/drivers/xen/swiotlb-xen.c

  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		swiotlb_tbl_unmap_single(dev, map, size, dir);
283		dev_addr = 0;
284	}
285	return dev_addr;
286}
287EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
288
289/*
290 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
291 * match what was provided for in a previous xen_swiotlb_map_page call.  All
292 * other usages are undefined.
293 *
294 * After this call, reads by the cpu to the buffer are guaranteed to see
295 * whatever the device wrote there.
296 */
297static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
298			     size_t size, enum dma_data_direction dir)
299{
300	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
301
302	BUG_ON(dir == DMA_NONE);
303
304	/* NOTE: We use dev_addr here, not paddr! */
305	if (is_xen_swiotlb_buffer(dev_addr)) {
306		swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
307		return;
308	}
309
310	if (dir != DMA_FROM_DEVICE)
311		return;
312
313	/*
314	 * phys_to_virt doesn't work with hihgmem page but we could
315	 * call dma_mark_clean() with hihgmem page here. However, we
316	 * are fine since dma_mark_clean() is null on POWERPC. We can
317	 * make dma_mark_clean() take a physical address if necessary.
318	 */
319	dma_mark_clean(phys_to_virt(paddr), size);
320}
321
322void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
323			    size_t size, enum dma_data_direction dir,
324			    struct dma_attrs *attrs)
325{
326	xen_unmap_single(hwdev, dev_addr, size, dir);
327}
328EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
329
330/*
331 * Make physical memory consistent for a single streaming mode DMA translation
332 * after a transfer.
333 *
334 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
335 * using the cpu, yet do not wish to teardown the dma mapping, you must
336 * call this function before doing so.  At the next point you give the dma
337 * address back to the card, you must first perform a
338 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
339 */
340static void
341xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
342			size_t size, enum dma_data_direction dir,
343			enum dma_sync_target target)
344{
345	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
346
347	BUG_ON(dir == DMA_NONE);
348
349	/* NOTE: We use dev_addr here, not paddr! */
350	if (is_xen_swiotlb_buffer(dev_addr)) {
351		swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
352				       target);
353		return;
354	}
355
356	if (dir != DMA_FROM_DEVICE)
357		return;
358
359	dma_mark_clean(phys_to_virt(paddr), size);
360}
361
362void
363xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
364				size_t size, enum dma_data_direction dir)
365{
366	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
367}
368EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
369
370void
371xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
372				   size_t size, enum dma_data_direction dir)
373{
374	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
375}
376EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
377
378/*
379 * Map a set of buffers described by scatterlist in streaming mode for DMA.
380 * This is the scatter-gather version of the above xen_swiotlb_map_page
381 * interface.  Here the scatter gather list elements are each tagged with the
382 * appropriate dma address and length.  They are obtained via
383 * sg_dma_{address,length}(SG).
384 *
385 * NOTE: An implementation may be able to use a smaller number of
386 *       DMA address/length pairs than there are SG table elements.
387 *       (for example via virtual mapping capabilities)
388 *       The routine returns the number of addr/length pairs actually
389 *       used, at most nents.
390 *
391 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
392 * same here.
393 */
394int
395xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
396			 int nelems, enum dma_data_direction dir,
397			 struct dma_attrs *attrs)
398{
399	struct scatterlist *sg;
400	int i;
401
402	BUG_ON(dir == DMA_NONE);
403
404	for_each_sg(sgl, sg, nelems, i) {
405		phys_addr_t paddr = sg_phys(sg);
406		dma_addr_t dev_addr = xen_phys_to_bus(paddr);
407
408		if (swiotlb_force ||
409		    !dma_capable(hwdev, dev_addr, sg->length) ||
410		    range_straddles_page_boundary(paddr, sg->length)) {
411			void *map = swiotlb_tbl_map_single(hwdev,
412							   start_dma_addr,
413							   sg_phys(sg),
414							   sg->length, dir);
415			if (!map) {
416				/* Don't panic here, we expect map_sg users
417				   to do proper error handling. */
418				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
419							   attrs);
420				sgl[0].dma_length = 0;
421				return DMA_ERROR_CODE;
422			}
423			sg->dma_address = xen_virt_to_bus(map);
424		} else
425			sg->dma_address = dev_addr;
426		sg->dma_length = sg->length;
427	}
428	return nelems;
429}
430EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
431
432int
433xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
434		   enum dma_data_direction dir)
435{
436	return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
437}
438EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
439
440/*
441 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
442 * concerning calls here are the same as for swiotlb_unmap_page() above.
443 */
444void
445xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
446			   int nelems, enum dma_data_direction dir,
447			   struct dma_attrs *attrs)
448{
449	struct scatterlist *sg;
450	int i;
451
452	BUG_ON(dir == DMA_NONE);
453
454	for_each_sg(sgl, sg, nelems, i)
455		xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
456
457}
458EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
459
460void
461xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
462		     enum dma_data_direction dir)
463{
464	return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
465}
466EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
467
468/*
469 * Make physical memory consistent for a set of streaming mode DMA translations
470 * after a transfer.
471 *
472 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
473 * and usage.
474 */
475static void
476xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
477		    int nelems, enum dma_data_direction dir,
478		    enum dma_sync_target target)
479{
480	struct scatterlist *sg;
481	int i;
482
483	for_each_sg(sgl, sg, nelems, i)
484		xen_swiotlb_sync_single(hwdev, sg->dma_address,
485					sg->dma_length, dir, target);
486}
487
488void
489xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
490			    int nelems, enum dma_data_direction dir)
491{
492	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
493}
494EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
495
496void
497xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
498			       int nelems, enum dma_data_direction dir)
499{
500	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
501}
502EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
503
504int
505xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
506{
507	return !dma_addr;
508}
509EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
510
511/*
512 * Return whether the given device DMA address mask can be supported
513 * properly.  For example, if your device can only drive the low 24-bits
514 * during bus mastering, then you would pass 0x00ffffff as the mask to
515 * this function.
516 */
517int
518xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
519{
520	return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
521}
522EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);