fs/fuse/virtio_fs.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * virtio-fs: Virtio Filesystem
   4  * Copyright (C) 2018 Red Hat, Inc.
   5  */
   6
   7 #include <linux/fs.h>
   8 #include <linux/dax.h>
   9 #include <linux/pci.h>
  10 #include <linux/pfn_t.h>
  11 #include <linux/memremap.h>
  12 #include <linux/module.h>
  13 #include <linux/virtio.h>
  14 #include <linux/virtio_fs.h>
  15 #include <linux/delay.h>
  16 #include <linux/fs_context.h>
  17 #include <linux/fs_parser.h>
  18 #include <linux/highmem.h>
  19 #include <linux/uio.h>
  20 #include "fuse_i.h"
  21
  22 /* Used to help calculate the FUSE connection's max_pages limit for a request's
  23  * size. Parts of the struct fuse_req are sliced into scattergather lists in
  24  * addition to the pages used, so this can help account for that overhead.
  25  */
  26 #define FUSE_HEADER_OVERHEAD    4
  27
  28 /* List of virtio-fs device instances and a lock for the list. Also provides
  29  * mutual exclusion in device removal and mounting path
  30  */
  31 static DEFINE_MUTEX(virtio_fs_mutex);
  32 static LIST_HEAD(virtio_fs_instances);
  33
  34 enum {
  35         VQ_HIPRIO,
  36         VQ_REQUEST
  37 };
  38
  39 #define VQ_NAME_LEN     24
  40
  41 /* Per-virtqueue state */
  42 struct virtio_fs_vq {
  43         spinlock_t lock;
  44         struct virtqueue *vq;     /* protected by ->lock */
  45         struct work_struct done_work;
  46         struct list_head queued_reqs;
  47         struct list_head end_reqs;      /* End these requests */
  48         struct delayed_work dispatch_work;
  49         struct fuse_dev *fud;
  50         bool connected;
  51         long in_flight;
  52         struct completion in_flight_zero; /* No inflight requests */
  53         char name[VQ_NAME_LEN];
  54 } ____cacheline_aligned_in_smp;
  55
  56 /* A virtio-fs device instance */
  57 struct virtio_fs {
  58         struct kref refcount;
  59         struct list_head list;    /* on virtio_fs_instances */
  60         char *tag;
  61         struct virtio_fs_vq *vqs;
  62         unsigned int nvqs;               /* number of virtqueues */
  63         unsigned int num_request_queues; /* number of request queues */
  64         struct dax_device *dax_dev;
  65
  66         /* DAX memory window where file contents are mapped */
  67         void *window_kaddr;
  68         phys_addr_t window_phys_addr;
  69         size_t window_len;
  70 };
  71
  72 struct virtio_fs_forget_req {
  73         struct fuse_in_header ih;
  74         struct fuse_forget_in arg;
  75 };
  76
  77 struct virtio_fs_forget {
  78         /* This request can be temporarily queued on virt queue */
  79         struct list_head list;
  80         struct virtio_fs_forget_req req;
  81 };
  82
  83 struct virtio_fs_req_work {
  84         struct fuse_req *req;
  85         struct virtio_fs_vq *fsvq;
  86         struct work_struct done_work;
  87 };
  88
  89 static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
  90                                  struct fuse_req *req, bool in_flight);
  91
  92 static const struct constant_table dax_param_enums[] = {
  93         {"always",      FUSE_DAX_ALWAYS },
  94         {"never",       FUSE_DAX_NEVER },
  95         {"inode",       FUSE_DAX_INODE_USER },
  96         {}
  97 };
  98
  99 enum {
 100         OPT_DAX,
 101         OPT_DAX_ENUM,
 102 };
 103
 104 static const struct fs_parameter_spec virtio_fs_parameters[] = {
 105         fsparam_flag("dax", OPT_DAX),
 106         fsparam_enum("dax", OPT_DAX_ENUM, dax_param_enums),
 107         {}
 108 };
 109
 110 static int virtio_fs_parse_param(struct fs_context *fsc,
 111                                  struct fs_parameter *param)
 112 {
 113         struct fs_parse_result result;
 114         struct fuse_fs_context *ctx = fsc->fs_private;
 115         int opt;
 116
 117         opt = fs_parse(fsc, virtio_fs_parameters, param, &result);
 118         if (opt < 0)
 119                 return opt;
 120
 121         switch (opt) {
 122         case OPT_DAX:
 123                 ctx->dax_mode = FUSE_DAX_ALWAYS;
 124                 break;
 125         case OPT_DAX_ENUM:
 126                 ctx->dax_mode = result.uint_32;
 127                 break;
 128         default:
 129                 return -EINVAL;
 130         }
 131
 132         return 0;
 133 }
 134
 135 static void virtio_fs_free_fsc(struct fs_context *fsc)
 136 {
 137         struct fuse_fs_context *ctx = fsc->fs_private;
 138
 139         kfree(ctx);
 140 }
 141
 142 static inline struct virtio_fs_vq *vq_to_fsvq(struct virtqueue *vq)
 143 {
 144         struct virtio_fs *fs = vq->vdev->priv;
 145
 146         return &fs->vqs[vq->index];
 147 }
 148
 149 /* Should be called with fsvq->lock held. */
 150 static inline void inc_in_flight_req(struct virtio_fs_vq *fsvq)
 151 {
 152         fsvq->in_flight++;
 153 }
 154
 155 /* Should be called with fsvq->lock held. */
 156 static inline void dec_in_flight_req(struct virtio_fs_vq *fsvq)
 157 {
 158         WARN_ON(fsvq->in_flight <= 0);
 159         fsvq->in_flight--;
 160         if (!fsvq->in_flight)
 161                 complete(&fsvq->in_flight_zero);
 162 }
 163
 164 static void release_virtio_fs_obj(struct kref *ref)
 165 {
 166         struct virtio_fs *vfs = container_of(ref, struct virtio_fs, refcount);
 167
 168         kfree(vfs->vqs);
 169         kfree(vfs);
 170 }
 171
 172 /* Make sure virtiofs_mutex is held */
 173 static void virtio_fs_put(struct virtio_fs *fs)
 174 {
 175         kref_put(&fs->refcount, release_virtio_fs_obj);
 176 }
 177
 178 static void virtio_fs_fiq_release(struct fuse_iqueue *fiq)
 179 {
 180         struct virtio_fs *vfs = fiq->priv;
 181
 182         mutex_lock(&virtio_fs_mutex);
 183         virtio_fs_put(vfs);
 184         mutex_unlock(&virtio_fs_mutex);
 185 }
 186
 187 static void virtio_fs_drain_queue(struct virtio_fs_vq *fsvq)
 188 {
 189         WARN_ON(fsvq->in_flight < 0);
 190
 191         /* Wait for in flight requests to finish.*/
 192         spin_lock(&fsvq->lock);
 193         if (fsvq->in_flight) {
 194                 /* We are holding virtio_fs_mutex. There should not be any
 195                  * waiters waiting for completion.
