_page call. All
422 * other usages are undefined.
423 *
424 * After this call, reads by the cpu to the buffer are guaranteed to see
425 * whatever the device wrote there.
426 */
427static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
428 size_t size, enum dma_data_direction dir,
429 unsigned long attrs)
430{
431 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
432
433 BUG_ON(dir == DMA_NONE);
434
435 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
436
437 /* NOTE: We use dev_addr here, not paddr! */
438 if (is_xen_swiotlb_buffer(dev_addr)) {
439 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
440 return;
441 }
442
443 if (dir != DMA_FROM_DEVICE)
444 return;
445
446 /*
447 * phys_to_virt doesn't work with hihgmem page but we could
448 * call dma_mark_clean() with hihgmem page here. However, we
449 * are fine since dma_mark_clean() is null on POWERPC. We can
450 * make dma_mark_clean() take a physical address if necessary.
451 */
452 dma_mark_clean(phys_to_virt(paddr), size);
453}
454
455static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
456 size_t size, enum dma_data_direction dir,
457 unsigned long attrs)
458{
459 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
460}
461
462/*
463 * Make physical memory consistent for a single streaming mode DMA translation
464 * after a transfer.
465 *
466 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
467 * using the cpu, yet do not wish to teardown the dma mapping, you must
468 * call this function before doing so. At the next point you give the dma
469 * address back to the card, you must first perform a
470 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
471 */
472static void
473xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
474 size_t size, enum dma_data_direction dir,
475 enum dma_sync_target target)
476{
477 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
478
479 BUG_ON(dir == DMA_NONE);
480
481 if (target == SYNC_FOR_CPU)
482 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
483
484 /* NOTE: We use dev_addr here, not paddr! */
485 if (is_xen_swiotlb_buffer(dev_addr))
486 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
487
488 if (target == SYNC_FOR_DEVICE)
489 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
490
491 if (dir != DMA_FROM_DEVICE)
492 return;
493
494 dma_mark_clean(phys_to_virt(paddr), size);
495}
496
497void
498xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
499 size_t size, enum dma_data_direction dir)
500{
501 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
502}
503
504void
505xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
506 size_t size, enum dma_data_direction dir)
507{
508 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
509}
510
511/*
512 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
513 * concerning calls here are the same as for swiotlb_unmap_page() above.
514 */
515static void
516xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
517 int nelems, enum dma_data_direction dir,
518 unsigned long attrs)
519{
520 struct scatterlist *sg;
521 int i;
522
523 BUG_ON(dir == DMA_NONE);
524
525 for_each_sg(sgl, sg, nelems, i)
526 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
527
528}
529
530/*
531 * Map a set of buffers described by scatterlist in streaming mode for DMA.
532 * This is the scatter-gather version of the above xen_swiotlb_map_page
533 * interface. Here the scatter gather list elements are each tagged with the
534 * appropriate dma address and length. They are obtained via
535 * sg_dma_{address,length}(SG).
536 *
537 * NOTE: An implementation may be able to use a smaller number of
538 * DMA address/length pairs than there are SG table elements.
539 * (for example via virtual mapping capabilities)
540 * The routine returns the number of addr/length pairs actually
541 * used, at most nents.
542 *
543 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
544 * same here.
