| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * VIRTIO based driver for vDPA device |
| 4 | * |
| 5 | * Copyright (c) 2020, Red Hat. All rights reserved. |
| 6 | * Author: Jason Wang <jasowang@redhat.com> |
| 7 | * |
| 8 | */ |
| 9 | |
| 10 | #include <linux/init.h> |
| 11 | #include <linux/module.h> |
| 12 | #include <linux/device.h> |
| 13 | #include <linux/kernel.h> |
| 14 | #include <linux/slab.h> |
| 15 | #include <linux/uuid.h> |
| 16 | #include <linux/group_cpus.h> |
| 17 | #include <linux/virtio.h> |
| 18 | #include <linux/vdpa.h> |
| 19 | #include <linux/virtio_config.h> |
| 20 | #include <linux/virtio_ring.h> |
| 21 | |
| 22 | #define MOD_VERSION "0.1" |
| 23 | #define MOD_AUTHOR "Jason Wang <jasowang@redhat.com>" |
| 24 | #define MOD_DESC "vDPA bus driver for virtio devices" |
| 25 | #define MOD_LICENSE "GPL v2" |
| 26 | |
| 27 | struct virtio_vdpa_device { |
| 28 | struct virtio_device vdev; |
| 29 | struct vdpa_device *vdpa; |
| 30 | u64 features; |
| 31 | }; |
| 32 | |
| 33 | static inline struct virtio_vdpa_device * |
| 34 | to_virtio_vdpa_device(struct virtio_device *dev) |
| 35 | { |
| 36 | return container_of(dev, struct virtio_vdpa_device, vdev); |
| 37 | } |
| 38 | |
| 39 | static struct vdpa_device *vd_get_vdpa(struct virtio_device *vdev) |
| 40 | { |
| 41 | return to_virtio_vdpa_device(dev: vdev)->vdpa; |
| 42 | } |
| 43 | |
| 44 | static void virtio_vdpa_get(struct virtio_device *vdev, unsigned int offset, |
| 45 | void *buf, unsigned int len) |
| 46 | { |
| 47 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 48 | |
| 49 | vdpa_get_config(vdev: vdpa, offset, buf, len); |
| 50 | } |
| 51 | |
| 52 | static void virtio_vdpa_set(struct virtio_device *vdev, unsigned int offset, |
| 53 | const void *buf, unsigned int len) |
| 54 | { |
| 55 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 56 | |
| 57 | vdpa_set_config(dev: vdpa, offset, buf, length: len); |
| 58 | } |
| 59 | |
| 60 | static u32 virtio_vdpa_generation(struct virtio_device *vdev) |
| 61 | { |
| 62 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 63 | const struct vdpa_config_ops *ops = vdpa->config; |
| 64 | |
| 65 | if (ops->get_generation) |
| 66 | return ops->get_generation(vdpa); |
| 67 | |
| 68 | return 0; |
| 69 | } |
| 70 | |
| 71 | static u8 virtio_vdpa_get_status(struct virtio_device *vdev) |
| 72 | { |
| 73 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 74 | const struct vdpa_config_ops *ops = vdpa->config; |
| 75 | |
| 76 | return ops->get_status(vdpa); |
| 77 | } |
| 78 | |
| 79 | static void virtio_vdpa_set_status(struct virtio_device *vdev, u8 status) |
| 80 | { |
| 81 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 82 | |
| 83 | vdpa_set_status(vdev: vdpa, status); |
| 84 | } |
| 85 | |
| 86 | static void virtio_vdpa_reset(struct virtio_device *vdev) |
| 87 | { |
| 88 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 89 | |
