| 1 |
// SPDX-License-Identifier: GPL-2.0 |
| 2 |
/* |
| 3 |
* FUSE: Filesystem in Userspace |
| 4 |
* Copyright (c) 2023-2024 DataDirect Networks. |
| 5 |
*/ |
| 6 |
|
| 7 |
#include "fuse_i.h" |
| 8 |
#include "dev_uring_i.h" |
| 9 |
#include "fuse_dev_i.h" |
| 10 |
#include "fuse_trace.h" |
| 11 |
|
| 12 |
#include <linux/fs.h> |
| 13 |
#include <linux/io_uring/cmd.h> |
| 14 |
|
| 15 |
static bool __read_mostly enable_uring; |
| 16 |
module_param(enable_uring, bool, 0644); |
| 17 |
MODULE_PARM_DESC(enable_uring, |
| 18 |
"Enable userspace communication through io-uring"); |
| 19 |
|
| 20 |
#define FUSE_URING_IOV_SEGS 2 /* header and payload */ |
| 21 |
|
| 22 |
|
| 23 |
bool fuse_uring_enabled(void) |
| 24 |
{ |
| 25 |
return enable_uring; |
| 26 |
} |
| 27 |
|
| 28 |
struct fuse_uring_pdu { |
| 29 |
struct fuse_ring_ent *ent; |
| 30 |
}; |
| 31 |
|
| 32 |
static const struct fuse_iqueue_ops fuse_io_uring_ops; |
| 33 |
|
| 34 |
static void uring_cmd_set_ring_ent(struct io_uring_cmd *cmd, |
| 35 |
struct fuse_ring_ent *ring_ent) |
| 36 |
{ |
| 37 |
struct fuse_uring_pdu *pdu = |
| 38 |
io_uring_cmd_to_pdu(cmd, struct fuse_uring_pdu); |
| 39 |
|
| 40 |
pdu->ent = ring_ent; |
| 41 |
} |
| 42 |
|
| 43 |
static struct fuse_ring_ent *uring_cmd_to_ring_ent(struct io_uring_cmd *cmd) |
| 44 |
{ |
| 45 |
struct fuse_uring_pdu *pdu = |
| 46 |
io_uring_cmd_to_pdu(cmd, struct fuse_uring_pdu); |
| 47 |
|
| 48 |
return pdu->ent; |
| 49 |
} |
| 50 |
|
| 51 |
static void fuse_uring_flush_bg(struct fuse_ring_queue *queue) |
| 52 |
{ |
| 53 |
struct fuse_ring *ring = queue->ring; |
| 54 |
struct fuse_conn *fc = ring->fc; |
| 55 |
|
| 56 |
lockdep_assert_held(&queue->lock); |
| 57 |
lockdep_assert_held(&fc->bg_lock); |
| 58 |
|
| 59 |
/* |
| 60 |
* Allow one bg request per queue, ignoring global fc limits. |
| 61 |
* This prevents a single queue from consuming all resources and |
| 62 |
* eliminates the need for remote queue wake-ups when global |
| 63 |
* limits are met but this queue has no more waiting requests. |
| 64 |
*/ |
| 65 |
while ((fc->active_background < fc->max_background || |
| 66 |
!queue->active_background) && |
| 67 |
(!list_empty(&queue->fuse_req_bg_queue))) { |
| 68 |
struct fuse_req *req; |
| 69 |
|
| 70 |
req = list_first_entry(&queue->fuse_req_bg_queue, |
| 71 |
struct fuse_req, list); |
| 72 |
fc->active_background++; |
| 73 |
queue->active_background++; |
| 74 |
|
| 75 |
list_move_tail(&req->list, &queue->fuse_req_queue); |
| 76 |
} |
| 77 |
} |
| 78 |
|
| 79 |
static void fuse_uring_req_end(struct fuse_ring_ent *ent, struct fuse_req *req, |
| 80 |
int error) |
| 81 |
{ |
| 82 |
struct fuse_ring_queue *queue = ent->queue; |
| 83 |
struct fuse_ring *ring = queue->ring; |
| 84 |
struct fuse_conn *fc = ring->fc; |
| 85 |
|
| 86 |
lockdep_assert_not_held(&queue->lock); |
| 87 |
spin_lock(&queue->lock); |
| 88 |
ent->fuse_req = NULL; |
| 89 |
list_del_init(&req->list); |
| 90 |
if (test_bit(FR_BACKGROUND, &req->flags)) { |
| 91 |
queue->active_background--; |
| 92 |
spin_lock(&fc->bg_lock); |
| 93 |
fuse_uring_flush_bg(queue); |
| 94 |
spin_unlock(&fc->bg_lock); |
| 95 |
} |
| 96 |
|
| 97 |
spin_unlock(&queue->lock); |
| 98 |
|
| 99 |
if (error) |
| 100 |
req->out.h.error = error; |
| 101 |
|
| 102 |
clear_bit(FR_SENT, &req->flags); |
| 103 |
fuse_request_end(req); |
| 104 |
} |
| 105 |
|
| 106 |
/* Abort all list queued request on the given ring queue */ |
| 107 |
static void fuse_uring_abort_end_queue_requests(struct fuse_ring_queue *queue) |
| 108 |
{ |
| 109 |
struct fuse_req *req; |
| 110 |
LIST_HEAD(req_list); |
| 111 |
|
| 112 |
spin_lock(&queue->lock); |
| 113 |
list_for_each_entry(req, &queue->fuse_req_queue, list) |
| 114 |
clear_bit(FR_PENDING, &req->flags); |
| 115 |
list_splice_init(&queue->fuse_req_queue, &req_list); |
| 116 |
spin_unlock(&queue->lock); |
| 117 |
|
| 118 |
/* must not hold queue lock to avoid order issues with fi->lock */ |
| 119 |
fuse_dev_end_requests(&req_list); |
| 120 |
} |
| 121 |
|
| 122 |
void fuse_uring_abort_end_requests(struct fuse_ring *ring) |
| 123 |
{ |
| 124 |
int qid; |
| 125 |
struct fuse_ring_queue *queue; |
| 126 |
struct fuse_conn *fc = ring->fc; |
| 127 |
|
| 128 |
for (qid = 0; qid < ring->nr_queues; qid++) { |
| 129 |
queue = READ_ONCE(ring->queues[qid]); |
| 130 |
if (!queue) |
| 131 |
continue; |
| 132 |
|
| 133 |
queue->stopped = true; |
| 134 |
|
| 135 |
WARN_ON_ONCE(ring->fc->max_background != UINT_MAX); |
| 136 |
spin_lock(&queue->lock); |
| 137 |
spin_lock(&fc->bg_lock); |
| 138 |
fuse_uring_flush_bg(queue); |
| 139 |
spin_unlock(&fc->bg_lock); |
| 140 |
spin_unlock(&queue->lock); |
| 141 |
fuse_uring_abort_end_queue_requests(queue); |
| 142 |
} |
| 143 |
} |
| 144 |
|
| 145 |
static bool ent_list_request_expired(struct fuse_conn *fc, struct list_head *list) |
| 146 |
{ |
| 147 |
struct fuse_ring_ent *ent; |
| 148 |
struct fuse_req *req; |
| 149 |
|
| 150 |
ent = list_first_entry_or_null(list, struct fuse_ring_ent, list); |
| 151 |
if (!ent) |
| 152 |
return false; |
| 153 |
|
| 154 |
req = ent->fuse_req; |
| 155 |
|
| 156 |
return time_is_before_jiffies(req->create_time + |
| 157 |
fc->timeout.req_timeout); |
| 158 |
} |
| 159 |
|
| 160 |
bool fuse_uring_request_expired(struct fuse_conn *fc) |
| 161 |
{ |
| 162 |
struct fuse_ring *ring = fc->ring; |
| 163 |
struct fuse_ring_queue *queue; |
| 164 |
int qid; |
| 165 |
|
| 166 |
if (!ring) |
| 167 |
return false; |
| 168 |
|
| 169 |
for (qid = 0; qid < ring->nr_queues; qid++) { |
| 170 |
queue = READ_ONCE(ring->queues[qid]); |
| 171 |
if (!queue) |
| 172 |
continue; |
| 173 |
|
| 174 |
spin_lock(&queue->lock); |
| 175 |
if (fuse_request_expired(fc, &queue->fuse_req_queue) || |
