Merge tag 'block-5.11-2021-01-10' of git://git.kernel.dk/linux-block
[linux/fpc-iii.git] / drivers / xen / swiotlb-xen.c
blob2b385c1b4a99cbaadf9f5fec4c0d822b96f95da1
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright 2010
4 * by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
6 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
8 * PV guests under Xen are running in an non-contiguous memory architecture.
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).
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.
27 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
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>
38 #include <asm/dma-mapping.h>
39 #include <asm/xen/page-coherent.h>
41 #include <trace/events/swiotlb.h>
42 #define MAX_DMA_BITS 32
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.
49 static char *xen_io_tlb_start, *xen_io_tlb_end;
50 static unsigned long xen_io_tlb_nslabs;
52 * Quick lookup value of the bus address of the IOTLB.
55 static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr)
57 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
58 phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT;
60 baddr |= paddr & ~XEN_PAGE_MASK;
61 return baddr;
64 static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr)
66 return phys_to_dma(dev, xen_phys_to_bus(dev, paddr));
69 static inline phys_addr_t xen_bus_to_phys(struct device *dev,
70 phys_addr_t baddr)
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);
76 return paddr;
79 static inline phys_addr_t xen_dma_to_phys(struct device *dev,
80 dma_addr_t dma_addr)
82 return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr));
85 static inline dma_addr_t xen_virt_to_bus(struct device *dev, void *address)
87 return xen_phys_to_dma(dev, virt_to_phys(address));
90 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
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);
95 next_bfn = pfn_to_bfn(xen_pfn);
97 for (i = 1; i < nr_pages; i++)
98 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
99 return 1;
101 return 0;
104 static int is_xen_swiotlb_buffer(struct device *dev, dma_addr_t dma_addr)
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;
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.
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);
118 return 0;
121 static int
122 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
124 int i, rc;
125 int dma_bits;
126 dma_addr_t dma_handle;
127 phys_addr_t p = virt_to_phys(buf);
129 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
131 i = 0;
132 do {
133 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
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;
144 i += slabs;
145 } while (i < nslabs);
146 return 0;
148 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
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;
156 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
159 enum xen_swiotlb_err {
160 XEN_SWIOTLB_UNKNOWN = 0,
161 XEN_SWIOTLB_ENOMEM,
162 XEN_SWIOTLB_EFIXUP
165 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
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;
178 return "";
180 int __ref xen_swiotlb_init(int verbose, bool early)
182 unsigned long bytes, order;
183 int rc = -ENOMEM;
184 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
185 unsigned int repeat = 3;
187 xen_io_tlb_nslabs = swiotlb_nr_tbl();
188 retry:
189 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
190 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
193 * IO TLB memory already allocated. Just use it.
195 if (io_tlb_start != 0) {
196 xen_io_tlb_start = phys_to_virt(io_tlb_start);
197 goto end;
201 * Get IO TLB memory from any location.
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--;
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;
225 if (!xen_io_tlb_start) {
226 m_ret = XEN_SWIOTLB_ENOMEM;
227 goto error;
230 * And replace that memory with pages under 4GB.
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;
243 m_ret = XEN_SWIOTLB_EFIXUP;
244 goto error;
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);
254 end:
255 xen_io_tlb_end = xen_io_tlb_start + bytes;
256 if (!rc)
257 swiotlb_set_max_segment(PAGE_SIZE);
259 return rc;
260 error:
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;
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;
276 static void *
277 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
278 dma_addr_t *dma_handle, gfp_t flags,
279 unsigned long attrs)
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;
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.
293 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
295 /* Convert the size to actually allocated. */
296 size = 1UL << (order + XEN_PAGE_SHIFT);
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.
303 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
305 if (!ret)
306 return ret;
308 if (hwdev && hwdev->coherent_dma_mask)
309 dma_mask = hwdev->coherent_dma_mask;
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;
326 *dma_handle = phys_to_dma(hwdev, *dma_handle);
327 SetPageXenRemapped(virt_to_page(ret));
329 memset(ret, 0, size);
330 return ret;
333 static void
334 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
335 dma_addr_t dev_addr, unsigned long attrs)
337 int order = get_order(size);
338 phys_addr_t phys;
339 u64 dma_mask = DMA_BIT_MASK(32);
340 struct page *page;
342 if (hwdev && hwdev->coherent_dma_mask)
343 dma_mask = hwdev->coherent_dma_mask;
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);
349 /* Convert the size to actually allocated. */
350 size = 1UL << (order + XEN_PAGE_SHIFT);
352 if (is_vmalloc_addr(vaddr))
353 page = vmalloc_to_page(vaddr);
354 else
355 page = virt_to_page(vaddr);
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);
362 xen_free_coherent_pages(hwdev, size, vaddr, phys_to_dma(hwdev, phys),
363 attrs);
367 * Map a single buffer of the indicated size for DMA in streaming mode. The
368 * physical address to use is returned.
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.
373 static 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)
378 phys_addr_t map, phys = page_to_phys(page) + offset;
379 dma_addr_t dev_addr = xen_phys_to_dma(dev, phys);
381 BUG_ON(dir == DMA_NONE);
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.
