fs/orangefs/orangefs-bufmap.c

   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 (unlikely(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         int i;
 172
 173         for (i = 0; i < bufmap->page_count; i++)
 174                 put_page(bufmap->page_array[i]);
 175 }
 176
 177 static void
 178 orangefs_bufmap_free(struct orangefs_bufmap *bufmap)
 179 {
 180         kfree(bufmap->page_array);
 181         kfree(bufmap->desc_array);
 182         kfree(bufmap->buffer_index_array);
 183         kfree(bufmap);
 184 }
 185
 186 /*
 187  * XXX: Can the size and shift change while the caller gives up the
 188  * XXX: lock between calling this and doing something useful?
 189  */
 190
 191 int orangefs_bufmap_size_query(void)
 192 {
 193         struct orangefs_bufmap *bufmap;
 194         int size = 0;
 195         spin_lock(&orangefs_bufmap_lock);
 196         bufmap = __orangefs_bufmap;
 197         if (bufmap)
 198                 size = bufmap->desc_size;
 199         spin_unlock(&orangefs_bufmap_lock);
 200         return size;
 201 }
 202
 203 int orangefs_bufmap_shift_query(void)
 204 {
 205         struct orangefs_bufmap *bufmap;
 206         int shift = 0;
 207         spin_lock(&orangefs_bufmap_lock);
 208         bufmap = __orangefs_bufmap;
 209         if (bufmap)
 210                 shift = bufmap->desc_shift;
 211         spin_unlock(&orangefs_bufmap_lock);
 212         return shift;
 213 }
 214
 215 static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
 216 static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq);
 217
 218 static struct orangefs_bufmap *
 219 orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc)
 220 {
 221         struct orangefs_bufmap *bufmap;
 222
 223         bufmap = kzalloc(sizeof(*bufmap), GFP_KERNEL);
 224         if (!bufmap)
 225                 goto out;
 226
 227         bufmap->total_size = user_desc->total_size;
 228         bufmap->desc_count = user_desc->count;
 229         bufmap->desc_size = user_desc->size;
 230         bufmap->desc_shift = ilog2(bufmap->desc_size);
 231
 232         bufmap->buffer_index_array =
 233                 kzalloc(DIV_ROUND_UP(bufmap->desc_count, BITS_PER_LONG), GFP_KERNEL);
 234         if (!bufmap->buffer_index_array)
 235                 goto out_free_bufmap;
 236
 237         bufmap->desc_array =
 238                 kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
 239                         GFP_KERNEL);
 240         if (!bufmap->desc_array)
 241                 goto out_free_index_array;
 242
 243         bufmap->page_count = bufmap->total_size / PAGE_SIZE;
 244
 245         /* allocate storage to track our page mappings */
 246         bufmap->page_array =
 247                 kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
 248         if (!bufmap->page_array)
 249                 goto out_free_desc_array;
 250
 251         return bufmap;
 252
 253 out_free_desc_array:
 254         kfree(bufmap->desc_array);
 255 out_free_index_array:
 256         kfree(bufmap->buffer_index_array);
 257 out_free_bufmap:
 258         kfree(bufmap);
 259 out:
 260         return NULL;
 261 }
 262
 263 static int
 264 orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
 265                 struct ORANGEFS_dev_map_desc *user_desc)
 266 {
 267         int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
 268         int offset = 0, ret, i;
 269
 270         /* map the pages */
 271         ret = get_user_pages_fast((unsigned long)user_desc->ptr,
 272                              bufmap->page_count, 1, bufmap->page_array);
 273
 274         if (ret < 0)
 275                 return ret;
 276
 277         if (ret != bufmap->page_count) {
 278                 gossip_err("orangefs error: asked for %d pages, only got %d.\n",
 279                                 bufmap->page_count, ret);
 280
 281                 for (i = 0; i < ret; i++) {
 282                         SetPageError(bufmap->page_array[i]);
 283                         put_page(bufmap->page_array[i]);
 284                 }
 285                 return -ENOMEM;
 286         }
 287
 288         /*
 289          * ideally we want to get kernel space pointers for each page, but
 290          * we can't kmap that many pages at once if highmem is being used.
