The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/lib/swiotlb.c

Version: -  FREEBSD  -  FREEBSD-13-STABLE  -  FREEBSD-13-0  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  l41  -  OPENBSD  -  linux-2.6  -  MK84  -  PLAN9  -  xnu-8792 
SearchContext: -  none  -  3  -  10 

    1 /*
    2  * Dynamic DMA mapping support.
    3  *
    4  * This implementation is a fallback for platforms that do not support
    5  * I/O TLBs (aka DMA address translation hardware).
    6  * Copyright (C) 2000 Asit Mallick <Asit.K.Mallick@intel.com>
    7  * Copyright (C) 2000 Goutham Rao <goutham.rao@intel.com>
    8  * Copyright (C) 2000, 2003 Hewlett-Packard Co
    9  *      David Mosberger-Tang <davidm@hpl.hp.com>
   10  *
   11  * 03/05/07 davidm      Switch from PCI-DMA to generic device DMA API.
   12  * 00/12/13 davidm      Rename to swiotlb.c and add mark_clean() to avoid
   13  *                      unnecessary i-cache flushing.
   14  * 04/07/.. ak          Better overflow handling. Assorted fixes.
   15  * 05/09/10 linville    Add support for syncing ranges, support syncing for
   16  *                      DMA_BIDIRECTIONAL mappings, miscellaneous cleanup.
   17  * 08/12/11 beckyb      Add highmem support
   18  */
   19 
   20 #include <linux/cache.h>
   21 #include <linux/dma-mapping.h>
   22 #include <linux/mm.h>
   23 #include <linux/export.h>
   24 #include <linux/spinlock.h>
   25 #include <linux/string.h>
   26 #include <linux/swiotlb.h>
   27 #include <linux/pfn.h>
   28 #include <linux/types.h>
   29 #include <linux/ctype.h>
   30 #include <linux/highmem.h>
   31 #include <linux/gfp.h>
   32 
   33 #include <asm/io.h>
   34 #include <asm/dma.h>
   35 #include <asm/scatterlist.h>
   36 
   37 #include <linux/init.h>
   38 #include <linux/bootmem.h>
   39 #include <linux/iommu-helper.h>
   40 
   41 #define OFFSET(val,align) ((unsigned long)      \
   42                            ( (val) & ( (align) - 1)))
   43 
   44 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
   45 
   46 /*
   47  * Minimum IO TLB size to bother booting with.  Systems with mainly
   48  * 64bit capable cards will only lightly use the swiotlb.  If we can't
   49  * allocate a contiguous 1MB, we're probably in trouble anyway.
   50  */
   51 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
   52 
   53 int swiotlb_force;
   54 
   55 /*
   56  * Used to do a quick range check in swiotlb_tbl_unmap_single and
   57  * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
   58  * API.
   59  */
   60 static phys_addr_t io_tlb_start, io_tlb_end;
   61 
   62 /*
   63  * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
   64  * io_tlb_end.  This is command line adjustable via setup_io_tlb_npages.
   65  */
   66 static unsigned long io_tlb_nslabs;
   67 
   68 /*
   69  * When the IOMMU overflows we return a fallback buffer. This sets the size.
   70  */
   71 static unsigned long io_tlb_overflow = 32*1024;
   72 
   73 static phys_addr_t io_tlb_overflow_buffer;
   74 
   75 /*
   76  * This is a free list describing the number of free entries available from
   77  * each index
   78  */
   79 static unsigned int *io_tlb_list;
   80 static unsigned int io_tlb_index;
   81 
   82 /*
   83  * We need to save away the original address corresponding to a mapped entry
   84  * for the sync operations.