545 */
546static int
547xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
548 int nelems, enum dma_data_direction dir,
549 unsigned long attrs)
550{
551 struct scatterlist *sg;
552 int i;
553
554 BUG_ON(dir == DMA_NONE);
555
556 for_each_sg(sgl, sg, nelems, i) {
557 phys_addr_t paddr = sg_phys(sg);
558 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
559
560 if (swiotlb_force == SWIOTLB_FORCE ||
561 xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
562 !dma_capable(hwdev, dev_addr, sg->length) ||
563 range_straddles_page_boundary(paddr, sg->length)) {
564 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
565 start_dma_addr,
566 sg_phys(sg),
567 sg->length,
568 dir, attrs);
569 if (map == SWIOTLB_MAP_ERROR) {
570 dev_warn(hwdev, "swiotlb buffer is full\n");
571 /* Don't panic here, we expect map_sg users
572 to do proper error handling. */
573 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
574 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
575 attrs);
576 sg_dma_len(sgl) = 0;
577 return 0;
578 }
579 dev_addr = xen_phys_to_bus(map);
580 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
581 dev_addr,
582 map & ~PAGE_MASK,
583 sg->length,
584 dir,
585 attrs);
586 sg->dma_address = dev_addr;
587 } else {
588 /* we are not interested in the dma_addr returned by
589 * xen_dma_map_page, only in the potential cache flushes executed
590 * by the function. */
591 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
592 dev_addr,
593 paddr & ~PAGE_MASK,
594 sg->length,
595 dir,
596 attrs);
597 sg->dma_address = dev_addr;
598 }
599 sg_dma_len(sg) = sg->length;
600 }
601 return nelems;
602}
603
604/*
605 * Make physical memory consistent for a set of streaming mode DMA translations
606 * after a transfer.
607 *
608 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
609 * and usage.
610 */
611static void
612xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
613 int nelems, enum dma_data_direction dir,
614 enum dma_sync_target target)
615{
616 struct scatterlist *sg;
617 int i;
618
619 for_each_sg(sgl, sg, nelems, i)
620 xen_swiotlb_sync_single(hwdev, sg->dma_address,
621 sg_dma_len(sg), dir, target);
622}
623
624static void
625xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
626 int nelems, enum dma_data_direction dir)
627{
628 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
629}
630
631static void
632xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
633 int nelems, enum dma_data_direction dir)
634{
635 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
636}
637
638/*
639 * Return whether the given device DMA address mask can be supported
640 * properly. For example, if your device can only drive the low 24-bits
641 * during bus mastering, then you would pass 0x00ffffff as the mask to
642 * this function.
643 */
644static int
645xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
646{
647 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
648}
649
650/*
651 * Create userspace mapping for the DMA-coherent memory.
652 * This function should be called with the pages from the current domain only,
653 * passing pages mapped from other domains would lead to memory corruption.
654 */
655static int
656xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
657 void *cpu_addr, dma_addr_t dma_addr, size_t size,
658 unsigned long attrs)
659{
660#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
661 if (xen_get_dma_ops(dev)->mmap)
662 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
663 dma_addr, size, attrs);
664#endif
665 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size);
666}
667
668/*
669 * This function should be called with the pages from the current domain only,
670 * passing pages mapped from other domains would lead to memory corruption.
671 */
672static int
673xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
674 void *cpu_addr, dma_addr_t handle, size_t size,
675 unsigned long attrs)
676{
677#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
678 if (xen_get_dma_ops(dev)->get_sgtable) {
679#if 0
680 /*
681 * This check verifies that the page belongs to the current domain and
682 * is not one mapped from another domain.
683 * This check is for debug only, and should not go to production build
684 */
685 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
686 BUG_ON (!page_is_ram(bfn));
687#endif
688 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
689 handle, size, attrs);
690 }
691#endif
692 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size);
693}
694
695static int xen_swiotlb_mapping_error(struct device *dev, dma_addr_t dma_addr)
696{
697 return dma_addr == XEN_SWIOTLB_ERROR_CODE;
698}
699
700const struct dma_map_ops xen_swiotlb_dma_ops = {
701 .alloc = xen_swiotlb_alloc_coherent,
702 .free = xen_swiotlb_free_coherent,
703 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
704 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
705 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
706 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
707 .map_sg = xen_swiotlb_map_sg_attrs,
708 .unmap_sg = xen_swiotlb_unmap_sg_attrs,
709 .map_page = xen_swiotlb_map_page,
710 .unmap_page = xen_swiotlb_unmap_page,
711 .dma_supported = xen_swiotlb_dma_supported,
712 .mmap = xen_swiotlb_dma_mmap,
713 .get_sgtable = xen_swiotlb_get_sgtable,
714 .mapping_error = xen_swiotlb_mapping_error,
715};