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 (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         kfree(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 int orangefs_bufmap_shift_query(void)
 201 {
 202         struct orangefs_bufmap *bufmap;
 203         int shift = 0;
 204         spin_lock(&orangefs_bufmap_lock);
 205         bufmap = __orangefs_bufmap;
 206         if (bufmap)
 207                 shift = bufmap->desc_shift;
 208         spin_unlock(&orangefs_bufmap_lock);
 209         return shift;
 210 }
 211
 212 static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
 213 static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq);
 214
 215 static struct orangefs_bufmap *
 216 orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc)
 217 {
 218         struct orangefs_bufmap *bufmap;
 219
 220         bufmap = kzalloc(sizeof(*bufmap), GFP_KERNEL);
 221         if (!bufmap)
 222                 goto out;
 223
 224         bufmap->total_size = user_desc->total_size;
 225         bufmap->desc_count = user_desc->count;
 226         bufmap->desc_size = user_desc->size;
 227         bufmap->desc_shift = ilog2(bufmap->desc_size);
 228
 229         bufmap->buffer_index_array =
 230                 kzalloc(DIV_ROUND_UP(bufmap->desc_count, BITS_PER_LONG), GFP_KERNEL);
 231         if (!bufmap->buffer_index_array)
 232                 goto out_free_bufmap;
 233
 234         bufmap->desc_array =
 235                 kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
 236                         GFP_KERNEL);
 237         if (!bufmap->desc_array)
 238                 goto out_free_index_array;
 239
 240         bufmap->page_count = bufmap->total_size / PAGE_SIZE;
 241
 242         /* allocate storage to track our page mappings */
 243         bufmap->page_array =
 244                 kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
 245         if (!bufmap->page_array)
 246                 goto out_free_desc_array;
 247
 248         return bufmap;
 249
 250 out_free_desc_array:
 251         kfree(bufmap->desc_array);
 252 out_free_index_array:
 253         kfree(bufmap->buffer_index_array);
 254 out_free_bufmap:
 255         kfree(bufmap);
 256 out:
 257         return NULL;
 258 }
 259
 260 static int
 261 orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
 262                 struct ORANGEFS_dev_map_desc *user_desc)
 263 {
 264         int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
 265         int offset = 0, ret, i;
 266
 267         /* map the pages */
 268         ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
 269                              bufmap->page_count, FOLL_WRITE, bufmap->page_array);
 270
 271         if (ret < 0)
 272                 return ret;
 273
 274         if (ret != bufmap->page_count) {
 275                 gossip_err("orangefs error: asked for %d pages, only got %d.\n",
 276                                 bufmap->page_count, ret);
 277
 278                 for (i = 0; i < ret; i++) {
 279                         SetPageError(bufmap->page_array[i]);
 280                         unpin_user_page(bufmap->page_array[i]);
 281                 }
 282                 return -ENOMEM;
 283         }
 284
 285         /*
 286          * ideally we want to get kernel space pointers for each page, but
 287          * we can't kmap that many pages at once if highmem is being used.
 288          * so instead, we just kmap/kunmap the page address each time the
 289          * kaddr is needed.
 290          */
 291         for (i = 0; i < bufmap->page_count; i++)
 292                 flush_dcache_page(bufmap->page_array[i]);
 293
 294         /* build a list of available descriptors */
 295         for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
 296                 bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
 297                 bufmap->desc_array[i].array_count = pages_per_desc;
 298                 bufmap->desc_array[i].uaddr =
 299                     (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
 300                 offset += pages_per_desc;
 301         }
 302
 303         return 0;
 304 }
 305
 306 /*
 307  * orangefs_bufmap_initialize()
 308  *
 309  * initializes the mapped buffer interface
 310  *
 311  * returns 0 on success, -errno on failure
 312  */
 313 int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
 314 {
 315         struct orangefs_bufmap *bufmap;
 316         int ret = -EINVAL;
 317
 318         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 319                      "orangefs_bufmap_initialize: called (ptr ("
 320                      "%p) sz (%d) cnt(%d).\n",
 321                      user_desc->ptr,
 322                      user_desc->size,
 323                      user_desc->count);
 324
 325         if (user_desc->total_size < 0 ||
 326             user_desc->size < 0 ||
 327             user_desc->count < 0)
 328                 goto out;
 329
 330         /*
 331          * sanity check alignment and size of buffer that caller wants to
 332          * work with
 333          */
 334         if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
 335             (unsigned long)user_desc->ptr) {
 336                 gossip_err("orangefs error: memory alignment (front). %p\n",
 337                            user_desc->ptr);
 338                 goto out;
 339         }
 340
 341         if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
 342             != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
 343                 gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
 344                            user_desc->ptr,
 345                            user_desc->total_size);
 346                 goto out;
 347         }
 348
 349         if (user_desc->total_size != (user_desc->size * user_desc->count)) {
 350                 gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
 351                            user_desc->total_size,
 352                            user_desc->size,
 353                            user_desc->count);
 354                 goto out;
 355         }
 356
 357         if ((user_desc->size % PAGE_SIZE) != 0) {
 358                 gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
