| 1 |
// SPDX-License-Identifier: GPL-2.0 |
| 2 |
/* |
| 3 |
* (C) 2001 Clemson University and The University of Chicago |
| 4 |
* |
| 5 |
* See COPYING in top-level directory. |
| 6 |
*/ |
| 7 |
#include "protocol.h" |
| 8 |
#include "orangefs-kernel.h" |
| 9 |
#include "orangefs-bufmap.h" |
| 10 |
|
| 11 |
struct slot_map { |
| 12 |
int c; |
| 13 |
wait_queue_head_t q; |
| 14 |
int count; |
| 15 |
unsigned long *map; |
| 16 |
}; |
| 17 |
|
| 18 |
static struct slot_map rw_map = { |
| 19 |
.c = -1, |
| 20 |
.q = __WAIT_QUEUE_HEAD_INITIALIZER(rw_map.q) |
| 21 |
}; |
| 22 |
static struct slot_map readdir_map = { |
| 23 |
.c = -1, |
| 24 |
.q = __WAIT_QUEUE_HEAD_INITIALIZER(readdir_map.q) |
| 25 |
}; |
| 26 |
|
| 27 |
|
| 28 |
static void install(struct slot_map *m, int count, unsigned long *map) |
| 29 |
{ |
| 30 |
spin_lock(&m->q.lock); |
| 31 |
m->c = m->count = count; |
| 32 |
m->map = map; |
| 33 |
wake_up_all_locked(&m->q); |
| 34 |
spin_unlock(&m->q.lock); |
| 35 |
} |
| 36 |
|
| 37 |
static void mark_killed(struct slot_map *m) |
| 38 |
{ |
| 39 |
spin_lock(&m->q.lock); |
| 40 |
m->c -= m->count + 1; |
| 41 |
spin_unlock(&m->q.lock); |
| 42 |
} |
| 43 |
|
| 44 |
static void run_down(struct slot_map *m) |
| 45 |
{ |
| 46 |
DEFINE_WAIT(wait); |
| 47 |
spin_lock(&m->q.lock); |
| 48 |
if (m->c != -1) { |
| 49 |
for (;;) { |
| 50 |
if (likely(list_empty(&wait.entry))) |
| 51 |
__add_wait_queue_entry_tail(&m->q, &wait); |
| 52 |
set_current_state(TASK_UNINTERRUPTIBLE); |
| 53 |
|
| 54 |
if (m->c == -1) |
| 55 |
break; |
| 56 |
|
| 57 |
spin_unlock(&m->q.lock); |
| 58 |
schedule(); |
| 59 |
spin_lock(&m->q.lock); |
| 60 |
} |
| 61 |
__remove_wait_queue(&m->q, &wait); |
| 62 |
__set_current_state(TASK_RUNNING); |
| 63 |
} |
| 64 |
m->map = NULL; |
| 65 |
spin_unlock(&m->q.lock); |
| 66 |
} |
| 67 |
|
| 68 |
static void put(struct slot_map *m, int slot) |
| 69 |
{ |
| 70 |
int v; |
| 71 |
spin_lock(&m->q.lock); |
| 72 |
__clear_bit(slot, m->map); |
| 73 |
v = ++m->c; |
| 74 |
if (v > 0) |
| 75 |
wake_up_locked(&m->q); |
| 76 |
if (unlikely(v == -1)) /* finished dying */ |
| 77 |
wake_up_all_locked(&m->q); |
| 78 |
spin_unlock(&m->q.lock); |
| 79 |
} |
| 80 |
|
| 81 |
static int wait_for_free(struct slot_map *m) |
| 82 |
{ |
| 83 |
long left = slot_timeout_secs * HZ; |
| 84 |
DEFINE_WAIT(wait); |
| 85 |
|
| 86 |
do { |
| 87 |
long n = left, t; |
| 88 |
if (likely(list_empty(&wait.entry))) |
| 89 |
__add_wait_queue_entry_tail_exclusive(&m->q, &wait); |
| 90 |
set_current_state(TASK_INTERRUPTIBLE); |
| 91 |
|
| 92 |
if (m->c > 0) |
| 93 |
break; |
| 94 |
|
| 95 |
if (m->c < 0) { |
| 96 |
/* we are waiting for map to be installed */ |
| 97 |
/* it would better be there soon, or we go away */ |
| 98 |
