hw/ppc/spapr_pci_nvlink2.c

   1 /*
   2  * QEMU sPAPR PCI for NVLink2 pass through
   3  *
   4  * Copyright (c) 2019 Alexey Kardashevskiy, IBM Corporation.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24 #include "qemu/osdep.h"
  25 #include "qapi/error.h"
  26 #include "qemu-common.h"
  27 #include "hw/pci/pci.h"
  28 #include "hw/pci-host/spapr.h"
  29 #include "hw/ppc/spapr_numa.h"
  30 #include "qemu/error-report.h"
  31 #include "hw/ppc/fdt.h"
  32 #include "hw/pci/pci_bridge.h"
  33
  34 #define PHANDLE_PCIDEV(phb, pdev)    (0x12000000 | \
  35                                      (((phb)->index) << 16) | ((pdev)->devfn))
  36 #define PHANDLE_GPURAM(phb, n)       (0x110000FF | ((n) << 8) | \
  37                                      (((phb)->index) << 16))
  38 #define PHANDLE_NVLINK(phb, gn, nn)  (0x00130000 | (((phb)->index) << 8) | \
  39                                      ((gn) << 4) | (nn))
  40
  41 typedef struct SpaprPhbPciNvGpuSlot {
  42         uint64_t tgt;
  43         uint64_t gpa;
  44         unsigned numa_id;
  45         PCIDevice *gpdev;
  46         int linknum;
  47         struct {
  48             uint64_t atsd_gpa;
  49             PCIDevice *npdev;
  50             uint32_t link_speed;
  51         } links[NVGPU_MAX_LINKS];
  52 } SpaprPhbPciNvGpuSlot;
  53
  54 struct SpaprPhbPciNvGpuConfig {
  55     uint64_t nv2_ram_current;
  56     uint64_t nv2_atsd_current;
  57     int num; /* number of non empty (i.e. tgt!=0) entries in slots[] */
  58     SpaprPhbPciNvGpuSlot slots[NVGPU_MAX_NUM];
  59     Error *err;
  60 };
  61
  62 static SpaprPhbPciNvGpuSlot *
  63 spapr_nvgpu_get_slot(SpaprPhbPciNvGpuConfig *nvgpus, uint64_t tgt)
  64 {
  65     int i;
  66
  67     /* Search for partially collected "slot" */
  68     for (i = 0; i < nvgpus->num; ++i) {
  69         if (nvgpus->slots[i].tgt == tgt) {
  70             return &nvgpus->slots[i];
  71         }
  72     }
  73
  74     if (nvgpus->num == ARRAY_SIZE(nvgpus->slots)) {
  75         return NULL;
  76     }
  77
  78     i = nvgpus->num;
  79     nvgpus->slots[i].tgt = tgt;
  80     ++nvgpus->num;
  81
  82     return &nvgpus->slots[i];
  83 }
  84
  85 static void spapr_pci_collect_nvgpu(SpaprPhbPciNvGpuConfig *nvgpus,
  86                                     PCIDevice *pdev, uint64_t tgt,
  87                                     MemoryRegion *mr, Error **errp)
  88 {
  89     MachineState *machine = MACHINE(qdev_get_machine());
  90     SpaprMachineState *spapr = SPAPR_MACHINE(machine);
  91     SpaprPhbPciNvGpuSlot *nvslot = spapr_nvgpu_get_slot(nvgpus, tgt);
  92
  93     if (!nvslot) {
  94         error_setg(errp, "Found too many GPUs per vPHB");
  95         return;
  96     }
  97     g_assert(!nvslot->gpdev);
  98     nvslot->gpdev = pdev;
  99
 100     nvslot->gpa = nvgpus->nv2_ram_current;
 101     nvgpus->nv2_ram_current += memory_region_size(mr);
 102     nvslot->numa_id = spapr->gpu_numa_id;
 103     ++spapr->gpu_numa_id;
 104 }
 105
 106 static void spapr_pci_collect_nvnpu(SpaprPhbPciNvGpuConfig *nvgpus,
 107                                     PCIDevice *pdev, uint64_t tgt,
 108                                     MemoryRegion *mr, Error **errp)
 109 {
 110     SpaprPhbPciNvGpuSlot *nvslot = spapr_nvgpu_get_slot(nvgpus, tgt);
 111     int j;
 112
 113     if (!nvslot) {
 114         error_setg(errp, "Found too many NVLink bridges per vPHB");
 115         return;
 116     }
 117
 118     j = nvslot->linknum;
 119     if (j == ARRAY_SIZE(nvslot->links)) {
 120         error_setg(errp, "Found too many NVLink bridges per GPU");
