swiotlb-xen.c

  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
 40#include <trace/events/swiotlb.h>
 41#define MAX_DMA_BITS 32
 42
 43/*
 44 * Quick lookup value of the bus address of the IOTLB.
 45 */
 46
 47static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr)
 48{
 49	unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
 50	phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT;
 51
 52	baddr |= paddr & ~XEN_PAGE_MASK;
 53	return baddr;
 54}
 55
 56static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr)
 57{
 58	return phys_to_dma(dev, xen_phys_to_bus(dev, paddr));
 59}
 60
 61static inline phys_addr_t xen_bus_to_phys(struct device *dev,
 62					  phys_addr_t baddr)
 63{
 64	unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
 65	phys_addr_t paddr = (xen_pfn << XEN_PAGE_SHIFT) |
 66			    (baddr & ~XEN_PAGE_MASK);
 67
 68	return paddr;
 69}
 70
 71static inline phys_addr_t xen_dma_to_phys(struct device *dev,
 72					  dma_addr_t dma_addr)
 73{
 74	return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr));
 75}
 76
 77static inline bool range_requires_alignment(phys_addr_t p, size_t size)
 78{
 79	phys_addr_t algn = 1ULL << (get_order(size) + PAGE_SHIFT);
 80	phys_addr_t bus_addr = pfn_to_bfn(XEN_PFN_DOWN(p)) << XEN_PAGE_SHIFT;
 81
 82	return IS_ALIGNED(p, algn) && !IS_ALIGNED(bus_addr, algn);
 83}
 84
 85static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
 86{
 87	unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p);
 88	unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size);
 89
 90	next_bfn = pfn_to_bfn(xen_pfn);
 91
 92	for (i = 1; i < nr_pages; i++)
 93		if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
 94			return 1;
 95
 96	return 0;
 97}
 98
 99static struct io_tlb_pool *xen_swiotlb_find_pool(struct device *dev,
100						 dma_addr_t dma_addr)
101{
102	unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr));
103	unsigned long xen_pfn = bfn_to_local_pfn(bfn);
104	phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT;
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_DOWN(paddr)))
111		return swiotlb_find_pool(dev, paddr);
112	return NULL;
113}
114
115#ifdef CONFIG_X86
116int __init xen_swiotlb_fixup(void *buf, unsigned long nslabs)
117{
118	int rc;
119	unsigned int order = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT);
120	unsigned int i, dma_bits = order + PAGE_SHIFT;
121	dma_addr_t dma_handle;
122	phys_addr_t p = virt_to_phys(buf);
123
124	BUILD_BUG_ON(IO_TLB_SEGSIZE & (IO_TLB_SEGSIZE - 1));
125	BUG_ON(nslabs % IO_TLB_SEGSIZE);
126
127	i = 0;
128	do {
129		do {
130			rc = xen_create_contiguous_region(
131				p + (i << IO_TLB_SHIFT), order,
132				dma_bits, &dma_handle);
133		} while (rc && dma_bits++ < MAX_DMA_BITS);
134		if (rc)
135			return rc;
136
137		i += IO_TLB_SEGSIZE;
138	} while (i < nslabs);
139	return 0;
140}
141
142static void *
143xen_swiotlb_alloc_coherent(struct device *dev, size_t size,
144		dma_addr_t *dma_handle, gfp_t flags, unsigned long attrs)
145{
146	u64 dma_mask = dev->coherent_dma_mask;
147	int order = get_order(size);
148	phys_addr_t phys;
149	void *ret;
150
151	/* Align the allocation to the Xen page size */
152	size = ALIGN(size, XEN_PAGE_SIZE);
153
154	ret = (void *)__get_free_pages(flags, get_order(size));
155	if (!ret)
156		return ret;
157	phys = virt_to_phys(ret);
158
159	*dma_handle = xen_phys_to_dma(dev, phys);
160	if (*dma_handle + size - 1 > dma_mask ||
161	    range_straddles_page_boundary(phys, size) ||
162	    range_requires_alignment(phys, size)) {
163		if (xen_create_contiguous_region(phys, order, fls64(dma_mask),
164				dma_handle) != 0)
165			goto out_free_pages;
166		SetPageXenRemapped(virt_to_page(ret));
167	}
168
169	memset(ret, 0, size);
170	return ret;
171
172out_free_pages:
173	free_pages((unsigned long)ret, get_order(size));
174	return NULL;
175}
176
177static void
178xen_swiotlb_free_coherent(struct device *dev, size_t size, void *vaddr,
179		dma_addr_t dma_handle, unsigned long attrs)
180{
181	phys_addr_t phys = virt_to_phys(vaddr);
182	int order = get_order(size);
183
184	/* Convert the size to actually allocated. */
185	size = ALIGN(size, XEN_PAGE_SIZE);
186
187	if (WARN_ON_ONCE(dma_handle + size - 1 > dev->coherent_dma_mask) ||
188	    WARN_ON_ONCE(range_straddles_page_boundary(phys, size) ||
189			 range_requires_alignment(phys, size)))
190	    	return;
191
192	if (TestClearPageXenRemapped(virt_to_page(vaddr)))
193		xen_destroy_contiguous_region(phys, order);
194	free_pages((unsigned long)vaddr, get_order(size));
195}
196#endif /* CONFIG_X86 */
197
198/*
199 * Map a single buffer of the indicated size for DMA in streaming mode.  The
200 * physical address to use is returned.