| 176 |
fuse_request_expired(fc, &queue->fuse_req_bg_queue) || |
| 177 |
ent_list_request_expired(fc, &queue->ent_w_req_queue) || |
| 178 |
ent_list_request_expired(fc, &queue->ent_in_userspace)) { |
| 179 |
spin_unlock(&queue->lock); |
| 180 |
return true; |
| 181 |
} |
| 182 |
spin_unlock(&queue->lock); |
| 183 |
} |
| 184 |
|
| 185 |
return false; |
| 186 |
} |
| 187 |
|
| 188 |
void fuse_uring_destruct(struct fuse_conn *fc) |
| 189 |
{ |
| 190 |
struct fuse_ring *ring = fc->ring; |
| 191 |
int qid; |
| 192 |
|
| 193 |
if (!ring) |
| 194 |
return; |
| 195 |
|
| 196 |
for (qid = 0; qid < ring->nr_queues; qid++) { |
| 197 |
struct fuse_ring_queue *queue = ring->queues[qid]; |
| 198 |
struct fuse_ring_ent *ent, *next; |
| 199 |
|
| 200 |
if (!queue) |
| 201 |
continue; |
| 202 |
|
| 203 |
WARN_ON(!list_empty(&queue->ent_avail_queue)); |
| 204 |
WARN_ON(!list_empty(&queue->ent_w_req_queue)); |
| 205 |
WARN_ON(!list_empty(&queue->ent_commit_queue)); |
| 206 |
WARN_ON(!list_empty(&queue->ent_in_userspace)); |
| 207 |
|
| 208 |
list_for_each_entry_safe(ent, next, &queue->ent_released, |
| 209 |
list) { |
| 210 |
list_del_init(&ent->list); |
| 211 |
kfree(ent); |
| 212 |
} |
| 213 |
|
| 214 |
kfree(queue->fpq.processing); |
| 215 |
kfree(queue); |
| 216 |
ring->queues[qid] = NULL; |
| 217 |
} |
| 218 |
|
| 219 |
kfree(ring->queues); |
| 220 |
kfree(ring); |
| 221 |
fc->ring = NULL; |
| 222 |
} |
| 223 |
|
| 224 |
/* |
| 225 |
* Basic ring setup for this connection based on the provided configuration |
| 226 |
*/ |
| 227 |
static struct fuse_ring *fuse_uring_create(struct fuse_conn *fc) |
| 228 |
{ |
| 229 |
struct fuse_ring *ring; |
| 230 |
size_t nr_queues = num_possible_cpus(); |
| 231 |
struct fuse_ring *res = NULL; |
| 232 |
size_t max_payload_size; |
| 233 |
|
| 234 |
ring = kzalloc_obj(*fc->ring, GFP_KERNEL_ACCOUNT); |
| 235 |
if (!ring) |
| 236 |
return NULL; |
| 237 |
|
| 238 |
ring->queues = kzalloc_objs(struct fuse_ring_queue *, nr_queues, |
| 239 |
GFP_KERNEL_ACCOUNT); |
| 240 |
if (!ring->queues) |
| 241 |
goto out_err; |
| 242 |
|
| 243 |
max_payload_size = max(FUSE_MIN_READ_BUFFER, fc->max_write); |
| 244 |
max_payload_size = max(max_payload_size, fc->max_pages * PAGE_SIZE); |
| 245 |
|
| 246 |
spin_lock(&fc->lock); |
| 247 |
if (fc->ring) { |
| 248 |
/* race, another thread created the ring in the meantime */ |
| 249 |
spin_unlock(&fc->lock); |
| 250 |
res = fc->ring; |
| 251 |
goto out_err; |
| 252 |
} |
| 253 |
|
| 254 |
init_waitqueue_head(&ring->stop_waitq); |
| 255 |
|
| 256 |
ring->nr_queues = nr_queues; |
| 257 |
ring->fc = fc; |
| 258 |
ring->max_payload_sz = max_payload_size; |
| 259 |
smp_store_release(&fc->ring, ring); |
| 260 |
|
| 261 |
spin_unlock(&fc->lock); |
| 262 |
return ring; |
| 263 |
|
| 264 |
out_err: |
| 265 |
kfree(ring->queues); |
| 266 |
kfree(ring); |
| 267 |
return res; |
| 268 |
} |
| 269 |
|
| 270 |
static struct fuse_ring_queue *fuse_uring_create_queue(struct fuse_ring *ring, |
| 271 |
int qid) |
| 272 |
{ |
| 273 |
struct fuse_conn *fc = ring->fc; |
| 274 |
struct fuse_ring_queue *queue; |
| 275 |
struct list_head *pq; |
| 276 |
|
| 277 |
queue = kzalloc_obj(*queue, GFP_KERNEL_ACCOUNT); |
| 278 |
if (!queue) |
| 279 |
return NULL; |
| 280 |
pq = kzalloc_objs(struct list_head, FUSE_PQ_HASH_SIZE); |
| 281 |
if (!pq) { |
| 282 |
kfree(queue); |
| 283 |
return NULL; |
| 284 |
} |
| 285 |
|
| 286 |
queue->qid = qid; |
| 287 |
queue->ring = ring; |
| 288 |
spin_lock_init(&queue->lock); |
| 289 |
|
| 290 |
INIT_LIST_HEAD(&queue->ent_avail_queue); |
| 291 |
INIT_LIST_HEAD(&queue->ent_commit_queue); |
| 292 |
INIT_LIST_HEAD(&queue->ent_w_req_queue); |
| 293 |
INIT_LIST_HEAD(&queue->ent_in_userspace); |
| 294 |
INIT_LIST_HEAD(&queue->fuse_req_queue); |
| 295 |
INIT_LIST_HEAD(&queue->fuse_req_bg_queue); |
| 296 |
INIT_LIST_HEAD(&queue->ent_released); |
| 297 |
|
| 298 |
queue->fpq.processing = pq; |
| 299 |
fuse_pqueue_init(&queue->fpq); |
| 300 |
|
| 301 |
spin_lock(&fc->lock); |
| 302 |
if (ring->queues[qid]) { |
| 303 |
spin_unlock(&fc->lock); |
| 304 |
kfree(queue->fpq.processing); |
| 305 |
kfree(queue); |
| 306 |
return ring->queues[qid]; |
| 307 |
} |
| 308 |
|
| 309 |
/* |
| 310 |
* write_once and lock as the caller mostly doesn't take the lock at all |
| 311 |
*/ |
| 312 |
WRITE_ONCE(ring->queues[qid], queue); |
| 313 |
spin_unlock(&fc->lock); |
| 314 |
|
| 315 |
return queue; |
| 316 |
} |
| 317 |
|
| 318 |
static void fuse_uring_stop_fuse_req_end(struct fuse_req *req) |
| 319 |
{ |
| 320 |
clear_bit(FR_SENT, &req->flags); |
| 321 |
req->out.h.error = -ECONNABORTED; |
| 322 |
fuse_request_end(req); |
| 323 |
} |
| 324 |
|
| 325 |
/* |
| 326 |
* Release a request/entry on connection tear down |
| 327 |
*/ |
| 328 |
static void fuse_uring_entry_teardown(struct fuse_ring_ent *ent) |
| 329 |
{ |
| 330 |
struct fuse_req *req; |
| 331 |
struct io_uring_cmd *cmd; |
| 332 |
|
| 333 |
struct fuse_ring_queue *queue = ent->queue; |
| 334 |
|
| 335 |
spin_lock(&queue->lock); |
| 336 |
cmd = ent->cmd; |
| 337 |
ent->cmd = NULL; |
| 338 |
req = ent->fuse_req; |
| 339 |
ent->fuse_req = NULL; |
| 340 |
if (req) { |
| 341 |
/* remove entry from queue->fpq->processing */ |
| 342 |
list_del_init(&req->list); |
| 343 |
} |
| 344 |
|
| 345 |
/* |
| 346 |
* The entry must not be freed immediately, due to access of direct |
| 347 |
* pointer access of entries through IO_URING_F_CANCEL - there is a risk |
| 348 |
* of race between daemon termination (which triggers IO_URING_F_CANCEL |
| 349 |
* and accesses entries without checking the list state first |
| 350 |
*/ |
| 351 |
list_move(&ent->list, &queue->ent_released); |
| 352 |