if (n > ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ) |
| 99 |
n = ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ; |
| 100 |
} |
| 101 |
spin_unlock(&m->q.lock); |
| 102 |
t = schedule_timeout(n); |
| 103 |
spin_lock(&m->q.lock); |
| 104 |
if (unlikely(!t) && n != left && m->c < 0) |
| 105 |
left = t; |
| 106 |
else |
| 107 |
left = t + (left - n); |
| 108 |
if (signal_pending(current)) |
| 109 |
left = -EINTR; |
| 110 |
} while (left > 0); |
| 111 |
|
| 112 |
if (!list_empty(&wait.entry)) |
| 113 |
list_del(&wait.entry); |
| 114 |
else if (left <= 0 && waitqueue_active(&m->q)) |
| 115 |
__wake_up_locked_key(&m->q, TASK_INTERRUPTIBLE, NULL); |
| 116 |
__set_current_state(TASK_RUNNING); |
| 117 |
|
| 118 |
if (likely(left > 0)) |
| 119 |
return 0; |
| 120 |
|
| 121 |
return left < 0 ? -EINTR : -ETIMEDOUT; |
| 122 |
} |
| 123 |
|
| 124 |
static int get(struct slot_map *m) |
| 125 |
{ |
| 126 |
int res = 0; |
| 127 |
spin_lock(&m->q.lock); |
| 128 |
if (unlikely(m->c <= 0)) |
| 129 |
res = wait_for_free(m); |
| 130 |
if (likely(!res)) { |
| 131 |
m->c--; |
| 132 |
res = find_first_zero_bit(m->map, m->count); |
| 133 |
__set_bit(res, m->map); |
| 134 |
} |
| 135 |
spin_unlock(&m->q.lock); |
| 136 |
return res; |
| 137 |
} |
| 138 |
|
| 139 |
/* used to describe mapped buffers */ |
| 140 |
struct orangefs_bufmap_desc { |
| 141 |
void __user *uaddr; /* user space address pointer */ |
| 142 |
struct page **page_array; /* array of mapped pages */ |
| 143 |
int array_count; /* size of above arrays */ |
| 144 |
struct list_head list_link; |
| 145 |
}; |
| 146 |
|
| 147 |
static struct orangefs_bufmap { |
| 148 |
int desc_size; |
| 149 |
int desc_shift; |
| 150 |
int desc_count; |
| 151 |
int total_size; |
| 152 |
int page_count; |
| 153 |
|
| 154 |
struct page **page_array; |
| 155 |
struct orangefs_bufmap_desc *desc_array; |
| 156 |
|
| 157 |
/* array to track usage of buffer descriptors */ |
| 158 |
unsigned long *buffer_index_array; |
| 159 |
|
| 160 |
/* array to track usage of buffer descriptors for readdir */ |
| 161 |
#define N DIV_ROUND_UP(ORANGEFS_READDIR_DEFAULT_DESC_COUNT, BITS_PER_LONG) |
| 162 |
unsigned long readdir_index_array[N]; |
| 163 |
#undef N |
| 164 |
} *__orangefs_bufmap; |
| 165 |
|
| 166 |
static DEFINE_SPINLOCK(orangefs_bufmap_lock); |
| 167 |
|
| 168 |
static void |
| 169 |
orangefs_bufmap_unmap(struct orangefs_bufmap *bufmap) |
| 170 |
{ |
| 171 |
unpin_user_pages(bufmap->page_array, bufmap->page_count); |
| 172 |
} |
| 173 |
|
| 174 |
static void |
| 175 |
orangefs_bufmap_free(struct orangefs_bufmap *bufmap) |
| 176 |
{ |
| 177 |
kfree(bufmap->page_array); |
| 178 |
kfree(bufmap->desc_array); |
| 179 |
bitmap_free(bufmap->buffer_index_array); |
| 180 |
kfree(bufmap); |
| 181 |
} |
| 182 |
|
| 183 |
