linux/drivers/xen/swiotlb-xen.c
<<
>>
Prefs
   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
  56static char *xen_io_tlb_start, *xen_io_tlb_end;
  57static unsigned long xen_io_tlb_nslabs;
  58/*
  59 * Quick lookup value of the bus address of the IOTLB.
  60 */
  61
  62static u64 start_dma_addr;
  63
  64/*
  65 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
  66 * can be 32bit when dma_addr_t is 64bit leading to a loss in
  67 * information if the shift is done before casting to 64bit.
  68 */
  69static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
  70{
  71        unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
  72        dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
  73
  74        dma |= paddr & ~XEN_PAGE_MASK;
  75
  76        return dma;
  77}
  78
  79static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
  80{
  81        unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
  82        dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
  83        phys_addr_t paddr = dma;
  84
  85        paddr |= baddr & ~XEN_PAGE_MASK;
  86
  87        return paddr;
  88}
  89
  90static inline dma_addr_t xen_virt_to_bus(void *address)
  91{
  92        return xen_phys_to_bus(virt_to_phys(address));
  93}
  94
  95static int check_pages_physically_contiguous(unsigned long xen_pfn,
  96                                             unsigned int offset,
  97                                             size_t length)
  98{
  99        unsigned long next_bfn;
 100        int i;
 101        int nr_pages;
 102
 103        next_bfn = pfn_to_bfn(xen_pfn);
 104        nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
 105
 106        for (i = 1; i < nr_pages; i++) {
 107                if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
 108                        return 0;
 109        }
 110        return 1;
 111}
 112
 113static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
 114{
 115        unsigned long xen_pfn = XEN_PFN_DOWN(p);
 116        unsigned int offset = p & ~XEN_PAGE_MASK;
 117
 118        if (offset + size <= XEN_PAGE_SIZE)
 119                return 0;
 120        if (check_pages_physically_contiguous(xen_pfn, offset, size))
 121                return 0;
 122        return 1;
 123}
 124
 125static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
 126{
 127        unsigned long bfn = XEN_PFN_DOWN(dma_addr);
 128        unsigned long xen_pfn = bfn_to_local_pfn(bfn);
 129        phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
 130
 131        /* If the address is outside our domain, it CAN
 132         * have the same virtual address as another address
 133         * in our domain. Therefore _only_ check address within our domain.
 134         */
 135        if (pfn_valid(PFN_DOWN(paddr))) {
 136                return paddr >= virt_to_phys(xen_io_tlb_start) &&
 137                       paddr < virt_to_phys(xen_io_tlb_end);
 138        }
 139        return 0;
 140}
 141
 142static int max_dma_bits = 32;
 143
 144static int
 145xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
 146{
 147        int i, rc;
 148        int dma_bits;
 149        dma_addr_t dma_handle;
 150        phys_addr_t p = virt_to_phys(buf);
 151
 152        dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
 153
 154        i = 0;
 155        do {
 156                int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
 157
 158                do {
 159                        rc = xen_create_contiguous_region(
 160                                p + (i << IO_TLB_SHIFT),
 161                                get_order(slabs << IO_TLB_SHIFT),
 162                                dma_bits, &dma_handle);
 163                } while (rc && dma_bits++ < max_dma_bits);
 164                if (rc)
 165                        return rc;
 166
 167                i += slabs;
 168        } while (i < nslabs);
 169        return 0;
 170}
 171static unsigned long xen_set_nslabs(unsigned long nr_tbl)
 172{
 173        if (!nr_tbl) {
 174                xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
 175                xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
 176        } else
 177                xen_io_tlb_nslabs = nr_tbl;
 178
 179        return xen_io_tlb_nslabs << IO_TLB_SHIFT;
 180}
 181
 182enum xen_swiotlb_err {
 183        XEN_SWIOTLB_UNKNOWN = 0,
 184        XEN_SWIOTLB_ENOMEM,
 185        XEN_SWIOTLB_EFIXUP
 186};
 187
 188static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
 189{
 190        switch (err) {
 191        case XEN_SWIOTLB_ENOMEM:
 192                return "Cannot allocate Xen-SWIOTLB buffer\n";
 193        case XEN_SWIOTLB_EFIXUP:
 194                return "Failed to get contiguous memory for DMA from Xen!\n"\
 195                    "You either: don't have the permissions, do not have"\
 196                    " enough free memory under 4GB, or the hypervisor memory"\
 197                    " is too fragmented!";
 198        default:
 199                break;
 200        }
 201        return "";
 202}
 203int __ref xen_swiotlb_init(int verbose, bool early)
 204{
 205        unsigned long bytes, order;
 206        int rc = -ENOMEM;
 207        enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
 208        unsigned int repeat = 3;
 209
 210        xen_io_tlb_nslabs = swiotlb_nr_tbl();
 211retry:
 212        bytes = xen_set_nslabs(xen_io_tlb_nslabs);
 213        order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
 214        /*
 215         * Get IO TLB memory from any location.
