drivers/gpu/drm/amd/amdkfd/kfd_device.c

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include <linux/bsearch.h>
  24 #include <linux/pci.h>
  25 #include <linux/slab.h>
  26 #include "kfd_priv.h"
  27 #include "kfd_device_queue_manager.h"
  28 #include "kfd_pm4_headers_vi.h"
  29 #include "cwsr_trap_handler.h"
  30 #include "kfd_iommu.h"
  31 #include "amdgpu_amdkfd.h"
  32
  33 #define MQD_SIZE_ALIGNED 768
  34
  35 /*
  36  * kfd_locked is used to lock the kfd driver during suspend or reset
  37  * once locked, kfd driver will stop any further GPU execution.
  38  * create process (open) will return -EAGAIN.
  39  */
  40 static atomic_t kfd_locked = ATOMIC_INIT(0);
  41
  42 #ifdef KFD_SUPPORT_IOMMU_V2
  43 static const struct kfd_device_info kaveri_device_info = {
  44         .asic_family = CHIP_KAVERI,
  45         .max_pasid_bits = 16,
  46         /* max num of queues for KV.TODO should be a dynamic value */
  47         .max_no_of_hqd  = 24,
  48         .doorbell_size  = 4,
  49         .ih_ring_entry_size = 4 * sizeof(uint32_t),
  50         .event_interrupt_class = &event_interrupt_class_cik,
  51         .num_of_watch_points = 4,
  52         .mqd_size_aligned = MQD_SIZE_ALIGNED,
  53         .supports_cwsr = false,
  54         .needs_iommu_device = true,
  55         .needs_pci_atomics = false,
  56         .num_sdma_engines = 2,
  57         .num_sdma_queues_per_engine = 2,
  58 };
  59
  60 static const struct kfd_device_info carrizo_device_info = {
  61         .asic_family = CHIP_CARRIZO,
  62         .max_pasid_bits = 16,
  63         /* max num of queues for CZ.TODO should be a dynamic value */
  64         .max_no_of_hqd  = 24,
  65         .doorbell_size  = 4,
  66         .ih_ring_entry_size = 4 * sizeof(uint32_t),
  67         .event_interrupt_class = &event_interrupt_class_cik,
  68         .num_of_watch_points = 4,
  69         .mqd_size_aligned = MQD_SIZE_ALIGNED,
  70         .supports_cwsr = true,
  71         .needs_iommu_device = true,
  72         .needs_pci_atomics = false,
  73         .num_sdma_engines = 2,
  74         .num_sdma_queues_per_engine = 2,
  75 };
  76
  77 static const struct kfd_device_info raven_device_info = {
  78         .asic_family = CHIP_RAVEN,
  79         .max_pasid_bits = 16,
  80         .max_no_of_hqd  = 24,
  81         .doorbell_size  = 8,
  82         .ih_ring_entry_size = 8 * sizeof(uint32_t),
  83         .event_interrupt_class = &event_interrupt_class_v9,
  84         .num_of_watch_points = 4,
  85         .mqd_size_aligned = MQD_SIZE_ALIGNED,
  86         .supports_cwsr = true,
  87         .needs_iommu_device = true,
  88         .needs_pci_atomics = true,
  89         .num_sdma_engines = 1,
  90         .num_sdma_queues_per_engine = 2,
  91 };
  92 #endif
  93
  94 static const struct kfd_device_info hawaii_device_info = {
  95         .asic_family = CHIP_HAWAII,
  96         .max_pasid_bits = 16,
  97         /* max num of queues for KV.TODO should be a dynamic value */
  98         .max_no_of_hqd  = 24,
  99         .doorbell_size  = 4,
 100         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 101         .event_interrupt_class = &event_interrupt_class_cik,
 102         .num_of_watch_points = 4,
 103         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 104         .supports_cwsr = false,
 105         .needs_iommu_device = false,
 106         .needs_pci_atomics = false,
 107         .num_sdma_engines = 2,
 108         .num_sdma_queues_per_engine = 2,
 109 };
 110
 111 static const struct kfd_device_info tonga_device_info = {
 112         .asic_family = CHIP_TONGA,
 113         .max_pasid_bits = 16,
 114         .max_no_of_hqd  = 24,
 115         .doorbell_size  = 4,
 116         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 117         .event_interrupt_class = &event_interrupt_class_cik,
 118         .num_of_watch_points = 4,
 119         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 120         .supports_cwsr = false,
 121         .needs_iommu_device = false,
 122         .needs_pci_atomics = true,
 123         .num_sdma_engines = 2,
 124         .num_sdma_queues_per_engine = 2,
 125 };
 126
 127 static const struct kfd_device_info fiji_device_info = {
 128         .asic_family = CHIP_FIJI,
 129         .max_pasid_bits = 16,
 130         .max_no_of_hqd  = 24,
 131         .doorbell_size  = 4,
 132         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 133         .event_interrupt_class = &event_interrupt_class_cik,
 134         .num_of_watch_points = 4,
 135         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 136         .supports_cwsr = true,
 137         .needs_iommu_device = false,
 138         .needs_pci_atomics = true,
 139         .num_sdma_engines = 2,
 140         .num_sdma_queues_per_engine = 2,
 141 };
 142
 143 static const struct kfd_device_info fiji_vf_device_info = {