| 90 | vdpa_reset(vdev: vdpa, flags: 0); |
| 91 | } |
| 92 | |
| 93 | static bool virtio_vdpa_notify(struct virtqueue *vq) |
| 94 | { |
| 95 | struct vdpa_device *vdpa = vd_get_vdpa(vdev: vq->vdev); |
| 96 | const struct vdpa_config_ops *ops = vdpa->config; |
| 97 | |
| 98 | ops->kick_vq(vdpa, vq->index); |
| 99 | |
| 100 | return true; |
| 101 | } |
| 102 | |
| 103 | static bool virtio_vdpa_notify_with_data(struct virtqueue *vq) |
| 104 | { |
| 105 | struct vdpa_device *vdpa = vd_get_vdpa(vdev: vq->vdev); |
| 106 | const struct vdpa_config_ops *ops = vdpa->config; |
| 107 | u32 data = vring_notification_data(vq: vq); |
| 108 | |
| 109 | ops->kick_vq_with_data(vdpa, data); |
| 110 | |
| 111 | return true; |
| 112 | } |
| 113 | |
| 114 | static irqreturn_t virtio_vdpa_config_cb(void *private) |
| 115 | { |
| 116 | struct virtio_vdpa_device *vd_dev = private; |
| 117 | |
| 118 | virtio_config_changed(dev: &vd_dev->vdev); |
| 119 | |
| 120 | return IRQ_HANDLED; |
| 121 | } |
| 122 | |
| 123 | static irqreturn_t virtio_vdpa_virtqueue_cb(void *private) |
| 124 | { |
| 125 | struct virtqueue *vq = private; |
| 126 | |
| 127 | return vring_interrupt(irq: 0, vq: vq); |
| 128 | } |
| 129 | |
| 130 | static struct virtqueue * |
| 131 | virtio_vdpa_setup_vq(struct virtio_device *vdev, unsigned int index, |
| 132 | void (*callback)(struct virtqueue *vq), |
| 133 | const char *name, bool ctx) |
| 134 | { |
| 135 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 136 | const struct vdpa_config_ops *ops = vdpa->config; |
| 137 | bool (*notify)(struct virtqueue *vq) = virtio_vdpa_notify; |
| 138 | struct vdpa_callback cb; |
| 139 | struct virtqueue *vq; |
| 140 | u64 desc_addr, driver_addr, device_addr; |
| 141 | union virtio_map map = {0}; |
| 142 | /* Assume split virtqueue, switch to packed if necessary */ |
| 143 | struct vdpa_vq_state state = {0}; |
| 144 | u32 align, max_num, min_num = 1; |
| 145 | bool may_reduce_num = true; |
| 146 | int err; |
| 147 | |
| 148 | if (!name) |
| 149 | return NULL; |
| 150 | |
| 151 | if (index >= vdpa->nvqs) |
| 152 | return ERR_PTR(error: -ENOENT); |
| 153 | |
| 154 | /* We cannot accept VIRTIO_F_NOTIFICATION_DATA without kick_vq_with_data */ |
| 155 | if (__virtio_test_bit(vdev, VIRTIO_F_NOTIFICATION_DATA)) { |
| 156 | if (ops->kick_vq_with_data) |
| 157 | notify = virtio_vdpa_notify_with_data; |
| 158 | else |
| 159 | __virtio_clear_bit(vdev, VIRTIO_F_NOTIFICATION_DATA); |
| 160 | } |
| 161 | |
| 162 | /* Queue shouldn't already be set up. */ |
| 163 | if (ops->get_vq_ready(vdpa, index)) |
| 164 | return ERR_PTR(error: -ENOENT); |
| 165 | |
| 166 | if (ops->get_vq_size) |
| 167 | max_num = ops->get_vq_size(vdpa, index); |
| 168 | else |
| 169 | max_num = ops->get_vq_num_max(vdpa); |
| 170 | |
| 171 | if (max_num == 0) { |
| 172 | err = -ENOENT; |
| 173 | goto error_new_virtqueue; |
| 174 | } |