 216         */
 217        if (early)
 218                xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
 219        else {
 220#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
 221#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
 222                while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
 223                        xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
 224                        if (xen_io_tlb_start)
 225                                break;
 226                        order--;
 227                }
 228                if (order != get_order(bytes)) {
 229                        pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
 230                                (PAGE_SIZE << order) >> 20);
 231                        xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
 232                        bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
 233                }
 234        }
 235        if (!xen_io_tlb_start) {
 236                m_ret = XEN_SWIOTLB_ENOMEM;
 237                goto error;
 238        }
 239        xen_io_tlb_end = xen_io_tlb_start + bytes;
 240        /*
 241         * And replace that memory with pages under 4GB.
 242         */
 243        rc = xen_swiotlb_fixup(xen_io_tlb_start,
 244                               bytes,
 245                               xen_io_tlb_nslabs);
 246        if (rc) {
 247                if (early)
 248                        free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
 249                else {
 250                        free_pages((unsigned long)xen_io_tlb_start, order);
 251                        xen_io_tlb_start = NULL;
 252                }
 253                m_ret = XEN_SWIOTLB_EFIXUP;
 254                goto error;
 255        }
 256        start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
 257        if (early) {
 258                if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
 259                         verbose))
 260                        panic("Cannot allocate SWIOTLB buffer");
 261                rc = 0;
 262        } else
 263                rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
 264
 265        if (!rc)
 266                swiotlb_set_max_segment(PAGE_SIZE);
 267
 268        return rc;
 269error:
 270        if (repeat--) {
 271                xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
 272                                        (xen_io_tlb_nslabs >> 1));
 273                pr_info("Lowering to %luMB\n",
 274                        (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
 275                goto retry;
 276        }
 277        pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
 278        if (early)
 279                panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
 280        else
 281                free_pages((unsigned long)xen_io_tlb_start, order);
 282        return rc;
 283}
 284
 285static void *
 286xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 287                           dma_addr_t *dma_handle, gfp_t flags,
 288                           unsigned long attrs)
 289{
 290        void *ret;
 291        int order = get_order(size);
 292        u64 dma_mask = DMA_BIT_MASK(32);
 293        phys_addr_t phys;
 294        dma_addr_t dev_addr;
 295
 296        /*
 297        * Ignore region specifiers - the kernel's ideas of
 298        * pseudo-phys memory layout has nothing to do with the
 299        * machine physical layout.  We can't allocate highmem
 300        * because we can't return a pointer to it.
 301        */
 302        flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
 303
 304        /* On ARM this function returns an ioremap'ped virtual address for
 305         * which virt_to_phys doesn't return the corresponding physical
 306         * address. In fact on ARM virt_to_phys only works for kernel direct
 307         * mapped RAM memory. Also see comment below.
 308         */
 309        ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
 310
 311        if (!ret)
 312                return ret;
 313
 314        if (hwdev && hwdev->coherent_dma_mask)
 315                dma_mask = hwdev->coherent_dma_mask;
 316
 317        /* At this point dma_handle is the physical address, next we are
 318         * going to set it to the machine address.