 144         .asic_family = CHIP_FIJI,
 145         .max_pasid_bits = 16,
 146         .max_no_of_hqd  = 24,
 147         .doorbell_size  = 4,
 148         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 149         .event_interrupt_class = &event_interrupt_class_cik,
 150         .num_of_watch_points = 4,
 151         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 152         .supports_cwsr = true,
 153         .needs_iommu_device = false,
 154         .needs_pci_atomics = false,
 155         .num_sdma_engines = 2,
 156         .num_sdma_queues_per_engine = 2,
 157 };
 158
 159
 160 static const struct kfd_device_info polaris10_device_info = {
 161         .asic_family = CHIP_POLARIS10,
 162         .max_pasid_bits = 16,
 163         .max_no_of_hqd  = 24,
 164         .doorbell_size  = 4,
 165         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 166         .event_interrupt_class = &event_interrupt_class_cik,
 167         .num_of_watch_points = 4,
 168         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 169         .supports_cwsr = true,
 170         .needs_iommu_device = false,
 171         .needs_pci_atomics = true,
 172         .num_sdma_engines = 2,
 173         .num_sdma_queues_per_engine = 2,
 174 };
 175
 176 static const struct kfd_device_info polaris10_vf_device_info = {
 177         .asic_family = CHIP_POLARIS10,
 178         .max_pasid_bits = 16,
 179         .max_no_of_hqd  = 24,
 180         .doorbell_size  = 4,
 181         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 182         .event_interrupt_class = &event_interrupt_class_cik,
 183         .num_of_watch_points = 4,
 184         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 185         .supports_cwsr = true,
 186         .needs_iommu_device = false,
 187         .needs_pci_atomics = false,
 188         .num_sdma_engines = 2,
 189         .num_sdma_queues_per_engine = 2,
 190 };
 191
 192 static const struct kfd_device_info polaris11_device_info = {
 193         .asic_family = CHIP_POLARIS11,
 194         .max_pasid_bits = 16,
 195         .max_no_of_hqd  = 24,
 196         .doorbell_size  = 4,
 197         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 198         .event_interrupt_class = &event_interrupt_class_cik,
 199         .num_of_watch_points = 4,
 200         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 201         .supports_cwsr = true,
 202         .needs_iommu_device = false,
 203         .needs_pci_atomics = true,
 204         .num_sdma_engines = 2,
 205         .num_sdma_queues_per_engine = 2,
 206 };
 207
 208 static const struct kfd_device_info polaris12_device_info = {
 209         .asic_family = CHIP_POLARIS12,
 210         .max_pasid_bits = 16,
 211         .max_no_of_hqd  = 24,
 212         .doorbell_size  = 4,
 213         .ih_ring_entry_size = 4 * sizeof(uint32_t),
 214         .event_interrupt_class = &event_interrupt_class_cik,
 215         .num_of_watch_points = 4,
 216         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 217         .supports_cwsr = true,
 218         .needs_iommu_device = false,
 219         .needs_pci_atomics = true,
 220         .num_sdma_engines = 2,
 221         .num_sdma_queues_per_engine = 2,
 222 };
 223
 224 static const struct kfd_device_info vega10_device_info = {
 225         .asic_family = CHIP_VEGA10,
 226         .max_pasid_bits = 16,
 227         .max_no_of_hqd  = 24,
 228         .doorbell_size  = 8,
 229         .ih_ring_entry_size = 8 * sizeof(uint32_t),
 230         .event_interrupt_class = &event_interrupt_class_v9,
 231         .num_of_watch_points = 4,
 232         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 233         .supports_cwsr = true,
 234         .needs_iommu_device = false,
 235         .needs_pci_atomics = false,
 236         .num_sdma_engines = 2,
 237         .num_sdma_queues_per_engine = 2,
 238 };
 239
 240 static const struct kfd_device_info vega10_vf_device_info = {
 241         .asic_family = CHIP_VEGA10,
 242         .max_pasid_bits = 16,
 243         .max_no_of_hqd  = 24,
 244         .doorbell_size  = 8,
 245         .ih_ring_entry_size = 8 * sizeof(uint32_t),
 246         .event_interrupt_class = &event_interrupt_class_v9,
 247         .num_of_watch_points = 4,
 248         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 249         .supports_cwsr = true,
 250         .needs_iommu_device = false,
 251         .needs_pci_atomics = false,
 252         .num_sdma_engines = 2,
 253         .num_sdma_queues_per_engine = 2,
 254 };
 255
 256 static const struct kfd_device_info vega12_device_info = {
 257         .asic_family = CHIP_VEGA12,
 258         .max_pasid_bits = 16,
 259         .max_no_of_hqd  = 24,
 260         .doorbell_size  = 8,
 261         .ih_ring_entry_size = 8 * sizeof(uint32_t),
 262         .event_interrupt_class = &event_interrupt_class_v9,
 263         .num_of_watch_points = 4,