   85  */
   86 static phys_addr_t *io_tlb_orig_addr;
   87 
   88 /*
   89  * Protect the above data structures in the map and unmap calls
   90  */
   91 static DEFINE_SPINLOCK(io_tlb_lock);
   92 
   93 static int late_alloc;
   94 
   95 static int __init
   96 setup_io_tlb_npages(char *str)
   97 {
   98         if (isdigit(*str)) {
   99                 io_tlb_nslabs = simple_strtoul(str, &str, 0);
  100                 /* avoid tail segment of size < IO_TLB_SEGSIZE */
  101                 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
  102         }
  103         if (*str == ',')
  104                 ++str;
  105         if (!strcmp(str, "force"))
  106                 swiotlb_force = 1;
  107 
  108         return 1;
  109 }
  110 __setup("swiotlb=", setup_io_tlb_npages);
  111 /* make io_tlb_overflow tunable too? */
  112 
  113 unsigned long swiotlb_nr_tbl(void)
  114 {
  115         return io_tlb_nslabs;
  116 }
  117 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
  118 /* Note that this doesn't work with highmem page */
  119 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
  120                                       volatile void *address)
  121 {
  122         return phys_to_dma(hwdev, virt_to_phys(address));
  123 }
  124 
  125 void swiotlb_print_info(void)
  126 {
  127         unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
  128         unsigned char *vstart, *vend;
  129 
  130         vstart = phys_to_virt(io_tlb_start);
  131         vend = phys_to_virt(io_tlb_end);
  132 
  133         printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
  134                (unsigned long long)io_tlb_start,
  135                (unsigned long long)io_tlb_end,
  136                bytes >> 20, vstart, vend - 1);
  137 }
  138 
  139 void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
  140 {
  141         void *v_overflow_buffer;
  142         unsigned long i, bytes;
  143 
  144         bytes = nslabs << IO_TLB_SHIFT;
  145 
  146         io_tlb_nslabs = nslabs;
  147         io_tlb_start = __pa(tlb);
  148         io_tlb_end = io_tlb_start + bytes;
  149 
  150         /*
  151          * Get the overflow emergency buffer
  152          */
  153         v_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
  154         if (!v_overflow_buffer)
  155                 panic("Cannot allocate SWIOTLB overflow buffer!\n");
  156 
  157         io_tlb_overflow_buffer = __pa(v_overflow_buffer);
  158 
  159         /*
  160          * Allocate and initialize the free list array.  This array is used
  161          * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
  162          * between io_tlb_start and io_tlb_end.
  163          */
  164         io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
  165         for (i = 0; i < io_tlb_nslabs; i++)
  166                 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
  167         io_tlb_index = 0;
  168         io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
  169 
  170         if (verbose)
  171                 swiotlb_print_info();
  172 }
  173 
  174 /*
  175  * Statically reserve bounce buffer space and initialize bounce buffer data
  176  * structures for the software IO TLB used to implement the DMA API.
  177  */
  178 static void __init
  179 swiotlb_init_with_default_size(size_t default_size, int verbose)
  180 {
  181         unsigned char *vstart;
  182         unsigned long bytes;
  183 
  184         if (!io_tlb_nslabs) {
  185                 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
  186                 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
  187         }
  188 
  189         bytes = io_tlb_nslabs << IO_TLB_SHIFT;
  190 
  191         /*
  192          * Get IO TLB memory from the low pages
  193          */
  194         vstart = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
  195         if (!vstart)
  196                 panic("Cannot allocate SWIOTLB buffer");
  197 
  198         swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose);
  199 }
  200 
  201 void __init
  202 swiotlb_init(int verbose)
  203 {
  204         swiotlb_init_with_default_size(64 * (1<<20), verbose);  /* default to 64MB */
  205 }
  206 
  207 /*
  208  * Systems with larger DMA zones (those that don't support ISA) can
  209  * initialize the swiotlb later using the slab allocator if needed.
  210  * This should be just like above, but with some error catching.