 319         * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
 320         * to *dma_handle. */
 321        phys = *dma_handle;
 322        dev_addr = xen_phys_to_bus(phys);
 323        if (((dev_addr + size - 1 <= dma_mask)) &&
 324            !range_straddles_page_boundary(phys, size))
 325                *dma_handle = dev_addr;
 326        else {
 327                if (xen_create_contiguous_region(phys, order,
 328                                                 fls64(dma_mask), dma_handle) != 0) {
 329                        xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
 330                        return NULL;
 331                }
 332        }
 333        memset(ret, 0, size);
 334        return ret;
 335}
 336
 337static void
 338xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
 339                          dma_addr_t dev_addr, unsigned long attrs)
 340{
 341        int order = get_order(size);
 342        phys_addr_t phys;
 343        u64 dma_mask = DMA_BIT_MASK(32);
 344
 345        if (hwdev && hwdev->coherent_dma_mask)
 346                dma_mask = hwdev->coherent_dma_mask;
 347
 348        /* do not use virt_to_phys because on ARM it doesn't return you the
 349         * physical address */
 350        phys = xen_bus_to_phys(dev_addr);
 351
 352        if (((dev_addr + size - 1 <= dma_mask)) ||
 353            range_straddles_page_boundary(phys, size))
 354                xen_destroy_contiguous_region(phys, order);
 355
 356        xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
 357}
 358
 359/*
 360 * Map a single buffer of the indicated size for DMA in streaming mode.  The
 361 * physical address to use is returned.
 362 *
 363 * Once the device is given the dma address, the device owns this memory until
 364 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
 365 */
 366static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
 367                                unsigned long offset, size_t size,
 368                                enum dma_data_direction dir,
 369                                unsigned long attrs)
 370{
 371        phys_addr_t map, phys = page_to_phys(page) + offset;
 372        dma_addr_t dev_addr = xen_phys_to_bus(phys);
 373
 374        BUG_ON(dir == DMA_NONE);
 375        /*
 376         * If the address happens to be in the device's DMA window,
 377         * we can safely return the device addr and not worry about bounce
 378         * buffering it.
 379         */
 380        if (dma_capable(dev, dev_addr, size) &&
 381            !range_straddles_page_boundary(phys, size) &&
 382                !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
 383                (swiotlb_force != SWIOTLB_FORCE)) {
 384                /* we are not interested in the dma_addr returned by
 385                 * xen_dma_map_page, only in the potential cache flushes executed
 386                 * by the function. */
 387                xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
 388                return dev_addr;
 389        }
 390
 391        /*
 392         * Oh well, have to allocate and map a bounce buffer.
 393         */
 394        trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
 395
 396        map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
 397                                     attrs);
 398        if (map == (phys_addr_t)DMA_MAPPING_ERROR)
 399                return DMA_MAPPING_ERROR;
 400
 401        dev_addr = xen_phys_to_bus(map);
 402        xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
 403                                        dev_addr, map & ~PAGE_MASK, size, dir, attrs);
 404
 405        /*
 406         * Ensure that the address returned is DMA'ble
 407         */
 408        if (dma_capable(dev, dev_addr, size))
 409                return dev_addr;
 410
 411        attrs |= DMA_ATTR_SKIP_CPU_SYNC;
 412        swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
 413
 414        return DMA_MAPPING_ERROR;
 415}
 416
 417/*
 418 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
 419 * match what was provided for in a previous xen_swiotlb_map_page call.  All
 420 * other usages are undefined.
 421 *
 422 * After this call, reads by the cpu to the buffer are guaranteed to see
 423 * whatever the device wrote there.