 196                  */
 197                 reinit_completion(&fsvq->in_flight_zero);
 198                 spin_unlock(&fsvq->lock);
 199                 wait_for_completion(&fsvq->in_flight_zero);
 200         } else {
 201                 spin_unlock(&fsvq->lock);
 202         }
 203
 204         flush_work(&fsvq->done_work);
 205         flush_delayed_work(&fsvq->dispatch_work);
 206 }
 207
 208 static void virtio_fs_drain_all_queues_locked(struct virtio_fs *fs)
 209 {
 210         struct virtio_fs_vq *fsvq;
 211         int i;
 212
 213         for (i = 0; i < fs->nvqs; i++) {
 214                 fsvq = &fs->vqs[i];
 215                 virtio_fs_drain_queue(fsvq);
 216         }
 217 }
 218
 219 static void virtio_fs_drain_all_queues(struct virtio_fs *fs)
 220 {
 221         /* Provides mutual exclusion between ->remove and ->kill_sb
 222          * paths. We don't want both of these draining queue at the
 223          * same time. Current completion logic reinits completion
 224          * and that means there should not be any other thread
 225          * doing reinit or waiting for completion already.
 226          */
 227         mutex_lock(&virtio_fs_mutex);
 228         virtio_fs_drain_all_queues_locked(fs);
 229         mutex_unlock(&virtio_fs_mutex);
 230 }
 231
 232 static void virtio_fs_start_all_queues(struct virtio_fs *fs)
 233 {
 234         struct virtio_fs_vq *fsvq;
 235         int i;
 236
 237         for (i = 0; i < fs->nvqs; i++) {
 238                 fsvq = &fs->vqs[i];
 239                 spin_lock(&fsvq->lock);
 240                 fsvq->connected = true;
 241                 spin_unlock(&fsvq->lock);
 242         }
 243 }
 244
 245 /* Add a new instance to the list or return -EEXIST if tag name exists*/
 246 static int virtio_fs_add_instance(struct virtio_fs *fs)
 247 {
 248         struct virtio_fs *fs2;
 249         bool duplicate = false;
 250
 251         mutex_lock(&virtio_fs_mutex);
 252
 253         list_for_each_entry(fs2, &virtio_fs_instances, list) {
 254                 if (strcmp(fs->tag, fs2->tag) == 0)
 255                         duplicate = true;
 256         }
 257
 258         if (!duplicate)
 259                 list_add_tail(&fs->list, &virtio_fs_instances);
 260
 261         mutex_unlock(&virtio_fs_mutex);
 262
 263         if (duplicate)
 264                 return -EEXIST;
 265         return 0;
 266 }
 267
 268 /* Return the virtio_fs with a given tag, or NULL */
 269 static struct virtio_fs *virtio_fs_find_instance(const char *tag)
 270 {
 271         struct virtio_fs *fs;
 272
 273         mutex_lock(&virtio_fs_mutex);
 274
 275         list_for_each_entry(fs, &virtio_fs_instances, list) {
 276                 if (strcmp(fs->tag, tag) == 0) {
 277                         kref_get(&fs->refcount);
 278                         goto found;
 279                 }
 280         }
 281
 282         fs = NULL; /* not found */
 283
 284 found:
 285         mutex_unlock(&virtio_fs_mutex);
 286
 287         return fs;
 288 }
 289
 290 static void virtio_fs_free_devs(struct virtio_fs *fs)
 291 {
 292         unsigned int i;
 293
 294         for (i = 0; i < fs->nvqs; i++) {
 295                 struct virtio_fs_vq *fsvq = &fs->vqs[i];
 296
 297                 if (!fsvq->fud)
 298                         continue;
 299
 300                 fuse_dev_free(fsvq->fud);
 301                 fsvq->fud = NULL;
 302         }
 303 }
 304
 305 /* Read filesystem name from virtio config into fs->tag (must kfree()). */
 306 static int virtio_fs_read_tag(struct virtio_device *vdev, struct virtio_fs *fs)
 307 {
 308         char tag_buf[sizeof_field(struct virtio_fs_config, tag)];
 309         char *end;
 310         size_t len;
 311
 312         virtio_cread_bytes(vdev, offsetof(struct virtio_fs_config, tag),
 313                            &tag_buf, sizeof(tag_buf));
 314         end = memchr(tag_buf, '\0', sizeof(tag_buf));
 315         if (end == tag_buf)
 316                 return -EINVAL; /* empty tag */
 317         if (!end)
 318                 end = &tag_buf[sizeof(tag_buf)];
 319
 320         len = end - tag_buf;
 321         fs->tag = devm_kmalloc(&vdev->dev, len + 1, GFP_KERNEL);
 322         if (!fs->tag)
 323                 return -ENOMEM;
 324         memcpy(fs->tag, tag_buf, len);
 325         fs->tag[len] = '\0';
 326         return 0;
 327 }
 328
 329 /* Work function for hiprio completion */
 330 static void virtio_fs_hiprio_done_work(struct work_struct *work)
 331 {
 332         struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
 333                                                  done_work);
 334         struct virtqueue *vq = fsvq->vq;
 335
 336         /* Free completed FUSE_FORGET requests */
 337         spin_lock(&fsvq->lock);
 338         do {
 339                 unsigned int len;
 340                 void *req;
 341
 342                 virtqueue_disable_cb(vq);
 343
 344                 while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
 345                         kfree(req);
 346                         dec_in_flight_req(fsvq);
 347                 }
 348         } while (!virtqueue_enable_cb(vq) && likely(!virtqueue_is_broken(vq)));
 349         spin_unlock(&fsvq->lock);
 350 }
 351
 352 static void virtio_fs_request_dispatch_work(struct work_struct *work)
 353 {
 354         struct fuse_req *req;
 355         struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
 356                                                  dispatch_work.work);
 357         int ret;
 358
 359         pr_debug("virtio-fs: worker %s called.\n", __func__);
 360         while (1) {
 361                 spin_lock(&fsvq->lock);
 362                 req = list_first_entry_or_null(&fsvq->end_reqs, struct fuse_req,
 363                                                list);
 364                 if (!req) {
 365                         spin_unlock(&fsvq->lock);
 366                         break;
 367                 }
 368
 369                 list_del_init(&req->list);
 370                 spin_unlock(&fsvq->lock);
 371                 fuse_request_end(req);
 372         }
 373
 374         /* Dispatch pending requests */
 375         while (1) {
 376                 spin_lock(&fsvq->lock);
 377                 req = list_first_entry_or_null(&fsvq->queued_reqs,
 378                                                struct fuse_req, list);
 379                 if (!req) {
 380                         spin_unlock(&fsvq->lock);
 381                         return;
 382                 }
 383                 list_del_init(&req->list);
 384                 spin_unlock(&fsvq->lock);
 385
 386                 ret = virtio_fs_enqueue_req(fsvq, req, true);
 387                 if (ret < 0) {
 388                         if (ret == -ENOMEM || ret == -ENOSPC) {
 389                                 spin_lock(&fsvq->lock);
 390                                 list_add_tail(&req->list, &fsvq->queued_reqs);
 391                                 schedule_delayed_work(&fsvq->dispatch_work,
 392                                                       msecs_to_jiffies(1));
 393                                 spin_unlock(&fsvq->lock);
 394                                 return;
 395                         }
 396                         req->out.h.error = ret;
 397                         spin_lock(&fsvq->lock);
 398                         dec_in_flight_req(fsvq);
 399                         spin_unlock(&fsvq->lock);
 400                         pr_err("virtio-fs: virtio_fs_enqueue_req() failed %d\n",
 401                                ret);
 402                         fuse_request_end(req);
 403                 }
 404         }
 405 }
 406
 407 /*
 408  * Returns 1 if queue is full and sender should wait a bit before sending
 409  * next request, 0 otherwise.