  211  */
  212 int
  213 swiotlb_late_init_with_default_size(size_t default_size)
  214 {
  215         unsigned long bytes, req_nslabs = io_tlb_nslabs;
  216         unsigned char *vstart = NULL;
  217         unsigned int order;
  218         int rc = 0;
  219 
  220         if (!io_tlb_nslabs) {
  221                 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
  222                 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
  223         }
  224 
  225         /*
  226          * Get IO TLB memory from the low pages
  227          */
  228         order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
  229         io_tlb_nslabs = SLABS_PER_PAGE << order;
  230         bytes = io_tlb_nslabs << IO_TLB_SHIFT;
  231 
  232         while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
  233                 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
  234                                                   order);
  235                 if (vstart)
  236                         break;
  237                 order--;
  238         }
  239 
  240         if (!vstart) {
  241                 io_tlb_nslabs = req_nslabs;
  242                 return -ENOMEM;
  243         }
  244         if (order != get_order(bytes)) {
  245                 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
  246                        "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
  247                 io_tlb_nslabs = SLABS_PER_PAGE << order;
  248         }
  249         rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
  250         if (rc)
  251                 free_pages((unsigned long)vstart, order);
  252         return rc;
  253 }
  254 
  255 int
  256 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
  257 {
  258         unsigned long i, bytes;
  259         unsigned char *v_overflow_buffer;
  260 
  261         bytes = nslabs << IO_TLB_SHIFT;
  262 
  263         io_tlb_nslabs = nslabs;
  264         io_tlb_start = virt_to_phys(tlb);
  265         io_tlb_end = io_tlb_start + bytes;
  266 
  267         memset(tlb, 0, bytes);
  268 
  269         /*
  270          * Get the overflow emergency buffer
  271          */
  272         v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
  273                                                      get_order(io_tlb_overflow));
  274         if (!v_overflow_buffer)
  275                 goto cleanup2;
  276 
  277         io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
  278 
  279         /*
  280          * Allocate and initialize the free list array.  This array is used
  281          * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
  282          * between io_tlb_start and io_tlb_end.
  283          */
  284         io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
  285                                       get_order(io_tlb_nslabs * sizeof(int)));
  286         if (!io_tlb_list)
  287                 goto cleanup3;
  288 
  289         for (i = 0; i < io_tlb_nslabs; i++)
  290                 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
  291         io_tlb_index = 0;
  292 
  293         io_tlb_orig_addr = (phys_addr_t *)
  294                 __get_free_pages(GFP_KERNEL,
  295                                  get_order(io_tlb_nslabs *
  296                                            sizeof(phys_addr_t)));
  297         if (!io_tlb_orig_addr)
  298                 goto cleanup4;
  299 
  300         memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
  301 
  302         swiotlb_print_info();
  303 
  304         late_alloc = 1;
  305 
  306         return 0;
  307 
  308 cleanup4:
  309         free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
  310                                                          sizeof(int)));
  311         io_tlb_list = NULL;
  312 cleanup3:
  313         free_pages((unsigned long)v_overflow_buffer,
  314                    get_order(io_tlb_overflow));
  315         io_tlb_overflow_buffer = 0;
  316 cleanup2:
  317         io_tlb_end = 0;
  318         io_tlb_start = 0;
  319         io_tlb_nslabs = 0;
  320         return -ENOMEM;
  321 }
  322 
  323 void __init swiotlb_free(void)
  324 {
  325         if (!io_tlb_orig_addr)
  326                 return;
  327 
  328         if (late_alloc) {
  329                 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
  330                            get_order(io_tlb_overflow));
  331                 free_pages((unsigned long)io_tlb_orig_addr,
  332                            get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
  333                 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
  334                                                                  sizeof(int)));
  335                 free_pages((unsigned long)phys_to_virt(io_tlb_start),
  336                            get_order(io_tlb_nslabs << IO_TLB_SHIFT));
  337         } else {
  338                 free_bootmem_late(io_tlb_overflow_buffer,
