hw/vfio/spapr.c

   1 /*
   2  * DMA memory preregistration
   3  *
   4  * Authors:
   5  *  Alexey Kardashevskiy <aik@ozlabs.ru>
   6  *
   7  * This work is licensed under the terms of the GNU GPL, version 2.  See
   8  * the COPYING file in the top-level directory.
   9  */
  10
  11 #include "qemu/osdep.h"
  12 #include "cpu.h"
  13 #include <sys/ioctl.h>
  14 #include <linux/vfio.h>
  15
  16 #include "hw/vfio/vfio-common.h"
  17 #include "hw/hw.h"
  18 #include "exec/ram_addr.h"
  19 #include "qemu/error-report.h"
  20 #include "trace.h"
  21
  22 static bool vfio_prereg_listener_skipped_section(MemoryRegionSection *section)
  23 {
  24     if (memory_region_is_iommu(section->mr)) {
  25         hw_error("Cannot possibly preregister IOMMU memory");
  26     }
  27
  28     return !memory_region_is_ram(section->mr) ||
  29             memory_region_is_ram_device(section->mr);
  30 }
  31
  32 static void *vfio_prereg_gpa_to_vaddr(MemoryRegionSection *section, hwaddr gpa)
  33 {
  34     return memory_region_get_ram_ptr(section->mr) +
  35         section->offset_within_region +
  36         (gpa - section->offset_within_address_space);
  37 }
  38
  39 static void vfio_prereg_listener_region_add(MemoryListener *listener,
  40                                             MemoryRegionSection *section)
  41 {
  42     VFIOContainer *container = container_of(listener, VFIOContainer,
  43                                             prereg_listener);
  44     const hwaddr gpa = section->offset_within_address_space;
  45     hwaddr end;
  46     int ret;
  47     hwaddr page_mask = qemu_real_host_page_mask;
  48     struct vfio_iommu_spapr_register_memory reg = {
  49         .argsz = sizeof(reg),
  50         .flags = 0,
  51     };
  52
  53     if (vfio_prereg_listener_skipped_section(section)) {
  54         trace_vfio_prereg_listener_region_add_skip(
  55                 section->offset_within_address_space,
  56                 section->offset_within_address_space +
  57                 int128_get64(int128_sub(section->size, int128_one())));
  58         return;
  59     }
  60
  61     if (unlikely((section->offset_within_address_space & ~page_mask) ||
  62                  (section->offset_within_region & ~page_mask) ||
  63                  (int128_get64(section->size) & ~page_mask))) {
  64         error_report("%s received unaligned region", __func__);
  65         return;
  66     }
  67
  68     end = section->offset_within_address_space + int128_get64(section->size);
  69     if (gpa >= end) {
  70         return;
  71     }
  72
  73     memory_region_ref(section->mr);
  74
  75     reg.vaddr = (uintptr_t) vfio_prereg_gpa_to_vaddr(section, gpa);
  76     reg.size = end - gpa;
  77
  78     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_REGISTER_MEMORY, &reg);
  79     trace_vfio_prereg_register(reg.vaddr, reg.size, ret ? -errno : 0);
  80     if (ret) {
  81         /*
  82          * On the initfn path, store the first error in the container so we
  83          * can gracefully fail.  Runtime, there's not much we can do other
  84          * than throw a hardware error.
  85          */
  86         if (!container->initialized) {
  87             if (!container->error) {
  88                 container->error = ret;
  89             }
  90         } else {
  91             hw_error("vfio: Memory registering failed, unable to continue");
  92         }
  93     }
  94 }
  95
  96 static void vfio_prereg_listener_region_del(MemoryListener *listener,
  97                                             MemoryRegionSection *section)
  98 {
  99     VFIOContainer *container = container_of(listener, VFIOContainer,
 100                                             prereg_listener);
 101     const hwaddr gpa = section->offset_within_address_space;
 102     hwaddr end;
 103     int ret;
 104     hwaddr page_mask = qemu_real_host_page_mask;
 105     struct vfio_iommu_spapr_register_memory reg = {
 106         .argsz = sizeof(reg),
 107         .flags = 0,
 108     };
 109
 110     if (vfio_prereg_listener_skipped_section(section)) {
 111         trace_vfio_prereg_listener_region_del_skip(
 112                 section->offset_within_address_space,
 113                 section->offset_within_address_space +
 114                 int128_get64(int128_sub(section->size, int128_one())));
 115         return;
 116     }
 117
 118     if (unlikely((section->offset_within_address_space & ~page_mask) ||
 119                  (section->offset_within_region & ~page_mask) ||
 120                  (int128_get64(section->size) & ~page_mask))) {
 121         error_report("%s received unaligned region", __func__);
 122         return;
 123     }
 124
 125     end = section->offset_within_address_space + int128_get64(section->size);
 126     if (gpa >= end) {
 127         return;
 128     }
 129
 130     reg.vaddr = (uintptr_t) vfio_prereg_gpa_to_vaddr(section, gpa);
 131     reg.size = end - gpa;
 132
 133     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, &reg);
 134     trace_vfio_prereg_unregister(reg.vaddr, reg.size, ret ? -errno : 0);
 135 }
 136
 137 const MemoryListener vfio_prereg_listener = {
 138     .region_add = vfio_prereg_listener_region_add,
 139     .region_del = vfio_prereg_listener_region_del,
 140 };
 141
 142 int vfio_spapr_create_window(VFIOContainer *container,
 143                              MemoryRegionSection *section,
 144                              hwaddr *pgsize)
 145 {
 146     int ret = 0;
 147     IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
 148     uint64_t pagesize = memory_region_iommu_get_min_page_size(iommu_mr);
 149     unsigned entries, bits_total, bits_per_level, max_levels;
 150     struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) };
 151     long rampagesize = qemu_minrampagesize();
 152
 153     /*
 154      * The host might not support the guest supported IOMMU page size,
 155      * so we will use smaller physical IOMMU pages to back them.