 410  */
 411 static int send_forget_request(struct virtio_fs_vq *fsvq,
 412                                struct virtio_fs_forget *forget,
 413                                bool in_flight)
 414 {
 415         struct scatterlist sg;
 416         struct virtqueue *vq;
 417         int ret = 0;
 418         bool notify;
 419         struct virtio_fs_forget_req *req = &forget->req;
 420
 421         spin_lock(&fsvq->lock);
 422         if (!fsvq->connected) {
 423                 if (in_flight)
 424                         dec_in_flight_req(fsvq);
 425                 kfree(forget);
 426                 goto out;
 427         }
 428
 429         sg_init_one(&sg, req, sizeof(*req));
 430         vq = fsvq->vq;
 431         dev_dbg(&vq->vdev->dev, "%s\n", __func__);
 432
 433         ret = virtqueue_add_outbuf(vq, &sg, 1, forget, GFP_ATOMIC);
 434         if (ret < 0) {
 435                 if (ret == -ENOMEM || ret == -ENOSPC) {
 436                         pr_debug("virtio-fs: Could not queue FORGET: err=%d. Will try later\n",
 437                                  ret);
 438                         list_add_tail(&forget->list, &fsvq->queued_reqs);
 439                         schedule_delayed_work(&fsvq->dispatch_work,
 440                                               msecs_to_jiffies(1));
 441                         if (!in_flight)
 442                                 inc_in_flight_req(fsvq);
 443                         /* Queue is full */
 444                         ret = 1;
 445                 } else {
 446                         pr_debug("virtio-fs: Could not queue FORGET: err=%d. Dropping it.\n",
 447                                  ret);
 448                         kfree(forget);
 449                         if (in_flight)
 450                                 dec_in_flight_req(fsvq);
 451                 }
 452                 goto out;
 453         }
 454
 455         if (!in_flight)
 456                 inc_in_flight_req(fsvq);
 457         notify = virtqueue_kick_prepare(vq);
 458         spin_unlock(&fsvq->lock);
 459
 460         if (notify)
 461                 virtqueue_notify(vq);
 462         return ret;
 463 out:
 464         spin_unlock(&fsvq->lock);
 465         return ret;
 466 }
 467
 468 static void virtio_fs_hiprio_dispatch_work(struct work_struct *work)
 469 {
 470         struct virtio_fs_forget *forget;
 471         struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
 472                                                  dispatch_work.work);
 473         pr_debug("virtio-fs: worker %s called.\n", __func__);
 474         while (1) {
 475                 spin_lock(&fsvq->lock);
 476                 forget = list_first_entry_or_null(&fsvq->queued_reqs,
 477                                         struct virtio_fs_forget, list);
 478                 if (!forget) {
 479                         spin_unlock(&fsvq->lock);
 480                         return;
 481                 }
 482
 483                 list_del(&forget->list);
 484                 spin_unlock(&fsvq->lock);
 485                 if (send_forget_request(fsvq, forget, true))
 486                         return;
 487         }
 488 }
 489
 490 /* Allocate and copy args into req->argbuf */
 491 static int copy_args_to_argbuf(struct fuse_req *req)
 492 {
 493         struct fuse_args *args = req->args;
 494         unsigned int offset = 0;
 495         unsigned int num_in;
 496         unsigned int num_out;
 497         unsigned int len;
 498         unsigned int i;
 499
 500         num_in = args->in_numargs - args->in_pages;
 501         num_out = args->out_numargs - args->out_pages;
 502         len = fuse_len_args(num_in, (struct fuse_arg *) args->in_args) +
 503               fuse_len_args(num_out, args->out_args);
 504
 505         req->argbuf = kmalloc(len, GFP_ATOMIC);
 506         if (!req->argbuf)
 507                 return -ENOMEM;
 508
 509         for (i = 0; i < num_in; i++) {
 510                 memcpy(req->argbuf + offset,
 511                        args->in_args[i].value,
 512                        args->in_args[i].size);
 513                 offset += args->in_args[i].size;
 514         }
 515
 516         return 0;
 517 }
 518
 519 /* Copy args out of and free req->argbuf */
 520 static void copy_args_from_argbuf(struct fuse_args *args, struct fuse_req *req)
 521 {
 522         unsigned int remaining;
 523         unsigned int offset;
 524         unsigned int num_in;
 525         unsigned int num_out;
 526         unsigned int i;
 527
 528         remaining = req->out.h.len - sizeof(req->out.h);
 529         num_in = args->in_numargs - args->in_pages;
 530         num_out = args->out_numargs - args->out_pages;
 531         offset = fuse_len_args(num_in, (struct fuse_arg *)args->in_args);
 532
 533         for (i = 0; i < num_out; i++) {
 534                 unsigned int argsize = args->out_args[i].size;
 535
 536                 if (args->out_argvar &&
 537                     i == args->out_numargs - 1 &&
 538                     argsize > remaining) {
 539                         argsize = remaining;
 540                 }
 541
 542                 memcpy(args->out_args[i].value, req->argbuf + offset, argsize);
 543                 offset += argsize;
 544
 545                 if (i != args->out_numargs - 1)
 546                         remaining -= argsize;
 547         }
 548
 549         /* Store the actual size of the variable-length arg */
 550         if (args->out_argvar)
 551                 args->out_args[args->out_numargs - 1].size = remaining;
 552
 553         kfree(req->argbuf);
 554         req->argbuf = NULL;
 555 }
 556
 557 /* Work function for request completion */
 558 static void virtio_fs_request_complete(struct fuse_req *req,
 559                                        struct virtio_fs_vq *fsvq)
 560 {
 561         struct fuse_pqueue *fpq = &fsvq->fud->pq;
 562         struct fuse_args *args;
 563         struct fuse_args_pages *ap;
 564         unsigned int len, i, thislen;
 565         struct page *page;
 566
 567         /*
 568          * TODO verify that server properly follows FUSE protocol
 569          * (oh.uniq, oh.len)
 570          */
 571         args = req->args;
 572         copy_args_from_argbuf(args, req);
 573
 574         if (args->out_pages && args->page_zeroing) {
 575                 len = args->out_args[args->out_numargs - 1].size;
 576                 ap = container_of(args, typeof(*ap), args);
 577                 for (i = 0; i < ap->num_pages; i++) {
 578                         thislen = ap->descs[i].length;
 579                         if (len < thislen) {
 580                                 WARN_ON(ap->descs[i].offset);