/* |
| 184 |
* XXX: Can the size and shift change while the caller gives up the |
| 185 |
* XXX: lock between calling this and doing something useful? |
| 186 |
*/ |
| 187 |
|
| 188 |
int orangefs_bufmap_size_query(void) |
| 189 |
{ |
| 190 |
struct orangefs_bufmap *bufmap; |
| 191 |
int size = 0; |
| 192 |
spin_lock(&orangefs_bufmap_lock); |
| 193 |
bufmap = __orangefs_bufmap; |
| 194 |
if (bufmap) |
| 195 |
size = bufmap->desc_size; |
| 196 |
spin_unlock(&orangefs_bufmap_lock); |
| 197 |
return size; |
| 198 |
} |
| 199 |
|
| 200 |
static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq); |
| 201 |
static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq); |
| 202 |
|
| 203 |
static struct orangefs_bufmap * |
| 204 |
orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc) |
| 205 |
{ |
| 206 |
struct orangefs_bufmap *bufmap; |
| 207 |
|
| 208 |
bufmap = kzalloc_obj(*bufmap); |
| 209 |
if (!bufmap) |
| 210 |
goto out; |
| 211 |
|
| 212 |
bufmap->total_size = user_desc->total_size; |
| 213 |
bufmap->desc_count = user_desc->count; |
| 214 |
bufmap->desc_size = user_desc->size; |
| 215 |
bufmap->desc_shift = ilog2(bufmap->desc_size); |
| 216 |
|
| 217 |
bufmap->buffer_index_array = bitmap_zalloc(bufmap->desc_count, GFP_KERNEL); |
| 218 |
if (!bufmap->buffer_index_array) |
| 219 |
goto out_free_bufmap; |
| 220 |
|
| 221 |
bufmap->desc_array = |
| 222 |
kzalloc_objs(struct orangefs_bufmap_desc, bufmap->desc_count); |
| 223 |
if (!bufmap->desc_array) |
| 224 |
goto out_free_index_array; |
| 225 |
|
| 226 |
bufmap->page_count = bufmap->total_size / PAGE_SIZE; |
| 227 |
|
| 228 |
/* allocate storage to track our page mappings */ |
| 229 |
bufmap->page_array = |
| 230 |
kzalloc_objs(struct page *, bufmap->page_count); |
| 231 |
if (!bufmap->page_array) |
| 232 |
goto out_free_desc_array; |
| 233 |
|
| 234 |
return bufmap; |
| 235 |
|
| 236 |
out_free_desc_array: |
| 237 |
kfree(bufmap->desc_array); |
| 238 |
out_free_index_array: |
| 239 |
bitmap_free(bufmap->buffer_index_array); |
| 240 |
out_free_bufmap: |
| 241 |
kfree(bufmap); |
| 242 |
out: |
| 243 |
return NULL; |
| 244 |
} |
| 245 |
|
| 246 |
static int |
| 247 |
orangefs_bufmap_map(struct orangefs_bufmap *bufmap, |
| 248 |
struct ORANGEFS_dev_map_desc *user_desc) |
| 249 |
{ |
| 250 |
int pages_per_desc = bufmap->desc_size / PAGE_SIZE; |
| 251 |
int offset = 0, ret, i; |
| 252 |
|
| 253 |
/* map the pages */ |
| 254 |
ret = pin_user_pages_fast((unsigned long)user_desc->ptr, |
| 255 |
bufmap->page_count, FOLL_WRITE, bufmap->page_array); |
| 256 |
|
| 257 |
if (ret < 0) |
| 258 |
return ret; |
| 259 |
|
| 260 |
if (ret != bufmap->page_count) { |
| 261 |
gossip_err("orangefs error: asked for %d pages, only got %d.\n", |
| 262 |
bufmap->page_count, ret); |
