fs/orangefs/orangefs-bufmap.c

   1 /*
   2  * (C) 2001 Clemson University and The University of Chicago
   3  *
   4  * See COPYING in top-level directory.
   5  */
   6 #include "protocol.h"
   7 #include "orangefs-kernel.h"
   8 #include "orangefs-bufmap.h"
   9
  10 struct slot_map {
  11         int c;
  12         wait_queue_head_t q;
  13         int count;
  14         unsigned long *map;
  15 };
  16
  17 static struct slot_map rw_map = {
  18         .c = -1,
  19         .q = __WAIT_QUEUE_HEAD_INITIALIZER(rw_map.q)
  20 };
  21 static struct slot_map readdir_map = {
  22         .c = -1,
  23         .q = __WAIT_QUEUE_HEAD_INITIALIZER(readdir_map.q)
  24 };
  25
  26
  27 static void install(struct slot_map *m, int count, unsigned long *map)
  28 {
  29         spin_lock(&m->q.lock);
  30         m->c = m->count = count;
  31         m->map = map;
  32         wake_up_all_locked(&m->q);
  33         spin_unlock(&m->q.lock);
  34 }
  35
  36 static void mark_killed(struct slot_map *m)
  37 {
  38         spin_lock(&m->q.lock);
  39         m->c -= m->count + 1;
  40         spin_unlock(&m->q.lock);
  41 }
  42
  43 static void run_down(struct slot_map *m)
  44 {
  45         DEFINE_WAIT(wait);
  46         spin_lock(&m->q.lock);
  47         if (m->c != -1) {
  48                 for (;;) {
  49                         if (likely(list_empty(&wait.task_list)))
  50                                 __add_wait_queue_tail(&m->q, &wait);
  51                         set_current_state(TASK_UNINTERRUPTIBLE);
  52
  53                         if (m->c == -1)
  54                                 break;
  55
  56                         spin_unlock(&m->q.lock);
  57                         schedule();
  58                         spin_lock(&m->q.lock);
  59                 }
  60                 __remove_wait_queue(&m->q, &wait);
  61                 __set_current_state(TASK_RUNNING);
  62         }
  63         m->map = NULL;
  64         spin_unlock(&m->q.lock);
  65 }
  66
  67 static void put(struct slot_map *m, int slot)
  68 {
  69         int v;
  70         spin_lock(&m->q.lock);
  71         __clear_bit(slot, m->map);
  72         v = ++m->c;
  73         if (unlikely(v == 1))   /* no free slots -> one free slot */
  74                 wake_up_locked(&m->q);
  75         else if (unlikely(v == -1))     /* finished dying */
  76                 wake_up_all_locked(&m->q);
  77         spin_unlock(&m->q.lock);
  78 }
  79
  80 static int wait_for_free(struct slot_map *m)
  81 {
  82         long left = slot_timeout_secs * HZ;
  83         DEFINE_WAIT(wait);
  84
  85         do {
  86                 long n = left, t;
  87                 if (likely(list_empty(&wait.task_list)))
  88                         __add_wait_queue_tail_exclusive(&m->q, &wait);
  89                 set_current_state(TASK_INTERRUPTIBLE);
  90
  91                 if (m->c > 0)
  92                         break;
  93
  94                 if (m->c < 0) {
  95                         /* we are waiting for map to be installed */
  96                         /* it would better be there soon, or we go away */
  97                         if (n > ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ)
  98                                 n = ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ;
  99                 }
 100                 spin_unlock(&m->q.lock);
 101                 t = schedule_timeout(n);
 102                 spin_lock(&m->q.lock);
 103                 if (unlikely(!t) && n != left && m->c < 0)
 104                         left = t;
 105                 else
 106                         left = t + (left - n);
 107                 if (unlikely(signal_pending(current)))
 108                         left = -EINTR;
 109         } while (left > 0);
 110
 111         if (!list_empty(&wait.task_list))
 112                 list_del(&wait.task_list);
 113         else if (left <= 0 && waitqueue_active(&m->q))
 114                 __wake_up_locked_key(&m->q, TASK_INTERRUPTIBLE, NULL);
 115         __set_current_state(TASK_RUNNING);
 116
 117         if (likely(left > 0))
 118                 return 0;
 119
 120         return left < 0 ? -EINTR : -ETIMEDOUT;
 121 }
 122
 123 static int get(struct slot_map *m)
 124 {
 125         int res = 0;
 126         spin_lock(&m->q.lock);
 127         if (unlikely(m->c <= 0))
 128                 res = wait_for_free(m);
 129         if (likely(!res)) {
 130                 m->c--;
 131                 res = find_first_zero_bit(m->map, m->count);
 132                 __set_bit(res, m->map);
 133         }
 134         spin_unlock(&m->q.lock);
 135         return res;
 136 }
 137
 138 /* used to describe mapped buffers */
 139 struct orangefs_bufmap_desc {
 140         void *uaddr;                    /* user space address pointer */