 581                                 page = ap->pages[i];
 582                                 zero_user_segment(page, len, thislen);
 583                                 len = 0;
 584                         } else {
 585                                 len -= thislen;
 586                         }
 587                 }
 588         }
 589
 590         spin_lock(&fpq->lock);
 591         clear_bit(FR_SENT, &req->flags);
 592         spin_unlock(&fpq->lock);
 593
 594         fuse_request_end(req);
 595         spin_lock(&fsvq->lock);
 596         dec_in_flight_req(fsvq);
 597         spin_unlock(&fsvq->lock);
 598 }
 599
 600 static void virtio_fs_complete_req_work(struct work_struct *work)
 601 {
 602         struct virtio_fs_req_work *w =
 603                 container_of(work, typeof(*w), done_work);
 604
 605         virtio_fs_request_complete(w->req, w->fsvq);
 606         kfree(w);
 607 }
 608
 609 static void virtio_fs_requests_done_work(struct work_struct *work)
 610 {
 611         struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
 612                                                  done_work);
 613         struct fuse_pqueue *fpq = &fsvq->fud->pq;
 614         struct virtqueue *vq = fsvq->vq;
 615         struct fuse_req *req;
 616         struct fuse_req *next;
 617         unsigned int len;
 618         LIST_HEAD(reqs);
 619
 620         /* Collect completed requests off the virtqueue */
 621         spin_lock(&fsvq->lock);
 622         do {
 623                 virtqueue_disable_cb(vq);
 624
 625                 while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
 626                         spin_lock(&fpq->lock);
 627                         list_move_tail(&req->list, &reqs);
 628                         spin_unlock(&fpq->lock);
 629                 }
 630         } while (!virtqueue_enable_cb(vq) && likely(!virtqueue_is_broken(vq)));
 631         spin_unlock(&fsvq->lock);
 632
 633         /* End requests */
 634         list_for_each_entry_safe(req, next, &reqs, list) {
 635                 list_del_init(&req->list);
 636
 637                 /* blocking async request completes in a worker context */
 638                 if (req->args->may_block) {
 639                         struct virtio_fs_req_work *w;
 640
 641                         w = kzalloc(sizeof(*w), GFP_NOFS | __GFP_NOFAIL);
 642                         INIT_WORK(&w->done_work, virtio_fs_complete_req_work);
 643                         w->fsvq = fsvq;
 644                         w->req = req;
 645                         schedule_work(&w->done_work);
 646                 } else {
 647                         virtio_fs_request_complete(req, fsvq);
 648                 }
 649         }
 650 }
 651
 652 /* Virtqueue interrupt handler */
 653 static void virtio_fs_vq_done(struct virtqueue *vq)
 654 {
 655         struct virtio_fs_vq *fsvq = vq_to_fsvq(vq);
 656
 657         dev_dbg(&vq->vdev->dev, "%s %s\n", __func__, fsvq->name);
 658
 659         schedule_work(&fsvq->done_work);
 660 }
 661
 662 static void virtio_fs_init_vq(struct virtio_fs_vq *fsvq, char *name,
 663                               int vq_type)
 664 {
 665         strscpy(fsvq->name, name, VQ_NAME_LEN);
 666         spin_lock_init(&fsvq->lock);
 667         INIT_LIST_HEAD(&fsvq->queued_reqs);
 668         INIT_LIST_HEAD(&fsvq->end_reqs);
 669         init_completion(&fsvq->in_flight_zero);
 670
 671         if (vq_type == VQ_REQUEST) {
 672                 INIT_WORK(&fsvq->done_work, virtio_fs_requests_done_work);
 673                 INIT_DELAYED_WORK(&fsvq->dispatch_work,
 674                                   virtio_fs_request_dispatch_work);
 675         } else {
 676                 INIT_WORK(&fsvq->done_work, virtio_fs_hiprio_done_work);
 677                 INIT_DELAYED_WORK(&fsvq->dispatch_work,
 678                                   virtio_fs_hiprio_dispatch_work);
 679         }
 680 }
 681
 682 /* Initialize virtqueues */
 683 static int virtio_fs_setup_vqs(struct virtio_device *vdev,
 684                                struct virtio_fs *fs)
 685 {
 686         struct virtqueue **vqs;
 687         vq_callback_t **callbacks;
 688         const char **names;
 689         unsigned int i;
 690         int ret = 0;
 691
 692         virtio_cread_le(vdev, struct virtio_fs_config, num_request_queues,
 693                         &fs->num_request_queues);
 694         if (fs->num_request_queues == 0)
 695                 return -EINVAL;
 696
 697         fs->nvqs = VQ_REQUEST + fs->num_request_queues;
 698         fs->vqs = kcalloc(fs->nvqs, sizeof(fs->vqs[VQ_HIPRIO]), GFP_KERNEL);
 699         if (!fs->vqs)
 700                 return -ENOMEM;
 701
 702         vqs = kmalloc_array(fs->nvqs, sizeof(vqs[VQ_HIPRIO]), GFP_KERNEL);
 703         callbacks = kmalloc_array(fs->nvqs, sizeof(callbacks[VQ_HIPRIO]),
 704                                         GFP_KERNEL);
 705         names = kmalloc_array(fs->nvqs, sizeof(names[VQ_HIPRIO]), GFP_KERNEL);
 706         if (!vqs || !callbacks || !names) {
 707                 ret = -ENOMEM;
 708                 goto out;
 709         }
 710
 711         /* Initialize the hiprio/forget request virtqueue */
 712         callbacks[VQ_HIPRIO] = virtio_fs_vq_done;
 713         virtio_fs_init_vq(&fs->vqs[VQ_HIPRIO], "hiprio", VQ_HIPRIO);
 714         names[VQ_HIPRIO] = fs->vqs[VQ_HIPRIO].name;
 715
 716         /* Initialize the requests virtqueues */
 717         for (i = VQ_REQUEST; i < fs->nvqs; i++) {
 718                 char vq_name[VQ_NAME_LEN];
 719
 720                 snprintf(vq_name, VQ_NAME_LEN, "requests.%u", i - VQ_REQUEST);
 721                 virtio_fs_init_vq(&fs->vqs[i], vq_name, VQ_REQUEST);
 722                 callbacks[i] = virtio_fs_vq_done;
 723                 names[i] = fs->vqs[i].name;
 724         }
 725
 726         ret = virtio_find_vqs(vdev, fs->nvqs, vqs, callbacks, names, NULL);
 727         if (ret < 0)
 728                 goto out;
 729
 730         for (i = 0; i < fs->nvqs; i++)
 731                 fs->vqs[i].vq = vqs[i];
 732
 733         virtio_fs_start_all_queues(fs);
 734 out:
 735         kfree(names);
 736         kfree(callbacks);
 737         kfree(vqs);
 738         if (ret)
 739                 kfree(fs->vqs);
 740         return ret;
 741 }
 742
 743 /* Free virtqueues (device must already be reset) */
 744 static void virtio_fs_cleanup_vqs(struct virtio_device *vdev)
 745 {
 746         vdev->config->del_vqs(vdev);
 747 }
 748
 749 /* Map a window offset to a page frame number.  The window offset will have
 750  * been produced by .iomap_begin(), which maps a file offset to a window
 751  * offset.