 424 */
 425static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
 426                             size_t size, enum dma_data_direction dir,
 427                             unsigned long attrs)
 428{
 429        phys_addr_t paddr = xen_bus_to_phys(dev_addr);
 430
 431        BUG_ON(dir == DMA_NONE);
 432
 433        xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
 434
 435        /* NOTE: We use dev_addr here, not paddr! */
 436        if (is_xen_swiotlb_buffer(dev_addr))
 437                swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
 438}
 439
 440static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
 441                            size_t size, enum dma_data_direction dir,
 442                            unsigned long attrs)
 443{
 444        xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
 445}
 446
 447/*
 448 * Make physical memory consistent for a single streaming mode DMA translation
 449 * after a transfer.
 450 *
 451 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
 452 * using the cpu, yet do not wish to teardown the dma mapping, you must
 453 * call this function before doing so.  At the next point you give the dma
 454 * address back to the card, you must first perform a
 455 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
 456 */
 457static void
 458xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
 459                        size_t size, enum dma_data_direction dir,
 460                        enum dma_sync_target target)
 461{
 462        phys_addr_t paddr = xen_bus_to_phys(dev_addr);
 463
 464        BUG_ON(dir == DMA_NONE);
 465
 466        if (target == SYNC_FOR_CPU)
 467                xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
 468
 469        /* NOTE: We use dev_addr here, not paddr! */
 470        if (is_xen_swiotlb_buffer(dev_addr))
 471                swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
 472
 473        if (target == SYNC_FOR_DEVICE)
 474                xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
 475}
 476
 477void
 478xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
 479                                size_t size, enum dma_data_direction dir)
 480{
 481        xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
 482}
 483
 484void
 485xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
 486                                   size_t size, enum dma_data_direction dir)
 487{
 488        xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
 489}
 490
 491/*
 492 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
 493 * concerning calls here are the same as for swiotlb_unmap_page() above.
 494 */
 495static void
 496xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 497                           int nelems, enum dma_data_direction dir,
 498                           unsigned long attrs)
 499{
 500        struct scatterlist *sg;
 501        int i;
 502
 503        BUG_ON(dir == DMA_NONE);
 504
 505        for_each_sg(sgl, sg, nelems, i)
 506                xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
 507
 508}
 509
 510/*
 511 * Map a set of buffers described by scatterlist in streaming mode for DMA.
 512 * This is the scatter-gather version of the above xen_swiotlb_map_page
 513 * interface.  Here the scatter gather list elements are each tagged with the
 514 * appropriate dma address and length.  They are obtained via
 515 * sg_dma_{address,length}(SG).
 516 *
 517 * NOTE: An implementation may be able to use a smaller number of
 518 *       DMA address/length pairs than there are SG table elements.
 519 *       (for example via virtual mapping capabilities)
 520 *       The routine returns the number of addr/length pairs actually
 521 *       used, at most nents.
 522 *
 523 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
 524 * same here.
 525 */
 526static int
 527xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 528                         int nelems, enum dma_data_direction dir,
 529                         unsigned long attrs)
 530{
 531        struct scatterlist *sg;
 532        int i;
 533
 534        BUG_ON(dir == DMA_NONE);
 535
 536        for_each_sg(sgl, sg, nelems, i) {
 537                phys_addr_t paddr = sg_phys(sg);
 538                dma_addr_t dev_addr = xen_phys_to_bus(paddr);
 539
 540                if (swiotlb_force == SWIOTLB_FORCE ||
 541                    xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
 542                    !dma_capable(hwdev, dev_addr, sg->length) ||
 543                    range_straddles_page_boundary(paddr, sg->length)) {
 544                        phys_addr_t map = swiotlb_tbl_map_single(hwdev,
 545                                                                 start_dma_addr,
 546                                                                 sg_phys(sg),
 547                                                                 sg->length,
 548                                                                 dir, attrs);
 549                        if (map == DMA_MAPPING_ERROR) {
 550                                dev_warn(hwdev, "swiotlb buffer is full\n");
 551                                /* Don't panic here, we expect map_sg users
 552                                   to do proper error handling. */
 553                                attrs |= DMA_ATTR_SKIP_CPU_SYNC;
 554                                xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
 555                                                           attrs);
 556                                sg_dma_len(sgl) = 0;
 557                                return 0;
 558                        }
 559                        dev_addr = xen_phys_to_bus(map);
 560                        xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
 561                                                dev_addr,
 562                                                map & ~PAGE_MASK,
 563                                                sg->length,
 564                                                dir,
 565                                                attrs);
 566                        sg->dma_address = dev_addr;
 567                } else {
 568                        /* we are not interested in the dma_addr returned by
 569                         * xen_dma_map_page, only in the potential cache flushes executed
 570                         * by the function. */
 571                        xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
 572                                                dev_addr,
 573                                                paddr & ~PAGE_MASK,
 574                                                sg->length,
 575                                                dir,
 576                                                attrs);
 577                        sg->dma_address = dev_addr;
 578                }
 579                sg_dma_len(sg) = sg->length;
 580        }
 581        return nelems;
 582}
 583
 584/*
 585 * Make physical memory consistent for a set of streaming mode DMA translations
 586 * after a transfer.