 264         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 265         .supports_cwsr = true,
 266         .needs_iommu_device = false,
 267         .needs_pci_atomics = false,
 268         .num_sdma_engines = 2,
 269         .num_sdma_queues_per_engine = 2,
 270 };
 271
 272 static const struct kfd_device_info vega20_device_info = {
 273         .asic_family = CHIP_VEGA20,
 274         .max_pasid_bits = 16,
 275         .max_no_of_hqd  = 24,
 276         .doorbell_size  = 8,
 277         .ih_ring_entry_size = 8 * sizeof(uint32_t),
 278         .event_interrupt_class = &event_interrupt_class_v9,
 279         .num_of_watch_points = 4,
 280         .mqd_size_aligned = MQD_SIZE_ALIGNED,
 281         .supports_cwsr = true,
 282         .needs_iommu_device = false,
 283         .needs_pci_atomics = false,
 284         .num_sdma_engines = 2,
 285         .num_sdma_queues_per_engine = 8,
 286 };
 287
 288 struct kfd_deviceid {
 289         unsigned short did;
 290         const struct kfd_device_info *device_info;
 291 };
 292
 293 static const struct kfd_deviceid supported_devices[] = {
 294 #ifdef KFD_SUPPORT_IOMMU_V2
 295         { 0x1304, &kaveri_device_info },        /* Kaveri */
 296         { 0x1305, &kaveri_device_info },        /* Kaveri */
 297         { 0x1306, &kaveri_device_info },        /* Kaveri */
 298         { 0x1307, &kaveri_device_info },        /* Kaveri */
 299         { 0x1309, &kaveri_device_info },        /* Kaveri */
 300         { 0x130A, &kaveri_device_info },        /* Kaveri */
 301         { 0x130B, &kaveri_device_info },        /* Kaveri */
 302         { 0x130C, &kaveri_device_info },        /* Kaveri */
 303         { 0x130D, &kaveri_device_info },        /* Kaveri */
 304         { 0x130E, &kaveri_device_info },        /* Kaveri */
 305         { 0x130F, &kaveri_device_info },        /* Kaveri */
 306         { 0x1310, &kaveri_device_info },        /* Kaveri */
 307         { 0x1311, &kaveri_device_info },        /* Kaveri */
 308         { 0x1312, &kaveri_device_info },        /* Kaveri */
 309         { 0x1313, &kaveri_device_info },        /* Kaveri */
 310         { 0x1315, &kaveri_device_info },        /* Kaveri */
 311         { 0x1316, &kaveri_device_info },        /* Kaveri */
 312         { 0x1317, &kaveri_device_info },        /* Kaveri */
 313         { 0x1318, &kaveri_device_info },        /* Kaveri */
 314         { 0x131B, &kaveri_device_info },        /* Kaveri */
 315         { 0x131C, &kaveri_device_info },        /* Kaveri */
 316         { 0x131D, &kaveri_device_info },        /* Kaveri */
 317         { 0x9870, &carrizo_device_info },       /* Carrizo */
 318         { 0x9874, &carrizo_device_info },       /* Carrizo */
 319         { 0x9875, &carrizo_device_info },       /* Carrizo */
 320         { 0x9876, &carrizo_device_info },       /* Carrizo */
 321         { 0x9877, &carrizo_device_info },       /* Carrizo */
 322         { 0x15DD, &raven_device_info },         /* Raven */
 323         { 0x15D8, &raven_device_info },         /* Raven */
 324 #endif
 325         { 0x67A0, &hawaii_device_info },        /* Hawaii */
 326         { 0x67A1, &hawaii_device_info },        /* Hawaii */
 327         { 0x67A2, &hawaii_device_info },        /* Hawaii */
 328         { 0x67A8, &hawaii_device_info },        /* Hawaii */
 329         { 0x67A9, &hawaii_device_info },        /* Hawaii */
 330         { 0x67AA, &hawaii_device_info },        /* Hawaii */
 331         { 0x67B0, &hawaii_device_info },        /* Hawaii */
 332         { 0x67B1, &hawaii_device_info },        /* Hawaii */
 333         { 0x67B8, &hawaii_device_info },        /* Hawaii */
 334         { 0x67B9, &hawaii_device_info },        /* Hawaii */
 335         { 0x67BA, &hawaii_device_info },        /* Hawaii */
 336         { 0x67BE, &hawaii_device_info },        /* Hawaii */
 337         { 0x6920, &tonga_device_info },         /* Tonga */
 338         { 0x6921, &tonga_device_info },         /* Tonga */
 339         { 0x6928, &tonga_device_info },         /* Tonga */
 340         { 0x6929, &tonga_device_info },         /* Tonga */
 341         { 0x692B, &tonga_device_info },         /* Tonga */
 342         { 0x6938, &tonga_device_info },         /* Tonga */
 343         { 0x6939, &tonga_device_info },         /* Tonga */
 344         { 0x7300, &fiji_device_info },          /* Fiji */
 345         { 0x730F, &fiji_vf_device_info },       /* Fiji vf*/
 346         { 0x67C0, &polaris10_device_info },     /* Polaris10 */
 347         { 0x67C1, &polaris10_device_info },     /* Polaris10 */
 348         { 0x67C2, &polaris10_device_info },     /* Polaris10 */
 349         { 0x67C4, &polaris10_device_info },     /* Polaris10 */
 350         { 0x67C7, &polaris10_device_info },     /* Polaris10 */
 351         { 0x67C8, &polaris10_device_info },     /* Polaris10 */
 352         { 0x67C9, &polaris10_device_info },     /* Polaris10 */