201 *
202 * Once the device is given the dma address, the device owns this memory until
203 * either xen_swiotlb_unmap_phys or xen_swiotlb_dma_sync_single is performed.
204 */
205static dma_addr_t xen_swiotlb_map_phys(struct device *dev, phys_addr_t phys,
206				size_t size, enum dma_data_direction dir,
207				unsigned long attrs)
208{
209	dma_addr_t dev_addr;
210	phys_addr_t map;
211
212	BUG_ON(dir == DMA_NONE);
213
214	if (attrs & DMA_ATTR_MMIO) {
215		if (unlikely(!dma_capable(dev, phys, size, false))) {
216			dev_err_once(
217				dev,
218				"DMA addr %pa+%zu overflow (mask %llx, bus limit %llx).\n",
219				&phys, size, *dev->dma_mask,
220				dev->bus_dma_limit);
221			WARN_ON_ONCE(1);
222			return DMA_MAPPING_ERROR;
223		}
224		return phys;
225	}
226
227	dev_addr = xen_phys_to_dma(dev, phys);
228
229	/*
230	 * If the address happens to be in the device's DMA window,
231	 * we can safely return the device addr and not worry about bounce
232	 * buffering it.
233	 */
234	if (dma_capable(dev, dev_addr, size, true) &&
235	    !dma_kmalloc_needs_bounce(dev, size, dir) &&
236	    !range_straddles_page_boundary(phys, size) &&
237		!xen_arch_need_swiotlb(dev, phys, dev_addr) &&
238		!is_swiotlb_force_bounce(dev))
239		goto done;
240
241	/*
242	 * Oh well, have to allocate and map a bounce buffer.
243	 */
244	trace_swiotlb_bounced(dev, dev_addr, size);
245
246	map = swiotlb_tbl_map_single(dev, phys, size, 0, dir, attrs);
247	if (map == (phys_addr_t)DMA_MAPPING_ERROR)
248		return DMA_MAPPING_ERROR;
249
250	phys = map;
251	dev_addr = xen_phys_to_dma(dev, map);
252
253	/*
254	 * Ensure that the address returned is DMA'ble
255	 */
256	if (unlikely(!dma_capable(dev, dev_addr, size, true))) {
257		__swiotlb_tbl_unmap_single(dev, map, size, dir,
258				attrs | DMA_ATTR_SKIP_CPU_SYNC,
259				swiotlb_find_pool(dev, map));
260		return DMA_MAPPING_ERROR;
261	}
262
263done:
264	if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
265		if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr))))
266			arch_sync_dma_for_device(phys, size, dir);
267		else
268			xen_dma_sync_for_device(dev, dev_addr, size, dir);
269	}
270	return dev_addr;
271}
272
273/*
274 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
275 * match what was provided for in a previous xen_swiotlb_map_phys call.  All
276 * other usages are undefined.
277 *
278 * After this call, reads by the cpu to the buffer are guaranteed to see
279 * whatever the device wrote there.