 587 *
 588 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
 589 * and usage.
 590 */
 591static void
 592xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
 593                    int nelems, enum dma_data_direction dir,
 594                    enum dma_sync_target target)
 595{
 596        struct scatterlist *sg;
 597        int i;
 598
 599        for_each_sg(sgl, sg, nelems, i)
 600                xen_swiotlb_sync_single(hwdev, sg->dma_address,
 601                                        sg_dma_len(sg), dir, target);
 602}
 603
 604static void
 605xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
 606                            int nelems, enum dma_data_direction dir)
 607{
 608        xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
 609}
 610
 611static void
 612xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
 613                               int nelems, enum dma_data_direction dir)
 614{
 615        xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
 616}
 617
 618/*
 619 * Return whether the given device DMA address mask can be supported
 620 * properly.  For example, if your device can only drive the low 24-bits
 621 * during bus mastering, then you would pass 0x00ffffff as the mask to
 622 * this function.
 623 */
 624static int
 625xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
 626{
 627        return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
 628}
 629
 630/*
 631 * Create userspace mapping for the DMA-coherent memory.
 632 * This function should be called with the pages from the current domain only,
 633 * passing pages mapped from other domains would lead to memory corruption.
 634 */
 635static int
 636xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
 637                     void *cpu_addr, dma_addr_t dma_addr, size_t size,
 638                     unsigned long attrs)
 639{
 640#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
 641        if (xen_get_dma_ops(dev)->mmap)
 642                return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
 643                                                    dma_addr, size, attrs);
 644#endif
 645        return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
 646}
 647
 648/*
 649 * This function should be called with the pages from the current domain only,
 650 * passing pages mapped from other domains would lead to memory corruption.
 651 */
 652static int
 653xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
 654                        void *cpu_addr, dma_addr_t handle, size_t size,
 655                        unsigned long attrs)
 656{
 657#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
 658        if (xen_get_dma_ops(dev)->get_sgtable) {
 659#if 0
 660        /*
 661         * This check verifies that the page belongs to the current domain and
 662         * is not one mapped from another domain.
 663         * This check is for debug only, and should not go to production build
 664         */
 665                unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
 666                BUG_ON (!page_is_ram(bfn));
 667#endif
 668                return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
 669                                                           handle, size, attrs);
 670        }
 671#endif
 672        return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
 673}
 674
 675const struct dma_map_ops xen_swiotlb_dma_ops = {
 676        .alloc = xen_swiotlb_alloc_coherent,
 677        .free = xen_swiotlb_free_coherent,
 678        .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
 679        .sync_single_for_device = xen_swiotlb_sync_single_for_device,
 680        .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
 681        .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
 682        .map_sg = xen_swiotlb_map_sg_attrs,
 683        .unmap_sg = xen_swiotlb_unmap_sg_attrs,
 684        .map_page = xen_swiotlb_map_page,
 685        .unmap_page = xen_swiotlb_unmap_page,
 686        .dma_supported = xen_swiotlb_dma_supported,
 687        .mmap = xen_swiotlb_dma_mmap,
 688        .get_sgtable = xen_swiotlb_get_sgtable,
 689};
 690