 353         { 0x67CA, &polaris10_device_info },     /* Polaris10 */
 354         { 0x67CC, &polaris10_device_info },     /* Polaris10 */
 355         { 0x67CF, &polaris10_device_info },     /* Polaris10 */
 356         { 0x67D0, &polaris10_vf_device_info },  /* Polaris10 vf*/
 357         { 0x67DF, &polaris10_device_info },     /* Polaris10 */
 358         { 0x67E0, &polaris11_device_info },     /* Polaris11 */
 359         { 0x67E1, &polaris11_device_info },     /* Polaris11 */
 360         { 0x67E3, &polaris11_device_info },     /* Polaris11 */
 361         { 0x67E7, &polaris11_device_info },     /* Polaris11 */
 362         { 0x67E8, &polaris11_device_info },     /* Polaris11 */
 363         { 0x67E9, &polaris11_device_info },     /* Polaris11 */
 364         { 0x67EB, &polaris11_device_info },     /* Polaris11 */
 365         { 0x67EF, &polaris11_device_info },     /* Polaris11 */
 366         { 0x67FF, &polaris11_device_info },     /* Polaris11 */
 367         { 0x6980, &polaris12_device_info },     /* Polaris12 */
 368         { 0x6981, &polaris12_device_info },     /* Polaris12 */
 369         { 0x6985, &polaris12_device_info },     /* Polaris12 */
 370         { 0x6986, &polaris12_device_info },     /* Polaris12 */
 371         { 0x6987, &polaris12_device_info },     /* Polaris12 */
 372         { 0x6995, &polaris12_device_info },     /* Polaris12 */
 373         { 0x6997, &polaris12_device_info },     /* Polaris12 */
 374         { 0x699F, &polaris12_device_info },     /* Polaris12 */
 375         { 0x6860, &vega10_device_info },        /* Vega10 */
 376         { 0x6861, &vega10_device_info },        /* Vega10 */
 377         { 0x6862, &vega10_device_info },        /* Vega10 */
 378         { 0x6863, &vega10_device_info },        /* Vega10 */
 379         { 0x6864, &vega10_device_info },        /* Vega10 */
 380         { 0x6867, &vega10_device_info },        /* Vega10 */
 381         { 0x6868, &vega10_device_info },        /* Vega10 */
 382         { 0x6869, &vega10_device_info },        /* Vega10 */
 383         { 0x686A, &vega10_device_info },        /* Vega10 */
 384         { 0x686B, &vega10_device_info },        /* Vega10 */
 385         { 0x686C, &vega10_vf_device_info },     /* Vega10  vf*/
 386         { 0x686D, &vega10_device_info },        /* Vega10 */
 387         { 0x686E, &vega10_device_info },        /* Vega10 */
 388         { 0x686F, &vega10_device_info },        /* Vega10 */
 389         { 0x687F, &vega10_device_info },        /* Vega10 */
 390         { 0x69A0, &vega12_device_info },        /* Vega12 */
 391         { 0x69A1, &vega12_device_info },        /* Vega12 */
 392         { 0x69A2, &vega12_device_info },        /* Vega12 */
 393         { 0x69A3, &vega12_device_info },        /* Vega12 */
 394         { 0x69AF, &vega12_device_info },        /* Vega12 */
 395         { 0x66a0, &vega20_device_info },        /* Vega20 */
 396         { 0x66a1, &vega20_device_info },        /* Vega20 */
 397         { 0x66a2, &vega20_device_info },        /* Vega20 */
 398         { 0x66a3, &vega20_device_info },        /* Vega20 */
 399         { 0x66a4, &vega20_device_info },        /* Vega20 */
 400         { 0x66a7, &vega20_device_info },        /* Vega20 */
 401         { 0x66af, &vega20_device_info }         /* Vega20 */
 402 };
 403
 404 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
 405                                 unsigned int chunk_size);
 406 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
 407
 408 static int kfd_resume(struct kfd_dev *kfd);
 409
 410 static const struct kfd_device_info *lookup_device_info(unsigned short did)
 411 {
 412         size_t i;
 413
 414         for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
 415                 if (supported_devices[i].did == did) {
 416                         WARN_ON(!supported_devices[i].device_info);
 417                         return supported_devices[i].device_info;
 418                 }
 419         }
 420
 421         dev_warn(kfd_device, "DID %04x is missing in supported_devices\n",
 422                  did);
 423
 424         return NULL;
 425 }
 426
 427 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
 428         struct pci_dev *pdev, const struct kfd2kgd_calls *f2g)
 429 {
 430         struct kfd_dev *kfd;
 431         int ret;
 432         const struct kfd_device_info *device_info =
 433                                         lookup_device_info(pdev->device);
 434
 435         if (!device_info) {
 436                 dev_err(kfd_device, "kgd2kfd_probe failed\n");
 437                 return NULL;
 438         }
 439
 440         kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
 441         if (!kfd)
 442                 return NULL;
 443
 444         /* Allow BIF to recode atomics to PCIe 3.0 AtomicOps.
 445          * 32 and 64-bit requests are possible and must be
 446          * supported.