 141         struct page **page_array;       /* array of mapped pages */
 142         int array_count;                /* size of above arrays */
 143         struct list_head list_link;
 144 };
 145
 146 static struct orangefs_bufmap {
 147         int desc_size;
 148         int desc_shift;
 149         int desc_count;
 150         int total_size;
 151         int page_count;
 152
 153         struct page **page_array;
 154         struct orangefs_bufmap_desc *desc_array;
 155
 156         /* array to track usage of buffer descriptors */
 157         unsigned long *buffer_index_array;
 158
 159         /* array to track usage of buffer descriptors for readdir */
 160 #define N DIV_ROUND_UP(ORANGEFS_READDIR_DEFAULT_DESC_COUNT, BITS_PER_LONG)
 161         unsigned long readdir_index_array[N];
 162 #undef N
 163 } *__orangefs_bufmap;
 164
 165 static DEFINE_SPINLOCK(orangefs_bufmap_lock);
 166
 167 static void
 168 orangefs_bufmap_unmap(struct orangefs_bufmap *bufmap)
 169 {
 170         int i;
 171
 172         for (i = 0; i < bufmap->page_count; i++)
 173                 put_page(bufmap->page_array[i]);
 174 }
 175
 176 static void
 177 orangefs_bufmap_free(struct orangefs_bufmap *bufmap)
 178 {
 179         kfree(bufmap->page_array);
 180         kfree(bufmap->desc_array);
 181         kfree(bufmap->buffer_index_array);
 182         kfree(bufmap);
 183 }
 184
 185 /*
 186  * XXX: Can the size and shift change while the caller gives up the
 187  * XXX: lock between calling this and doing something useful?
 188  */
 189
 190 int orangefs_bufmap_size_query(void)
 191 {
 192         struct orangefs_bufmap *bufmap;
 193         int size = 0;
 194         spin_lock(&orangefs_bufmap_lock);
 195         bufmap = __orangefs_bufmap;
 196         if (bufmap)
 197                 size = bufmap->desc_size;
 198         spin_unlock(&orangefs_bufmap_lock);
 199         return size;
 200 }
 201
 202 int orangefs_bufmap_shift_query(void)
 203 {
 204         struct orangefs_bufmap *bufmap;
 205         int shift = 0;
 206         spin_lock(&orangefs_bufmap_lock);
 207         bufmap = __orangefs_bufmap;
 208         if (bufmap)
 209                 shift = bufmap->desc_shift;
 210         spin_unlock(&orangefs_bufmap_lock);
 211         return shift;
 212 }
 213
 214 static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
 215 static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq);
 216
 217 /*
 218  * orangefs_get_bufmap_init
 219  *
 220  * If bufmap_init is 1, then the shared memory system, including the
 221  * buffer_index_array, is available.  Otherwise, it is not.
 222  *
 223  * returns the value of bufmap_init
 224  */
 225 int orangefs_get_bufmap_init(void)
 226 {
 227         return __orangefs_bufmap ? 1 : 0;
 228 }
 229
 230
 231 static struct orangefs_bufmap *
 232 orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc)
 233 {
 234         struct orangefs_bufmap *bufmap;
 235
 236         bufmap = kzalloc(sizeof(*bufmap), GFP_KERNEL);
 237         if (!bufmap)
 238                 goto out;
 239
 240         bufmap->total_size = user_desc->total_size;
 241         bufmap->desc_count = user_desc->count;
 242         bufmap->desc_size = user_desc->size;
 243         bufmap->desc_shift = ilog2(bufmap->desc_size);
 244
 245         bufmap->buffer_index_array =
 246                 kzalloc(DIV_ROUND_UP(bufmap->desc_count, BITS_PER_LONG), GFP_KERNEL);
 247         if (!bufmap->buffer_index_array) {
 248                 gossip_err("orangefs: could not allocate %d buffer indices\n",
 249                                 bufmap->desc_count);
 250                 goto out_free_bufmap;
 251         }
 252
 253         bufmap->desc_array =
 254                 kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
 255                         GFP_KERNEL);
 256         if (!bufmap->desc_array) {
 257                 gossip_err("orangefs: could not allocate %d descriptors\n",
 258                                 bufmap->desc_count);
 259                 goto out_free_index_array;
 260         }
 261
 262         bufmap->page_count = bufmap->total_size / PAGE_SIZE;
 263
 264         /* allocate storage to track our page mappings */
 265         bufmap->page_array =
 266                 kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
 267         if (!bufmap->page_array)
 268                 goto out_free_desc_array;
 269
 270         return bufmap;
 271
 272 out_free_desc_array:
 273         kfree(bufmap->desc_array);
 274 out_free_index_array:
 275         kfree(bufmap->buffer_index_array);
 276 out_free_bufmap:
 277         kfree(bufmap);
 278 out:
 279         return NULL;
 280 }
 281
 282 static int
 283 orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
 284                 struct ORANGEFS_dev_map_desc *user_desc)
 285 {