  339                                   PAGE_ALIGN(io_tlb_overflow));
  340                 free_bootmem_late(__pa(io_tlb_orig_addr),
  341                                   PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
  342                 free_bootmem_late(__pa(io_tlb_list),
  343                                   PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
  344                 free_bootmem_late(io_tlb_start,
  345                                   PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
  346         }
  347         io_tlb_nslabs = 0;
  348 }
  349 
  350 static int is_swiotlb_buffer(phys_addr_t paddr)
  351 {
  352         return paddr >= io_tlb_start && paddr < io_tlb_end;
  353 }
  354 
  355 /*
  356  * Bounce: copy the swiotlb buffer back to the original dma location
  357  */
  358 static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
  359                            size_t size, enum dma_data_direction dir)
  360 {
  361         unsigned long pfn = PFN_DOWN(orig_addr);
  362         unsigned char *vaddr = phys_to_virt(tlb_addr);
  363 
  364         if (PageHighMem(pfn_to_page(pfn))) {
  365                 /* The buffer does not have a mapping.  Map it in and copy */
  366                 unsigned int offset = orig_addr & ~PAGE_MASK;
  367                 char *buffer;
  368                 unsigned int sz = 0;
  369                 unsigned long flags;
  370 
  371                 while (size) {
  372                         sz = min_t(size_t, PAGE_SIZE - offset, size);
  373 
  374                         local_irq_save(flags);
  375                         buffer = kmap_atomic(pfn_to_page(pfn));
  376                         if (dir == DMA_TO_DEVICE)
  377                                 memcpy(vaddr, buffer + offset, sz);
  378                         else
  379                                 memcpy(buffer + offset, vaddr, sz);
  380                         kunmap_atomic(buffer);
  381                         local_irq_restore(flags);
  382 
  383                         size -= sz;
  384                         pfn++;
  385                         vaddr += sz;
  386                         offset = 0;
  387                 }
  388         } else if (dir == DMA_TO_DEVICE) {
  389                 memcpy(vaddr, phys_to_virt(orig_addr), size);
  390         } else {
  391                 memcpy(phys_to_virt(orig_addr), vaddr, size);
  392         }
  393 }
  394 
  395 phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
  396                                    dma_addr_t tbl_dma_addr,
  397                                    phys_addr_t orig_addr, size_t size,
  398                                    enum dma_data_direction dir)
  399 {
  400         unsigned long flags;
  401         phys_addr_t tlb_addr;
  402         unsigned int nslots, stride, index, wrap;
  403         int i;
  404         unsigned long mask;
  405         unsigned long offset_slots;
  406         unsigned long max_slots;
  407 
  408         mask = dma_get_seg_boundary(hwdev);
  409 
  410         tbl_dma_addr &= mask;
  411 
  412         offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
  413 
  414         /*
  415          * Carefully handle integer overflow which can occur when mask == ~0UL.
  416          */
  417         max_slots = mask + 1
  418                     ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
  419                     : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
  420 
  421         /*
  422          * For mappings greater than a page, we limit the stride (and
  423          * hence alignment) to a page size.
  424          */
  425         nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
  426         if (size > PAGE_SIZE)
  427                 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
  428         else
  429                 stride = 1;
  430 
  431         BUG_ON(!nslots);
  432 
  433         /*
  434          * Find suitable number of IO TLB entries size that will fit this
  435          * request and allocate a buffer from that IO TLB pool.
  436          */
  437         spin_lock_irqsave(&io_tlb_lock, flags);
  438         index = ALIGN(io_tlb_index, stride);
  439         if (index >= io_tlb_nslabs)
  440                 index = 0;
  441         wrap = index;
  442 
  443         do {
  444                 while (iommu_is_span_boundary(index, nslots, offset_slots,
  445                                               max_slots)) {
  446                         index += stride;
  447                         if (index >= io_tlb_nslabs)
  448                                 index = 0;
  449                         if (index == wrap)
  450                                 goto not_found;
  451                 }
  452 
  453                 /*
  454                  * If we find a slot that indicates we have 'nslots' number of
  455                  * contiguous buffers, we allocate the buffers from that slot
  456                  * and mark the entries as '' indicating unavailable.