 447          */
 448         ret = pci_enable_atomic_ops_to_root(pdev,
 449                         PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
 450                         PCI_EXP_DEVCAP2_ATOMIC_COMP64);
 451         if (device_info->needs_pci_atomics && ret < 0) {
 452                 dev_info(kfd_device,
 453                          "skipped device %x:%x, PCI rejects atomics\n",
 454                          pdev->vendor, pdev->device);
 455                 kfree(kfd);
 456                 return NULL;
 457         } else if (!ret)
 458                 kfd->pci_atomic_requested = true;
 459
 460         kfd->kgd = kgd;
 461         kfd->device_info = device_info;
 462         kfd->pdev = pdev;
 463         kfd->init_complete = false;
 464         kfd->kfd2kgd = f2g;
 465
 466         mutex_init(&kfd->doorbell_mutex);
 467         memset(&kfd->doorbell_available_index, 0,
 468                 sizeof(kfd->doorbell_available_index));
 469
 470         return kfd;
 471 }
 472
 473 static void kfd_cwsr_init(struct kfd_dev *kfd)
 474 {
 475         if (cwsr_enable && kfd->device_info->supports_cwsr) {
 476                 if (kfd->device_info->asic_family < CHIP_VEGA10) {
 477                         BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE);
 478                         kfd->cwsr_isa = cwsr_trap_gfx8_hex;
 479                         kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
 480                 } else {
 481                         BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex) > PAGE_SIZE);
 482                         kfd->cwsr_isa = cwsr_trap_gfx9_hex;
 483                         kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_hex);
 484                 }
 485
 486                 kfd->cwsr_enabled = true;
 487         }
 488 }
 489
 490 bool kgd2kfd_device_init(struct kfd_dev *kfd,
 491                          const struct kgd2kfd_shared_resources *gpu_resources)
 492 {
 493         unsigned int size;
 494
 495         kfd->mec_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
 496                         KGD_ENGINE_MEC1);
 497         kfd->sdma_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
 498                         KGD_ENGINE_SDMA1);
 499         kfd->shared_resources = *gpu_resources;
 500
 501         kfd->vm_info.first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1;
 502         kfd->vm_info.last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1;
 503         kfd->vm_info.vmid_num_kfd = kfd->vm_info.last_vmid_kfd
 504                         - kfd->vm_info.first_vmid_kfd + 1;
 505
 506         /* Verify module parameters regarding mapped process number*/
 507         if ((hws_max_conc_proc < 0)
 508                         || (hws_max_conc_proc > kfd->vm_info.vmid_num_kfd)) {
 509                 dev_err(kfd_device,
 510                         "hws_max_conc_proc %d must be between 0 and %d, use %d instead\n",
 511                         hws_max_conc_proc, kfd->vm_info.vmid_num_kfd,
 512                         kfd->vm_info.vmid_num_kfd);
 513                 kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd;
 514         } else
 515                 kfd->max_proc_per_quantum = hws_max_conc_proc;
 516
 517         /* calculate max size of mqds needed for queues */
 518         size = max_num_of_queues_per_device *
 519                         kfd->device_info->mqd_size_aligned;
 520
 521         /*
 522          * calculate max size of runlist packet.
 523          * There can be only 2 packets at once
 524          */
 525         size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_mes_map_process) +
 526                 max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues)
 527                 + sizeof(struct pm4_mes_runlist)) * 2;
 528
 529         /* Add size of HIQ & DIQ */
 530         size += KFD_KERNEL_QUEUE_SIZE * 2;
 531
 532         /* add another 512KB for all other allocations on gart (HPD, fences) */
 533         size += 512 * 1024;
 534
 535         if (amdgpu_amdkfd_alloc_gtt_mem(
 536                         kfd->kgd, size, &kfd->gtt_mem,
 537                         &kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr,
 538                         false)) {
 539                 dev_err(kfd_device, "Could not allocate %d bytes\n", size);
 540                 goto out;
 541         }
 542
 543         dev_info(kfd_device, "Allocated %d bytes on gart\n", size);
 544
 545         /* Initialize GTT sa with 512 byte chunk size */
 546         if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
 547                 dev_err(kfd_device, "Error initializing gtt sub-allocator\n");
 548                 goto kfd_gtt_sa_init_error;
 549         }
 550
 551         if (kfd_doorbell_init(kfd)) {
 552                 dev_err(kfd_device,
 553                         "Error initializing doorbell aperture\n");
 554                 goto kfd_doorbell_error;
 555         }
 556
 557         if (kfd->kfd2kgd->get_hive_id)
 558                 kfd->hive_id = kfd->kfd2kgd->get_hive_id(kfd->kgd);
 559
 560         if (kfd_topology_add_device(kfd)) {
 561                 dev_err(kfd_device, "Error adding device to topology\n");
 562                 goto kfd_topology_add_device_error;
 563         }
 564
 565         if (kfd_interrupt_init(kfd)) {
 566                 dev_err(kfd_device, "Error initializing interrupts\n");
 567                 goto kfd_interrupt_error;
 568         }
 569
 570         kfd->dqm = device_queue_manager_init(kfd);
 571         if (!kfd->dqm) {
 572                 dev_err(kfd_device, "Error initializing queue manager\n");
 573                 goto device_queue_manager_error;
 574         }
 575
 576         if (kfd_iommu_device_init(kfd)) {
 577                 dev_err(kfd_device, "Error initializing iommuv2\n");
 578                 goto device_iommu_error;
 579         }
 580
