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 "qemu/error-report.h"
  19 #include "trace.h"
  20
  21 static bool vfio_prereg_listener_skipped_section(MemoryRegionSection *section)
  22 {
  23     if (memory_region_is_iommu(section->mr)) {
  24         hw_error("Cannot possibly preregister IOMMU memory");
  25     }
  26
  27     return !memory_region_is_ram(section->mr) ||
  28             memory_region_is_ram_device(section->mr);
  29 }
  30
  31 static void *vfio_prereg_gpa_to_vaddr(MemoryRegionSection *section, hwaddr gpa)
  32 {
  33     return memory_region_get_ram_ptr(section->mr) +
  34         section->offset_within_region +
  35         (gpa - section->offset_within_address_space);
  36 }
  37
  38 static void vfio_prereg_listener_region_add(MemoryListener *listener,
  39                                             MemoryRegionSection *section)
  40 {
  41     VFIOContainer *container = container_of(listener, VFIOContainer,
  42                                             prereg_listener);
  43     const hwaddr gpa = section->offset_within_address_space;
  44     hwaddr end;
  45     int ret;
  46     hwaddr page_mask = qemu_real_host_page_mask;
  47     struct vfio_iommu_spapr_register_memory reg = {
  48         .argsz = sizeof(reg),
  49         .flags = 0,
  50     };
  51
  52     if (vfio_prereg_listener_skipped_section(section)) {
  53         trace_vfio_prereg_listener_region_add_skip(
  54                 section->offset_within_address_space,
  55                 section->offset_within_address_space +
  56                 int128_get64(int128_sub(section->size, int128_one())));
  57         return;
  58     }
  59
  60     if (unlikely((section->offset_within_address_space & ~page_mask) ||
  61                  (section->offset_within_region & ~page_mask) ||
  62                  (int128_get64(section->size) & ~page_mask))) {
  63         error_report("%s received unaligned region", __func__);
  64         return;
  65     }
  66
  67     end = section->offset_within_address_space + int128_get64(section->size);
  68     if (gpa >= end) {
  69         return;
  70     }
  71
  72     memory_region_ref(section->mr);
  73
  74     reg.vaddr = (uintptr_t) vfio_prereg_gpa_to_vaddr(section, gpa);
  75     reg.size = end - gpa;
  76
  77     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_REGISTER_MEMORY, &reg);
  78     trace_vfio_prereg_register(reg.vaddr, reg.size, ret ? -errno : 0);
  79     if (ret) {
  80         /*
  81          * On the initfn path, store the first error in the container so we
  82          * can gracefully fail.  Runtime, there's not much we can do other
  83          * than throw a hardware error.
  84          */
  85         if (!container->initialized) {
  86             if (!container->error) {
  87                 container->error = ret;
  88             }
  89         } else {
  90             hw_error("vfio: Memory registering failed, unable to continue");
  91         }
  92     }
  93 }
  94
  95 static void vfio_prereg_listener_region_del(MemoryListener *listener,
  96                                             MemoryRegionSection *section)
  97 {
  98     VFIOContainer *container = container_of(listener, VFIOContainer,
  99                                             prereg_listener);
 100     const hwaddr gpa = section->offset_within_address_space;
 101     hwaddr end;
 102     int ret;
 103     hwaddr page_mask = qemu_real_host_page_mask;
 104     struct vfio_iommu_spapr_register_memory reg = {
 105         .argsz = sizeof(reg),
 106         .flags = 0,
 107     };
 108
 109     if (vfio_prereg_listener_skipped_section(section)) {
 110         trace_vfio_prereg_listener_region_del_skip(
 111                 section->offset_within_address_space,
 112                 section->offset_within_address_space +
 113                 int128_get64(int128_sub(section->size, int128_one())));
 114         return;
 115     }
 116
 117     if (unlikely((section->offset_within_address_space & ~page_mask) ||
 118                  (section->offset_within_region & ~page_mask) ||
 119                  (int128_get64(section->size) & ~page_mask))) {
 120         error_report("%s received unaligned region", __func__);
 121         return;
 122     }
 123
 124     end = section->offset_within_address_space + int128_get64(section->size);
 125     if (gpa >= end) {
 126         return;
 127     }
 128
 129     reg.vaddr = (uintptr_t) vfio_prereg_gpa_to_vaddr(section, gpa);
 130     reg.size = end - gpa;
 131
 132     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, &reg);
 133     trace_vfio_prereg_unregister(reg.vaddr, reg.size, ret ? -errno : 0);
 134 }
 135
 136 const MemoryListener vfio_prereg_listener = {
 137     .region_add = vfio_prereg_listener_region_add,
 138     .region_del = vfio_prereg_listener_region_del,
 139 };
 140
 141 int vfio_spapr_create_window(VFIOContainer *container,
 142                              MemoryRegionSection *section,
 143                              hwaddr *pgsize)
 144 {
 145     int ret;
 146     unsigned pagesize = memory_region_iommu_get_min_page_size(section->mr);
 147     unsigned entries, pages;
 148     struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) };
 149
 150     /*
 151      * FIXME: For VFIO iommu types which have KVM acceleration to
 152      * avoid bouncing all map/unmaps through qemu this way, this
 153      * would be the right place to wire that up (tell the KVM
 154      * device emulation the VFIO iommu handles to use).
 155      */
 156     create.window_size = int128_get64(section->size);
 157     create.page_shift = ctz64(pagesize);
 158     /*
 159      * SPAPR host supports multilevel TCE tables, there is some
 160      * heuristic to decide how many levels we want for our table:
 161      * 0..64 = 1; 65..4096 = 2; 4097..262144 = 3; 262145.. = 4
 162      */
 163     entries = create.window_size >> create.page_shift;
 164     pages = MAX((entries * sizeof(uint64_t)) / getpagesize(), 1);
 165     pages = MAX(pow2ceil(pages) - 1, 1); /* Round up */
 166     create.levels = ctz64(pages) / 6 + 1;
 167
 168     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
 169     if (ret) {
 170         error_report("Failed to create a window, ret = %d (%m)", ret);
 171         return -errno;
 172     }
 173
 174     if (create.start_addr != section->offset_within_address_space) {
 175         vfio_spapr_remove_window(container, create.start_addr);
 176
 177         error_report("Host doesn't support DMA window at %"HWADDR_PRIx", must be %"PRIx64,
 178                      section->offset_within_address_space,
 179                      (uint64_t)create.start_addr);
 180         return -EINVAL;
 181     }
 182     trace_vfio_spapr_create_window(create.page_shift,
 183                                    create.window_size,
 184                                    create.start_addr);
 185     *pgsize = pagesize;
 186
 187     return 0;
 188 }
 189
 190 int vfio_spapr_remove_window(VFIOContainer *container,
 191                              hwaddr offset_within_address_space)
 192 {
 193     struct vfio_iommu_spapr_tce_remove remove = {
 194         .argsz = sizeof(remove),
 195         .start_addr = offset_within_address_space,
 196     };
 197     int ret;
 198
 199     ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
 200     if (ret) {
 201         error_report("Failed to remove window at %"PRIx64,
 202                      (uint64_t)remove.start_addr);
 203         return -errno;
 204     }
 205
 206     trace_vfio_spapr_remove_window(offset_within_address_space);
 207
 208     return 0;
 209 }