  457                  */
  458                 if (io_tlb_list[index] >= nslots) {
  459                         int count = 0;
  460 
  461                         for (i = index; i < (int) (index + nslots); i++)
  462                                 io_tlb_list[i] = 0;
  463                         for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
  464                                 io_tlb_list[i] = ++count;
  465                         tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
  466 
  467                         /*
  468                          * Update the indices to avoid searching in the next
  469                          * round.
  470                          */
  471                         io_tlb_index = ((index + nslots) < io_tlb_nslabs
  472                                         ? (index + nslots) : 0);
  473 
  474                         goto found;
  475                 }
  476                 index += stride;
  477                 if (index >= io_tlb_nslabs)
  478                         index = 0;
  479         } while (index != wrap);
  480 
  481 not_found:
  482         spin_unlock_irqrestore(&io_tlb_lock, flags);
  483         return SWIOTLB_MAP_ERROR;
  484 found:
  485         spin_unlock_irqrestore(&io_tlb_lock, flags);
  486 
  487         /*
  488          * Save away the mapping from the original address to the DMA address.
  489          * This is needed when we sync the memory.  Then we sync the buffer if
  490          * needed.
  491          */
  492         for (i = 0; i < nslots; i++)
  493                 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
  494         if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
  495                 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
  496 
  497         return tlb_addr;
  498 }
  499 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
  500 
  501 /*
  502  * Allocates bounce buffer and returns its kernel virtual address.
  503  */
  504 
  505 phys_addr_t map_single(struct device *hwdev, phys_addr_t phys, size_t size,
  506                        enum dma_data_direction dir)
  507 {
  508         dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
  509 
  510         return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
  511 }
  512 
  513 /*
  514  * dma_addr is the kernel virtual address of the bounce buffer to unmap.
  515  */
  516 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
  517                               size_t size, enum dma_data_direction dir)
  518 {
  519         unsigned long flags;
  520         int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
  521         int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
  522         phys_addr_t orig_addr = io_tlb_orig_addr[index];
  523 
  524         /*
  525          * First, sync the memory before unmapping the entry
  526          */
  527         if (orig_addr && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
  528                 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
  529 
  530         /*
  531          * Return the buffer to the free list by setting the corresponding
  532          * entries to indicate the number of contiguous entries available.
  533          * While returning the entries to the free list, we merge the entries
  534          * with slots below and above the pool being returned.
  535          */
  536         spin_lock_irqsave(&io_tlb_lock, flags);
  537         {
  538                 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
  539                          io_tlb_list[index + nslots] : 0);
  540                 /*
  541                  * Step 1: return the slots to the free list, merging the
  542                  * slots with superceeding slots
  543                  */
  544                 for (i = index + nslots - 1; i >= index; i--)
  545                         io_tlb_list[i] = ++count;
  546                 /*
  547                  * Step 2: merge the returned slots with the preceding slots,
  548                  * if available (non zero)
  549                  */
  550                 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
  551                         io_tlb_list[i] = ++count;
  552         }
  553         spin_unlock_irqrestore(&io_tlb_lock, flags);
  554 }
  555 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
  556 
  557 void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
  558                              size_t size, enum dma_data_direction dir,
  559                              enum dma_sync_target target)
  560 {
  561         int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
  562         phys_addr_t orig_addr = io_tlb_orig_addr[index];
  563 
  564         orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
  565 
  566         switch (target) {
  567         case SYNC_FOR_CPU:
  568                 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
  569                         swiotlb_bounce(orig_addr, tlb_addr,
  570                                        size, DMA_FROM_DEVICE);
  571                 else
  572                         BUG_ON(dir != DMA_TO_DEVICE);
  573                 break;
  574         case SYNC_FOR_DEVICE:
  575                 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
  576                         swiotlb_bounce(orig_addr, tlb_addr,
  577                                        size, DMA_TO_DEVICE);
  578                 else
  579                         BUG_ON(dir != DMA_FROM_DEVICE);
  580                 break;
  581         default:
  582                 BUG();
  583         }
  584 }
  585 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
  586 
  587 void *
  588 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
  589                        dma_addr_t *dma_handle, gfp_t flags)
  590 {
  591         dma_addr_t dev_addr;
  592         void *ret;
  593         int order = get_order(size);
  594         u64 dma_mask = DMA_BIT_MASK(32);
  595 
  596         if (hwdev && hwdev->coherent_dma_mask)
  597                 dma_mask = hwdev->coherent_dma_mask;
  598 
  599         ret = (void *)__get_free_pages(flags, order);
  600         if (ret) {
  601                 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
  602                 if (dev_addr + size - 1 > dma_mask) {
  603                         /*
  604                          * The allocated memory isn't reachable by the device.