 581         kfd_cwsr_init(kfd);
 582
 583         if (kfd_resume(kfd))
 584                 goto kfd_resume_error;
 585
 586         kfd->dbgmgr = NULL;
 587
 588         kfd->init_complete = true;
 589         dev_info(kfd_device, "added device %x:%x\n", kfd->pdev->vendor,
 590                  kfd->pdev->device);
 591
 592         pr_debug("Starting kfd with the following scheduling policy %d\n",
 593                 kfd->dqm->sched_policy);
 594
 595         goto out;
 596
 597 kfd_resume_error:
 598 device_iommu_error:
 599         device_queue_manager_uninit(kfd->dqm);
 600 device_queue_manager_error:
 601         kfd_interrupt_exit(kfd);
 602 kfd_interrupt_error:
 603         kfd_topology_remove_device(kfd);
 604 kfd_topology_add_device_error:
 605         kfd_doorbell_fini(kfd);
 606 kfd_doorbell_error:
 607         kfd_gtt_sa_fini(kfd);
 608 kfd_gtt_sa_init_error:
 609         amdgpu_amdkfd_free_gtt_mem(kfd->kgd, kfd->gtt_mem);
 610         dev_err(kfd_device,
 611                 "device %x:%x NOT added due to errors\n",
 612                 kfd->pdev->vendor, kfd->pdev->device);
 613 out:
 614         return kfd->init_complete;
 615 }
 616
 617 void kgd2kfd_device_exit(struct kfd_dev *kfd)
 618 {
 619         if (kfd->init_complete) {
 620                 kgd2kfd_suspend(kfd);
 621                 device_queue_manager_uninit(kfd->dqm);
 622                 kfd_interrupt_exit(kfd);
 623                 kfd_topology_remove_device(kfd);
 624                 kfd_doorbell_fini(kfd);
 625                 kfd_gtt_sa_fini(kfd);
 626                 amdgpu_amdkfd_free_gtt_mem(kfd->kgd, kfd->gtt_mem);
 627         }
 628
 629         kfree(kfd);
 630 }
 631
 632 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
 633 {
 634         if (!kfd->init_complete)
 635                 return 0;
 636         kgd2kfd_suspend(kfd);
 637
 638         /* hold dqm->lock to prevent further execution*/
 639         dqm_lock(kfd->dqm);
 640
 641         kfd_signal_reset_event(kfd);
 642         return 0;
 643 }
 644
 645 /*
 646  * Fix me. KFD won't be able to resume existing process for now.
 647  * We will keep all existing process in a evicted state and
 648  * wait the process to be terminated.
 649  */
 650
 651 int kgd2kfd_post_reset(struct kfd_dev *kfd)
 652 {
 653         int ret, count;
 654
 655         if (!kfd->init_complete)
 656                 return 0;
 657
 658         dqm_unlock(kfd->dqm);
 659
 660         ret = kfd_resume(kfd);
 661         if (ret)
 662                 return ret;
 663         count = atomic_dec_return(&kfd_locked);
 664         WARN_ONCE(count != 0, "KFD reset ref. error");
 665         return 0;
 666 }
 667
 668 bool kfd_is_locked(void)
 669 {
 670         return  (atomic_read(&kfd_locked) > 0);
 671 }
 672
 673 void kgd2kfd_suspend(struct kfd_dev *kfd)
 674 {
 675         if (!kfd->init_complete)
 676                 return;
 677
 678         /* For first KFD device suspend all the KFD processes */
 679         if (atomic_inc_return(&kfd_locked) == 1)
 680                 kfd_suspend_all_processes();
 681
 682         kfd->dqm->ops.stop(kfd->dqm);
 683
 684         kfd_iommu_suspend(kfd);
 685 }
 686
 687 int kgd2kfd_resume(struct kfd_dev *kfd)
 688 {
 689         int ret, count;
 690
 691         if (!kfd->init_complete)
 692                 return 0;
 693
 694         ret = kfd_resume(kfd);
 695         if (ret)
 696                 return ret;
 697
 698         count = atomic_dec_return(&kfd_locked);
 699         WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
 700         if (count == 0)
 701                 ret = kfd_resume_all_processes();
 702
 703         return ret;
 704 }
 705
 706 static int kfd_resume(struct kfd_dev *kfd)
 707 {
 708         int err = 0;
 709
 710         err = kfd_iommu_resume(kfd);
 711         if (err) {
 712                 dev_err(kfd_device,
 713                         "Failed to resume IOMMU for device %x:%x\n",
 714                         kfd->pdev->vendor, kfd->pdev->device);
 715                 return err;
 716         }
 717
 718         err = kfd->dqm->ops.start(kfd->dqm);
 719         if (err) {
 720                 dev_err(kfd_device,
 721                         "Error starting queue manager for device %x:%x\n",
 722                         kfd->pdev->vendor, kfd->pdev->device);
 723                 goto dqm_start_error;
 724         }
 725
 726         return err;
 727
 728 dqm_start_error:
 729         kfd_iommu_suspend(kfd);
 730         return err;
 731 }
 732
 733 /* This is called directly from KGD at ISR. */
 734 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 735 {
 736         uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE];
 737         bool is_patched = false;
 738         unsigned long flags;
 739
 740         if (!kfd->init_complete)
 741                 return;
 742
 743         if (kfd->device_info->ih_ring_entry_size > sizeof(patched_ihre)) {
 744                 dev_err_once(kfd_device, "Ring entry too small\n");
 745                 return;
 746         }
 747
 748         spin_lock_irqsave(&kfd->interrupt_lock, flags);
 749
 750         if (kfd->interrupts_active
 751             && interrupt_is_wanted(kfd, ih_ring_entry,
 752                                    patched_ihre, &is_patched)
 753             && enqueue_ih_ring_entry(kfd,
 754                                      is_patched ? patched_ihre : ih_ring_entry))
 755                 queue_work(kfd->ih_wq, &kfd->interrupt_work);
 756
 757         spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
 758 }
 759
 760 int kgd2kfd_quiesce_mm(struct mm_struct *mm)
 761 {
 762         struct kfd_process *p;
 763         int r;
 764
 765         /* Because we are called from arbitrary context (workqueue) as opposed
 766          * to process context, kfd_process could attempt to exit while we are
 767          * running so the lookup function increments the process ref count.