 121         return;
 122     }
 123     ++nvslot->linknum;
 124
 125     g_assert(!nvslot->links[j].npdev);
 126     nvslot->links[j].npdev = pdev;
 127     nvslot->links[j].atsd_gpa = nvgpus->nv2_atsd_current;
 128     nvgpus->nv2_atsd_current += memory_region_size(mr);
 129     nvslot->links[j].link_speed =
 130         object_property_get_uint(OBJECT(pdev), "nvlink2-link-speed", NULL);
 131 }
 132
 133 static void spapr_phb_pci_collect_nvgpu(PCIBus *bus, PCIDevice *pdev,
 134                                         void *opaque)
 135 {
 136     PCIBus *sec_bus;
 137     Object *po = OBJECT(pdev);
 138     uint64_t tgt = object_property_get_uint(po, "nvlink2-tgt", NULL);
 139
 140     if (tgt) {
 141         Error *local_err = NULL;
 142         SpaprPhbPciNvGpuConfig *nvgpus = opaque;
 143         Object *mr_gpu = object_property_get_link(po, "nvlink2-mr[0]", NULL);
 144         Object *mr_npu = object_property_get_link(po, "nvlink2-atsd-mr[0]",
 145                                                   NULL);
 146
 147         g_assert(mr_gpu || mr_npu);
 148         if (mr_gpu) {
 149             spapr_pci_collect_nvgpu(nvgpus, pdev, tgt, MEMORY_REGION(mr_gpu),
 150                                     &local_err);
 151         } else {
 152             spapr_pci_collect_nvnpu(nvgpus, pdev, tgt, MEMORY_REGION(mr_npu),
 153                                     &local_err);
 154         }
 155         error_propagate(&nvgpus->err, local_err);
 156     }
 157     if ((pci_default_read_config(pdev, PCI_HEADER_TYPE, 1) !=
 158          PCI_HEADER_TYPE_BRIDGE)) {
 159         return;
 160     }
 161
 162     sec_bus = pci_bridge_get_sec_bus(PCI_BRIDGE(pdev));
 163     if (!sec_bus) {
 164         return;
 165     }
 166
 167     pci_for_each_device_under_bus(sec_bus, spapr_phb_pci_collect_nvgpu, opaque);
 168 }
 169
 170 void spapr_phb_nvgpu_setup(SpaprPhbState *sphb, Error **errp)
 171 {
 172     int i, j, valid_gpu_num;
 173     PCIBus *bus;
 174
 175     /* Search for GPUs and NPUs */
 176     if (!sphb->nv2_gpa_win_addr || !sphb->nv2_atsd_win_addr) {
 177         return;
 178     }
 179
 180     sphb->nvgpus = g_new0(SpaprPhbPciNvGpuConfig, 1);
 181     sphb->nvgpus->nv2_ram_current = sphb->nv2_gpa_win_addr;
 182     sphb->nvgpus->nv2_atsd_current = sphb->nv2_atsd_win_addr;
 183
 184     bus = PCI_HOST_BRIDGE(sphb)->bus;
 185     pci_for_each_device_under_bus(bus, spapr_phb_pci_collect_nvgpu,
 186                                   sphb->nvgpus);
 187
 188     if (sphb->nvgpus->err) {
 189         error_propagate(errp, sphb->nvgpus->err);
 190         sphb->nvgpus->err = NULL;
 191         goto cleanup_exit;
 192     }
 193
 194     /* Add found GPU RAM and ATSD MRs if found */
 195     for (i = 0, valid_gpu_num = 0; i < sphb->nvgpus->num; ++i) {
 196         Object *nvmrobj;
 197         SpaprPhbPciNvGpuSlot *nvslot = &sphb->nvgpus->slots[i];
 198
 199         if (!nvslot->gpdev) {
 200             continue;
 201         }
 202         nvmrobj = object_property_get_link(OBJECT(nvslot->gpdev),
 203                                            "nvlink2-mr[0]", NULL);
 204         /* ATSD is pointless without GPU RAM MR so skip those */
 205         if (!nvmrobj) {
 206             continue;
 207         }
 208
 209         ++valid_gpu_num;
 210         memory_region_add_subregion(get_system_memory(), nvslot->gpa,
 211                                     MEMORY_REGION(nvmrobj));
 212
 213         for (j = 0; j < nvslot->linknum; ++j) {
 214             Object *atsdmrobj;
 215