 286         int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
 287         int offset = 0, ret, i;
 288
 289         /* map the pages */
 290         ret = get_user_pages_fast((unsigned long)user_desc->ptr,
 291                              bufmap->page_count, 1, bufmap->page_array);
 292
 293         if (ret < 0)
 294                 return ret;
 295
 296         if (ret != bufmap->page_count) {
 297                 gossip_err("orangefs error: asked for %d pages, only got %d.\n",
 298                                 bufmap->page_count, ret);
 299
 300                 for (i = 0; i < ret; i++) {
 301                         SetPageError(bufmap->page_array[i]);
 302                         put_page(bufmap->page_array[i]);
 303                 }
 304                 return -ENOMEM;
 305         }
 306
 307         /*
 308          * ideally we want to get kernel space pointers for each page, but
 309          * we can't kmap that many pages at once if highmem is being used.
 310          * so instead, we just kmap/kunmap the page address each time the
 311          * kaddr is needed.
 312          */
 313         for (i = 0; i < bufmap->page_count; i++)
 314                 flush_dcache_page(bufmap->page_array[i]);
 315
 316         /* build a list of available descriptors */
 317         for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
 318                 bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
 319                 bufmap->desc_array[i].array_count = pages_per_desc;
 320                 bufmap->desc_array[i].uaddr =
 321                     (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
 322                 offset += pages_per_desc;
 323         }
 324
 325         return 0;
 326 }
 327
 328 /*
 329  * orangefs_bufmap_initialize()
 330  *
 331  * initializes the mapped buffer interface
 332  *
 333  * returns 0 on success, -errno on failure
 334  */
 335 int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
 336 {
 337         struct orangefs_bufmap *bufmap;
 338         int ret = -EINVAL;
 339
 340         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 341                      "orangefs_bufmap_initialize: called (ptr ("
 342                      "%p) sz (%d) cnt(%d).\n",
 343                      user_desc->ptr,
 344                      user_desc->size,
 345                      user_desc->count);
 346
 347         if (user_desc->total_size < 0 ||
 348             user_desc->size < 0 ||
 349             user_desc->count < 0)
 350                 goto out;
 351
 352         /*
 353          * sanity check alignment and size of buffer that caller wants to
 354          * work with
 355          */
 356         if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
 357             (unsigned long)user_desc->ptr) {
 358                 gossip_err("orangefs error: memory alignment (front). %p\n",
 359                            user_desc->ptr);
 360                 goto out;
 361         }
 362
 363         if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
 364             != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
 365                 gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
 366                            user_desc->ptr,
 367                            user_desc->total_size);
 368                 goto out;
 369         }
 370
 371         if (user_desc->total_size != (user_desc->size * user_desc->count)) {
 372                 gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
 373                            user_desc->total_size,
 374                            user_desc->size,
 375                            user_desc->count);
 376                 goto out;
 377         }
 378
 379         if ((user_desc->size % PAGE_SIZE) != 0) {
 380                 gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
 381                            user_desc->size);
 382                 goto out;
 383         }
 384
 385         ret = -ENOMEM;
 386         bufmap = orangefs_bufmap_alloc(user_desc);
 387         if (!bufmap)
 388                 goto out;
 389
 390         ret = orangefs_bufmap_map(bufmap, user_desc);
 391         if (ret)
 392                 goto out_free_bufmap;
 393
 394
 395         spin_lock(&orangefs_bufmap_lock);
 396         if (__orangefs_bufmap) {
 397                 spin_unlock(&orangefs_bufmap_lock);
 398                 gossip_err("orangefs: error: bufmap already initialized.\n");
 399                 ret = -EINVAL;
 400                 goto out_unmap_bufmap;
 401         }
 402         __orangefs_bufmap = bufmap;
 403         install(&rw_map,
 404                 bufmap->desc_count,
 405                 bufmap->buffer_index_array);
 406         install(&readdir_map,
 407                 ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
 408                 bufmap->readdir_index_array);