 752  */
 753 static long virtio_fs_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
 754                                     long nr_pages, enum dax_access_mode mode,
 755                                     void **kaddr, pfn_t *pfn)
 756 {
 757         struct virtio_fs *fs = dax_get_private(dax_dev);
 758         phys_addr_t offset = PFN_PHYS(pgoff);
 759         size_t max_nr_pages = fs->window_len / PAGE_SIZE - pgoff;
 760
 761         if (kaddr)
 762                 *kaddr = fs->window_kaddr + offset;
 763         if (pfn)
 764                 *pfn = phys_to_pfn_t(fs->window_phys_addr + offset,
 765                                         PFN_DEV | PFN_MAP);
 766         return nr_pages > max_nr_pages ? max_nr_pages : nr_pages;
 767 }
 768
 769 static int virtio_fs_zero_page_range(struct dax_device *dax_dev,
 770                                      pgoff_t pgoff, size_t nr_pages)
 771 {
 772         long rc;
 773         void *kaddr;
 774
 775         rc = dax_direct_access(dax_dev, pgoff, nr_pages, DAX_ACCESS, &kaddr,
 776                                NULL);
 777         if (rc < 0)
 778                 return rc;
 779         memset(kaddr, 0, nr_pages << PAGE_SHIFT);
 780         dax_flush(dax_dev, kaddr, nr_pages << PAGE_SHIFT);
 781         return 0;
 782 }
 783
 784 static const struct dax_operations virtio_fs_dax_ops = {
 785         .direct_access = virtio_fs_direct_access,
 786         .zero_page_range = virtio_fs_zero_page_range,
 787 };
 788
 789 static void virtio_fs_cleanup_dax(void *data)
 790 {
 791         struct dax_device *dax_dev = data;
 792
 793         kill_dax(dax_dev);
 794         put_dax(dax_dev);
 795 }
 796
 797 static int virtio_fs_setup_dax(struct virtio_device *vdev, struct virtio_fs *fs)
 798 {
 799         struct virtio_shm_region cache_reg;
 800         struct dev_pagemap *pgmap;
 801         bool have_cache;
 802
 803         if (!IS_ENABLED(CONFIG_FUSE_DAX))
 804                 return 0;
 805
 806         /* Get cache region */
 807         have_cache = virtio_get_shm_region(vdev, &cache_reg,
 808                                            (u8)VIRTIO_FS_SHMCAP_ID_CACHE);
 809         if (!have_cache) {
 810                 dev_notice(&vdev->dev, "%s: No cache capability\n", __func__);
 811                 return 0;
 812         }
 813
 814         if (!devm_request_mem_region(&vdev->dev, cache_reg.addr, cache_reg.len,
 815                                      dev_name(&vdev->dev))) {
 816                 dev_warn(&vdev->dev, "could not reserve region addr=0x%llx len=0x%llx\n",
 817                          cache_reg.addr, cache_reg.len);
 818                 return -EBUSY;
 819         }
 820
 821         dev_notice(&vdev->dev, "Cache len: 0x%llx @ 0x%llx\n", cache_reg.len,
 822                    cache_reg.addr);
 823
 824         pgmap = devm_kzalloc(&vdev->dev, sizeof(*pgmap), GFP_KERNEL);
 825         if (!pgmap)
 826                 return -ENOMEM;
 827
 828         pgmap->type = MEMORY_DEVICE_FS_DAX;
 829
 830         /* Ideally we would directly use the PCI BAR resource but
 831          * devm_memremap_pages() wants its own copy in pgmap.  So
 832          * initialize a struct resource from scratch (only the start
 833          * and end fields will be used).
 834          */
 835         pgmap->range = (struct range) {
 836                 .start = (phys_addr_t) cache_reg.addr,
 837                 .end = (phys_addr_t) cache_reg.addr + cache_reg.len - 1,
 838         };
 839         pgmap->nr_range = 1;
 840
 841         fs->window_kaddr = devm_memremap_pages(&vdev->dev, pgmap);
 842         if (IS_ERR(fs->window_kaddr))
 843                 return PTR_ERR(fs->window_kaddr);
 844
 845         fs->window_phys_addr = (phys_addr_t) cache_reg.addr;
 846         fs->window_len = (phys_addr_t) cache_reg.len;
 847
 848         dev_dbg(&vdev->dev, "%s: window kaddr 0x%px phys_addr 0x%llx len 0x%llx\n",
 849                 __func__, fs->window_kaddr, cache_reg.addr, cache_reg.len);
 850
 851         fs->dax_dev = alloc_dax(fs, &virtio_fs_dax_ops);
 852         if (IS_ERR(fs->dax_dev))
 853                 return PTR_ERR(fs->dax_dev);
 854
 855         return devm_add_action_or_reset(&vdev->dev, virtio_fs_cleanup_dax,
 856                                         fs->dax_dev);
 857 }
 858
 859 static int virtio_fs_probe(struct virtio_device *vdev)
 860 {
 861         struct virtio_fs *fs;
 862         int ret;
 863
 864         fs = kzalloc(sizeof(*fs), GFP_KERNEL);
 865         if (!fs)
 866                 return -ENOMEM;
 867         kref_init(&fs->refcount);
 868         vdev->priv = fs;
 869
 870         ret = virtio_fs_read_tag(vdev, fs);
 871         if (ret < 0)
 872                 goto out;
 873
 874         ret = virtio_fs_setup_vqs(vdev, fs);
 875         if (ret < 0)
 876                 goto out;
 877
 878         /* TODO vq affinity */
 879
 880         ret = virtio_fs_setup_dax(vdev, fs);
 881         if (ret < 0)
 882                 goto out_vqs;
 883
 884         /* Bring the device online in case the filesystem is mounted and
 885          * requests need to be sent before we return.
 886          */
 887         virtio_device_ready(vdev);
 888
 889         ret = virtio_fs_add_instance(fs);
 890         if (ret < 0)
 891                 goto out_vqs;
 892
 893         return 0;
 894
 895 out_vqs:
 896         virtio_reset_device(vdev);
 897         virtio_fs_cleanup_vqs(vdev);
 898         kfree(fs->vqs);
 899
 900 out:
 901         vdev->priv = NULL;
 902         kfree(fs);
 903         return ret;
 904 }
 905
 906 static void virtio_fs_stop_all_queues(struct virtio_fs *fs)
 907 {
 908         struct virtio_fs_vq *fsvq;
 909         int i;
 910
 911         for (i = 0; i < fs->nvqs; i++) {
 912                 fsvq = &fs->vqs[i];
 913                 spin_lock(&fsvq->lock);
 914                 fsvq->connected = false;
 915                 spin_unlock(&fsvq->lock);
 916         }
 917 }
 918
 919 static void virtio_fs_remove(struct virtio_device *vdev)
 920 {
 921         struct virtio_fs *fs = vdev->priv;
 922
 923         mutex_lock(&virtio_fs_mutex);
 924         /* This device is going away. No one should get new reference */
 925         list_del_init(&fs->list);
 926         virtio_fs_stop_all_queues(fs);
 927         virtio_fs_drain_all_queues_locked(fs);
 928         virtio_reset_device(vdev);
 929         virtio_fs_cleanup_vqs(vdev);
 930
 931         vdev->priv = NULL;
 932         /* Put device reference on virtio_fs object */
 933         virtio_fs_put(fs);
 934         mutex_unlock(&virtio_fs_mutex);
 935 }
 936
 937 #ifdef CONFIG_PM_SLEEP
 938 static int virtio_fs_freeze(struct virtio_device *vdev)
 939 {
 940         /* TODO need to save state here */
 941         pr_warn("virtio-fs: suspend/resume not yet supported\n");
 942         return -EOPNOTSUPP;
 943 }
 944
 945 static int virtio_fs_restore(struct virtio_device *vdev)
 946 {
 947          /* TODO need to restore state here */
 948         return 0;
 949 }
 950 #endif /* CONFIG_PM_SLEEP */
 951
 952 static const struct virtio_device_id id_table[] = {
 953         { VIRTIO_ID_FS, VIRTIO_DEV_ANY_ID },
 954         {},
 955 };
 956
 957 static const unsigned int feature_table[] = {};
 958
 959 static struct virtio_driver virtio_fs_driver = {
 960         .driver.name            = KBUILD_MODNAME,
 961         .driver.owner           = THIS_MODULE,
 962         .id_table               = id_table,
 963         .feature_table          = feature_table,
 964         .feature_table_size     = ARRAY_SIZE(feature_table),
 965         .probe                  = virtio_fs_probe,
 966         .remove                 = virtio_fs_remove,
 967 #ifdef CONFIG_PM_SLEEP
 968         .freeze                 = virtio_fs_freeze,
 969         .restore                = virtio_fs_restore,
 970 #endif
 971 };
 972
 973 static void virtio_fs_wake_forget_and_unlock(struct fuse_iqueue *fiq)
 974 __releases(fiq->lock)
 975 {
 976         struct fuse_forget_link *link;
 977         struct virtio_fs_forget *forget;
 978         struct virtio_fs_forget_req *req;
 979         struct virtio_fs *fs;
 980         struct virtio_fs_vq *fsvq;
 981         u64 unique;
 982
 983         link = fuse_dequeue_forget(fiq, 1, NULL);
 984         unique = fuse_get_unique(fiq);
 985
 986         fs = fiq->priv;
 987         fsvq = &fs->vqs[VQ_HIPRIO];
 988         spin_unlock(&fiq->lock);
 989
 990         /* Allocate a buffer for the request */
 991         forget = kmalloc(sizeof(*forget), GFP_NOFS | __GFP_NOFAIL);
 992         req = &forget->req;
 993
 994         req->ih = (struct fuse_in_header){
 995                 .opcode = FUSE_FORGET,
 996                 .nodeid = link->forget_one.nodeid,
 997                 .unique = unique,
 998                 .len = sizeof(*req),
 999         };
1000         req->arg = (struct fuse_forget_in){
1001                 .nlookup = link->forget_one.nlookup,
1002         };
1003
1004         send_forget_request(fsvq, forget, false);
1005         kfree(link);
1006 }
1007
1008 static void virtio_fs_wake_interrupt_and_unlock(struct fuse_iqueue *fiq)
1009 __releases(fiq->lock)
1010 {
1011         /*
1012          * TODO interrupts.