| 263 |
|
| 264 |
for (i = 0; i < ret; i++) |
| 265 |
unpin_user_page(bufmap->page_array[i]); |
| 266 |
return -ENOMEM; |
| 267 |
} |
| 268 |
|
| 269 |
/* |
| 270 |
* ideally we want to get kernel space pointers for each page, but |
| 271 |
* we can't kmap that many pages at once if highmem is being used. |
| 272 |
* so instead, we just kmap/kunmap the page address each time the |
| 273 |
* kaddr is needed. |
| 274 |
*/ |
| 275 |
for (i = 0; i < bufmap->page_count; i++) |
| 276 |
flush_dcache_page(bufmap->page_array[i]); |
| 277 |
|
| 278 |
/* build a list of available descriptors */ |
| 279 |
for (offset = 0, i = 0; i < bufmap->desc_count; i++) { |
| 280 |
bufmap->desc_array[i].page_array = &bufmap->page_array[offset]; |
| 281 |
bufmap->desc_array[i].array_count = pages_per_desc; |
| 282 |
bufmap->desc_array[i].uaddr = |
| 283 |
(user_desc->ptr + (i * pages_per_desc * PAGE_SIZE)); |
| 284 |
offset += pages_per_desc; |
| 285 |
} |
| 286 |
|
| 287 |
return 0; |
| 288 |
} |
| 289 |
|
| 290 |
/* |
| 291 |
* orangefs_bufmap_initialize() |
| 292 |
* |
| 293 |
* initializes the mapped buffer interface |
| 294 |
* |
| 295 |
* returns 0 on success, -errno on failure |
| 296 |
*/ |
| 297 |
int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc) |
| 298 |
{ |
| 299 |
struct orangefs_bufmap *bufmap; |
| 300 |
int ret = -EINVAL; |
| 301 |
|
| 302 |
gossip_debug(GOSSIP_BUFMAP_DEBUG, |
| 303 |
"orangefs_bufmap_initialize: called (ptr (" |
| 304 |
"%p) sz (%d) cnt(%d).\n", |
| 305 |
user_desc->ptr, |
| 306 |
user_desc->size, |
| 307 |
user_desc->count); |
| 308 |
|
| 309 |
if (user_desc->total_size < 0 || |
| 310 |
user_desc->size < 0 || |
| 311 |
user_desc->count < 0) |
| 312 |
goto out; |
| 313 |
|
| 314 |
/* |
| 315 |
* sanity check alignment and size of buffer that caller wants to |
| 316 |
* work with |
| 317 |
*/ |
| 318 |
if (PAGE_ALIGN((unsigned long)user_desc->ptr) != |
| 319 |
(unsigned long)user_desc->ptr) { |
| 320 |
gossip_err("orangefs error: memory alignment (front). %p\n", |
| 321 |
user_desc->ptr); |
| 322 |
goto out; |
| 323 |
} |
| 324 |
|
| 325 |
if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size)) |
| 326 |
!= (unsigned long)(user_desc->ptr + user_desc->total_size)) { |
| 327 |
gossip_err("orangefs error: memory alignment (back).(%p + %d)\n", |
| 328 |
user_desc->ptr, |
| 329 |
user_desc->total_size); |
| 330 |
goto out; |
| 331 |
} |
| 332 |
|
| 333 |
if (user_desc->total_size != (user_desc->size * user_desc->count)) { |
| 334 |
gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n", |
| 335 |
user_desc->total_size, |
| 336 |
user_desc->size, |
| 337 |
user_desc->count); |
| 338 |
goto out; |
| 339 |
} |
| 340 |
|
| 341 |
if ((user_desc->size % PAGE_SIZE) != 0) { |
| 342 |