ent->state = FRRS_RELEASED; |
| 353 |
spin_unlock(&queue->lock); |
| 354 |
|
| 355 |
if (cmd) |
| 356 |
io_uring_cmd_done(cmd, -ENOTCONN, IO_URING_F_UNLOCKED); |
| 357 |
|
| 358 |
if (req) |
| 359 |
fuse_uring_stop_fuse_req_end(req); |
| 360 |
} |
| 361 |
|
| 362 |
static void fuse_uring_stop_list_entries(struct list_head *head, |
| 363 |
struct fuse_ring_queue *queue, |
| 364 |
enum fuse_ring_req_state exp_state) |
| 365 |
{ |
| 366 |
struct fuse_ring *ring = queue->ring; |
| 367 |
struct fuse_ring_ent *ent, *next; |
| 368 |
ssize_t queue_refs = SSIZE_MAX; |
| 369 |
LIST_HEAD(to_teardown); |
| 370 |
|
| 371 |
spin_lock(&queue->lock); |
| 372 |
list_for_each_entry_safe(ent, next, head, list) { |
| 373 |
if (ent->state != exp_state) { |
| 374 |
pr_warn("entry teardown qid=%d state=%d expected=%d", |
| 375 |
queue->qid, ent->state, exp_state); |
| 376 |
continue; |
| 377 |
} |
| 378 |
|
| 379 |
ent->state = FRRS_TEARDOWN; |
| 380 |
list_move(&ent->list, &to_teardown); |
| 381 |
} |
| 382 |
spin_unlock(&queue->lock); |
| 383 |
|
| 384 |
/* no queue lock to avoid lock order issues */ |
| 385 |
list_for_each_entry_safe(ent, next, &to_teardown, list) { |
| 386 |
fuse_uring_entry_teardown(ent); |
| 387 |
queue_refs = atomic_dec_return(&ring->queue_refs); |
| 388 |
WARN_ON_ONCE(queue_refs < 0); |
| 389 |
} |
| 390 |
} |
| 391 |
|
| 392 |
static void fuse_uring_teardown_entries(struct fuse_ring_queue *queue) |
| 393 |
{ |
| 394 |
fuse_uring_stop_list_entries(&queue->ent_in_userspace, queue, |
| 395 |
FRRS_USERSPACE); |
| 396 |
fuse_uring_stop_list_entries(&queue->ent_avail_queue, queue, |
| 397 |
FRRS_AVAILABLE); |
| 398 |
} |
| 399 |
|
| 400 |
/* |
| 401 |
* Log state debug info |
| 402 |
*/ |
| 403 |
static void fuse_uring_log_ent_state(struct fuse_ring *ring) |
| 404 |
{ |
| 405 |
int qid; |
| 406 |
struct fuse_ring_ent *ent; |
| 407 |
|
| 408 |
for (qid = 0; qid < ring->nr_queues; qid++) { |
| 409 |
struct fuse_ring_queue *queue = ring->queues[qid]; |
| 410 |
|
| 411 |
if (!queue) |
| 412 |
continue; |
| 413 |
|
| 414 |
spin_lock(&queue->lock); |
| 415 |
/* |
| 416 |
* Log entries from the intermediate queue, the other queues |
| 417 |
* should be empty |
| 418 |
*/ |
| 419 |
list_for_each_entry(ent, &queue->ent_w_req_queue, list) { |
| 420 |
pr_info(" ent-req-queue ring=%p qid=%d ent=%p state=%d\n", |
| 421 |
ring, qid, ent, ent->state); |
| 422 |
} |
| 423 |
list_for_each_entry(ent, &queue->ent_commit_queue, list) { |
| 424 |
pr_info(" ent-commit-queue ring=%p qid=%d ent=%p state=%d\n", |
| 425 |
ring, qid, ent, ent->state); |
| 426 |
} |
| 427 |
spin_unlock(&queue->lock); |
| 428 |
} |
| 429 |
ring->stop_debug_log = 1; |
| 430 |
} |
| 431 |
|
| 432 |
static void fuse_uring_async_stop_queues(struct work_struct *work) |
| 433 |
{ |
| 434 |
int qid; |
| 435 |
struct fuse_ring *ring = |
| 436 |
container_of(work, struct fuse_ring, async_teardown_work.work); |
| 437 |
|
| 438 |
/* XXX code dup */ |
| 439 |
for (qid = 0; qid < ring->nr_queues; qid++) { |
| 440 |
struct fuse_ring_queue *queue = READ_ONCE(ring->queues[qid]); |
| 441 |
|
| 442 |
if (!queue) |
| 443 |
continue; |
| 444 |
|
| 445 |
fuse_uring_teardown_entries(queue); |
| 446 |
} |
| 447 |
|
| 448 |
/* |
| 449 |
* Some ring entries might be in the middle of IO operations, |
| 450 |
* i.e. in process to get handled by file_operations::uring_cmd |
| 451 |
* or on the way to userspace - we could handle that with conditions in |
| 452 |
* run time code, but easier/cleaner to have an async tear down handler |
| 453 |
* If there are still queue references left |
| 454 |
*/ |
| 455 |
if (atomic_read(&ring->queue_refs) > 0) { |
| 456 |
if (time_after(jiffies, |
| 457 |
ring->teardown_time + FUSE_URING_TEARDOWN_TIMEOUT)) |
| 458 |
fuse_uring_log_ent_state(ring); |
| 459 |
|
| 460 |
schedule_delayed_work(&ring->async_teardown_work, |
| 461 |
FUSE_URING_TEARDOWN_INTERVAL); |
| 462 |
} else { |
| 463 |
wake_up_all(&ring->stop_waitq); |
| 464 |
} |
| 465 |
} |
| 466 |
|
| 467 |
/* |
| 468 |
* Stop the ring queues |
| 469 |
*/ |
| 470 |
void fuse_uring_stop_queues(struct fuse_ring *ring) |
| 471 |
{ |
| 472 |
int qid; |
| 473 |
|
| 474 |
for (qid = 0; qid < ring->nr_queues; qid++) { |
| 475 |
struct fuse_ring_queue *queue = READ_ONCE(ring->queues[qid]); |
| 476 |
|
| 477 |
if (!queue) |
| 478 |
continue; |
| 479 |
|
| 480 |
fuse_uring_teardown_entries(queue); |
| 481 |
} |
| 482 |
|
| 483 |
if (atomic_read(&ring->queue_refs) > 0) { |
| 484 |
ring->teardown_time = jiffies; |
| 485 |
INIT_DELAYED_WORK(&ring->async_teardown_work, |
| 486 |
fuse_uring_async_stop_queues); |
| 487 |
schedule_delayed_work(&ring->async_teardown_work, |
| 488 |
FUSE_URING_TEARDOWN_INTERVAL); |
| 489 |
} else { |
| 490 |
wake_up_all(&ring->stop_waitq); |
| 491 |
} |
| 492 |
} |
| 493 |
|
| 494 |
/* |
| 495 |
* Handle IO_URING_F_CANCEL, typically should come on daemon termination. |
| 496 |
* |
| 497 |
* Releasing the last entry should trigger fuse_dev_release() if |
| 498 |
* the daemon was terminated |
| 499 |
*/ |
| 500 |
static void fuse_uring_cancel(struct io_uring_cmd *cmd, |
| 501 |
unsigned int issue_flags) |
| 502 |
{ |
| 503 |
struct fuse_ring_ent *ent = uring_cmd_to_ring_ent(cmd); |
| 504 |
struct fuse_ring_queue *queue; |
| 505 |
bool need_cmd_done = false; |
| 506 |
|
| 507 |
/* |
| 508 |
* direct access on ent - it must not be destructed as long as |
| 509 |
* IO_URING_F_CANCEL might come up |
| 510 |
*/ |
| 511 |
queue = ent->queue; |
| 512 |
spin_lock(&queue->lock); |
| 513 |
if (ent->state == FRRS_AVAILABLE) { |
| 514 |
ent->state = FRRS_USERSPACE; |
| 515 |
list_move_tail(&ent->list, &queue->ent_in_userspace); |
| 516 |
need_cmd_done = true; |