1013          *
1014          * Normal fs operations on a local filesystems aren't interruptible.
1015          * Exceptions are blocking lock operations; for example fcntl(F_SETLKW)
1016          * with shared lock between host and guest.
1017          */
1018         spin_unlock(&fiq->lock);
1019 }
1020
1021 /* Count number of scatter-gather elements required */
1022 static unsigned int sg_count_fuse_pages(struct fuse_page_desc *page_descs,
1023                                        unsigned int num_pages,
1024                                        unsigned int total_len)
1025 {
1026         unsigned int i;
1027         unsigned int this_len;
1028
1029         for (i = 0; i < num_pages && total_len; i++) {
1030                 this_len =  min(page_descs[i].length, total_len);
1031                 total_len -= this_len;
1032         }
1033
1034         return i;
1035 }
1036
1037 /* Return the number of scatter-gather list elements required */
1038 static unsigned int sg_count_fuse_req(struct fuse_req *req)
1039 {
1040         struct fuse_args *args = req->args;
1041         struct fuse_args_pages *ap = container_of(args, typeof(*ap), args);
1042         unsigned int size, total_sgs = 1 /* fuse_in_header */;
1043
1044         if (args->in_numargs - args->in_pages)
1045                 total_sgs += 1;
1046
1047         if (args->in_pages) {
1048                 size = args->in_args[args->in_numargs - 1].size;
1049                 total_sgs += sg_count_fuse_pages(ap->descs, ap->num_pages,
1050                                                  size);
1051         }
1052
1053         if (!test_bit(FR_ISREPLY, &req->flags))
1054                 return total_sgs;
1055
1056         total_sgs += 1 /* fuse_out_header */;
1057
1058         if (args->out_numargs - args->out_pages)
1059                 total_sgs += 1;
1060
1061         if (args->out_pages) {
1062                 size = args->out_args[args->out_numargs - 1].size;
1063                 total_sgs += sg_count_fuse_pages(ap->descs, ap->num_pages,
1064                                                  size);
1065         }
1066
1067         return total_sgs;
1068 }
1069
1070 /* Add pages to scatter-gather list and return number of elements used */
1071 static unsigned int sg_init_fuse_pages(struct scatterlist *sg,
1072                                        struct page **pages,
1073                                        struct fuse_page_desc *page_descs,
1074                                        unsigned int num_pages,
1075                                        unsigned int total_len)
1076 {
1077         unsigned int i;
1078         unsigned int this_len;
1079
1080         for (i = 0; i < num_pages && total_len; i++) {
1081                 sg_init_table(&sg[i], 1);
1082                 this_len =  min(page_descs[i].length, total_len);
1083                 sg_set_page(&sg[i], pages[i], this_len, page_descs[i].offset);
1084                 total_len -= this_len;
1085         }
1086
1087         return i;
1088 }
1089
1090 /* Add args to scatter-gather list and return number of elements used */
1091 static unsigned int sg_init_fuse_args(struct scatterlist *sg,
1092                                       struct fuse_req *req,
1093                                       struct fuse_arg *args,
1094                                       unsigned int numargs,
1095                                       bool argpages,
1096                                       void *argbuf,
1097                                       unsigned int *len_used)
1098 {
1099         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
1100         unsigned int total_sgs = 0;
1101         unsigned int len;
1102
1103         len = fuse_len_args(numargs - argpages, args);
1104         if (len)
1105                 sg_init_one(&sg[total_sgs++], argbuf, len);
1106
1107         if (argpages)
1108                 total_sgs += sg_init_fuse_pages(&sg[total_sgs],
1109                                                 ap->pages, ap->descs,
1110                                                 ap->num_pages,
1111                                                 args[numargs - 1].size);
1112
1113         if (len_used)
1114                 *len_used = len;
1115
1116         return total_sgs;
1117 }
1118
1119 /* Add a request to a virtqueue and kick the device */
1120 static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
1121                                  struct fuse_req *req, bool in_flight)
1122 {
1123         /* requests need at least 4 elements */
1124         struct scatterlist *stack_sgs[6];
1125         struct scatterlist stack_sg[ARRAY_SIZE(stack_sgs)];
1126         struct scatterlist **sgs = stack_sgs;
1127         struct scatterlist *sg = stack_sg;
1128         struct virtqueue *vq;
1129         struct fuse_args *args = req->args;
1130         unsigned int argbuf_used = 0;
1131         unsigned int out_sgs = 0;
1132         unsigned int in_sgs = 0;
1133         unsigned int total_sgs;
1134         unsigned int i;
1135         int ret;
1136         bool notify;
1137         struct fuse_pqueue *fpq;
1138
1139         /* Does the sglist fit on the stack? */
1140         total_sgs = sg_count_fuse_req(req);
1141         if (total_sgs > ARRAY_SIZE(stack_sgs)) {
1142                 sgs = kmalloc_array(total_sgs, sizeof(sgs[0]), GFP_ATOMIC);
1143                 sg = kmalloc_array(total_sgs, sizeof(sg[0]), GFP_ATOMIC);
1144                 if (!sgs || !sg) {
1145                         ret = -ENOMEM;
1146                         goto out;
1147                 }
1148         }
1149
1150         /* Use a bounce buffer since stack args cannot be mapped */
1151         ret = copy_args_to_argbuf(req);
1152         if (ret < 0)
1153                 goto out;
1154
1155         /* Request elements */
1156         sg_init_one(&sg[out_sgs++], &req->in.h, sizeof(req->in.h));
1157         out_sgs += sg_init_fuse_args(&sg[out_sgs], req,
1158                                      (struct fuse_arg *)args->in_args,
1159                                      args->in_numargs, args->in_pages,
1160                                      req->argbuf, &argbuf_used);
1161
1162         /* Reply elements */
1163         if (test_bit(FR_ISREPLY, &req->flags)) {
1164                 sg_init_one(&sg[out_sgs + in_sgs++],
1165                             &req->out.h, sizeof(req->out.h));
1166                 in_sgs += sg_init_fuse_args(&sg[out_sgs + in_sgs], req,
1167                                             args->out_args, args->out_numargs,
1168                                             args->out_pages,
1169                                             req->argbuf + argbuf_used, NULL);
1170         }
1171
1172         WARN_ON(out_sgs + in_sgs != total_sgs);
1173
1174         for (i = 0; i < total_sgs; i++)
1175                 sgs[i] = &sg[i];
1176
1177         spin_lock(&fsvq->lock);
1178
1179         if (!fsvq->connected) {
1180                 spin_unlock(&fsvq->lock);
1181                 ret = -ENOTCONN;
1182                 goto out;
1183         }
1184
1185         vq = fsvq->vq;