 216             atsdmrobj = object_property_get_link(OBJECT(nvslot->links[j].npdev),
 217                                                  "nvlink2-atsd-mr[0]", NULL);
 218             if (!atsdmrobj) {
 219                 continue;
 220             }
 221             memory_region_add_subregion(get_system_memory(),
 222                                         nvslot->links[j].atsd_gpa,
 223                                         MEMORY_REGION(atsdmrobj));
 224         }
 225     }
 226
 227     if (valid_gpu_num) {
 228         return;
 229     }
 230     /* We did not find any interesting GPU */
 231 cleanup_exit:
 232     g_free(sphb->nvgpus);
 233     sphb->nvgpus = NULL;
 234 }
 235
 236 void spapr_phb_nvgpu_free(SpaprPhbState *sphb)
 237 {
 238     int i, j;
 239
 240     if (!sphb->nvgpus) {
 241         return;
 242     }
 243
 244     for (i = 0; i < sphb->nvgpus->num; ++i) {
 245         SpaprPhbPciNvGpuSlot *nvslot = &sphb->nvgpus->slots[i];
 246         Object *nv_mrobj = object_property_get_link(OBJECT(nvslot->gpdev),
 247                                                     "nvlink2-mr[0]", NULL);
 248
 249         if (nv_mrobj) {
 250             memory_region_del_subregion(get_system_memory(),
 251                                         MEMORY_REGION(nv_mrobj));
 252         }
 253         for (j = 0; j < nvslot->linknum; ++j) {
 254             PCIDevice *npdev = nvslot->links[j].npdev;
 255             Object *atsd_mrobj;
 256             atsd_mrobj = object_property_get_link(OBJECT(npdev),
 257                                                   "nvlink2-atsd-mr[0]", NULL);
 258             if (atsd_mrobj) {
 259                 memory_region_del_subregion(get_system_memory(),
 260                                             MEMORY_REGION(atsd_mrobj));
 261             }
 262         }
 263     }
 264     g_free(sphb->nvgpus);
 265     sphb->nvgpus = NULL;
 266 }
 267
 268 void spapr_phb_nvgpu_populate_dt(SpaprPhbState *sphb, void *fdt, int bus_off,
 269                                  Error **errp)
 270 {
 271     int i, j, atsdnum = 0;
 272     uint64_t atsd[8]; /* The existing limitation of known guests */
 273
 274     if (!sphb->nvgpus) {
 275         return;
 276     }
 277
 278     for (i = 0; (i < sphb->nvgpus->num) && (atsdnum < ARRAY_SIZE(atsd)); ++i) {
 279         SpaprPhbPciNvGpuSlot *nvslot = &sphb->nvgpus->slots[i];
 280
 281         if (!nvslot->gpdev) {
 282             continue;
 283         }
 284         for (j = 0; j < nvslot->linknum; ++j) {
 285             if (!nvslot->links[j].atsd_gpa) {
 286                 continue;
 287             }
 288
 289             if (atsdnum == ARRAY_SIZE(atsd)) {
 290                 error_report("Only %"PRIuPTR" ATSD registers supported",
 291                              ARRAY_SIZE(atsd));
 292                 break;
 293             }
 294             atsd[atsdnum] = cpu_to_be64(nvslot->links[j].atsd_gpa);
 295             ++atsdnum;
 296         }
 297     }
 298
 299     if (!atsdnum) {
 300         error_setg(errp, "No ATSD registers found");
 301         return;
 302     }
 303
 304     if (!spapr_phb_eeh_available(sphb)) {
 305         /*
 306          * ibm,mmio-atsd contains ATSD registers; these belong to an NPU PHB
 307          * which we do not emulate as a separate device. Instead we put
 308          * ibm,mmio-atsd to the vPHB with GPU and make sure that we do not
 309          * put GPUs from different IOMMU groups to the same vPHB to ensure
 310          * that the guest will use ATSDs from the corresponding NPU.