| 175 | |
| 176 | if (ops->get_vq_num_min) |
| 177 | min_num = ops->get_vq_num_min(vdpa); |
| 178 | |
| 179 | may_reduce_num = (max_num != min_num); |
| 180 | |
| 181 | /* Create the vring */ |
| 182 | align = ops->get_vq_align(vdpa); |
| 183 | |
| 184 | if (ops->get_vq_map) |
| 185 | map = ops->get_vq_map(vdpa, index); |
| 186 | else |
| 187 | map = vdpa_get_map(vdev: vdpa); |
| 188 | |
| 189 | vq = vring_create_virtqueue_map(index, num: max_num, vring_align: align, vdev, |
| 190 | weak_barriers: true, may_reduce_num, ctx, |
| 191 | notify, callback, name, map); |
| 192 | if (!vq) { |
| 193 | err = -ENOMEM; |
| 194 | goto error_new_virtqueue; |
| 195 | } |
| 196 | |
| 197 | if (index == 0) |
| 198 | vdev->vmap = map; |
| 199 | |
| 200 | vq->num_max = max_num; |
| 201 | |
| 202 | /* Setup virtqueue callback */ |
| 203 | cb.callback = callback ? virtio_vdpa_virtqueue_cb : NULL; |
| 204 | cb.private = vq; |
| 205 | cb.trigger = NULL; |
| 206 | ops->set_vq_cb(vdpa, index, &cb); |
| 207 | ops->set_vq_num(vdpa, index, virtqueue_get_vring_size(vq)); |
| 208 | |
| 209 | desc_addr = virtqueue_get_desc_addr(vq); |
| 210 | driver_addr = virtqueue_get_avail_addr(vq); |
| 211 | device_addr = virtqueue_get_used_addr(vq); |
| 212 | |
| 213 | if (ops->set_vq_address(vdpa, index, |
| 214 | desc_addr, driver_addr, |
| 215 | device_addr)) { |
| 216 | err = -EINVAL; |
| 217 | goto err_vq; |
| 218 | } |
| 219 | |
| 220 | /* reset virtqueue state index */ |
| 221 | if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED)) { |
| 222 | struct vdpa_vq_state_packed *s = &state.packed; |
| 223 | |
| 224 | s->last_avail_counter = 1; |
| 225 | s->last_avail_idx = 0; |
| 226 | s->last_used_counter = 1; |
| 227 | s->last_used_idx = 0; |
| 228 | } |
| 229 | err = ops->set_vq_state(vdpa, index, &state); |
| 230 | if (err) |
| 231 | goto err_vq; |
| 232 | |
| 233 | ops->set_vq_ready(vdpa, index, 1); |
| 234 | |
| 235 | return vq; |
| 236 | |
| 237 | err_vq: |
| 238 | vring_del_virtqueue(vq); |
| 239 | error_new_virtqueue: |
| 240 | ops->set_vq_ready(vdpa, index, 0); |
| 241 | /* VDPA driver should make sure vq is stopeed here */ |
| 242 | WARN_ON(ops->get_vq_ready(vdpa, index)); |
| 243 | return ERR_PTR(error: err); |
| 244 | } |
| 245 | |
| 246 | static void virtio_vdpa_del_vq(struct virtqueue *vq) |
| 247 | { |
| 248 | struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vq->vdev); |
| 249 | struct vdpa_device *vdpa = vd_dev->vdpa; |
| 250 | const struct vdpa_config_ops *ops = vdpa->config; |
| 251 | unsigned int index = vq->index; |
| 252 | |
| 253 | /* Select and deactivate the queue (best effort) */ |
| 254 | ops->set_vq_ready(vdpa, index, 0); |
| 255 | |
| 256 | vring_del_virtqueue(vq); |
| 257 | } |
| 258 | |
| 259 | static void virtio_vdpa_del_vqs(struct virtio_device *vdev) |
| 260 | { |
| 261 | struct virtqueue *vq, *n; |
| 262 | |