1186         ret = virtqueue_add_sgs(vq, sgs, out_sgs, in_sgs, req, GFP_ATOMIC);
1187         if (ret < 0) {
1188                 spin_unlock(&fsvq->lock);
1189                 goto out;
1190         }
1191
1192         /* Request successfully sent. */
1193         fpq = &fsvq->fud->pq;
1194         spin_lock(&fpq->lock);
1195         list_add_tail(&req->list, fpq->processing);
1196         spin_unlock(&fpq->lock);
1197         set_bit(FR_SENT, &req->flags);
1198         /* matches barrier in request_wait_answer() */
1199         smp_mb__after_atomic();
1200
1201         if (!in_flight)
1202                 inc_in_flight_req(fsvq);
1203         notify = virtqueue_kick_prepare(vq);
1204
1205         spin_unlock(&fsvq->lock);
1206
1207         if (notify)
1208                 virtqueue_notify(vq);
1209
1210 out:
1211         if (ret < 0 && req->argbuf) {
1212                 kfree(req->argbuf);
1213                 req->argbuf = NULL;
1214         }
1215         if (sgs != stack_sgs) {
1216                 kfree(sgs);
1217                 kfree(sg);
1218         }
1219
1220         return ret;
1221 }
1222
1223 static void virtio_fs_wake_pending_and_unlock(struct fuse_iqueue *fiq)
1224 __releases(fiq->lock)
1225 {
1226         unsigned int queue_id = VQ_REQUEST; /* TODO multiqueue */
1227         struct virtio_fs *fs;
1228         struct fuse_req *req;
1229         struct virtio_fs_vq *fsvq;
1230         int ret;
1231
1232         WARN_ON(list_empty(&fiq->pending));
1233         req = list_last_entry(&fiq->pending, struct fuse_req, list);
1234         clear_bit(FR_PENDING, &req->flags);
1235         list_del_init(&req->list);
1236         WARN_ON(!list_empty(&fiq->pending));
1237         spin_unlock(&fiq->lock);
1238
1239         fs = fiq->priv;
1240
1241         pr_debug("%s: opcode %u unique %#llx nodeid %#llx in.len %u out.len %u\n",
1242                   __func__, req->in.h.opcode, req->in.h.unique,
1243                  req->in.h.nodeid, req->in.h.len,
1244                  fuse_len_args(req->args->out_numargs, req->args->out_args));
1245
1246         fsvq = &fs->vqs[queue_id];
1247         ret = virtio_fs_enqueue_req(fsvq, req, false);
1248         if (ret < 0) {
1249                 if (ret == -ENOMEM || ret == -ENOSPC) {
1250                         /*
1251                          * Virtqueue full. Retry submission from worker
1252                          * context as we might be holding fc->bg_lock.
1253                          */
1254                         spin_lock(&fsvq->lock);
1255                         list_add_tail(&req->list, &fsvq->queued_reqs);
1256                         inc_in_flight_req(fsvq);
1257                         schedule_delayed_work(&fsvq->dispatch_work,
1258                                                 msecs_to_jiffies(1));
1259                         spin_unlock(&fsvq->lock);
1260                         return;
1261                 }
1262                 req->out.h.error = ret;
1263                 pr_err("virtio-fs: virtio_fs_enqueue_req() failed %d\n", ret);
1264
1265                 /* Can't end request in submission context. Use a worker */
1266                 spin_lock(&fsvq->lock);
1267                 list_add_tail(&req->list, &fsvq->end_reqs);
1268                 schedule_delayed_work(&fsvq->dispatch_work, 0);
1269                 spin_unlock(&fsvq->lock);
1270                 return;
1271         }
1272 }
1273
1274 static const struct fuse_iqueue_ops virtio_fs_fiq_ops = {
1275         .wake_forget_and_unlock         = virtio_fs_wake_forget_and_unlock,
1276         .wake_interrupt_and_unlock      = virtio_fs_wake_interrupt_and_unlock,
1277         .wake_pending_and_unlock        = virtio_fs_wake_pending_and_unlock,
1278         .release                        = virtio_fs_fiq_release,
1279 };
1280
1281 static inline void virtio_fs_ctx_set_defaults(struct fuse_fs_context *ctx)
1282 {
1283         ctx->rootmode = S_IFDIR;
1284         ctx->default_permissions = 1;
1285         ctx->allow_other = 1;
1286         ctx->max_read = UINT_MAX;
1287         ctx->blksize = 512;
1288         ctx->destroy = true;
1289         ctx->no_control = true;
1290         ctx->no_force_umount = true;
1291 }
1292
1293 static int virtio_fs_fill_super(struct super_block *sb, struct fs_context *fsc)
1294 {
1295         struct fuse_mount *fm = get_fuse_mount_super(sb);
1296         struct fuse_conn *fc = fm->fc;
1297         struct virtio_fs *fs = fc->iq.priv;
1298         struct fuse_fs_context *ctx = fsc->fs_private;
1299         unsigned int i;
1300         int err;
1301
1302         virtio_fs_ctx_set_defaults(ctx);
1303         mutex_lock(&virtio_fs_mutex);
1304
1305         /* After holding mutex, make sure virtiofs device is still there.
1306          * Though we are holding a reference to it, drive ->remove might
1307          * still have cleaned up virtual queues. In that case bail out.
1308          */
1309         err = -EINVAL;
1310         if (list_empty(&fs->list)) {
1311                 pr_info("virtio-fs: tag <%s> not found\n", fs->tag);
1312                 goto err;
1313         }
1314
1315         err = -ENOMEM;
1316         /* Allocate fuse_dev for hiprio and notification queues */
1317         for (i = 0; i < fs->nvqs; i++) {
1318                 struct virtio_fs_vq *fsvq = &fs->vqs[i];
1319
1320                 fsvq->fud = fuse_dev_alloc();
1321                 if (!fsvq->fud)
1322                         goto err_free_fuse_devs;
1323         }
1324
1325         /* virtiofs allocates and installs its own fuse devices */
1326         ctx->fudptr = NULL;
1327         if (ctx->dax_mode != FUSE_DAX_NEVER) {
1328                 if (ctx->dax_mode == FUSE_DAX_ALWAYS && !fs->dax_dev) {
1329                         err = -EINVAL;
1330                         pr_err("virtio-fs: dax can't be enabled as filesystem"
1331                                " device does not support it.\n");
1332                         goto err_free_fuse_devs;
1333                 }
1334                 ctx->dax_dev = fs->dax_dev;
1335         }
1336         err = fuse_fill_super_common(sb, ctx);
1337         if (err < 0)
1338                 goto err_free_fuse_devs;
1339
1340         for (i = 0; i < fs->nvqs; i++) {
1341                 struct virtio_fs_vq *fsvq = &fs->vqs[i];
1342
1343                 fuse_dev_install(fsvq->fud, fc);
1344         }
1345
1346         /* Previous unmount will stop all queues. Start these again */
1347         virtio_fs_start_all_queues(fs);
1348         fuse_send_init(fm);
1349         mutex_unlock(&virtio_fs_mutex);
1350         return 0;
1351
1352 err_free_fuse_devs:
1353         virtio_fs_free_devs(fs);
1354 err:
1355         mutex_unlock(&virtio_fs_mutex);
1356         return err;
1357 }
1358
1359 static void virtio_fs_conn_destroy(struct fuse_mount *fm)
1360 {
1361         struct fuse_conn *fc = fm->fc;
1362         struct virtio_fs *vfs = fc->iq.priv;
1363         struct virtio_fs_vq *fsvq = &vfs->vqs[VQ_HIPRIO];
1364
1365         /* Stop dax worker. Soon evict_inodes() will be called which
1366          * will free all memory ranges belonging to all inodes.