gossip_err("orangefs error: bufmap size not page size divisible (%d).\n", |
| 343 |
user_desc->size); |
| 344 |
goto out; |
| 345 |
} |
| 346 |
|
| 347 |
ret = -ENOMEM; |
| 348 |
bufmap = orangefs_bufmap_alloc(user_desc); |
| 349 |
if (!bufmap) |
| 350 |
goto out; |
| 351 |
|
| 352 |
ret = orangefs_bufmap_map(bufmap, user_desc); |
| 353 |
if (ret) |
| 354 |
goto out_free_bufmap; |
| 355 |
|
| 356 |
|
| 357 |
spin_lock(&orangefs_bufmap_lock); |
| 358 |
if (__orangefs_bufmap) { |
| 359 |
spin_unlock(&orangefs_bufmap_lock); |
| 360 |
gossip_err("orangefs: error: bufmap already initialized.\n"); |
| 361 |
ret = -EINVAL; |
| 362 |
goto out_unmap_bufmap; |
| 363 |
} |
| 364 |
__orangefs_bufmap = bufmap; |
| 365 |
install(&rw_map, |
| 366 |
bufmap->desc_count, |
| 367 |
bufmap->buffer_index_array); |
| 368 |
install(&readdir_map, |
| 369 |
ORANGEFS_READDIR_DEFAULT_DESC_COUNT, |
| 370 |
bufmap->readdir_index_array); |
| 371 |
spin_unlock(&orangefs_bufmap_lock); |
| 372 |
|
| 373 |
gossip_debug(GOSSIP_BUFMAP_DEBUG, |
| 374 |
"orangefs_bufmap_initialize: exiting normally\n"); |
| 375 |
return 0; |
| 376 |
|
| 377 |
out_unmap_bufmap: |
| 378 |
orangefs_bufmap_unmap(bufmap); |
| 379 |
out_free_bufmap: |
| 380 |
orangefs_bufmap_free(bufmap); |
| 381 |
out: |
| 382 |
return ret; |
| 383 |
} |
| 384 |
|
| 385 |
/* |
| 386 |
* orangefs_bufmap_finalize() |
| 387 |
* |
| 388 |
* shuts down the mapped buffer interface and releases any resources |
| 389 |
* associated with it |
| 390 |
* |
| 391 |
* no return value |
| 392 |
*/ |
| 393 |
void orangefs_bufmap_finalize(void) |
| 394 |
{ |
| 395 |
struct orangefs_bufmap *bufmap = __orangefs_bufmap; |
| 396 |
if (!bufmap) |
| 397 |
return; |
| 398 |
gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n"); |
| 399 |
mark_killed(&rw_map); |
| 400 |
mark_killed(&readdir_map); |
| 401 |
gossip_debug(GOSSIP_BUFMAP_DEBUG, |
| 402 |
"orangefs_bufmap_finalize: exiting normally\n"); |
| 403 |
} |
| 404 |
|
| 405 |
void orangefs_bufmap_run_down(void) |
| 406 |
{ |
| 407 |
struct orangefs_bufmap *bufmap = __orangefs_bufmap; |
| 408 |
if (!bufmap) |
| 409 |
return; |
| 410 |
run_down(&rw_map); |
| 411 |
run_down(&readdir_map); |
| 412 |
spin_lock(&orangefs_bufmap_lock); |
| 413 |
__orangefs_bufmap = NULL; |
| 414 |
spin_unlock(&orangefs_bufmap_lock); |
| 415 |
orangefs_bufmap_unmap(bufmap); |
| 416 |
orangefs_bufmap_free(bufmap); |
| 417 |
} |
| 418 |
|
| 419 |
/* |
| 420 |
* orangefs_bufmap_get() |
| 421 |
* |
| 422 |
* gets a free mapped buffer descriptor, will sleep until one becomes |
| 423 |
* available if necessary |
| 424 |
* |
| 425 |
* returns slot on success, -errno on failure |
| 426 |
*/ |
| 427 |
int orangefs_bufmap_get(void) |
| 428 |
{ |
| 429 |
return get(&rw_map); |
| 430 |
} |
| 431 |
|
| 432 |