| 263 | list_for_each_entry_safe(vq, n, &vdev->vqs, list) |
| 264 | virtio_vdpa_del_vq(vq); |
| 265 | } |
| 266 | |
| 267 | static void default_calc_sets(struct irq_affinity *affd, unsigned int affvecs) |
| 268 | { |
| 269 | affd->nr_sets = 1; |
| 270 | affd->set_size[0] = affvecs; |
| 271 | } |
| 272 | |
| 273 | static struct cpumask * |
| 274 | create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd) |
| 275 | { |
| 276 | unsigned int affvecs = 0, curvec, usedvecs, i; |
| 277 | struct cpumask *masks = NULL; |
| 278 | |
| 279 | if (nvecs > affd->pre_vectors + affd->post_vectors) |
| 280 | affvecs = nvecs - affd->pre_vectors - affd->post_vectors; |
| 281 | |
| 282 | if (!affd->calc_sets) |
| 283 | affd->calc_sets = default_calc_sets; |
| 284 | |
| 285 | affd->calc_sets(affd, affvecs); |
| 286 | |
| 287 | if (!affvecs) |
| 288 | return NULL; |
| 289 | |
| 290 | masks = kcalloc(nvecs, sizeof(*masks), GFP_KERNEL); |
| 291 | if (!masks) |
| 292 | return NULL; |
| 293 | |
| 294 | /* Fill out vectors at the beginning that don't need affinity */ |
| 295 | for (curvec = 0; curvec < affd->pre_vectors; curvec++) |
| 296 | cpumask_setall(dstp: &masks[curvec]); |
| 297 | |
| 298 | for (i = 0, usedvecs = 0; i < affd->nr_sets; i++) { |
| 299 | unsigned int this_vecs = affd->set_size[i]; |
| 300 | unsigned int nr_masks; |
| 301 | int j; |
| 302 | struct cpumask *result = group_cpus_evenly(numgrps: this_vecs, nummasks: &nr_masks); |
| 303 | |
| 304 | if (!result) { |
| 305 | kfree(objp: masks); |
| 306 | return NULL; |
| 307 | } |
| 308 | |
| 309 | for (j = 0; j < nr_masks; j++) |
| 310 | cpumask_copy(dstp: &masks[curvec + j], srcp: &result[j]); |
| 311 | kfree(objp: result); |
| 312 | |
| 313 | curvec += nr_masks; |
| 314 | usedvecs += nr_masks; |
| 315 | } |
| 316 | |
| 317 | /* Fill out vectors at the end that don't need affinity */ |
| 318 | if (usedvecs >= affvecs) |
| 319 | curvec = affd->pre_vectors + affvecs; |
| 320 | else |
| 321 | curvec = affd->pre_vectors + usedvecs; |
| 322 | for (; curvec < nvecs; curvec++) |
| 323 | cpumask_setall(dstp: &masks[curvec]); |
| 324 | |
| 325 | return masks; |
| 326 | } |
| 327 | |
| 328 | static int virtio_vdpa_find_vqs(struct virtio_device *vdev, unsigned int nvqs, |
| 329 | struct virtqueue *vqs[], |
| 330 | struct virtqueue_info vqs_info[], |
| 331 | struct irq_affinity *desc) |
| 332 | { |
| 333 | struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vdev); |
| 334 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 335 | const struct vdpa_config_ops *ops = vdpa->config; |
| 336 | struct cpumask *masks; |
| 337 | struct vdpa_callback cb; |
| 338 | bool has_affinity = desc && ops->set_vq_affinity; |
| 339 | int i, err, queue_idx = 0; |
| 340 | |
| 341 | if (has_affinity) { |
| 342 | masks = create_affinity_masks(nvecs: nvqs, affd: desc); |
| 343 | if (!masks) |
| 344 | return -ENOMEM; |
| 345 | } |
| 346 | |