 768          */
 769         p = kfd_lookup_process_by_mm(mm);
 770         if (!p)
 771                 return -ESRCH;
 772
 773         r = kfd_process_evict_queues(p);
 774
 775         kfd_unref_process(p);
 776         return r;
 777 }
 778
 779 int kgd2kfd_resume_mm(struct mm_struct *mm)
 780 {
 781         struct kfd_process *p;
 782         int r;
 783
 784         /* Because we are called from arbitrary context (workqueue) as opposed
 785          * to process context, kfd_process could attempt to exit while we are
 786          * running so the lookup function increments the process ref count.
 787          */
 788         p = kfd_lookup_process_by_mm(mm);
 789         if (!p)
 790                 return -ESRCH;
 791
 792         r = kfd_process_restore_queues(p);
 793
 794         kfd_unref_process(p);
 795         return r;
 796 }
 797
 798 /** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will
 799  *   prepare for safe eviction of KFD BOs that belong to the specified
 800  *   process.
 801  *
 802  * @mm: mm_struct that identifies the specified KFD process
 803  * @fence: eviction fence attached to KFD process BOs
 804  *
 805  */
 806 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
 807                                                struct dma_fence *fence)
 808 {
 809         struct kfd_process *p;
 810         unsigned long active_time;
 811         unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
 812
 813         if (!fence)
 814                 return -EINVAL;
 815
 816         if (dma_fence_is_signaled(fence))
 817                 return 0;
 818
 819         p = kfd_lookup_process_by_mm(mm);
 820         if (!p)
 821                 return -ENODEV;
 822
 823         if (fence->seqno == p->last_eviction_seqno)
 824                 goto out;
 825
 826         p->last_eviction_seqno = fence->seqno;
 827
 828         /* Avoid KFD process starvation. Wait for at least
 829          * PROCESS_ACTIVE_TIME_MS before evicting the process again
 830          */
 831         active_time = get_jiffies_64() - p->last_restore_timestamp;
 832         if (delay_jiffies > active_time)
 833                 delay_jiffies -= active_time;
 834         else
 835                 delay_jiffies = 0;
 836
 837         /* During process initialization eviction_work.dwork is initialized
 838          * to kfd_evict_bo_worker
 839          */
 840         schedule_delayed_work(&p->eviction_work, delay_jiffies);
 841 out:
 842         kfd_unref_process(p);
 843         return 0;
 844 }
 845
 846 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
 847                                 unsigned int chunk_size)
 848 {
 849         unsigned int num_of_longs;
 850
 851         if (WARN_ON(buf_size < chunk_size))
 852                 return -EINVAL;
 853         if (WARN_ON(buf_size == 0))
 854                 return -EINVAL;
 855         if (WARN_ON(chunk_size == 0))
 856                 return -EINVAL;
 857
 858         kfd->gtt_sa_chunk_size = chunk_size;
 859         kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
 860
 861         num_of_longs = (kfd->gtt_sa_num_of_chunks + BITS_PER_LONG - 1) /
 862                 BITS_PER_LONG;
 863
 864         kfd->gtt_sa_bitmap = kcalloc(num_of_longs, sizeof(long), GFP_KERNEL);
 865
 866         if (!kfd->gtt_sa_bitmap)
 867                 return -ENOMEM;
 868
 869         pr_debug("gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
 870                         kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
 871
 872         mutex_init(&kfd->gtt_sa_lock);
 873
 874         return 0;
 875
 876 }
 877
 878 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
 879 {
 880         mutex_destroy(&kfd->gtt_sa_lock);
 881         kfree(kfd->gtt_sa_bitmap);
 882 }
 883
 884 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
 885                                                 unsigned int bit_num,
 886                                                 unsigned int chunk_size)
 887 {
 888         return start_addr + bit_num * chunk_size;
 889 }
 890
 891 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
 892                                                 unsigned int bit_num,
 893                                                 unsigned int chunk_size)
 894 {
 895         return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
 896 }
 897
 898 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
 899                         struct kfd_mem_obj **mem_obj)
 900 {
 901         unsigned int found, start_search, cur_size;
 902
 903         if (size == 0)
 904                 return -EINVAL;
 905
 906         if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
 907                 return -ENOMEM;
 908
 909         *mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
 910         if (!(*mem_obj))
 911                 return -ENOMEM;
 912
 913         pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size);
 914
 915         start_search = 0;
 916
 917         mutex_lock(&kfd->gtt_sa_lock);
 918