  605                          */
  606                         free_pages((unsigned long) ret, order);
  607                         ret = NULL;
  608                 }
  609         }
  610         if (!ret) {
  611                 /*
  612                  * We are either out of memory or the device can't DMA to
  613                  * GFP_DMA memory; fall back on map_single(), which
  614                  * will grab memory from the lowest available address range.
  615                  */
  616                 phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
  617                 if (paddr == SWIOTLB_MAP_ERROR)
  618                         return NULL;
  619 
  620                 ret = phys_to_virt(paddr);
  621                 dev_addr = phys_to_dma(hwdev, paddr);
  622 
  623                 /* Confirm address can be DMA'd by device */
  624                 if (dev_addr + size - 1 > dma_mask) {
  625                         printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
  626                                (unsigned long long)dma_mask,
  627                                (unsigned long long)dev_addr);
  628 
  629                         /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
  630                         swiotlb_tbl_unmap_single(hwdev, paddr,
  631                                                  size, DMA_TO_DEVICE);
  632                         return NULL;
  633                 }
  634         }
  635 
  636         *dma_handle = dev_addr;
  637         memset(ret, 0, size);
  638 
  639         return ret;
  640 }
  641 EXPORT_SYMBOL(swiotlb_alloc_coherent);
  642 
  643 void
  644 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
  645                       dma_addr_t dev_addr)
  646 {
  647         phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
  648 
  649         WARN_ON(irqs_disabled());
  650         if (!is_swiotlb_buffer(paddr))
  651                 free_pages((unsigned long)vaddr, get_order(size));
  652         else
  653                 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
  654                 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
  655 }
  656 EXPORT_SYMBOL(swiotlb_free_coherent);
  657 
  658 static void
  659 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
  660              int do_panic)
  661 {
  662         /*
  663          * Ran out of IOMMU space for this operation. This is very bad.
  664          * Unfortunately the drivers cannot handle this operation properly.
  665          * unless they check for dma_mapping_error (most don't)
  666          * When the mapping is small enough return a static buffer to limit
  667          * the damage, or panic when the transfer is too big.
  668          */
  669         printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
  670                "device %s\n", size, dev ? dev_name(dev) : "?");
  671 
  672         if (size <= io_tlb_overflow || !do_panic)
  673                 return;
  674 
  675         if (dir == DMA_BIDIRECTIONAL)
  676                 panic("DMA: Random memory could be DMA accessed\n");
  677         if (dir == DMA_FROM_DEVICE)
  678                 panic("DMA: Random memory could be DMA written\n");
  679         if (dir == DMA_TO_DEVICE)
  680                 panic("DMA: Random memory could be DMA read\n");
  681 }
  682 
  683 /*
  684  * Map a single buffer of the indicated size for DMA in streaming mode.  The
  685  * physical address to use is returned.
  686  *
  687  * Once the device is given the dma address, the device owns this memory until
  688  * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
  689  */
  690 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
  691                             unsigned long offset, size_t size,
  692                             enum dma_data_direction dir,
  693                             struct dma_attrs *attrs)
  694 {
  695         phys_addr_t map, phys = page_to_phys(page) + offset;
  696         dma_addr_t dev_addr = phys_to_dma(dev, phys);
  697 
  698         BUG_ON(dir == DMA_NONE);
  699         /*
  700          * If the address happens to be in the device's DMA window,
  701          * we can safely return the device addr and not worry about bounce
  702          * buffering it.