| 347 | for (i = 0; i < nvqs; ++i) { |
| 348 | struct virtqueue_info *vqi = &vqs_info[i]; |
| 349 | |
| 350 | if (!vqi->name) { |
| 351 | vqs[i] = NULL; |
| 352 | continue; |
| 353 | } |
| 354 | |
| 355 | vqs[i] = virtio_vdpa_setup_vq(vdev, index: queue_idx++, callback: vqi->callback, |
| 356 | name: vqi->name, ctx: vqi->ctx); |
| 357 | if (IS_ERR(ptr: vqs[i])) { |
| 358 | err = PTR_ERR(ptr: vqs[i]); |
| 359 | goto err_setup_vq; |
| 360 | } |
| 361 | |
| 362 | if (has_affinity) |
| 363 | ops->set_vq_affinity(vdpa, i, &masks[i]); |
| 364 | } |
| 365 | |
| 366 | cb.callback = virtio_vdpa_config_cb; |
| 367 | cb.private = vd_dev; |
| 368 | ops->set_config_cb(vdpa, &cb); |
| 369 | if (has_affinity) |
| 370 | kfree(objp: masks); |
| 371 | |
| 372 | return 0; |
| 373 | |
| 374 | err_setup_vq: |
| 375 | virtio_vdpa_del_vqs(vdev); |
| 376 | if (has_affinity) |
| 377 | kfree(objp: masks); |
| 378 | return err; |
| 379 | } |
| 380 | |
| 381 | static u64 virtio_vdpa_get_features(struct virtio_device *vdev) |
| 382 | { |
| 383 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 384 | const struct vdpa_config_ops *ops = vdpa->config; |
| 385 | |
| 386 | return ops->get_device_features(vdpa); |
| 387 | } |
| 388 | |
| 389 | static int virtio_vdpa_finalize_features(struct virtio_device *vdev) |
| 390 | { |
| 391 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 392 | |
| 393 | /* Give virtio_ring a chance to accept features. */ |
| 394 | vring_transport_features(vdev); |
| 395 | |
| 396 | return vdpa_set_features(vdev: vdpa, features: vdev->features); |
| 397 | } |
| 398 | |
| 399 | static const char *virtio_vdpa_bus_name(struct virtio_device *vdev) |
| 400 | { |
| 401 | struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vdev); |
| 402 | struct vdpa_device *vdpa = vd_dev->vdpa; |
| 403 | |
| 404 | return dev_name(dev: &vdpa->dev); |
| 405 | } |
| 406 | |
| 407 | static int virtio_vdpa_set_vq_affinity(struct virtqueue *vq, |
| 408 | const struct cpumask *cpu_mask) |
| 409 | { |
| 410 | struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vq->vdev); |
| 411 | struct vdpa_device *vdpa = vd_dev->vdpa; |
| 412 | const struct vdpa_config_ops *ops = vdpa->config; |
| 413 | unsigned int index = vq->index; |
| 414 | |
| 415 | if (ops->set_vq_affinity) |
| 416 | return ops->set_vq_affinity(vdpa, index, cpu_mask); |
| 417 | |
| 418 | return 0; |
| 419 | } |
| 420 | |
| 421 | static const struct cpumask * |
| 422 | virtio_vdpa_get_vq_affinity(struct virtio_device *vdev, int index) |
| 423 | { |
| 424 | struct vdpa_device *vdpa = vd_get_vdpa(vdev); |
| 425 | const struct vdpa_config_ops *ops = vdpa->config; |
| 426 | |
| 427 | if (ops->get_vq_affinity) |
| 428 | return ops->get_vq_affinity(vdpa, index); |
| 429 | |
| 430 | return NULL; |
| 431 | } |
| 432 | |
| 433 | static const struct virtio_config_ops virtio_vdpa_config_ops = { |
| 434 | .get = virtio_vdpa_get, |