| 517 |
ent->cmd = NULL; |
| 518 |
} |
| 519 |
spin_unlock(&queue->lock); |
| 520 |
|
| 521 |
if (need_cmd_done) { |
| 522 |
/* no queue lock to avoid lock order issues */ |
| 523 |
io_uring_cmd_done(cmd, -ENOTCONN, issue_flags); |
| 524 |
} |
| 525 |
} |
| 526 |
|
| 527 |
static void fuse_uring_prepare_cancel(struct io_uring_cmd *cmd, int issue_flags, |
| 528 |
struct fuse_ring_ent *ring_ent) |
| 529 |
{ |
| 530 |
uring_cmd_set_ring_ent(cmd, ring_ent); |
| 531 |
io_uring_cmd_mark_cancelable(cmd, issue_flags); |
| 532 |
} |
| 533 |
|
| 534 |
/* |
| 535 |
* Checks for errors and stores it into the request |
| 536 |
*/ |
| 537 |
static int fuse_uring_out_header_has_err(struct fuse_out_header *oh, |
| 538 |
struct fuse_req *req, |
| 539 |
struct fuse_conn *fc) |
| 540 |
{ |
| 541 |
int err; |
| 542 |
|
| 543 |
err = -EINVAL; |
| 544 |
if (oh->unique == 0) { |
| 545 |
/* Not supported through io-uring yet */ |
| 546 |
pr_warn_once("notify through fuse-io-uring not supported\n"); |
| 547 |
goto err; |
| 548 |
} |
| 549 |
|
| 550 |
if (oh->error <= -ERESTARTSYS || oh->error > 0) |
| 551 |
goto err; |
| 552 |
|
| 553 |
if (oh->error) { |
| 554 |
err = oh->error; |
| 555 |
goto err; |
| 556 |
} |
| 557 |
|
| 558 |
err = -ENOENT; |
| 559 |
if ((oh->unique & ~FUSE_INT_REQ_BIT) != req->in.h.unique) { |
| 560 |
pr_warn_ratelimited("unique mismatch, expected: %llu got %llu\n", |
| 561 |
req->in.h.unique, |
| 562 |
oh->unique & ~FUSE_INT_REQ_BIT); |
| 563 |
goto err; |
| 564 |
} |
| 565 |
|
| 566 |
/* |
| 567 |
* Is it an interrupt reply ID? |
| 568 |
* XXX: Not supported through fuse-io-uring yet, it should not even |
| 569 |
* find the request - should not happen. |
| 570 |
*/ |
| 571 |
WARN_ON_ONCE(oh->unique & FUSE_INT_REQ_BIT); |
| 572 |
|
| 573 |
err = 0; |
| 574 |
err: |
| 575 |
return err; |
| 576 |
} |
| 577 |
|
| 578 |
static int fuse_uring_copy_from_ring(struct fuse_ring *ring, |
| 579 |
struct fuse_req *req, |
| 580 |
struct fuse_ring_ent *ent) |
| 581 |
{ |
| 582 |
struct fuse_copy_state cs; |
| 583 |
struct fuse_args *args = req->args; |
| 584 |
struct iov_iter iter; |
| 585 |
int err; |
| 586 |
struct fuse_uring_ent_in_out ring_in_out; |
| 587 |
|
| 588 |
err = copy_from_user(&ring_in_out, &ent->headers->ring_ent_in_out, |
| 589 |
sizeof(ring_in_out)); |
| 590 |
if (err) |
| 591 |
return -EFAULT; |
| 592 |
|
| 593 |
err = import_ubuf(ITER_SOURCE, ent->payload, ring->max_payload_sz, |
| 594 |
&iter); |
| 595 |
if (err) |
| 596 |
return err; |
| 597 |
|
| 598 |
fuse_copy_init(&cs, false, &iter); |
| 599 |
cs.is_uring = true; |
| 600 |
cs.req = req; |
| 601 |
|
| 602 |
err = fuse_copy_out_args(&cs, args, ring_in_out.payload_sz); |
| 603 |
fuse_copy_finish(&cs); |
| 604 |
return err; |
| 605 |
} |
| 606 |
|
| 607 |
/* |
| 608 |
* Copy data from the req to the ring buffer |
| 609 |
*/ |
| 610 |
static int fuse_uring_args_to_ring(struct fuse_ring *ring, struct fuse_req *req, |
| 611 |
struct fuse_ring_ent *ent) |
| 612 |
{ |
| 613 |
struct fuse_copy_state cs; |
| 614 |
struct fuse_args *args = req->args; |
| 615 |
struct fuse_in_arg *in_args = args->in_args; |
| 616 |
int num_args = args->in_numargs; |
| 617 |
int err; |
| 618 |
struct iov_iter iter; |
| 619 |
struct fuse_uring_ent_in_out ent_in_out = { |
| 620 |
.flags = 0, |
| 621 |
.commit_id = req->in.h.unique, |
| 622 |
}; |
| 623 |
|
| 624 |
err = import_ubuf(ITER_DEST, ent->payload, ring->max_payload_sz, &iter); |
| 625 |
if (err) { |
| 626 |
pr_info_ratelimited("fuse: Import of user buffer failed\n"); |
| 627 |
return err; |
| 628 |
} |
| 629 |
|
| 630 |
fuse_copy_init(&cs, true, &iter); |
| 631 |
cs.is_uring = true; |
| 632 |
cs.req = req; |
| 633 |
|
| 634 |
if (num_args > 0) { |
| 635 |
/* |
| 636 |
* Expectation is that the first argument is the per op header. |
| 637 |
* Some op code have that as zero size. |
| 638 |
*/ |
| 639 |
if (args->in_args[0].size > 0) { |
| 640 |
err = copy_to_user(&ent->headers->op_in, in_args->value, |
| 641 |
in_args->size); |
| 642 |
if (err) { |
| 643 |
pr_info_ratelimited( |
| 644 |
"Copying the header failed.\n"); |
| 645 |
return -EFAULT; |
| 646 |
} |
| 647 |
} |
| 648 |
in_args++; |
| 649 |
num_args--; |
| 650 |
} |
| 651 |
|
| 652 |
/* copy the payload */ |
| 653 |
err = fuse_copy_args(&cs, num_args, args->in_pages, |
| 654 |
(struct fuse_arg *)in_args, 0); |
| 655 |
fuse_copy_finish(&cs); |
| 656 |
if (err) { |
| 657 |
pr_info_ratelimited("%s fuse_copy_args failed\n", __func__); |
| 658 |
return err; |
| 659 |
} |
| 660 |
|
| 661 |
ent_in_out.payload_sz = cs.ring.copied_sz; |
| 662 |
err = copy_to_user(&ent->headers->ring_ent_in_out, &ent_in_out, |
| 663 |
sizeof(ent_in_out)); |
| 664 |
return err ? -EFAULT : 0; |
| 665 |
} |
| 666 |
|
| 667 |
static int fuse_uring_copy_to_ring(struct fuse_ring_ent *ent, |
| 668 |
struct fuse_req *req) |
| 669 |
{ |
| 670 |
struct fuse_ring_queue *queue = ent->queue; |
| 671 |
struct fuse_ring *ring = queue->ring; |
| 672 |
int err; |
| 673 |
|
| 674 |
err = -EIO; |
| 675 |
if (WARN_ON(ent->state != FRRS_FUSE_REQ)) { |
| 676 |
pr_err("qid=%d ring-req=%p invalid state %d on send\n", |
| 677 |
queue->qid, ent, ent->state); |
| 678 |
return err; |
| 679 |
} |
| 680 |
|
| 681 |
err = -EINVAL; |
| 682 |
if (WARN_ON(req->in.h.unique == 0)) |
| 683 |
return err; |
| 684 |
|
| 685 |
/* copy the request */ |
| 686 |
err = fuse_uring_args_to_ring(ring, req, ent); |
| 687 |
if (unlikely(err)) { |
| 688 |
pr_info_ratelimited("Copy to ring failed: %d\n", err); |
| 689 |
return err; |
| 690 |
} |
| 691 |
|
| 692 |
/* copy fuse_in_header */ |
| 693 |
err = copy_to_user(&ent->headers->in_out, &req->in.h, |
| 694 |
sizeof(req->in.h)); |
| 695 |
if (err) { |
