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/amd-iommu.h>
  24 #include <linux/bsearch.h>
  25 #include <linux/pci.h>
  26 #include <linux/slab.h>
  27 #include "kfd_priv.h"
  28 #include "kfd_device_queue_manager.h"
  29
  30 #define MQD_SIZE_ALIGNED 768
  31
  32 static const struct kfd_device_info kaveri_device_info = {
  33         .max_pasid_bits = 16,
  34         .ih_ring_entry_size = 4 * sizeof(uint32_t),
  35         .mqd_size_aligned = MQD_SIZE_ALIGNED
  36 };
  37
  38 struct kfd_deviceid {
  39         unsigned short did;
  40         const struct kfd_device_info *device_info;
  41 };
  42
  43 /* Please keep this sorted by increasing device id. */
  44 static const struct kfd_deviceid supported_devices[] = {
  45         { 0x1304, &kaveri_device_info },        /* Kaveri */
  46         { 0x1305, &kaveri_device_info },        /* Kaveri */
  47         { 0x1306, &kaveri_device_info },        /* Kaveri */
  48         { 0x1307, &kaveri_device_info },        /* Kaveri */
  49         { 0x1309, &kaveri_device_info },        /* Kaveri */
  50         { 0x130A, &kaveri_device_info },        /* Kaveri */
  51         { 0x130B, &kaveri_device_info },        /* Kaveri */
  52         { 0x130C, &kaveri_device_info },        /* Kaveri */
  53         { 0x130D, &kaveri_device_info },        /* Kaveri */
  54         { 0x130E, &kaveri_device_info },        /* Kaveri */
  55         { 0x130F, &kaveri_device_info },        /* Kaveri */
  56         { 0x1310, &kaveri_device_info },        /* Kaveri */
  57         { 0x1311, &kaveri_device_info },        /* Kaveri */
  58         { 0x1312, &kaveri_device_info },        /* Kaveri */
  59         { 0x1313, &kaveri_device_info },        /* Kaveri */
  60         { 0x1315, &kaveri_device_info },        /* Kaveri */
  61         { 0x1316, &kaveri_device_info },        /* Kaveri */
  62         { 0x1317, &kaveri_device_info },        /* Kaveri */
  63         { 0x1318, &kaveri_device_info },        /* Kaveri */
  64         { 0x131B, &kaveri_device_info },        /* Kaveri */
  65         { 0x131C, &kaveri_device_info },        /* Kaveri */
  66         { 0x131D, &kaveri_device_info },        /* Kaveri */
  67 };
  68
  69 static const struct kfd_device_info *lookup_device_info(unsigned short did)
  70 {
  71         size_t i;
  72
  73         for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
  74                 if (supported_devices[i].did == did) {
  75                         BUG_ON(supported_devices[i].device_info == NULL);
  76                         return supported_devices[i].device_info;
  77                 }
  78         }
  79
  80         return NULL;
  81 }
  82
  83 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev)
  84 {
  85         struct kfd_dev *kfd;
  86
  87         const struct kfd_device_info *device_info =
  88                                         lookup_device_info(pdev->device);
  89
  90         if (!device_info)
  91                 return NULL;
  92
  93         kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
  94         if (!kfd)
  95                 return NULL;
  96
  97         kfd->kgd = kgd;
  98         kfd->device_info = device_info;
  99         kfd->pdev = pdev;
 100         kfd->init_complete = false;
 101
 102         return kfd;
 103 }
 104
 105 static bool device_iommu_pasid_init(struct kfd_dev *kfd)
 106 {
 107         const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP |
 108                                         AMD_IOMMU_DEVICE_FLAG_PRI_SUP |
 109                                         AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
 110
 111         struct amd_iommu_device_info iommu_info;
 112         unsigned int pasid_limit;
 113         int err;
 114
 115         err = amd_iommu_device_info(kfd->pdev, &iommu_info);
 116         if (err < 0) {
 117                 dev_err(kfd_device,
 118                         "error getting iommu info. is the iommu enabled?\n");
 119                 return false;
 120         }
 121
 122         if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) {
 123                 dev_err(kfd_device, "error required iommu flags ats(%i), pri(%i), pasid(%i)\n",
 124                        (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0,
 125                        (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0,
 126                        (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) != 0);
 127                 return false;
 128         }
 129
 130         pasid_limit = min_t(unsigned int,
 131                         (unsigned int)1 << kfd->device_info->max_pasid_bits,
 132                         iommu_info.max_pasids);
 133         /*
 134          * last pasid is used for kernel queues doorbells
 135          * in the future the last pasid might be used for a kernel thread.