280 */
281static void xen_swiotlb_unmap_phys(struct device *hwdev, dma_addr_t dev_addr,
282		size_t size, enum dma_data_direction dir, unsigned long attrs)
283{
284	phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr);
285	struct io_tlb_pool *pool;
286
287	BUG_ON(dir == DMA_NONE);
288
289	if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
290		if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr))))
291			arch_sync_dma_for_cpu(paddr, size, dir);
292		else
293			xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir);
294	}
295
296	/* NOTE: We use dev_addr here, not paddr! */
297	pool = xen_swiotlb_find_pool(hwdev, dev_addr);
298	if (pool)
299		__swiotlb_tbl_unmap_single(hwdev, paddr, size, dir,
300					   attrs, pool);
301}
302
303static void
304xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
305		size_t size, enum dma_data_direction dir)
306{
307	phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
308	struct io_tlb_pool *pool;
309
310	if (!dev_is_dma_coherent(dev)) {
311		if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
312			arch_sync_dma_for_cpu(paddr, size, dir);
313		else
314			xen_dma_sync_for_cpu(dev, dma_addr, size, dir);
315	}
316
317	pool = xen_swiotlb_find_pool(dev, dma_addr);
318	if (pool)
319		__swiotlb_sync_single_for_cpu(dev, paddr, size, dir, pool);
320}
321
322static void
323xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
324		size_t size, enum dma_data_direction dir)
325{
326	phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
327	struct io_tlb_pool *pool;
328
329	pool = xen_swiotlb_find_pool(dev, dma_addr);
330	if (pool)
331		__swiotlb_sync_single_for_device(dev, paddr, size, dir, pool);
332
333	if (!dev_is_dma_coherent(dev)) {
334		if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
335			arch_sync_dma_for_device(paddr, size, dir);
336		else
337			xen_dma_sync_for_device(dev, dma_addr, size, dir);
338	}
339}
340
341/*
342 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
343 * concerning calls here are the same as for swiotlb_unmap_phys() above.
344 */
345static void
346xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
347		enum dma_data_direction dir, unsigned long attrs)
348{
349	struct scatterlist *sg;
350	int i;
351
352	BUG_ON(dir == DMA_NONE);
353
354	for_each_sg(sgl, sg, nelems, i)
355		xen_swiotlb_unmap_phys(hwdev, sg->dma_address, sg_dma_len(sg),
356				dir, attrs);
357
358}
359
360static int
361xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
362		enum dma_data_direction dir, unsigned long attrs)
363{
364	struct scatterlist *sg;
365	int i;
366
367	BUG_ON(dir == DMA_NONE);
368
369	for_each_sg(sgl, sg, nelems, i) {
370		sg->dma_address = xen_swiotlb_map_phys(dev, sg_phys(sg),
371				sg->length, dir, attrs);
372		if (sg->dma_address == DMA_MAPPING_ERROR)
373			goto out_unmap;
374		sg_dma_len(sg) = sg->length;
375	}
376
377	return nelems;
378out_unmap:
379	xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
380	sg_dma_len(sgl) = 0;
381	return -EIO;
382}
383
384static void
385xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
386			    int nelems, enum dma_data_direction dir)
387{
388	struct scatterlist *sg;
389	int i;
390
391	for_each_sg(sgl, sg, nelems, i) {
392		xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
393				sg->length, dir);
394	}
395}
396
397static void
398xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
399			       int nelems, enum dma_data_direction dir)
400{
401	struct scatterlist *sg;
402	int i;
403
404	for_each_sg(sgl, sg, nelems, i) {
405		xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
406				sg->length, dir);
407	}
408}
409
410/*
411 * Return whether the given device DMA address mask can be supported
412 * properly.  For example, if your device can only drive the low 24-bits
413 * during bus mastering, then you would pass 0x00ffffff as the mask to
414 * this function.
415 */
416static int
417xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
418{
419	return xen_phys_to_dma(hwdev, default_swiotlb_limit()) <= mask;
420}
421
422const struct dma_map_ops xen_swiotlb_dma_ops = {
423#ifdef CONFIG_X86
424	.alloc = xen_swiotlb_alloc_coherent,
425	.free = xen_swiotlb_free_coherent,
426#else
427	.alloc = dma_direct_alloc,
428	.free = dma_direct_free,
429#endif
430	.sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
431	.sync_single_for_device = xen_swiotlb_sync_single_for_device,
432	.sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
433	.sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
434	.map_sg = xen_swiotlb_map_sg,
435	.unmap_sg = xen_swiotlb_unmap_sg,
436	.map_phys = xen_swiotlb_map_phys,
437	.unmap_phys = xen_swiotlb_unmap_phys,
438	.dma_supported = xen_swiotlb_dma_supported,
439	.mmap = dma_common_mmap,
440	.get_sgtable = dma_common_get_sgtable,
441	.alloc_pages_op = dma_common_alloc_pages,
442	.free_pages = dma_common_free_pages,
443	.max_mapping_size = swiotlb_max_mapping_size,
444};