| 696 |
err = -EFAULT; |
| 697 |
return err; |
| 698 |
} |
| 699 |
|
| 700 |
return 0; |
| 701 |
} |
| 702 |
|
| 703 |
static int fuse_uring_prepare_send(struct fuse_ring_ent *ent, |
| 704 |
struct fuse_req *req) |
| 705 |
{ |
| 706 |
int err; |
| 707 |
|
| 708 |
err = fuse_uring_copy_to_ring(ent, req); |
| 709 |
if (!err) |
| 710 |
set_bit(FR_SENT, &req->flags); |
| 711 |
else |
| 712 |
fuse_uring_req_end(ent, req, err); |
| 713 |
|
| 714 |
return err; |
| 715 |
} |
| 716 |
|
| 717 |
/* |
| 718 |
* Write data to the ring buffer and send the request to userspace, |
| 719 |
* userspace will read it |
| 720 |
* This is comparable with classical read(/dev/fuse) |
| 721 |
*/ |
| 722 |
static int fuse_uring_send_next_to_ring(struct fuse_ring_ent *ent, |
| 723 |
struct fuse_req *req, |
| 724 |
unsigned int issue_flags) |
| 725 |
{ |
| 726 |
struct fuse_ring_queue *queue = ent->queue; |
| 727 |
int err; |
| 728 |
struct io_uring_cmd *cmd; |
| 729 |
|
| 730 |
err = fuse_uring_prepare_send(ent, req); |
| 731 |
if (err) |
| 732 |
return err; |
| 733 |
|
| 734 |
spin_lock(&queue->lock); |
| 735 |
cmd = ent->cmd; |
| 736 |
ent->cmd = NULL; |
| 737 |
ent->state = FRRS_USERSPACE; |
| 738 |
list_move_tail(&ent->list, &queue->ent_in_userspace); |
| 739 |
spin_unlock(&queue->lock); |
| 740 |
|
| 741 |
io_uring_cmd_done(cmd, 0, issue_flags); |
| 742 |
return 0; |
| 743 |
} |
| 744 |
|
| 745 |
/* |
| 746 |
* Make a ring entry available for fuse_req assignment |
| 747 |
*/ |
| 748 |
static void fuse_uring_ent_avail(struct fuse_ring_ent *ent, |
| 749 |
struct fuse_ring_queue *queue) |
| 750 |
{ |
| 751 |
WARN_ON_ONCE(!ent->cmd); |
| 752 |
list_move(&ent->list, &queue->ent_avail_queue); |
| 753 |
ent->state = FRRS_AVAILABLE; |
| 754 |
} |
| 755 |
|
| 756 |
/* Used to find the request on SQE commit */ |
| 757 |
static void fuse_uring_add_to_pq(struct fuse_ring_ent *ent, |
| 758 |
struct fuse_req *req) |
| 759 |
{ |
| 760 |
struct fuse_ring_queue *queue = ent->queue; |
| 761 |
struct fuse_pqueue *fpq = &queue->fpq; |
| 762 |
unsigned int hash; |
| 763 |
|
| 764 |
req->ring_entry = ent; |
| 765 |
hash = fuse_req_hash(req->in.h.unique); |
| 766 |
list_move_tail(&req->list, &fpq->processing[hash]); |
| 767 |
} |
| 768 |
|
| 769 |
/* |
| 770 |
* Assign a fuse queue entry to the given entry |
| 771 |
*/ |
| 772 |
static void fuse_uring_add_req_to_ring_ent(struct fuse_ring_ent *ent, |
| 773 |
struct fuse_req *req) |
| 774 |
{ |
| 775 |
struct fuse_ring_queue *queue = ent->queue; |
| 776 |
|
| 777 |
lockdep_assert_held(&queue->lock); |
| 778 |
|
| 779 |
if (WARN_ON_ONCE(ent->state != FRRS_AVAILABLE && |
| 780 |
ent->state != FRRS_COMMIT)) { |
| 781 |
pr_warn("%s qid=%d state=%d\n", __func__, ent->queue->qid, |
| 782 |
ent->state); |
| 783 |
} |
| 784 |
|
| 785 |
clear_bit(FR_PENDING, &req->flags); |
| 786 |
ent->fuse_req = req; |
| 787 |
ent->state = FRRS_FUSE_REQ; |
| 788 |
list_move_tail(&ent->list, &queue->ent_w_req_queue); |
| 789 |
fuse_uring_add_to_pq(ent, req); |
| 790 |
} |
| 791 |
|
| 792 |
/* Fetch the next fuse request if available */ |
| 793 |
static struct fuse_req *fuse_uring_ent_assign_req(struct fuse_ring_ent *ent) |
| 794 |
__must_hold(&queue->lock) |
| 795 |
{ |
| 796 |
struct fuse_req *req; |
| 797 |
struct fuse_ring_queue *queue = ent->queue; |
| 798 |
struct list_head *req_queue = &queue->fuse_req_queue; |
| 799 |
|
| 800 |
lockdep_assert_held(&queue->lock); |
| 801 |
|
| 802 |
/* get and assign the next entry while it is still holding the lock */ |
| 803 |
req = list_first_entry_or_null(req_queue, struct fuse_req, list); |
| 804 |
if (req) |
| 805 |
fuse_uring_add_req_to_ring_ent(ent, req); |
| 806 |
|
| 807 |
return req; |
| 808 |
} |
| 809 |
|
| 810 |
/* |
| 811 |
* Read data from the ring buffer, which user space has written to |
| 812 |
* This is comparible with handling of classical write(/dev/fuse). |
| 813 |
* Also make the ring request available again for new fuse requests. |
| 814 |
*/ |
| 815 |
static void fuse_uring_commit(struct fuse_ring_ent *ent, struct fuse_req *req, |
| 816 |
unsigned int issue_flags) |
| 817 |
{ |
| 818 |
struct fuse_ring *ring = ent->queue->ring; |
| 819 |
struct fuse_conn *fc = ring->fc; |
| 820 |
ssize_t err = 0; |
| 821 |
|
| 822 |
err = copy_from_user(&req->out.h, &ent->headers->in_out, |
| 823 |
sizeof(req->out.h)); |
| 824 |
if (err) { |
| 825 |
req->out.h.error = -EFAULT; |
| 826 |
goto out; |
| 827 |
} |
| 828 |
|
| 829 |
err = fuse_uring_out_header_has_err(&req->out.h, req, fc); |
| 830 |
if (err) { |
| 831 |
/* req->out.h.error already set */ |
| 832 |
goto out; |
| 833 |
} |
| 834 |
|
| 835 |
err = fuse_uring_copy_from_ring(ring, req, ent); |
| 836 |
out: |
| 837 |
fuse_uring_req_end(ent, req, err); |
| 838 |
} |
| 839 |
|
| 840 |
/* |
| 841 |
* Get the next fuse req and send it |
| 842 |
*/ |
| 843 |
static void fuse_uring_next_fuse_req(struct fuse_ring_ent *ent, |
| 844 |
struct fuse_ring_queue *queue, |
| 845 |
unsigned int issue_flags) |
| 846 |
{ |
| 847 |
int err; |
| 848 |
struct fuse_req *req; |
| 849 |
|
| 850 |
retry: |
| 851 |
spin_lock(&queue->lock); |
| 852 |
fuse_uring_ent_avail(ent, queue); |
| 853 |
req = fuse_uring_ent_assign_req(ent); |
| 854 |
spin_unlock(&queue->lock); |
| 855 |
|
| 856 |
if (req) { |
| 857 |
err = fuse_uring_send_next_to_ring(ent, req, issue_flags); |
| 858 |
if (err) |
| 859 |
goto retry; |
| 860 |
} |
| 861 |
} |
| 862 |
|
| 863 |
static int fuse_ring_ent_set_commit(struct fuse_ring_ent *ent) |
| 864 |
{ |
| 865 |
struct fuse_ring_queue *queue = ent->queue; |
| 866 |
|
| 867 |
lockdep_assert_held(&queue->lock); |
| 868 |
|
| 869 |
if (WARN_ON_ONCE(ent->state != FRRS_USERSPACE)) |
| 870 |
