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 <linux/export.h>
 39#include <xen/swiotlb-xen.h>
 40#include <xen/page.h>
 41#include <xen/xen-ops.h>
 42#include <xen/hvc-console.h>
 43/*
 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.
 47 */
 48
 49static char *xen_io_tlb_start, *xen_io_tlb_end;
 50static unsigned long xen_io_tlb_nslabs;
 51/*
 52 * Quick lookup value of the bus address of the IOTLB.
 53 */
 54
 55u64 start_dma_addr;
 56
 57static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
 58{
 59	return phys_to_machine(XPADDR(paddr)).maddr;
 60}
 61
 62static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
 63{
 64	return machine_to_phys(XMADDR(baddr)).paddr;
 65}
 66
 67static dma_addr_t xen_virt_to_bus(void *address)
 68{
 69	return xen_phys_to_bus(virt_to_phys(address));
 70}
 71
 72static int check_pages_physically_contiguous(unsigned long pfn,
 73					     unsigned int offset,
 74					     size_t length)
 75{
 76	unsigned long next_mfn;
 77	int i;
 78	int nr_pages;
 79
 80	next_mfn = pfn_to_mfn(pfn);
 81	nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
 82
 83	for (i = 1; i < nr_pages; i++) {
 84		if (pfn_to_mfn(++pfn) != ++next_mfn)
 85			return 0;
 86	}
 87	return 1;
 88}
 89
 90static int range_straddles_page_boundary(phys_addr_t p, size_t size)
 91{
 92	unsigned long pfn = PFN_DOWN(p);
 93	unsigned int offset = p & ~PAGE_MASK;
 94
 95	if (offset + size <= PAGE_SIZE)
 96		return 0;
 97	if (check_pages_physically_contiguous(pfn, offset, size))
 98		return 0;
 99	return 1;
100}
101
102static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
103{
104	unsigned long mfn = PFN_DOWN(dma_addr);
105	unsigned long pfn = mfn_to_local_pfn(mfn);
106	phys_addr_t paddr;
107
108	/* If the address is outside our domain, it CAN
109	 * have the same virtual address as another address
110	 * in our domain. Therefore _only_ check address within our domain.
111	 */
112	if (pfn_valid(pfn)) {
113		paddr = PFN_PHYS(pfn);
114		return paddr >= virt_to_phys(xen_io_tlb_start) &&
115		       paddr < virt_to_phys(xen_io_tlb_end);
116	}
117	return 0;
118}
119
120static int max_dma_bits = 32;
121
122static int
123xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
124{
125	int i, rc;
126	int dma_bits;
127
128	dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
129
130	i = 0;
131	do {
132		int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
133
134		do {
135			rc = xen_create_contiguous_region(
136				(unsigned long)buf + (i << IO_TLB_SHIFT),
137				get_order(slabs << IO_TLB_SHIFT),
138				dma_bits);
139		} while (rc && dma_bits++ < max_dma_bits);
140		if (rc)
141			return rc;
142
143		i += slabs;
144	} while (i < nslabs);
145	return 0;
146}
147
148void __init xen_swiotlb_init(int verbose)
149{
150	unsigned long bytes;
151	int rc = -ENOMEM;
152	unsigned long nr_tbl;
153	char *m = NULL;
154	unsigned int repeat = 3;
155
156	nr_tbl = swiotlb_nr_tbl();
157	if (nr_tbl)
158		xen_io_tlb_nslabs = nr_tbl;
159	else {
160		xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
161		xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
162	}
163retry:
164	bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
165
166	/*
167	 * Get IO TLB memory from any location.
168	 */
169	xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
170	if (!xen_io_tlb_start) {
171		m = "Cannot allocate Xen-SWIOTLB buffer!\n";
172		goto error;
173	}
174	xen_io_tlb_end = xen_io_tlb_start + bytes;
175	/*
176	 * And replace that memory with pages under 4GB.
177	 */
178	rc = xen_swiotlb_fixup(xen_io_tlb_start,
179			       bytes,
180			       xen_io_tlb_nslabs);
181	if (rc) {
182		free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
183		m = "Failed to get contiguous memory for DMA from Xen!\n"\
184		    "You either: don't have the permissions, do not have"\
185		    " enough free memory under 4GB, or the hypervisor memory"\
186		    "is too fragmented!";
187		goto error;
188	}
189	start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
190	swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
191
192	return;
193error:
194	if (repeat--) {
195		xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
196					(xen_io_tlb_nslabs >> 1));
197		printk(KERN_INFO "Xen-SWIOTLB: Lowering to %luMB\n",
198		      (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
199		goto retry;
200	}
201	xen_raw_printk("%s (rc:%d)", m, rc);
202	panic("%s (rc:%d)", m, rc);
203}
204
205void *
206xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
207			   dma_addr_t *dma_handle, gfp_t flags,
208			   struct dma_attrs *attrs)
209{
210	void *ret;
211	int order = get_order(size);
212	u64 dma_mask = DMA_BIT_MASK(32);
213	unsigned long vstart;
214	phys_addr_t phys;
215	dma_addr_t dev_addr;
216
217	/*
218	* Ignore region specifiers - the kernel's ideas of
219	* pseudo-phys memory layout has nothing to do with the
220	* machine physical layout.  We can't allocate highmem
221	* because we can't return a pointer to it.