| 435 | .set = virtio_vdpa_set, |
| 436 | .generation = virtio_vdpa_generation, |
| 437 | .get_status = virtio_vdpa_get_status, |
| 438 | .set_status = virtio_vdpa_set_status, |
| 439 | .reset = virtio_vdpa_reset, |
| 440 | .find_vqs = virtio_vdpa_find_vqs, |
| 441 | .del_vqs = virtio_vdpa_del_vqs, |
| 442 | .get_features = virtio_vdpa_get_features, |
| 443 | .finalize_features = virtio_vdpa_finalize_features, |
| 444 | .bus_name = virtio_vdpa_bus_name, |
| 445 | .set_vq_affinity = virtio_vdpa_set_vq_affinity, |
| 446 | .get_vq_affinity = virtio_vdpa_get_vq_affinity, |
| 447 | }; |
| 448 | |
| 449 | static void virtio_vdpa_release_dev(struct device *_d) |
| 450 | { |
| 451 | struct virtio_device *vdev = |
| 452 | container_of(_d, struct virtio_device, dev); |
| 453 | struct virtio_vdpa_device *vd_dev = |
| 454 | container_of(vdev, struct virtio_vdpa_device, vdev); |
| 455 | |
| 456 | kfree(objp: vd_dev); |
| 457 | } |
| 458 | |
| 459 | static int virtio_vdpa_probe(struct vdpa_device *vdpa) |
| 460 | { |
| 461 | const struct vdpa_config_ops *ops = vdpa->config; |
| 462 | struct virtio_vdpa_device *vd_dev, *reg_dev = NULL; |
| 463 | int ret = -EINVAL; |
| 464 | |
| 465 | vd_dev = kzalloc(sizeof(*vd_dev), GFP_KERNEL); |
| 466 | if (!vd_dev) |
| 467 | return -ENOMEM; |
| 468 | |
| 469 | vd_dev->vdev.dev.parent = vdpa->map ? &vdpa->dev : |
| 470 | vdpa_get_map(vdev: vdpa).dma_dev; |
| 471 | vd_dev->vdev.dev.release = virtio_vdpa_release_dev; |
| 472 | vd_dev->vdev.config = &virtio_vdpa_config_ops; |
| 473 | vd_dev->vdev.map = vdpa->map; |
| 474 | vd_dev->vdpa = vdpa; |
| 475 | |
| 476 | vd_dev->vdev.id.device = ops->get_device_id(vdpa); |
| 477 | if (vd_dev->vdev.id.device == 0) |
| 478 | goto err; |
| 479 | |
| 480 | vd_dev->vdev.id.vendor = ops->get_vendor_id(vdpa); |
| 481 | ret = register_virtio_device(dev: &vd_dev->vdev); |
| 482 | reg_dev = vd_dev; |
| 483 | if (ret) |
| 484 | goto err; |
| 485 | |
| 486 | vdpa_set_drvdata(vdev: vdpa, data: vd_dev); |
| 487 | |
| 488 | return 0; |
| 489 | |
| 490 | err: |
| 491 | if (reg_dev) |
| 492 | put_device(dev: &vd_dev->vdev.dev); |
| 493 | else |
| 494 | kfree(objp: vd_dev); |
| 495 | return ret; |
| 496 | } |
| 497 | |
| 498 | static void virtio_vdpa_remove(struct vdpa_device *vdpa) |
| 499 | { |
| 500 | struct virtio_vdpa_device *vd_dev = vdpa_get_drvdata(vdev: vdpa); |
| 501 | |
| 502 | unregister_virtio_device(dev: &vd_dev->vdev); |
| 503 | } |
| 504 | |
| 505 | static struct vdpa_driver virtio_vdpa_driver = { |
| 506 | .driver = { |
| 507 | .name = "virtio_vdpa" , |
| 508 | }, |
| 509 | .probe = virtio_vdpa_probe, |
| 510 | .remove = virtio_vdpa_remove, |
| 511 | }; |
| 512 | |
| 513 | module_vdpa_driver(virtio_vdpa_driver); |
| 514 | |
| 515 | MODULE_VERSION(MOD_VERSION); |
| 516 | MODULE_LICENSE(MOD_LICENSE); |
| 517 | MODULE_AUTHOR(MOD_AUTHOR); |
| 518 | MODULE_DESCRIPTION(MOD_DESC); |
| 519 | |