return -EIO; |
| 871 |
|
| 872 |
ent->state = FRRS_COMMIT; |
| 873 |
list_move(&ent->list, &queue->ent_commit_queue); |
| 874 |
|
| 875 |
return 0; |
| 876 |
} |
| 877 |
|
| 878 |
/* FUSE_URING_CMD_COMMIT_AND_FETCH handler */ |
| 879 |
static int fuse_uring_commit_fetch(struct io_uring_cmd *cmd, int issue_flags, |
| 880 |
struct fuse_conn *fc) |
| 881 |
{ |
| 882 |
const struct fuse_uring_cmd_req *cmd_req = io_uring_sqe128_cmd(cmd->sqe, |
| 883 |
struct fuse_uring_cmd_req); |
| 884 |
struct fuse_ring_ent *ent; |
| 885 |
int err; |
| 886 |
struct fuse_ring *ring = fc->ring; |
| 887 |
struct fuse_ring_queue *queue; |
| 888 |
uint64_t commit_id = READ_ONCE(cmd_req->commit_id); |
| 889 |
unsigned int qid = READ_ONCE(cmd_req->qid); |
| 890 |
struct fuse_pqueue *fpq; |
| 891 |
struct fuse_req *req; |
| 892 |
|
| 893 |
err = -ENOTCONN; |
| 894 |
if (!ring) |
| 895 |
return err; |
| 896 |
|
| 897 |
if (qid >= ring->nr_queues) |
| 898 |
return -EINVAL; |
| 899 |
|
| 900 |
queue = ring->queues[qid]; |
| 901 |
if (!queue) |
| 902 |
return err; |
| 903 |
fpq = &queue->fpq; |
| 904 |
|
| 905 |
if (!READ_ONCE(fc->connected) || READ_ONCE(queue->stopped)) |
| 906 |
return err; |
| 907 |
|
| 908 |
spin_lock(&queue->lock); |
| 909 |
/* Find a request based on the unique ID of the fuse request |
| 910 |
* This should get revised, as it needs a hash calculation and list |
| 911 |
* search. And full struct fuse_pqueue is needed (memory overhead). |
| 912 |
* As well as the link from req to ring_ent. |
| 913 |
*/ |
| 914 |
req = fuse_request_find(fpq, commit_id); |
| 915 |
err = -ENOENT; |
| 916 |
if (!req) { |
| 917 |
pr_info("qid=%d commit_id %llu not found\n", queue->qid, |
| 918 |
commit_id); |
| 919 |
spin_unlock(&queue->lock); |
| 920 |
return err; |
| 921 |
} |
| 922 |
list_del_init(&req->list); |
| 923 |
ent = req->ring_entry; |
| 924 |
req->ring_entry = NULL; |
| 925 |
|
| 926 |
err = fuse_ring_ent_set_commit(ent); |
| 927 |
if (err != 0) { |
| 928 |
pr_info_ratelimited("qid=%d commit_id %llu state %d", |
| 929 |
queue->qid, commit_id, ent->state); |
| 930 |
spin_unlock(&queue->lock); |
| 931 |
req->out.h.error = err; |
| 932 |
clear_bit(FR_SENT, &req->flags); |
| 933 |
fuse_request_end(req); |
| 934 |
return err; |
| 935 |
} |
| 936 |
|
| 937 |
ent->cmd = cmd; |
| 938 |
spin_unlock(&queue->lock); |
| 939 |
|
| 940 |
/* without the queue lock, as other locks are taken */ |
| 941 |
fuse_uring_prepare_cancel(cmd, issue_flags, ent); |
| 942 |
fuse_uring_commit(ent, req, issue_flags); |
| 943 |
|
| 944 |
/* |
| 945 |
* Fetching the next request is absolutely required as queued |
| 946 |
* fuse requests would otherwise not get processed - committing |
| 947 |
* and fetching is done in one step vs legacy fuse, which has separated |
| 948 |
* read (fetch request) and write (commit result). |
| 949 |
*/ |
| 950 |
fuse_uring_next_fuse_req(ent, queue, issue_flags); |
| 951 |
return 0; |
| 952 |
} |
| 953 |
|
| 954 |
static bool is_ring_ready(struct fuse_ring *ring, int current_qid) |
| 955 |
{ |
| 956 |
int qid; |
| 957 |
struct fuse_ring_queue *queue; |
| 958 |
bool ready = true; |
| 959 |
|
| 960 |
for (qid = 0; qid < ring->nr_queues && ready; qid++) { |
| 961 |
if (current_qid == qid) |
| 962 |
continue; |
| 963 |
|
| 964 |
queue = ring->queues[qid]; |
| 965 |
if (!queue) { |
| 966 |
ready = false; |
| 967 |
break; |
| 968 |
} |
| 969 |
|
| 970 |
spin_lock(&queue->lock); |
| 971 |
if (list_empty(&queue->ent_avail_queue)) |
| 972 |
ready = false; |
| 973 |
spin_unlock(&queue->lock); |
| 974 |
} |
| 975 |
|
| 976 |
return ready; |
| 977 |
} |
| 978 |
|
| 979 |
/* |
| 980 |
* fuse_uring_req_fetch command handling |
| 981 |
*/ |
| 982 |
static void fuse_uring_do_register(struct fuse_ring_ent *ent, |
| 983 |
struct io_uring_cmd *cmd, |
| 984 |
unsigned int issue_flags) |
| 985 |
{ |
| 986 |
struct fuse_ring_queue *queue = ent->queue; |
| 987 |
struct fuse_ring *ring = queue->ring; |
| 988 |
struct fuse_conn *fc = ring->fc; |
| 989 |
struct fuse_iqueue *fiq = &fc->iq; |
| 990 |
|
| 991 |
fuse_uring_prepare_cancel(cmd, issue_flags, ent); |
| 992 |
|
| 993 |
spin_lock(&queue->lock); |
| 994 |
ent->cmd = cmd; |
| 995 |
fuse_uring_ent_avail(ent, queue); |
| 996 |
spin_unlock(&queue->lock); |
| 997 |
|
| 998 |
if (!ring->ready) { |
| 999 |
bool ready = is_ring_ready(ring, queue->qid); |
| 1000 |
|
| 1001 |
if (ready) { |
| 1002 |
WRITE_ONCE(fiq->ops, &fuse_io_uring_ops); |
| 1003 |
WRITE_ONCE(ring->ready, true); |
| 1004 |
wake_up_all(&fc->blocked_waitq); |
| 1005 |
} |
| 1006 |
} |
| 1007 |
} |
| 1008 |
|
| 1009 |
/* |
| 1010 |
* sqe->addr is a ptr to an iovec array, iov[0] has the headers, iov[1] |
| 1011 |
* the payload |
| 1012 |
*/ |
| 1013 |
static int fuse_uring_get_iovec_from_sqe(const struct io_uring_sqe *sqe, |
| 1014 |
struct iovec iov[FUSE_URING_IOV_SEGS]) |
| 1015 |
{ |
| 1016 |
struct iovec __user *uiov = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
| 1017 |
struct iov_iter iter; |
| 1018 |
ssize_t ret; |
| 1019 |
|
| 1020 |
if (sqe->len != FUSE_URING_IOV_SEGS) |
| 1021 |
return -EINVAL; |
| 1022 |
|
| 1023 |
/* |
| 1024 |
* Direction for buffer access will actually be READ and WRITE, |
| 1025 |
* using write for the import should include READ access as well. |
| 1026 |
*/ |
| 1027 |
ret = import_iovec(WRITE, uiov, FUSE_URING_IOV_SEGS, |
| 1028 |
FUSE_URING_IOV_SEGS, &iov, &iter); |
| 1029 |
if (ret < 0) |
| 1030 |
return ret; |
| 1031 |
|
| 1032 |
return 0; |
| 1033 |
} |
| 1034 |
|
| 1035 |
static struct fuse_ring_ent * |
| 1036 |
fuse_uring_create_ring_ent(struct io_uring_cmd *cmd, |
| 1037 |
struct fuse_ring_queue *queue) |
| 1038 |
{ |
| 1039 |
struct fuse_ring *ring = queue->ring; |