 311          */
 312         error_setg(errp, "ATSD requires separate vPHB per GPU IOMMU group");
 313         return;
 314     }
 315
 316     _FDT((fdt_setprop(fdt, bus_off, "ibm,mmio-atsd", atsd,
 317                       atsdnum * sizeof(atsd[0]))));
 318 }
 319
 320 void spapr_phb_nvgpu_ram_populate_dt(SpaprPhbState *sphb, void *fdt)
 321 {
 322     int i, j, linkidx, npuoff;
 323     char *npuname;
 324
 325     if (!sphb->nvgpus) {
 326         return;
 327     }
 328
 329     npuname = g_strdup_printf("npuphb%d", sphb->index);
 330     npuoff = fdt_add_subnode(fdt, 0, npuname);
 331     _FDT(npuoff);
 332     _FDT(fdt_setprop_cell(fdt, npuoff, "#address-cells", 1));
 333     _FDT(fdt_setprop_cell(fdt, npuoff, "#size-cells", 0));
 334     /* Advertise NPU as POWER9 so the guest can enable NPU2 contexts */
 335     _FDT((fdt_setprop_string(fdt, npuoff, "compatible", "ibm,power9-npu")));
 336     g_free(npuname);
 337
 338     for (i = 0, linkidx = 0; i < sphb->nvgpus->num; ++i) {
 339         for (j = 0; j < sphb->nvgpus->slots[i].linknum; ++j) {
 340             char *linkname = g_strdup_printf("link@%d", linkidx);
 341             int off = fdt_add_subnode(fdt, npuoff, linkname);
 342
 343             _FDT(off);
 344             /* _FDT((fdt_setprop_cell(fdt, off, "reg", linkidx))); */
 345             _FDT((fdt_setprop_string(fdt, off, "compatible",
 346                                      "ibm,npu-link")));
 347             _FDT((fdt_setprop_cell(fdt, off, "phandle",
 348                                    PHANDLE_NVLINK(sphb, i, j))));
 349             _FDT((fdt_setprop_cell(fdt, off, "ibm,npu-link-index", linkidx)));
 350             g_free(linkname);
 351             ++linkidx;
 352         }
 353     }
 354
 355     /* Add memory nodes for GPU RAM and mark them unusable */
 356     for (i = 0; i < sphb->nvgpus->num; ++i) {
 357         SpaprPhbPciNvGpuSlot *nvslot = &sphb->nvgpus->slots[i];
 358         Object *nv_mrobj = object_property_get_link(OBJECT(nvslot->gpdev),
 359                                                     "nvlink2-mr[0]",
 360                                                     &error_abort);
 361         uint64_t size = object_property_get_uint(nv_mrobj, "size", NULL);
 362         uint64_t mem_reg[2] = { cpu_to_be64(nvslot->gpa), cpu_to_be64(size) };
 363         char *mem_name = g_strdup_printf("memory@%"PRIx64, nvslot->gpa);
 364         int off = fdt_add_subnode(fdt, 0, mem_name);
 365
 366         _FDT(off);
 367         _FDT((fdt_setprop_string(fdt, off, "device_type", "memory")));
 368         _FDT((fdt_setprop(fdt, off, "reg", mem_reg, sizeof(mem_reg))));
 369
 370         spapr_numa_write_associativity_dt(SPAPR_MACHINE(qdev_get_machine()),
 371                                           fdt, off, nvslot->numa_id);
 372
 373         _FDT((fdt_setprop_string(fdt, off, "compatible",
 374                                  "ibm,coherent-device-memory")));