 136          */
 137         pasid_limit = min_t(unsigned int,
 138                                 pasid_limit,
 139                                 kfd->doorbell_process_limit - 1);
 140
 141         err = amd_iommu_init_device(kfd->pdev, pasid_limit);
 142         if (err < 0) {
 143                 dev_err(kfd_device, "error initializing iommu device\n");
 144                 return false;
 145         }
 146
 147         if (!kfd_set_pasid_limit(pasid_limit)) {
 148                 dev_err(kfd_device, "error setting pasid limit\n");
 149                 amd_iommu_free_device(kfd->pdev);
 150                 return false;
 151         }
 152
 153         return true;
 154 }
 155
 156 static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
 157 {
 158         struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);
 159
 160         if (dev)
 161                 kfd_unbind_process_from_device(dev, pasid);
 162 }
 163
 164 bool kgd2kfd_device_init(struct kfd_dev *kfd,
 165                          const struct kgd2kfd_shared_resources *gpu_resources)
 166 {
 167         unsigned int size;
 168
 169         kfd->shared_resources = *gpu_resources;
 170
 171         /* calculate max size of mqds needed for queues */
 172         size = max_num_of_processes *
 173                 max_num_of_queues_per_process *
 174                 kfd->device_info->mqd_size_aligned;
 175
 176         /* add another 512KB for all other allocations on gart */
 177         size += 512 * 1024;
 178
 179         if (kfd2kgd->init_sa_manager(kfd->kgd, size)) {
 180                 dev_err(kfd_device,
 181                         "Error initializing sa manager for device (%x:%x)\n",
 182                         kfd->pdev->vendor, kfd->pdev->device);
 183                 goto out;
 184         }
 185
 186         kfd_doorbell_init(kfd);
 187
 188         if (kfd_topology_add_device(kfd) != 0) {
 189                 dev_err(kfd_device,
 190                         "Error adding device (%x:%x) to topology\n",
 191                         kfd->pdev->vendor, kfd->pdev->device);
 192                 goto kfd_topology_add_device_error;
 193         }
 194
 195         if (kfd_interrupt_init(kfd)) {
 196                 dev_err(kfd_device,
 197                         "Error initializing interrupts for device (%x:%x)\n",
 198                         kfd->pdev->vendor, kfd->pdev->device);
 199                 goto kfd_interrupt_error;
 200         }
 201
 202         if (!device_iommu_pasid_init(kfd)) {
 203                 dev_err(kfd_device,
 204                         "Error initializing iommuv2 for device (%x:%x)\n",
 205                         kfd->pdev->vendor, kfd->pdev->device);
 206                 goto device_iommu_pasid_error;
 207         }
 208         amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
 209                                                 iommu_pasid_shutdown_callback);
 210
 211         kfd->dqm = device_queue_manager_init(kfd);
 212         if (!kfd->dqm) {
 213                 dev_err(kfd_device,
 214                         "Error initializing queue manager for device (%x:%x)\n",
 215                         kfd->pdev->vendor, kfd->pdev->device);
 216                 goto device_queue_manager_error;
 217         }
 218
 219         if (kfd->dqm->start(kfd->dqm) != 0) {
 220                 dev_err(kfd_device,
 221                         "Error starting queuen manager for device (%x:%x)\n",
 222                         kfd->pdev->vendor, kfd->pdev->device);
 223                 goto dqm_start_error;
 224         }
 225
 226         kfd->init_complete = true;
 227         dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,
 228                  kfd->pdev->device);
 229
 230         pr_debug("kfd: Starting kfd with the following scheduling policy %d\n",
 231                 sched_policy);
 232
 233         goto out;
 234
 235 dqm_start_error:
 236         device_queue_manager_uninit(kfd->dqm);
 237 device_queue_manager_error:
 238         amd_iommu_free_device(kfd->pdev);
 239 device_iommu_pasid_error:
 240         kfd_interrupt_exit(kfd);
 241 kfd_interrupt_error:
 242         kfd_topology_remove_device(kfd);
 243 kfd_topology_add_device_error:
 244         kfd2kgd->fini_sa_manager(kfd->kgd);
 245         dev_err(kfd_device,
 246                 "device (%x:%x) NOT added due to errors\n",
 247                 kfd->pdev->vendor, kfd->pdev->device);
 248 out:
 249         return kfd->init_complete;
 250 }
 251
 252 void kgd2kfd_device_exit(struct kfd_dev *kfd)
 253 {
 254         if (kfd->init_complete) {
 255                 device_queue_manager_uninit(kfd->dqm);
 256                 amd_iommu_free_device(kfd->pdev);
 257                 kfd_interrupt_exit(kfd);
 258                 kfd_topology_remove_device(kfd);
 259         }
 260
 261         kfree(kfd);
 262 }
 263
 264 void kgd2kfd_suspend(struct kfd_dev *kfd)
 265 {
 266         BUG_ON(kfd == NULL);
 267
 268         if (kfd->init_complete) {
 269                 kfd->dqm->stop(kfd->dqm);
 270                 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
 271                 amd_iommu_free_device(kfd->pdev);
 272         }
 273 }
 274
 275 int kgd2kfd_resume(struct kfd_dev *kfd)
 276 {
 277         unsigned int pasid_limit;
 278         int err;
 279
 280         BUG_ON(kfd == NULL);
 281
 282         pasid_limit = kfd_get_pasid_limit();
 283
 284         if (kfd->init_complete) {
 285                 err = amd_iommu_init_device(kfd->pdev, pasid_limit);
 286                 if (err < 0)
 287                         return -ENXIO;
 288                 amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
 289                                                 iommu_pasid_shutdown_callback);
 290                 kfd->dqm->start(kfd->dqm);
 291         }
 292
 293         return 0;
 294 }
 295
 296 /* This is called directly from KGD at ISR. */
 297 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 298 {
 299         if (kfd->init_complete) {
 300                 spin_lock(&kfd->interrupt_lock);
 301
 302                 if (kfd->interrupts_active
 303                     && enqueue_ih_ring_entry(kfd, ih_ring_entry))
 304                         schedule_work(&kfd->interrupt_work);
 305
 306                 spin_unlock(&kfd->interrupt_lock);
 307         }
 308 }