| 1040 |
struct fuse_ring_ent *ent; |
| 1041 |
size_t payload_size; |
| 1042 |
struct iovec iov[FUSE_URING_IOV_SEGS]; |
| 1043 |
int err; |
| 1044 |
|
| 1045 |
err = fuse_uring_get_iovec_from_sqe(cmd->sqe, iov); |
| 1046 |
if (err) { |
| 1047 |
pr_info_ratelimited("Failed to get iovec from sqe, err=%d\n", |
| 1048 |
err); |
| 1049 |
return ERR_PTR(err); |
| 1050 |
} |
| 1051 |
|
| 1052 |
err = -EINVAL; |
| 1053 |
if (iov[0].iov_len < sizeof(struct fuse_uring_req_header)) { |
| 1054 |
pr_info_ratelimited("Invalid header len %zu\n", iov[0].iov_len); |
| 1055 |
return ERR_PTR(err); |
| 1056 |
} |
| 1057 |
|
| 1058 |
payload_size = iov[1].iov_len; |
| 1059 |
if (payload_size < ring->max_payload_sz) { |
| 1060 |
pr_info_ratelimited("Invalid req payload len %zu\n", |
| 1061 |
payload_size); |
| 1062 |
return ERR_PTR(err); |
| 1063 |
} |
| 1064 |
|
| 1065 |
err = -ENOMEM; |
| 1066 |
ent = kzalloc_obj(*ent, GFP_KERNEL_ACCOUNT); |
| 1067 |
if (!ent) |
| 1068 |
return ERR_PTR(err); |
| 1069 |
|
| 1070 |
INIT_LIST_HEAD(&ent->list); |
| 1071 |
|
| 1072 |
ent->queue = queue; |
| 1073 |
ent->headers = iov[0].iov_base; |
| 1074 |
ent->payload = iov[1].iov_base; |
| 1075 |
|
| 1076 |
atomic_inc(&ring->queue_refs); |
| 1077 |
return ent; |
| 1078 |
} |
| 1079 |
|
| 1080 |
/* |
| 1081 |
* Register header and payload buffer with the kernel and puts the |
| 1082 |
* entry as "ready to get fuse requests" on the queue |
| 1083 |
*/ |
| 1084 |
static int fuse_uring_register(struct io_uring_cmd *cmd, |
| 1085 |
unsigned int issue_flags, struct fuse_conn *fc) |
| 1086 |
{ |
| 1087 |
const struct fuse_uring_cmd_req *cmd_req = io_uring_sqe128_cmd(cmd->sqe, |
| 1088 |
struct fuse_uring_cmd_req); |
| 1089 |
struct fuse_ring *ring = smp_load_acquire(&fc->ring); |
| 1090 |
struct fuse_ring_queue *queue; |
| 1091 |
struct fuse_ring_ent *ent; |
| 1092 |
int err; |
| 1093 |
unsigned int qid = READ_ONCE(cmd_req->qid); |
| 1094 |
|
| 1095 |
err = -ENOMEM; |
| 1096 |
if (!ring) { |
| 1097 |
ring = fuse_uring_create(fc); |
| 1098 |
if (!ring) |
| 1099 |
return err; |
| 1100 |
} |
| 1101 |
|
| 1102 |
if (qid >= ring->nr_queues) { |
| 1103 |
pr_info_ratelimited("fuse: Invalid ring qid %u\n", qid); |
| 1104 |
return -EINVAL; |
| 1105 |
} |
| 1106 |
|
| 1107 |
queue = ring->queues[qid]; |
| 1108 |
if (!queue) { |
| 1109 |
queue = fuse_uring_create_queue(ring, qid); |
| 1110 |
if (!queue) |
| 1111 |
return err; |
| 1112 |
} |
| 1113 |
|
| 1114 |
/* |
| 1115 |
* The created queue above does not need to be destructed in |
| 1116 |
* case of entry errors below, will be done at ring destruction time. |
| 1117 |
*/ |
| 1118 |
|
| 1119 |
ent = fuse_uring_create_ring_ent(cmd, queue); |
| 1120 |
if (IS_ERR(ent)) |
| 1121 |
return PTR_ERR(ent); |
| 1122 |
|
| 1123 |
fuse_uring_do_register(ent, cmd, issue_flags); |
| 1124 |
|
| 1125 |
return 0; |
| 1126 |
} |
| 1127 |
|
| 1128 |
/* |
| 1129 |
* Entry function from io_uring to handle the given passthrough command |
| 1130 |
* (op code IORING_OP_URING_CMD) |
| 1131 |
*/ |
| 1132 |
int fuse_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags) |
| 1133 |
{ |
| 1134 |
struct fuse_dev *fud; |
| 1135 |
struct fuse_conn *fc; |
| 1136 |
u32 cmd_op = cmd->cmd_op; |
| 1137 |
int err; |
| 1138 |
|
| 1139 |
if ((unlikely(issue_flags & IO_URING_F_CANCEL))) { |
| 1140 |
fuse_uring_cancel(cmd, issue_flags); |
| 1141 |
return 0; |
| 1142 |
} |
| 1143 |
|
| 1144 |
/* This extra SQE size holds struct fuse_uring_cmd_req */ |
| 1145 |
if (!(issue_flags & IO_URING_F_SQE128)) |
| 1146 |
return -EINVAL; |
| 1147 |
|
| 1148 |
fud = fuse_get_dev(cmd->file); |
| 1149 |
if (IS_ERR(fud)) { |
| 1150 |
pr_info_ratelimited("No fuse device found\n"); |
| 1151 |
return PTR_ERR(fud); |
| 1152 |
} |
| 1153 |
fc = fud->fc; |
| 1154 |
|
| 1155 |
/* Once a connection has io-uring enabled on it, it can't be disabled */ |
| 1156 |
if (!enable_uring && !fc->io_uring) { |
| 1157 |
pr_info_ratelimited("fuse-io-uring is disabled\n"); |
| 1158 |
return -EOPNOTSUPP; |
| 1159 |
} |
| 1160 |
|
| 1161 |
if (fc->aborted) |
| 1162 |
return -ECONNABORTED; |
| 1163 |
if (!fc->connected) |
| 1164 |
return -ENOTCONN; |
| 1165 |
|
| 1166 |
/* |
| 1167 |
* fuse_uring_register() needs the ring to be initialized, |
| 1168 |
* we need to know the max payload size |
| 1169 |
*/ |
| 1170 |
if (!fc->initialized) |
| 1171 |
return -EAGAIN; |
| 1172 |
|
| 1173 |
switch (cmd_op) { |
| 1174 |
case FUSE_IO_URING_CMD_REGISTER: |
| 1175 |
err = fuse_uring_register(cmd, issue_flags, fc); |
| 1176 |
if (err) { |
| 1177 |
pr_info_once("FUSE_IO_URING_CMD_REGISTER failed err=%d\n", |
| 1178 |
err); |
| 1179 |
fc->io_uring = 0; |
| 1180 |
wake_up_all(&fc->blocked_waitq); |
| 1181 |
return err; |
| 1182 |
} |
| 1183 |
break; |
| 1184 |
case FUSE_IO_URING_CMD_COMMIT_AND_FETCH: |
| 1185 |
err = fuse_uring_commit_fetch(cmd, issue_flags, fc); |
| 1186 |
if (err) { |
| 1187 |
pr_info_once("FUSE_IO_URING_COMMIT_AND_FETCH failed err=%d\n", |
| 1188 |
err); |
| 1189 |
return err; |
| 1190 |
} |
| 1191 |
break; |
| 1192 |
default: |
| 1193 |
return -EINVAL; |
| 1194 |
} |
| 1195 |
|
| 1196 |
return -EIOCBQUEUED; |
| 1197 |
} |
| 1198 |
|
| 1199 |
static void fuse_uring_send(struct fuse_ring_ent *ent, struct io_uring_cmd *cmd, |
| 1200 |
ssize_t ret, unsigned int issue_flags) |
| 1201 |
{ |
| 1202 |
struct fuse_ring_queue *queue = ent->queue; |
| 1203 |
|
| 1204 |
spin_lock(&queue->lock); |
| 1205 |
ent->state = FRRS_USERSPACE; |
| 1206 |
list_move_tail(&ent->list, &queue->ent_in_userspace); |
| 1207 |
ent->cmd = NULL; |
| 1208 |
spin_unlock(&queue->lock); |
| 1209 |
|
| 1210 |
io_uring_cmd_done(cmd, ret, issue_flags); |
| 1211 |
} |
| 1212 |