/* |
| 433 |
* orangefs_bufmap_put() |
| 434 |
* |
| 435 |
* returns a mapped buffer descriptor to the collection |
| 436 |
* |
| 437 |
* no return value |
| 438 |
*/ |
| 439 |
void orangefs_bufmap_put(int buffer_index) |
| 440 |
{ |
| 441 |
put(&rw_map, buffer_index); |
| 442 |
} |
| 443 |
|
| 444 |
/* |
| 445 |
* orangefs_readdir_index_get() |
| 446 |
* |
| 447 |
* gets a free descriptor, will sleep until one becomes |
| 448 |
* available if necessary. |
| 449 |
* Although the readdir buffers are not mapped into kernel space |
| 450 |
* we could do that at a later point of time. Regardless, these |
| 451 |
* indices are used by the client-core. |
| 452 |
* |
| 453 |
* returns slot on success, -errno on failure |
| 454 |
*/ |
| 455 |
int orangefs_readdir_index_get(void) |
| 456 |
{ |
| 457 |
return get(&readdir_map); |
| 458 |
} |
| 459 |
|
| 460 |
void orangefs_readdir_index_put(int buffer_index) |
| 461 |
{ |
| 462 |
put(&readdir_map, buffer_index); |
| 463 |
} |
| 464 |
|
| 465 |
/* |
| 466 |
* we've been handed an iovec, we need to copy it to |
| 467 |
* the shared memory descriptor at "buffer_index". |
| 468 |
*/ |
| 469 |
int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter, |
| 470 |
int buffer_index, |
| 471 |
size_t size) |
| 472 |
{ |
| 473 |
struct orangefs_bufmap_desc *to; |
| 474 |
int i; |
| 475 |
|
| 476 |
gossip_debug(GOSSIP_BUFMAP_DEBUG, |
| 477 |
"%s: buffer_index:%d: size:%zu:\n", |
| 478 |
__func__, buffer_index, size); |
| 479 |
|
| 480 |
to = &__orangefs_bufmap->desc_array[buffer_index]; |
| 481 |
for (i = 0; size; i++) { |
| 482 |
struct page *page = to->page_array[i]; |
| 483 |
size_t n = size; |
| 484 |
if (n > PAGE_SIZE) |
| 485 |
n = PAGE_SIZE; |
| 486 |
if (copy_page_from_iter(page, 0, n, iter) != n) |
| 487 |
return -EFAULT; |
| 488 |
size -= n; |
| 489 |
} |
| 490 |
return 0; |
| 491 |
} |
| 492 |
|
| 493 |
/* |
| 494 |
* we've been handed an iovec, we need to fill it from |
| 495 |
* the shared memory descriptor at "buffer_index". |
| 496 |
*/ |
| 497 |
int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter, |
| 498 |
int buffer_index, |
| 499 |
size_t size) |
| 500 |
{ |
| 501 |
struct orangefs_bufmap_desc *from; |
| 502 |
int i; |
| 503 |
|
| 504 |
from = &__orangefs_bufmap->desc_array[buffer_index]; |
| 505 |
gossip_debug(GOSSIP_BUFMAP_DEBUG, |
| 506 |
"%s: buffer_index:%d: size:%zu:\n", |
| 507 |
__func__, buffer_index, size); |
| 508 |
|
| 509 |
|
| 510 |
for (i = 0; size; i++) { |
| 511 |
struct page *page = from->page_array[i]; |
| 512 |
size_t n = size; |
| 513 |
if (n > PAGE_SIZE) |
| 514 |
n = PAGE_SIZE; |
| 515 |
n = copy_page_to_iter(page, 0, n, iter); |
| 516 |
if (!n) |
| 517 |
return -EFAULT; |
| 518 |
size -= n; |
| 519 |
} |
| 520 |
return 0; |
| 521 |
} |