222	*/
223	flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
224
225	if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
226		return ret;
227
228	vstart = __get_free_pages(flags, order);
229	ret = (void *)vstart;
230
231	if (!ret)
232		return ret;
233
234	if (hwdev && hwdev->coherent_dma_mask)
235		dma_mask = dma_alloc_coherent_mask(hwdev, flags);
236
237	phys = virt_to_phys(ret);
238	dev_addr = xen_phys_to_bus(phys);
239	if (((dev_addr + size - 1 <= dma_mask)) &&
240	    !range_straddles_page_boundary(phys, size))
241		*dma_handle = dev_addr;
242	else {
243		if (xen_create_contiguous_region(vstart, order,
244						 fls64(dma_mask)) != 0) {
245			free_pages(vstart, order);
246			return NULL;
247		}
248		*dma_handle = virt_to_machine(ret).maddr;
249	}
250	memset(ret, 0, size);
251	return ret;
252}
253EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
254
255void
256xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
257			  dma_addr_t dev_addr, struct dma_attrs *attrs)
258{
259	int order = get_order(size);
260	phys_addr_t phys;
261	u64 dma_mask = DMA_BIT_MASK(32);
262
263	if (dma_release_from_coherent(hwdev, order, vaddr))
264		return;
265
266	if (hwdev && hwdev->coherent_dma_mask)
267		dma_mask = hwdev->coherent_dma_mask;
268
269	phys = virt_to_phys(vaddr);
270
271	if (((dev_addr + size - 1 > dma_mask)) ||
272	    range_straddles_page_boundary(phys, size))
273		xen_destroy_contiguous_region((unsigned long)vaddr, order);
274
275	free_pages((unsigned long)vaddr, order);
276}
277EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
278
279
280/*
281 * Map a single buffer of the indicated size for DMA in streaming mode.  The
282 * physical address to use is returned.
283 *
284 * Once the device is given the dma address, the device owns this memory until
285 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
286 */
287dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
288				unsigned long offset, size_t size,
289				enum dma_data_direction dir,
290				struct dma_attrs *attrs)
291{
292	phys_addr_t phys = page_to_phys(page) + offset;
293	dma_addr_t dev_addr = xen_phys_to_bus(phys);
294	void *map;
295
296	BUG_ON(dir == DMA_NONE);
297	/*
298	 * If the address happens to be in the device's DMA window,
299	 * we can safely return the device addr and not worry about bounce
300	 * buffering it.
301	 */
302	if (dma_capable(dev, dev_addr, size) &&
303	    !range_straddles_page_boundary(phys, size) && !swiotlb_force)
304		return dev_addr;
305
306	/*
307	 * Oh well, have to allocate and map a bounce buffer.
308	 */
309	map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
310	if (!map)
311		return DMA_ERROR_CODE;
312
313	dev_addr = xen_virt_to_bus(map);
314
315	/*
316	 * Ensure that the address returned is DMA'ble
317	 */
318	if (!dma_capable(dev, dev_addr, size)) {
319		swiotlb_tbl_unmap_single(dev, map, size, dir);
320		dev_addr = 0;
321	}
322	return dev_addr;
323}
324EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
325
326/*
327 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
328 * match what was provided for in a previous xen_swiotlb_map_page call.  All
329 * other usages are undefined.
330 *
331 * After this call, reads by the cpu to the buffer are guaranteed to see
332 * whatever the device wrote there.
333 */
334static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
335			     size_t size, enum dma_data_direction dir)
336{
337	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
338
339	BUG_ON(dir == DMA_NONE);
340
341	/* NOTE: We use dev_addr here, not paddr! */
342	if (is_xen_swiotlb_buffer(dev_addr)) {
343		swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
344		return;
345	}
346
347	if (dir != DMA_FROM_DEVICE)
348		return;
349
350	/*
351	 * phys_to_virt doesn't work with hihgmem page but we could
352	 * call dma_mark_clean() with hihgmem page here. However, we
353	 * are fine since dma_mark_clean() is null on POWERPC. We can
354	 * make dma_mark_clean() take a physical address if necessary.