|
| 1213 |
/* |
| 1214 |
* This prepares and sends the ring request in fuse-uring task context. |
| 1215 |
* User buffers are not mapped yet - the application does not have permission |
| 1216 |
* to write to it - this has to be executed in ring task context. |
| 1217 |
*/ |
| 1218 |
static void fuse_uring_send_in_task(struct io_tw_req tw_req, io_tw_token_t tw) |
| 1219 |
{ |
| 1220 |
unsigned int issue_flags = IO_URING_CMD_TASK_WORK_ISSUE_FLAGS; |
| 1221 |
struct io_uring_cmd *cmd = io_uring_cmd_from_tw(tw_req); |
| 1222 |
struct fuse_ring_ent *ent = uring_cmd_to_ring_ent(cmd); |
| 1223 |
struct fuse_ring_queue *queue = ent->queue; |
| 1224 |
int err; |
| 1225 |
|
| 1226 |
if (!tw.cancel) { |
| 1227 |
err = fuse_uring_prepare_send(ent, ent->fuse_req); |
| 1228 |
if (err) { |
| 1229 |
fuse_uring_next_fuse_req(ent, queue, issue_flags); |
| 1230 |
return; |
| 1231 |
} |
| 1232 |
} else { |
| 1233 |
err = -ECANCELED; |
| 1234 |
} |
| 1235 |
|
| 1236 |
fuse_uring_send(ent, cmd, err, issue_flags); |
| 1237 |
} |
| 1238 |
|
| 1239 |
static struct fuse_ring_queue *fuse_uring_task_to_queue(struct fuse_ring *ring) |
| 1240 |
{ |
| 1241 |
unsigned int qid; |
| 1242 |
struct fuse_ring_queue *queue; |
| 1243 |
|
| 1244 |
qid = task_cpu(current); |
| 1245 |
|
| 1246 |
if (WARN_ONCE(qid >= ring->nr_queues, |
| 1247 |
"Core number (%u) exceeds nr queues (%zu)\n", qid, |
| 1248 |
ring->nr_queues)) |
| 1249 |
qid = 0; |
| 1250 |
|
| 1251 |
queue = ring->queues[qid]; |
| 1252 |
WARN_ONCE(!queue, "Missing queue for qid %d\n", qid); |
| 1253 |
|
| 1254 |
return queue; |
| 1255 |
} |
| 1256 |
|
| 1257 |
static void fuse_uring_dispatch_ent(struct fuse_ring_ent *ent) |
| 1258 |
{ |
| 1259 |
struct io_uring_cmd *cmd = ent->cmd; |
| 1260 |
|
| 1261 |
uring_cmd_set_ring_ent(cmd, ent); |
| 1262 |
io_uring_cmd_complete_in_task(cmd, fuse_uring_send_in_task); |
| 1263 |
} |
| 1264 |
|
| 1265 |
/* queue a fuse request and send it if a ring entry is available */ |
| 1266 |
void fuse_uring_queue_fuse_req(struct fuse_iqueue *fiq, struct fuse_req *req) |
| 1267 |
{ |
| 1268 |
struct fuse_conn *fc = req->fm->fc; |
| 1269 |
struct fuse_ring *ring = fc->ring; |
| 1270 |
struct fuse_ring_queue *queue; |
| 1271 |
struct fuse_ring_ent *ent = NULL; |
| 1272 |
int err; |
| 1273 |
|
| 1274 |
err = -EINVAL; |
| 1275 |
queue = fuse_uring_task_to_queue(ring); |
| 1276 |
if (!queue) |
| 1277 |
goto err; |
| 1278 |
|
| 1279 |
fuse_request_assign_unique(fiq, req); |
| 1280 |
|
| 1281 |
spin_lock(&queue->lock); |
| 1282 |
err = -ENOTCONN; |
| 1283 |
if (unlikely(queue->stopped)) |
| 1284 |
goto err_unlock; |
| 1285 |
|
| 1286 |
set_bit(FR_URING, &req->flags); |
| 1287 |
req->ring_queue = queue; |
| 1288 |
ent = list_first_entry_or_null(&queue->ent_avail_queue, |
| 1289 |
struct fuse_ring_ent, list); |
| 1290 |
if (ent) |
| 1291 |
fuse_uring_add_req_to_ring_ent(ent, req); |
| 1292 |
else |
| 1293 |
list_add_tail(&req->list, &queue->fuse_req_queue); |
| 1294 |
spin_unlock(&queue->lock); |
| 1295 |
|
| 1296 |
if (ent) |
| 1297 |
fuse_uring_dispatch_ent(ent); |
| 1298 |
|
| 1299 |
return; |
| 1300 |
|
| 1301 |
err_unlock: |
| 1302 |
spin_unlock(&queue->lock); |
| 1303 |
err: |
| 1304 |
req->out.h.error = err; |
| 1305 |
clear_bit(FR_PENDING, &req->flags); |
| 1306 |
fuse_request_end(req); |
| 1307 |
} |
| 1308 |
|
| 1309 |
bool fuse_uring_queue_bq_req(struct fuse_req *req) |
| 1310 |
{ |
| 1311 |
struct fuse_conn *fc = req->fm->fc; |
| 1312 |
struct fuse_ring *ring = fc->ring; |
| 1313 |
struct fuse_ring_queue *queue; |
| 1314 |
struct fuse_ring_ent *ent = NULL; |
| 1315 |
|
| 1316 |
queue = fuse_uring_task_to_queue(ring); |
| 1317 |
if (!queue) |
| 1318 |
return false; |
| 1319 |
|
| 1320 |
spin_lock(&queue->lock); |
| 1321 |
if (unlikely(queue->stopped)) { |
| 1322 |
spin_unlock(&queue->lock); |
| 1323 |
return false; |
| 1324 |
} |
| 1325 |
|
| 1326 |
set_bit(FR_URING, &req->flags); |
| 1327 |
req->ring_queue = queue; |
| 1328 |
list_add_tail(&req->list, &queue->fuse_req_bg_queue); |
| 1329 |
|
| 1330 |
ent = list_first_entry_or_null(&queue->ent_avail_queue, |
| 1331 |
struct fuse_ring_ent, list); |
| 1332 |
spin_lock(&fc->bg_lock); |
| 1333 |
fc->num_background++; |
| 1334 |
if (fc->num_background == fc->max_background) |
| 1335 |
fc->blocked = 1; |
| 1336 |
fuse_uring_flush_bg(queue); |
| 1337 |
spin_unlock(&fc->bg_lock); |
| 1338 |
|
| 1339 |
/* |
| 1340 |
* Due to bg_queue flush limits there might be other bg requests |
| 1341 |
* in the queue that need to be handled first. Or no further req |
| 1342 |
* might be available. |
| 1343 |
*/ |
| 1344 |
req = list_first_entry_or_null(&queue->fuse_req_queue, struct fuse_req, |
| 1345 |
list); |
| 1346 |
if (ent && req) { |
| 1347 |
fuse_uring_add_req_to_ring_ent(ent, req); |
| 1348 |
spin_unlock(&queue->lock); |
| 1349 |
|
| 1350 |
fuse_uring_dispatch_ent(ent); |
| 1351 |
} else { |
| 1352 |
spin_unlock(&queue->lock); |
| 1353 |
} |
| 1354 |
|
| 1355 |
return true; |
| 1356 |
} |
| 1357 |
|
| 1358 |
bool fuse_uring_remove_pending_req(struct fuse_req *req) |
| 1359 |
{ |
| 1360 |
struct fuse_ring_queue *queue = req->ring_queue; |
| 1361 |
|
| 1362 |
return fuse_remove_pending_req(req, &queue->lock); |
| 1363 |
} |
| 1364 |
|
| 1365 |
static const struct fuse_iqueue_ops fuse_io_uring_ops = { |
| 1366 |
/* should be send over io-uring as enhancement */ |
| 1367 |
.send_forget = fuse_dev_queue_forget, |
| 1368 |
|
| 1369 |
/* |
| 1370 |
* could be send over io-uring, but interrupts should be rare, |
| 1371 |
* no need to make the code complex |
| 1372 |
*/ |
| 1373 |
.send_interrupt = fuse_dev_queue_interrupt, |
| 1374 |
.send_req = fuse_uring_queue_fuse_req, |
| 1375 |
}; |