 409         spin_unlock(&orangefs_bufmap_lock);
 410
 411         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 412                      "orangefs_bufmap_initialize: exiting normally\n");
 413         return 0;
 414
 415 out_unmap_bufmap:
 416         orangefs_bufmap_unmap(bufmap);
 417 out_free_bufmap:
 418         orangefs_bufmap_free(bufmap);
 419 out:
 420         return ret;
 421 }
 422
 423 /*
 424  * orangefs_bufmap_finalize()
 425  *
 426  * shuts down the mapped buffer interface and releases any resources
 427  * associated with it
 428  *
 429  * no return value
 430  */
 431 void orangefs_bufmap_finalize(void)
 432 {
 433         struct orangefs_bufmap *bufmap = __orangefs_bufmap;
 434         if (!bufmap)
 435                 return;
 436         gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
 437         mark_killed(&rw_map);
 438         mark_killed(&readdir_map);
 439         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 440                      "orangefs_bufmap_finalize: exiting normally\n");
 441 }
 442
 443 void orangefs_bufmap_run_down(void)
 444 {
 445         struct orangefs_bufmap *bufmap = __orangefs_bufmap;
 446         if (!bufmap)
 447                 return;
 448         run_down(&rw_map);
 449         run_down(&readdir_map);
 450         spin_lock(&orangefs_bufmap_lock);
 451         __orangefs_bufmap = NULL;
 452         spin_unlock(&orangefs_bufmap_lock);
 453         orangefs_bufmap_unmap(bufmap);
 454         orangefs_bufmap_free(bufmap);
 455 }
 456
 457 /*
 458  * orangefs_bufmap_get()
 459  *
 460  * gets a free mapped buffer descriptor, will sleep until one becomes
 461  * available if necessary
 462  *
 463  * returns slot on success, -errno on failure
 464  */
 465 int orangefs_bufmap_get(void)
 466 {
 467         return get(&rw_map);
 468 }
 469
 470 /*
 471  * orangefs_bufmap_put()
 472  *
 473  * returns a mapped buffer descriptor to the collection
 474  *
 475  * no return value
 476  */
 477 void orangefs_bufmap_put(int buffer_index)
 478 {
 479         put(&rw_map, buffer_index);
 480 }
 481
 482 /*
 483  * orangefs_readdir_index_get()
 484  *
 485  * gets a free descriptor, will sleep until one becomes
 486  * available if necessary.
 487  * Although the readdir buffers are not mapped into kernel space
 488  * we could do that at a later point of time. Regardless, these
 489  * indices are used by the client-core.
 490  *
 491  * returns slot on success, -errno on failure
 492  */
 493 int orangefs_readdir_index_get(void)
 494 {
 495         return get(&readdir_map);
 496 }
 497
 498 void orangefs_readdir_index_put(int buffer_index)
 499 {
 500         put(&readdir_map, buffer_index);
 501 }
 502
 503 /*
 504  * we've been handed an iovec, we need to copy it to
 505  * the shared memory descriptor at "buffer_index".
 506  */
 507 int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
 508                                 int buffer_index,
 509                                 size_t size)
 510 {
 511         struct orangefs_bufmap_desc *to;
 512         int i;
 513
 514         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 515                      "%s: buffer_index:%d: size:%zu:\n",
 516                      __func__, buffer_index, size);
 517
 518         to = &__orangefs_bufmap->desc_array[buffer_index];
 519         for (i = 0; size; i++) {
 520                 struct page *page = to->page_array[i];
 521                 size_t n = size;
 522                 if (n > PAGE_SIZE)
 523                         n = PAGE_SIZE;
 524                 if (copy_page_from_iter(page, 0, n, iter) != n)
 525                         return -EFAULT;
 526                 size -= n;
 527         }
 528         return 0;
 529 }
 530
 531 /*
 532  * we've been handed an iovec, we need to fill it from
 533  * the shared memory descriptor at "buffer_index".
 534  */
 535 int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
 536                                     int buffer_index,
 537                                     size_t size)
 538 {
 539         struct orangefs_bufmap_desc *from;
 540         int i;
 541
 542         from = &__orangefs_bufmap->desc_array[buffer_index];
 543         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 544                      "%s: buffer_index:%d: size:%zu:\n",
 545                      __func__, buffer_index, size);
 546
 547
 548         for (i = 0; size; i++) {
 549                 struct page *page = from->page_array[i];
 550                 size_t n = size;
 551                 if (n > PAGE_SIZE)
 552                         n = PAGE_SIZE;
 553                 n = copy_page_to_iter(page, 0, n, iter);
 554                 if (!n)
 555                         return -EFAULT;
 556                 size -= n;
 557         }
 558         return 0;
 559 }