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;
394 * Oh well, have to allocate and map a bounce buffer.
396 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
398 map = swiotlb_tbl_map_single(dev, phys, size, size, dir, attrs);
399 if (map == (phys_addr_t)DMA_MAPPING_ERROR)
400 return DMA_MAPPING_ERROR;
402 phys = map;
403 dev_addr = xen_phys_to_dma(dev, map);
406 * Ensure that the address returned is DMA'ble
408 if (unlikely(!dma_capable(dev, dev_addr, size, true))) {
409 swiotlb_tbl_unmap_single(dev, map, size, size, dir,
410 attrs | DMA_ATTR_SKIP_CPU_SYNC);
411 return DMA_MAPPING_ERROR;
414 done:
415 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
416 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr))))
417 arch_sync_dma_for_device(phys, size, dir);
418 else
419 xen_dma_sync_for_device(dev, dev_addr, size, dir);
421 return dev_addr;
425 * Unmap a single streaming mode DMA translation. The dma_addr and size must
426 * match what was provided for in a previous xen_swiotlb_map_page call. All
427 * other usages are undefined.
429 * After this call, reads by the cpu to the buffer are guaranteed to see
430 * whatever the device wrote there.
432 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
433 size_t size, enum dma_data_direction dir, unsigned long attrs)
435 phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr);
437 BUG_ON(dir == DMA_NONE);
439 if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
440 if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr))))
441 arch_sync_dma_for_cpu(paddr, size, dir);
442 else
443 xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir);
446 /* NOTE: We use dev_addr here, not paddr! */
447 if (is_xen_swiotlb_buffer(hwdev, dev_addr))
448 swiotlb_tbl_unmap_single(hwdev, paddr, size, size, dir, attrs);
451 static void
452 xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
453 size_t size, enum dma_data_direction dir)
455 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
457 if (!dev_is_dma_coherent(dev)) {
458 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
459 arch_sync_dma_for_cpu(paddr, size, dir);
460 else
461 xen_dma_sync_for_cpu(dev, dma_addr, size, dir);
464 if (is_xen_swiotlb_buffer(dev, dma_addr))
465 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_CPU);
468 static void
469 xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
470 size_t size, enum dma_data_direction dir)
472 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
474 if (is_xen_swiotlb_buffer(dev, dma_addr))
475 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
477 if (!dev_is_dma_coherent(dev)) {
478 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
479 arch_sync_dma_for_device(paddr, size, dir);
480 else
481 xen_dma_sync_for_device(dev, dma_addr, size, dir);
486 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
487 * concerning calls here are the same as for swiotlb_unmap_page() above.
489 static void
490 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
491 enum dma_data_direction dir, unsigned long attrs)
493 struct scatterlist *sg;
494 int i;
496 BUG_ON(dir == DMA_NONE);
498 for_each_sg(sgl, sg, nelems, i)
499 xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg),
500 dir, attrs);
504 static int
505 xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
506 enum dma_data_direction dir, unsigned long attrs)
508 struct scatterlist *sg;
509 int i;
511 BUG_ON(dir == DMA_NONE);
513 for_each_sg(sgl, sg, nelems, i) {
514 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
515 sg->offset, sg->length, dir, attrs);
516 if (sg->dma_address == DMA_MAPPING_ERROR)
517 goto out_unmap;
518 sg_dma_len(sg) = sg->length;
521 return nelems;
522 out_unmap:
523 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
524 sg_dma_len(sgl) = 0;
525 return 0;
528 static void
529 xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
530 int nelems, enum dma_data_direction dir)
532 struct scatterlist *sg;
533 int i;
535 for_each_sg(sgl, sg, nelems, i) {
536 xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
537 sg->length, dir);
541 static void
542 xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
543 int nelems, enum dma_data_direction dir)
545 struct scatterlist *sg;
546 int i;
548 for_each_sg(sgl, sg, nelems, i) {
549 xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
550 sg->length, dir);
555 * Return whether the given device DMA address mask can be supported
556 * properly. For example, if your device can only drive the low 24-bits
557 * during bus mastering, then you would pass 0x00ffffff as the mask to
558 * this function.
560 static int
561 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
563 return xen_virt_to_bus(hwdev, xen_io_tlb_end - 1) <= mask;
566 const struct dma_map_ops xen_swiotlb_dma_ops = {
567 .alloc = xen_swiotlb_alloc_coherent,
568 .free = xen_swiotlb_free_coherent,
569 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
570 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
571 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
572 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
573 .map_sg = xen_swiotlb_map_sg,
574 .unmap_sg = xen_swiotlb_unmap_sg,
575 .map_page = xen_swiotlb_map_page,
576 .unmap_page = xen_swiotlb_unmap_page,
577 .dma_supported = xen_swiotlb_dma_supported,
578 .mmap = dma_common_mmap,
579 .get_sgtable = dma_common_get_sgtable,
580 .alloc_pages = dma_common_alloc_pages,
581 .free_pages = dma_common_free_pages,