 291          * so instead, we just kmap/kunmap the page address each time the
 292          * kaddr is needed.
 293          */
 294         for (i = 0; i < bufmap->page_count; i++)
 295                 flush_dcache_page(bufmap->page_array[i]);
 296
 297         /* build a list of available descriptors */
 298         for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
 299                 bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
 300                 bufmap->desc_array[i].array_count = pages_per_desc;
 301                 bufmap->desc_array[i].uaddr =
 302                     (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
 303                 offset += pages_per_desc;
 304         }
 305
 306         return 0;
 307 }
 308
 309 /*
 310  * orangefs_bufmap_initialize()
 311  *
 312  * initializes the mapped buffer interface
 313  *
 314  * returns 0 on success, -errno on failure
 315  */
 316 int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
 317 {
 318         struct orangefs_bufmap *bufmap;
 319         int ret = -EINVAL;
 320
 321         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 322                      "orangefs_bufmap_initialize: called (ptr ("
 323                      "%p) sz (%d) cnt(%d).\n",
 324                      user_desc->ptr,
 325                      user_desc->size,
 326                      user_desc->count);
 327
 328         if (user_desc->total_size < 0 ||
 329             user_desc->size < 0 ||
 330             user_desc->count < 0)
 331                 goto out;
 332
 333         /*
 334          * sanity check alignment and size of buffer that caller wants to
 335          * work with
 336          */
 337         if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
 338             (unsigned long)user_desc->ptr) {
 339                 gossip_err("orangefs error: memory alignment (front). %p\n",
 340                            user_desc->ptr);
 341                 goto out;
 342         }
 343
 344         if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
 345             != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
 346                 gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
 347                            user_desc->ptr,
 348                            user_desc->total_size);
 349                 goto out;
 350         }
 351
 352         if (user_desc->total_size != (user_desc->size * user_desc->count)) {
 353                 gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
 354                            user_desc->total_size,
 355                            user_desc->size,
 356                            user_desc->count);
 357                 goto out;
 358         }
 359
 360         if ((user_desc->size % PAGE_SIZE) != 0) {
 361                 gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
 362                            user_desc->size);
 363                 goto out;
 364         }
 365
 366         ret = -ENOMEM;
 367         bufmap = orangefs_bufmap_alloc(user_desc);
 368         if (!bufmap)
 369                 goto out;
 370
 371         ret = orangefs_bufmap_map(bufmap, user_desc);
 372         if (ret)
 373                 goto out_free_bufmap;
 374
 375
 376         spin_lock(&orangefs_bufmap_lock);
 377         if (__orangefs_bufmap) {
 378                 spin_unlock(&orangefs_bufmap_lock);
 379                 gossip_err("orangefs: error: bufmap already initialized.\n");
 380                 ret = -EINVAL;
 381                 goto out_unmap_bufmap;
 382         }
 383         __orangefs_bufmap = bufmap;
 384         install(&rw_map,
 385                 bufmap->desc_count,
 386                 bufmap->buffer_index_array);
 387         install(&readdir_map,
 388                 ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
 389                 bufmap->readdir_index_array);
 390         spin_unlock(&orangefs_bufmap_lock);
 391
 392         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 393                      "orangefs_bufmap_initialize: exiting normally\n");
 394         return 0;
 395
 396 out_unmap_bufmap:
 397         orangefs_bufmap_unmap(bufmap);
 398 out_free_bufmap:
 399         orangefs_bufmap_free(bufmap);
 400 out:
 401         return ret;
 402 }
 403
 404 /*
 405  * orangefs_bufmap_finalize()
 406  *