 359                            user_desc->size);
 360                 goto out;
 361         }
 362
 363         ret = -ENOMEM;
 364         bufmap = orangefs_bufmap_alloc(user_desc);
 365         if (!bufmap)
 366                 goto out;
 367
 368         ret = orangefs_bufmap_map(bufmap, user_desc);
 369         if (ret)
 370                 goto out_free_bufmap;
 371
 372
 373         spin_lock(&orangefs_bufmap_lock);
 374         if (__orangefs_bufmap) {
 375                 spin_unlock(&orangefs_bufmap_lock);
 376                 gossip_err("orangefs: error: bufmap already initialized.\n");
 377                 ret = -EINVAL;
 378                 goto out_unmap_bufmap;
 379         }
 380         __orangefs_bufmap = bufmap;
 381         install(&rw_map,
 382                 bufmap->desc_count,
 383                 bufmap->buffer_index_array);
 384         install(&readdir_map,
 385                 ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
 386                 bufmap->readdir_index_array);
 387         spin_unlock(&orangefs_bufmap_lock);
 388
 389         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 390                      "orangefs_bufmap_initialize: exiting normally\n");
 391         return 0;
 392
 393 out_unmap_bufmap:
 394         orangefs_bufmap_unmap(bufmap);
 395 out_free_bufmap:
 396         orangefs_bufmap_free(bufmap);
 397 out:
 398         return ret;
 399 }
 400
 401 /*
 402  * orangefs_bufmap_finalize()
 403  *
 404  * shuts down the mapped buffer interface and releases any resources
 405  * associated with it
 406  *
 407  * no return value
 408  */
 409 void orangefs_bufmap_finalize(void)
 410 {
 411         struct orangefs_bufmap *bufmap = __orangefs_bufmap;
 412         if (!bufmap)
 413                 return;
 414         gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
 415         mark_killed(&rw_map);
 416         mark_killed(&readdir_map);
 417         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 418                      "orangefs_bufmap_finalize: exiting normally\n");
 419 }
 420
 421 void orangefs_bufmap_run_down(void)
 422 {
 423         struct orangefs_bufmap *bufmap = __orangefs_bufmap;
 424         if (!bufmap)
 425                 return;
 426         run_down(&rw_map);
 427         run_down(&readdir_map);
 428         spin_lock(&orangefs_bufmap_lock);
 429         __orangefs_bufmap = NULL;
 430         spin_unlock(&orangefs_bufmap_lock);
 431         orangefs_bufmap_unmap(bufmap);
 432         orangefs_bufmap_free(bufmap);
 433 }
 434
 435 /*
 436  * orangefs_bufmap_get()
 437  *
 438  * gets a free mapped buffer descriptor, will sleep until one becomes
 439  * available if necessary
 440  *
 441  * returns slot on success, -errno on failure
 442  */
 443 int orangefs_bufmap_get(void)
 444 {
 445         return get(&rw_map);
 446 }
 447
 448 /*
 449  * orangefs_bufmap_put()
 450  *
 451  * returns a mapped buffer descriptor to the collection
 452  *
 453  * no return value
 454  */
 455 void orangefs_bufmap_put(int buffer_index)
 456 {
 457         put(&rw_map, buffer_index);
 458 }
 459
 460 /*
 461  * orangefs_readdir_index_get()
 462  *
 463  * gets a free descriptor, will sleep until one becomes
 464  * available if necessary.
 465  * Although the readdir buffers are not mapped into kernel space
 466  * we could do that at a later point of time. Regardless, these
 467  * indices are used by the client-core.
 468  *
 469  * returns slot on success, -errno on failure
 470  */
 471 int orangefs_readdir_index_get(void)
 472 {
 473         return get(&readdir_map);
 474 }
 475
 476 void orangefs_readdir_index_put(int buffer_index)
 477 {
 478         put(&readdir_map, buffer_index);
 479 }
 480
 481 /*
 482  * we've been handed an iovec, we need to copy it to
 483  * the shared memory descriptor at "buffer_index".
 484  */
 485 int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
 486                                 int buffer_index,
 487                                 size_t size)
 488 {
 489         struct orangefs_bufmap_desc *to;
 490         int i;
 491
 492         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 493                      "%s: buffer_index:%d: size:%zu:\n",
 494                      __func__, buffer_index, size);
 495
 496         to = &__orangefs_bufmap->desc_array[buffer_index];
 497         for (i = 0; size; i++) {
 498                 struct page *page = to->page_array[i];
 499                 size_t n = size;
 500                 if (n > PAGE_SIZE)
 501                         n = PAGE_SIZE;
 502                 if (copy_page_from_iter(page, 0, n, iter) != n)
 503                         return -EFAULT;
 504                 size -= n;
 505         }
 506         return 0;
 507 }
 508
 509 /*
 510  * we've been handed an iovec, we need to fill it from
 511  * the shared memory descriptor at "buffer_index".
 512  */
 513 int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
 514                                     int buffer_index,
 515                                     size_t size)
 516 {
 517         struct orangefs_bufmap_desc *from;
 518         int i;
 519
 520         from = &__orangefs_bufmap->desc_array[buffer_index];
 521         gossip_debug(GOSSIP_BUFMAP_DEBUG,
 522                      "%s: buffer_index:%d: size:%zu:\n",
 523                      __func__, buffer_index, size);
 524
 525
 526         for (i = 0; size; i++) {
 527                 struct page *page = from->page_array[i];
 528                 size_t n = size;
 529                 if (n > PAGE_SIZE)
 530                         n = PAGE_SIZE;
 531                 n = copy_page_to_iter(page, 0, n, iter);
 532                 if (!n)
 533                         return -EFAULT;
 534                 size -= n;
 535         }
 536         return 0;
 537 }
 538
 539 void orangefs_bufmap_page_fill(void *page_to,
 540                                 int buffer_index,
 541                                 int slot_index)
 542 {
 543         struct orangefs_bufmap_desc *from;
 544         void *page_from;
 545
 546         from = &__orangefs_bufmap->desc_array[buffer_index];
 547         page_from = kmap_atomic(from->page_array[slot_index]);
 548         memcpy(page_to, page_from, PAGE_SIZE);
 549         kunmap_atomic(page_from);
 550 }