 375
 376         mem_reg[1] = cpu_to_be64(0);
 377         _FDT((fdt_setprop(fdt, off, "linux,usable-memory", mem_reg,
 378                           sizeof(mem_reg))));
 379         _FDT((fdt_setprop_cell(fdt, off, "phandle",
 380                                PHANDLE_GPURAM(sphb, i))));
 381         g_free(mem_name);
 382     }
 383
 384 }
 385
 386 void spapr_phb_nvgpu_populate_pcidev_dt(PCIDevice *dev, void *fdt, int offset,
 387                                         SpaprPhbState *sphb)
 388 {
 389     int i, j;
 390
 391     if (!sphb->nvgpus) {
 392         return;
 393     }
 394
 395     for (i = 0; i < sphb->nvgpus->num; ++i) {
 396         SpaprPhbPciNvGpuSlot *nvslot = &sphb->nvgpus->slots[i];
 397
 398         /* Skip "slot" without attached GPU */
 399         if (!nvslot->gpdev) {
 400             continue;
 401         }
 402         if (dev == nvslot->gpdev) {
 403             uint32_t npus[nvslot->linknum];
 404
 405             for (j = 0; j < nvslot->linknum; ++j) {
 406                 PCIDevice *npdev = nvslot->links[j].npdev;
 407
 408                 npus[j] = cpu_to_be32(PHANDLE_PCIDEV(sphb, npdev));
 409             }
 410             _FDT(fdt_setprop(fdt, offset, "ibm,npu", npus,
 411                              j * sizeof(npus[0])));
 412             _FDT((fdt_setprop_cell(fdt, offset, "phandle",
 413                                    PHANDLE_PCIDEV(sphb, dev))));
 414             continue;
 415         }
 416
 417         for (j = 0; j < nvslot->linknum; ++j) {
 418             if (dev != nvslot->links[j].npdev) {
 419                 continue;
 420             }
 421
 422             _FDT((fdt_setprop_cell(fdt, offset, "phandle",
 423                                    PHANDLE_PCIDEV(sphb, dev))));
 424             _FDT(fdt_setprop_cell(fdt, offset, "ibm,gpu",
 425                                   PHANDLE_PCIDEV(sphb, nvslot->gpdev)));
 426             _FDT((fdt_setprop_cell(fdt, offset, "ibm,nvlink",
 427                                    PHANDLE_NVLINK(sphb, i, j))));
 428             /*
 429              * If we ever want to emulate GPU RAM at the same location as on
 430              * the host - here is the encoding GPA->TGT:
 431              *
 432              * gta  = ((sphb->nv2_gpa >> 42) & 0x1) << 42;
 433              * gta |= ((sphb->nv2_gpa >> 45) & 0x3) << 43;
 434              * gta |= ((sphb->nv2_gpa >> 49) & 0x3) << 45;
 435              * gta |= sphb->nv2_gpa & ((1UL << 43) - 1);
 436              */
 437             _FDT(fdt_setprop_cell(fdt, offset, "memory-region",
 438                                   PHANDLE_GPURAM(sphb, i)));
 439             _FDT(fdt_setprop_u64(fdt, offset, "ibm,device-tgt-addr",
 440                                  nvslot->tgt));
 441             _FDT(fdt_setprop_cell(fdt, offset, "ibm,nvlink-speed",
 442                                   nvslot->links[j].link_speed));
 443         }
 444     }
 445 }