  703          */
  704         if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
  705                 return dev_addr;
  706 
  707         /* Oh well, have to allocate and map a bounce buffer. */
  708         map = map_single(dev, phys, size, dir);
  709         if (map == SWIOTLB_MAP_ERROR) {
  710                 swiotlb_full(dev, size, dir, 1);
  711                 return phys_to_dma(dev, io_tlb_overflow_buffer);
  712         }
  713 
  714         dev_addr = phys_to_dma(dev, map);
  715 
  716         /* Ensure that the address returned is DMA'ble */
  717         if (!dma_capable(dev, dev_addr, size)) {
  718                 swiotlb_tbl_unmap_single(dev, map, size, dir);
  719                 return phys_to_dma(dev, io_tlb_overflow_buffer);
  720         }
  721 
  722         return dev_addr;
  723 }
  724 EXPORT_SYMBOL_GPL(swiotlb_map_page);
  725 
  726 /*
  727  * Unmap a single streaming mode DMA translation.  The dma_addr and size must
  728  * match what was provided for in a previous swiotlb_map_page call.  All
  729  * other usages are undefined.
  730  *
  731  * After this call, reads by the cpu to the buffer are guaranteed to see
  732  * whatever the device wrote there.
  733  */
  734 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
  735                          size_t size, enum dma_data_direction dir)
  736 {
  737         phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
  738 
  739         BUG_ON(dir == DMA_NONE);
  740 
  741         if (is_swiotlb_buffer(paddr)) {
  742                 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
  743                 return;
  744         }
  745 
  746         if (dir != DMA_FROM_DEVICE)
  747                 return;
  748 
  749         /*
  750          * phys_to_virt doesn't work with hihgmem page but we could
  751          * call dma_mark_clean() with hihgmem page here. However, we
  752          * are fine since dma_mark_clean() is null on POWERPC. We can
  753          * make dma_mark_clean() take a physical address if necessary.
  754          */
  755         dma_mark_clean(phys_to_virt(paddr), size);
  756 }
  757 
  758 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
  759                         size_t size, enum dma_data_direction dir,
  760                         struct dma_attrs *attrs)
  761 {
  762         unmap_single(hwdev, dev_addr, size, dir);
  763 }
  764 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
  765 
  766 /*
  767  * Make physical memory consistent for a single streaming mode DMA translation
  768  * after a transfer.
  769  *
  770  * If you perform a swiotlb_map_page() but wish to interrogate the buffer
  771  * using the cpu, yet do not wish to teardown the dma mapping, you must
  772  * call this function before doing so.  At the next point you give the dma
  773  * address back to the card, you must first perform a
  774  * swiotlb_dma_sync_for_device, and then the device again owns the buffer
  775  */
  776 static void
  777 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
  778                     size_t size, enum dma_data_direction dir,
  779                     enum dma_sync_target target)
  780 {
  781         phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
  782 
  783         BUG_ON(dir == DMA_NONE);
  784 
  785         if (is_swiotlb_buffer(paddr)) {
  786                 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
  787                 return;
  788         }
  789 
  790         if (dir != DMA_FROM_DEVICE)
  791                 return;
  792 
  793         dma_mark_clean(phys_to_virt(paddr), size);
  794 }
  795 
  796 void
  797 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
  798                             size_t size, enum dma_data_direction dir)
  799 {
  800         swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
  801 }
  802 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
  803 
  804 void
  805 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
  806                                size_t size, enum dma_data_direction dir)
  807 {
  808         swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
  809 }
  810 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
  811 
  812 /*
  813  * Map a set of buffers described by scatterlist in streaming mode for DMA.
  814  * This is the scatter-gather version of the above swiotlb_map_page
  815  * interface.  Here the scatter gather list elements are each tagged with the
  816  * appropriate dma address and length.  They are obtained via
  817  * sg_dma_{address,length}(SG).