355	 */
356	dma_mark_clean(phys_to_virt(paddr), size);
357}
358
359void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
360			    size_t size, enum dma_data_direction dir,
361			    struct dma_attrs *attrs)
362{
363	xen_unmap_single(hwdev, dev_addr, size, dir);
364}
365EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
366
367/*
368 * Make physical memory consistent for a single streaming mode DMA translation
369 * after a transfer.
370 *
371 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
372 * using the cpu, yet do not wish to teardown the dma mapping, you must
373 * call this function before doing so.  At the next point you give the dma
374 * address back to the card, you must first perform a
375 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
376 */
377static void
378xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
379			size_t size, enum dma_data_direction dir,
380			enum dma_sync_target target)
381{
382	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
383
384	BUG_ON(dir == DMA_NONE);
385
386	/* NOTE: We use dev_addr here, not paddr! */
387	if (is_xen_swiotlb_buffer(dev_addr)) {
388		swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
389				       target);
390		return;
391	}
392
393	if (dir != DMA_FROM_DEVICE)
394		return;
395
396	dma_mark_clean(phys_to_virt(paddr), size);
397}
398
399void
400xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
401				size_t size, enum dma_data_direction dir)
402{
403	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
404}
405EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
406
407void
408xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
409				   size_t size, enum dma_data_direction dir)
410{
411	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
412}
413EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
414
415/*
416 * Map a set of buffers described by scatterlist in streaming mode for DMA.
417 * This is the scatter-gather version of the above xen_swiotlb_map_page
418 * interface.  Here the scatter gather list elements are each tagged with the
419 * appropriate dma address and length.  They are obtained via
420 * sg_dma_{address,length}(SG).
421 *
422 * NOTE: An implementation may be able to use a smaller number of
423 *       DMA address/length pairs than there are SG table elements.
424 *       (for example via virtual mapping capabilities)
425 *       The routine returns the number of addr/length pairs actually
426 *       used, at most nents.
427 *
428 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
429 * same here.
430 */
431int
432xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
433			 int nelems, enum dma_data_direction dir,
434			 struct dma_attrs *attrs)
435{
436	struct scatterlist *sg;
437	int i;
438
439	BUG_ON(dir == DMA_NONE);
440
441	for_each_sg(sgl, sg, nelems, i) {
442		phys_addr_t paddr = sg_phys(sg);
443		dma_addr_t dev_addr = xen_phys_to_bus(paddr);
444
445		if (swiotlb_force ||
446		    !dma_capable(hwdev, dev_addr, sg->length) ||
447		    range_straddles_page_boundary(paddr, sg->length)) {
448			void *map = swiotlb_tbl_map_single(hwdev,
449							   start_dma_addr,
450							   sg_phys(sg),
451							   sg->length, dir);
452			if (!map) {
453				/* Don't panic here, we expect map_sg users
454				   to do proper error handling. */
455				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
456							   attrs);
457				sgl[0].dma_length = 0;
458				return DMA_ERROR_CODE;
459			}
460			sg->dma_address = xen_virt_to_bus(map);
461		} else
462			sg->dma_address = dev_addr;
463		sg->dma_length = sg->length;
464	}
465	return nelems;
466}
467EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
468
469int
470xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
471		   enum dma_data_direction dir)
472{
473	return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
474}
475EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
476
477/*
478 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
479 * concerning calls here are the same as for swiotlb_unmap_page() above.
480 */
481void
482xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
483			   int nelems, enum dma_data_direction dir,
484			   struct dma_attrs *attrs)
485{
486	struct scatterlist *sg;
487	int i;
488
489	BUG_ON(dir == DMA_NONE);
490
491	for_each_sg(sgl, sg, nelems, i)
492		xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
493
494}
495EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
496
497void
498xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
499		     enum dma_data_direction dir)
500{
501	return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
502}
503EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
504
505/*
506 * Make physical memory consistent for a set of streaming mode DMA translations
507 * after a transfer.
508 *
509 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
510 * and usage.
511 */
512static void
513xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
514		    int nelems, enum dma_data_direction dir,
515		    enum dma_sync_target target)
516{
517	struct scatterlist *sg;
518	int i;
519
520	for_each_sg(sgl, sg, nelems, i)
521		xen_swiotlb_sync_single(hwdev, sg->dma_address,
522					sg->dma_length, dir, target);
523}
524
525void
526xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
527			    int nelems, enum dma_data_direction dir)
528{
529	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
530}
531EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
532
533void
534xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
535			       int nelems, enum dma_data_direction dir)
536{
537	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
538}
539EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
540
541int
542xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
543{
544	return !dma_addr;
545}
546EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
547
548/*
549 * Return whether the given device DMA address mask can be supported
550 * properly.  For example, if your device can only drive the low 24-bits
551 * during bus mastering, then you would pass 0x00ffffff as the mask to
552 * this function.
553 */
554int
555xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
556{
557	return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
558}
559EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);