1367          */
1368         if (IS_ENABLED(CONFIG_FUSE_DAX))
1369                 fuse_dax_cancel_work(fc);
1370
1371         /* Stop forget queue. Soon destroy will be sent */
1372         spin_lock(&fsvq->lock);
1373         fsvq->connected = false;
1374         spin_unlock(&fsvq->lock);
1375         virtio_fs_drain_all_queues(vfs);
1376
1377         fuse_conn_destroy(fm);
1378
1379         /* fuse_conn_destroy() must have sent destroy. Stop all queues
1380          * and drain one more time and free fuse devices. Freeing fuse
1381          * devices will drop their reference on fuse_conn and that in
1382          * turn will drop its reference on virtio_fs object.
1383          */
1384         virtio_fs_stop_all_queues(vfs);
1385         virtio_fs_drain_all_queues(vfs);
1386         virtio_fs_free_devs(vfs);
1387 }
1388
1389 static void virtio_kill_sb(struct super_block *sb)
1390 {
1391         struct fuse_mount *fm = get_fuse_mount_super(sb);
1392         bool last;
1393
1394         /* If mount failed, we can still be called without any fc */
1395         if (sb->s_root) {
1396                 last = fuse_mount_remove(fm);
1397                 if (last)
1398                         virtio_fs_conn_destroy(fm);
1399         }
1400         kill_anon_super(sb);
1401         fuse_mount_destroy(fm);
1402 }
1403
1404 static int virtio_fs_test_super(struct super_block *sb,
1405                                 struct fs_context *fsc)
1406 {
1407         struct fuse_mount *fsc_fm = fsc->s_fs_info;
1408         struct fuse_mount *sb_fm = get_fuse_mount_super(sb);
1409
1410         return fsc_fm->fc->iq.priv == sb_fm->fc->iq.priv;
1411 }
1412
1413 static int virtio_fs_get_tree(struct fs_context *fsc)
1414 {
1415         struct virtio_fs *fs;
1416         struct super_block *sb;
1417         struct fuse_conn *fc = NULL;
1418         struct fuse_mount *fm;
1419         unsigned int virtqueue_size;
1420         int err = -EIO;
1421
1422         /* This gets a reference on virtio_fs object. This ptr gets installed
1423          * in fc->iq->priv. Once fuse_conn is going away, it calls ->put()
1424          * to drop the reference to this object.
1425          */
1426         fs = virtio_fs_find_instance(fsc->source);
1427         if (!fs) {
1428                 pr_info("virtio-fs: tag <%s> not found\n", fsc->source);
1429                 return -EINVAL;
1430         }
1431
1432         virtqueue_size = virtqueue_get_vring_size(fs->vqs[VQ_REQUEST].vq);
1433         if (WARN_ON(virtqueue_size <= FUSE_HEADER_OVERHEAD))
1434                 goto out_err;
1435
1436         err = -ENOMEM;
1437         fc = kzalloc(sizeof(struct fuse_conn), GFP_KERNEL);
1438         if (!fc)
1439                 goto out_err;
1440
1441         fm = kzalloc(sizeof(struct fuse_mount), GFP_KERNEL);
1442         if (!fm)
1443                 goto out_err;
1444
1445         fuse_conn_init(fc, fm, fsc->user_ns, &virtio_fs_fiq_ops, fs);
1446         fc->release = fuse_free_conn;
1447         fc->delete_stale = true;
1448         fc->auto_submounts = true;
1449         fc->sync_fs = true;
1450
1451         /* Tell FUSE to split requests that exceed the virtqueue's size */
1452         fc->max_pages_limit = min_t(unsigned int, fc->max_pages_limit,
1453                                     virtqueue_size - FUSE_HEADER_OVERHEAD);
1454
1455         fsc->s_fs_info = fm;
1456         sb = sget_fc(fsc, virtio_fs_test_super, set_anon_super_fc);
1457         if (fsc->s_fs_info)
1458                 fuse_mount_destroy(fm);
1459         if (IS_ERR(sb))
1460                 return PTR_ERR(sb);
1461
1462         if (!sb->s_root) {
1463                 err = virtio_fs_fill_super(sb, fsc);
1464                 if (err) {
1465                         deactivate_locked_super(sb);
1466                         return err;
1467                 }
1468
1469                 sb->s_flags |= SB_ACTIVE;
1470         }
1471
1472         WARN_ON(fsc->root);
1473         fsc->root = dget(sb->s_root);
1474         return 0;
1475
1476 out_err:
1477         kfree(fc);
1478         mutex_lock(&virtio_fs_mutex);
1479         virtio_fs_put(fs);
1480         mutex_unlock(&virtio_fs_mutex);
1481         return err;
1482 }
1483
1484 static const struct fs_context_operations virtio_fs_context_ops = {
1485         .free           = virtio_fs_free_fsc,
1486         .parse_param    = virtio_fs_parse_param,
1487         .get_tree       = virtio_fs_get_tree,
1488 };
1489
1490 static int virtio_fs_init_fs_context(struct fs_context *fsc)
1491 {
1492         struct fuse_fs_context *ctx;
1493
1494         if (fsc->purpose == FS_CONTEXT_FOR_SUBMOUNT)
1495                 return fuse_init_fs_context_submount(fsc);
1496
1497         ctx = kzalloc(sizeof(struct fuse_fs_context), GFP_KERNEL);
1498         if (!ctx)
1499                 return -ENOMEM;
1500         fsc->fs_private = ctx;
1501         fsc->ops = &virtio_fs_context_ops;
1502         return 0;
1503 }
1504
1505 static struct file_system_type virtio_fs_type = {
1506         .owner          = THIS_MODULE,
1507         .name           = "virtiofs",
1508         .init_fs_context = virtio_fs_init_fs_context,
1509         .kill_sb        = virtio_kill_sb,
1510 };
1511
1512 static int __init virtio_fs_init(void)
1513 {
1514         int ret;
1515
1516         ret = register_virtio_driver(&virtio_fs_driver);
1517         if (ret < 0)
1518                 return ret;
1519
1520         ret = register_filesystem(&virtio_fs_type);
1521         if (ret < 0) {
1522                 unregister_virtio_driver(&virtio_fs_driver);
1523                 return ret;
1524         }
1525
1526         return 0;
1527 }
1528 module_init(virtio_fs_init);
1529
1530 static void __exit virtio_fs_exit(void)
1531 {
1532         unregister_filesystem(&virtio_fs_type);
1533         unregister_virtio_driver(&virtio_fs_driver);
1534 }
1535 module_exit(virtio_fs_exit);
1536
1537 MODULE_AUTHOR("Stefan Hajnoczi <stefanha@redhat.com>");
1538 MODULE_DESCRIPTION("Virtio Filesystem");
1539 MODULE_LICENSE("GPL");
1540 MODULE_ALIAS_FS(KBUILD_MODNAME);
1541 MODULE_DEVICE_TABLE(virtio, id_table);