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