 919 kfd_gtt_restart_search:
 920         /* Find the first chunk that is free */
 921         found = find_next_zero_bit(kfd->gtt_sa_bitmap,
 922                                         kfd->gtt_sa_num_of_chunks,
 923                                         start_search);
 924
 925         pr_debug("Found = %d\n", found);
 926
 927         /* If there wasn't any free chunk, bail out */
 928         if (found == kfd->gtt_sa_num_of_chunks)
 929                 goto kfd_gtt_no_free_chunk;
 930
 931         /* Update fields of mem_obj */
 932         (*mem_obj)->range_start = found;
 933         (*mem_obj)->range_end = found;
 934         (*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
 935                                         kfd->gtt_start_gpu_addr,
 936                                         found,
 937                                         kfd->gtt_sa_chunk_size);
 938         (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
 939                                         kfd->gtt_start_cpu_ptr,
 940                                         found,
 941                                         kfd->gtt_sa_chunk_size);
 942
 943         pr_debug("gpu_addr = %p, cpu_addr = %p\n",
 944                         (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
 945
 946         /* If we need only one chunk, mark it as allocated and get out */
 947         if (size <= kfd->gtt_sa_chunk_size) {
 948                 pr_debug("Single bit\n");
 949                 set_bit(found, kfd->gtt_sa_bitmap);
 950                 goto kfd_gtt_out;
 951         }
 952
 953         /* Otherwise, try to see if we have enough contiguous chunks */
 954         cur_size = size - kfd->gtt_sa_chunk_size;
 955         do {
 956                 (*mem_obj)->range_end =
 957                         find_next_zero_bit(kfd->gtt_sa_bitmap,
 958                                         kfd->gtt_sa_num_of_chunks, ++found);
 959                 /*
 960                  * If next free chunk is not contiguous than we need to
 961                  * restart our search from the last free chunk we found (which
 962                  * wasn't contiguous to the previous ones
 963                  */
 964                 if ((*mem_obj)->range_end != found) {
 965                         start_search = found;
 966                         goto kfd_gtt_restart_search;
 967                 }
 968
 969                 /*
 970                  * If we reached end of buffer, bail out with error
 971                  */
 972                 if (found == kfd->gtt_sa_num_of_chunks)
 973                         goto kfd_gtt_no_free_chunk;
 974
 975                 /* Check if we don't need another chunk */
 976                 if (cur_size <= kfd->gtt_sa_chunk_size)
 977                         cur_size = 0;
 978                 else
 979                         cur_size -= kfd->gtt_sa_chunk_size;
 980
 981         } while (cur_size > 0);
 982
 983         pr_debug("range_start = %d, range_end = %d\n",
 984                 (*mem_obj)->range_start, (*mem_obj)->range_end);
 985
 986         /* Mark the chunks as allocated */
 987         for (found = (*mem_obj)->range_start;
 988                 found <= (*mem_obj)->range_end;
 989                 found++)
 990                 set_bit(found, kfd->gtt_sa_bitmap);
 991
 992 kfd_gtt_out:
 993         mutex_unlock(&kfd->gtt_sa_lock);
 994         return 0;
 995
 996 kfd_gtt_no_free_chunk:
 997         pr_debug("Allocation failed with mem_obj = %p\n", mem_obj);
 998         mutex_unlock(&kfd->gtt_sa_lock);
 999         kfree(mem_obj);
1000         return -ENOMEM;
1001 }
1002
1003 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
1004 {
1005         unsigned int bit;
1006
1007         /* Act like kfree when trying to free a NULL object */
1008         if (!mem_obj)
1009                 return 0;
1010
1011         pr_debug("Free mem_obj = %p, range_start = %d, range_end = %d\n",
1012                         mem_obj, mem_obj->range_start, mem_obj->range_end);
1013
1014         mutex_lock(&kfd->gtt_sa_lock);
1015
1016         /* Mark the chunks as free */
1017         for (bit = mem_obj->range_start;
1018                 bit <= mem_obj->range_end;
1019                 bit++)
1020                 clear_bit(bit, kfd->gtt_sa_bitmap);
1021
1022         mutex_unlock(&kfd->gtt_sa_lock);
1023
1024         kfree(mem_obj);
1025         return 0;
1026 }
1027
1028 #if defined(CONFIG_DEBUG_FS)
1029
1030 /* This function will send a package to HIQ to hang the HWS
1031  * which will trigger a GPU reset and bring the HWS back to normal state
1032  */
1033 int kfd_debugfs_hang_hws(struct kfd_dev *dev)
1034 {
1035         int r = 0;
1036
1037         if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
1038                 pr_err("HWS is not enabled");
1039                 return -EINVAL;
1040         }
1041
1042         r = pm_debugfs_hang_hws(&dev->dqm->packets);
1043         if (!r)
1044                 r = dqm_debugfs_execute_queues(dev->dqm);
1045
1046         return r;
1047 }
1048
1049 #endif