 407  * shuts down the mapped buffer interface and releases any resources
 408  * associated with it
 409  *
 410  * no return value
 411  */
 412 void orangefs_bufmap_finalize(void)
 413 {
 414         struct orangefs_bufmap *bufmap = __orangefs_bufmap;
 415         if (!bufmap)
 416                 return;
 417         gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
 418         mark_killed(&rw_map);
 419         mark_killed(&readdir_map);
 420         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 421                      "orangefs_bufmap_finalize: exiting normally\n");
 422 }
 423
 424 void orangefs_bufmap_run_down(void)
 425 {
 426         struct orangefs_bufmap *bufmap = __orangefs_bufmap;
 427         if (!bufmap)
 428                 return;
 429         run_down(&rw_map);
 430         run_down(&readdir_map);
 431         spin_lock(&orangefs_bufmap_lock);
 432         __orangefs_bufmap = NULL;
 433         spin_unlock(&orangefs_bufmap_lock);
 434         orangefs_bufmap_unmap(bufmap);
 435         orangefs_bufmap_free(bufmap);
 436 }
 437
 438 /*
 439  * orangefs_bufmap_get()
 440  *
 441  * gets a free mapped buffer descriptor, will sleep until one becomes
 442  * available if necessary
 443  *
 444  * returns slot on success, -errno on failure
 445  */
 446 int orangefs_bufmap_get(void)
 447 {
 448         return get(&rw_map);
 449 }
 450
 451 /*
 452  * orangefs_bufmap_put()
 453  *
 454  * returns a mapped buffer descriptor to the collection
 455  *
 456  * no return value
 457  */
 458 void orangefs_bufmap_put(int buffer_index)
 459 {
 460         put(&rw_map, buffer_index);
 461 }
 462
 463 /*
 464  * orangefs_readdir_index_get()
 465  *
 466  * gets a free descriptor, will sleep until one becomes
 467  * available if necessary.
 468  * Although the readdir buffers are not mapped into kernel space
 469  * we could do that at a later point of time. Regardless, these
 470  * indices are used by the client-core.
 471  *
 472  * returns slot on success, -errno on failure
 473  */
 474 int orangefs_readdir_index_get(void)
 475 {
 476         return get(&readdir_map);
 477 }
 478
 479 void orangefs_readdir_index_put(int buffer_index)
 480 {
 481         put(&readdir_map, buffer_index);
 482 }
 483
 484 /*
 485  * we've been handed an iovec, we need to copy it to
 486  * the shared memory descriptor at "buffer_index".
 487  */
 488 int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
 489                                 int buffer_index,
 490                                 size_t size)
 491 {
 492         struct orangefs_bufmap_desc *to;
 493         int i;
 494
 495         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 496                      "%s: buffer_index:%d: size:%zu:\n",
 497                      __func__, buffer_index, size);
 498
 499         to = &__orangefs_bufmap->desc_array[buffer_index];
 500         for (i = 0; size; i++) {
 501                 struct page *page = to->page_array[i];
 502                 size_t n = size;
 503                 if (n > PAGE_SIZE)
 504                         n = PAGE_SIZE;
 505                 if (copy_page_from_iter(page, 0, n, iter) != n)
 506                         return -EFAULT;
 507                 size -= n;
 508         }
 509         return 0;
 510 }
 511
 512 /*
 513  * we've been handed an iovec, we need to fill it from
 514  * the shared memory descriptor at "buffer_index".
 515  */
 516 int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
 517                                     int buffer_index,
 518                                     size_t size)
 519 {
 520         struct orangefs_bufmap_desc *from;
 521         int i;
 522
 523         from = &__orangefs_bufmap->desc_array[buffer_index];
 524         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 525                      "%s: buffer_index:%d: size:%zu:\n",
 526                      __func__, buffer_index, size);
 527
 528
 529         for (i = 0; size; i++) {
 530                 struct page *page = from->page_array[i];
 531                 size_t n = size;
 532                 if (n > PAGE_SIZE)
 533                         n = PAGE_SIZE;
 534                 n = copy_page_to_iter(page, 0, n, iter);
 535                 if (!n)
 536                         return -EFAULT;
 537                 size -= n;
 538         }
 539         return 0;
 540 }