  818  *
  819  * NOTE: An implementation may be able to use a smaller number of
  820  *       DMA address/length pairs than there are SG table elements.
  821  *       (for example via virtual mapping capabilities)
  822  *       The routine returns the number of addr/length pairs actually
  823  *       used, at most nents.
  824  *
  825  * Device ownership issues as mentioned above for swiotlb_map_page are the
  826  * same here.
  827  */
  828 int
  829 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
  830                      enum dma_data_direction dir, struct dma_attrs *attrs)
  831 {
  832         struct scatterlist *sg;
  833         int i;
  834 
  835         BUG_ON(dir == DMA_NONE);
  836 
  837         for_each_sg(sgl, sg, nelems, i) {
  838                 phys_addr_t paddr = sg_phys(sg);
  839                 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
  840 
  841                 if (swiotlb_force ||
  842                     !dma_capable(hwdev, dev_addr, sg->length)) {
  843                         phys_addr_t map = map_single(hwdev, sg_phys(sg),
  844                                                      sg->length, dir);
  845                         if (map == SWIOTLB_MAP_ERROR) {
  846                                 /* Don't panic here, we expect map_sg users
  847                                    to do proper error handling. */
  848                                 swiotlb_full(hwdev, sg->length, dir, 0);
  849                                 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
  850                                                        attrs);
  851                                 sgl[0].dma_length = 0;
  852                                 return 0;
  853                         }
  854                         sg->dma_address = phys_to_dma(hwdev, map);
  855                 } else
  856                         sg->dma_address = dev_addr;
  857                 sg->dma_length = sg->length;
  858         }
  859         return nelems;
  860 }
  861 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
  862 
  863 int
  864 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
  865                enum dma_data_direction dir)
  866 {
  867         return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
  868 }
  869 EXPORT_SYMBOL(swiotlb_map_sg);
  870 
  871 /*
  872  * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
  873  * concerning calls here are the same as for swiotlb_unmap_page() above.
  874  */
  875 void
  876 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
  877                        int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
  878 {
  879         struct scatterlist *sg;
  880         int i;
  881 
  882         BUG_ON(dir == DMA_NONE);
  883 
  884         for_each_sg(sgl, sg, nelems, i)
  885                 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
  886 
  887 }
  888 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
  889 
  890 void
  891 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
  892                  enum dma_data_direction dir)
  893 {
  894         return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
  895 }
  896 EXPORT_SYMBOL(swiotlb_unmap_sg);
  897 
  898 /*
  899  * Make physical memory consistent for a set of streaming mode DMA translations
  900  * after a transfer.
  901  *
  902  * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
  903  * and usage.
  904  */
  905 static void
  906 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
  907                 int nelems, enum dma_data_direction dir,
  908                 enum dma_sync_target target)
  909 {
  910         struct scatterlist *sg;
  911         int i;
  912 
  913         for_each_sg(sgl, sg, nelems, i)
  914                 swiotlb_sync_single(hwdev, sg->dma_address,
  915                                     sg->dma_length, dir, target);
  916 }
  917 
  918 void
  919 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
  920                         int nelems, enum dma_data_direction dir)
  921 {
  922         swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
  923 }
  924 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
  925 
  926 void
  927 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
  928                            int nelems, enum dma_data_direction dir)
  929 {
  930         swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
  931 }
  932 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
  933 
  934 int
  935 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
  936 {
  937         return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
  938 }
  939 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
  940 
  941 /*
  942  * Return whether the given device DMA address mask can be supported
  943  * properly.  For example, if your device can only drive the low 24-bits
  944  * during bus mastering, then you would pass 0x00ffffff as the mask to
  945  * this function.
  946  */
  947 int
  948 swiotlb_dma_supported(struct device *hwdev, u64 mask)
  949 {
  950         return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
  951 }
  952 EXPORT_SYMBOL(swiotlb_dma_supported);

Cache object: ecd96f611d46c0234e5f21928d383d03


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.