 156      */
 157     if (pagesize > rampagesize) {
 158         pagesize = rampagesize;
 159     }
 160     pagesize = 1ULL << (63 - clz64(container->pgsizes &
 161                                    (pagesize | (pagesize - 1))));
 162     if (!pagesize) {
 163         error_report("Host doesn't support page size 0x%"PRIx64
 164                      ", the supported mask is 0x%lx",
 165                      memory_region_iommu_get_min_page_size(iommu_mr),
 166                      container->pgsizes);
 167         return -EINVAL;
 168     }
 169
 170     /*
 171      * FIXME: For VFIO iommu types which have KVM acceleration to
 172      * avoid bouncing all map/unmaps through qemu this way, this
 173      * would be the right place to wire that up (tell the KVM
 174      * device emulation the VFIO iommu handles to use).
 175      */
 176     create.window_size = int128_get64(section->size);
 177     create.page_shift = ctz64(pagesize);
 178     /*
 179      * SPAPR host supports multilevel TCE tables. We try to guess optimal
 180      * levels number and if this fails (for example due to the host memory
 181      * fragmentation), we increase levels. The DMA address structure is:
 182      * rrrrrrrr rxxxxxxx xxxxxxxx xxxxxxxx  xxxxxxxx xxxxxxxx xxxxxxxx iiiiiiii
 183      * where:
 184      *   r = reserved (bits >= 55 are reserved in the existing hardware)
 185      *   i = IOMMU page offset (64K in this example)
 186      *   x = bits to index a TCE which can be split to equal chunks to index
 187      *      within the level.
 188      * The aim is to split "x" to smaller possible number of levels.
 189      */
 190     entries = create.window_size >> create.page_shift;
 191     /* bits_total is number of "x" needed */
 192     bits_total = ctz64(entries * sizeof(uint64_t));
 193     /*
 194      * bits_per_level is a safe guess of how much we can allocate per level:
 195      * 8 is the current minimum for CONFIG_FORCE_MAX_ZONEORDER and MAX_ORDER
 196      * is usually bigger than that.
 197      * Below we look at getpagesize() as TCEs are allocated from system pages.
 198      */
 199     bits_per_level = ctz64(getpagesize()) + 8;
 200     create.levels = bits_total / bits_per_level;
 201     if (bits_total % bits_per_level) {
 202         ++create.levels;
 203     }
 204     max_levels = (64 - create.page_shift) / ctz64(getpagesize());
 205     for ( ; create.levels <= max_levels; ++create.levels) {
 206         ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
 207         if (!ret) {
 208             break;
 209         }
 210     }
 211     if (ret) {
 212         error_report("Failed to create a window, ret = %d (%m)", ret);
 213         return -errno;
 214     }
 215
 216     if (create.start_addr != section->offset_within_address_space) {
 217         vfio_spapr_remove_window(container, create.start_addr);
 218
 219         error_report("Host doesn't support DMA window at %"HWADDR_PRIx", must be %"PRIx64,
 220                      section->offset_within_address_space,
 221                      (uint64_t)create.start_addr);
 222         return -EINVAL;
 223     }
 224     trace_vfio_spapr_create_window(create.page_shift,
 225                                    create.levels,
 226                                    create.window_size,
 227                                    create.start_addr);
 228     *pgsize = pagesize;
 229
 230     return 0;
 231 }
 232
 233 int vfio_spapr_remove_window(VFIOContainer *container,
 234                              hwaddr offset_within_address_space)
 235 {
 236     struct vfio_iommu_spapr_tce_remove remove = {
 237         .argsz = sizeof(remove),
 238         .start_addr = offset_within_address_space,
 239     };
 240     int ret;
 241
 242     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
 243     if (ret) {
 244         error_report("Failed to remove window at %"PRIx64,
 245                      (uint64_t)remove.start_addr);
 246         return -errno;
 247     }
 248
 249     trace_vfio_spapr_remove_window(offset_within_address_space);
 250
 251     return 0;
 252 }