xref: /linux/fs/orangefs/orangefs-bufmap.c (revision a36e9f5cfe9eb3a1dce8769c7058251c42705357)
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 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 = bitmap_zalloc(bufmap->desc_count, GFP_KERNEL);
230 	if (!bufmap->buffer_index_array)
231 		goto out_free_bufmap;
232 
233 	bufmap->desc_array =
234 		kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
235 			GFP_KERNEL);
236 	if (!bufmap->desc_array)
237 		goto out_free_index_array;
238 
239 	bufmap->page_count = bufmap->total_size / PAGE_SIZE;
240 
241 	/* allocate storage to track our page mappings */
242 	bufmap->page_array =
243 		kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
244 	if (!bufmap->page_array)
245 		goto out_free_desc_array;
246 
247 	return bufmap;
248 
249 out_free_desc_array:
250 	kfree(bufmap->desc_array);
251 out_free_index_array:
252 	bitmap_free(bufmap->buffer_index_array);
253 out_free_bufmap:
254 	kfree(bufmap);
255 out:
256 	return NULL;
257 }
258 
259 static int
260 orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
261 		struct ORANGEFS_dev_map_desc *user_desc)
262 {
263 	int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
264 	int offset = 0, ret, i;
265 
266 	/* map the pages */
267 	ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
268 			     bufmap->page_count, FOLL_WRITE, bufmap->page_array);
269 
270 	if (ret < 0)
271 		return ret;
272 
273 	if (ret != bufmap->page_count) {
274 		gossip_err("orangefs error: asked for %d pages, only got %d.\n",
275 				bufmap->page_count, ret);
276 
277 		for (i = 0; i < ret; i++)
278 			unpin_user_page(bufmap->page_array[i]);
279 		return -ENOMEM;
280 	}
281 
282 	/*
283 	 * ideally we want to get kernel space pointers for each page, but
284 	 * we can't kmap that many pages at once if highmem is being used.
285 	 * so instead, we just kmap/kunmap the page address each time the
286 	 * kaddr is needed.
287 	 */
288 	for (i = 0; i < bufmap->page_count; i++)
289 		flush_dcache_page(bufmap->page_array[i]);
290 
291 	/* build a list of available descriptors */
292 	for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
293 		bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
294 		bufmap->desc_array[i].array_count = pages_per_desc;
295 		bufmap->desc_array[i].uaddr =
296 		    (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
297 		offset += pages_per_desc;
298 	}
299 
300 	return 0;
301 }
302 
303 /*
304  * orangefs_bufmap_initialize()
305  *
306  * initializes the mapped buffer interface
307  *
308  * returns 0 on success, -errno on failure
309  */
310 int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
311 {
312 	struct orangefs_bufmap *bufmap;
313 	int ret = -EINVAL;
314 
315 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
316 		     "orangefs_bufmap_initialize: called (ptr ("
317 		     "%p) sz (%d) cnt(%d).\n",
318 		     user_desc->ptr,
319 		     user_desc->size,
320 		     user_desc->count);
321 
322 	if (user_desc->total_size < 0 ||
323 	    user_desc->size < 0 ||
324 	    user_desc->count < 0)
325 		goto out;
326 
327 	/*
328 	 * sanity check alignment and size of buffer that caller wants to
329 	 * work with
330 	 */
331 	if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
332 	    (unsigned long)user_desc->ptr) {
333 		gossip_err("orangefs error: memory alignment (front). %p\n",
334 			   user_desc->ptr);
335 		goto out;
336 	}
337 
338 	if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
339 	    != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
340 		gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
341 			   user_desc->ptr,
342 			   user_desc->total_size);
343 		goto out;
344 	}
345 
346 	if (user_desc->total_size != (user_desc->size * user_desc->count)) {
347 		gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
348 			   user_desc->total_size,
349 			   user_desc->size,
350 			   user_desc->count);
351 		goto out;
352 	}
353 
354 	if ((user_desc->size % PAGE_SIZE) != 0) {
355 		gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
356 			   user_desc->size);
357 		goto out;
358 	}
359 
360 	ret = -ENOMEM;
361 	bufmap = orangefs_bufmap_alloc(user_desc);
362 	if (!bufmap)
363 		goto out;
364 
365 	ret = orangefs_bufmap_map(bufmap, user_desc);
366 	if (ret)
367 		goto out_free_bufmap;
368 
369 
370 	spin_lock(&orangefs_bufmap_lock);
371 	if (__orangefs_bufmap) {
372 		spin_unlock(&orangefs_bufmap_lock);
373 		gossip_err("orangefs: error: bufmap already initialized.\n");
374 		ret = -EINVAL;
375 		goto out_unmap_bufmap;
376 	}
377 	__orangefs_bufmap = bufmap;
378 	install(&rw_map,
379 		bufmap->desc_count,
380 		bufmap->buffer_index_array);
381 	install(&readdir_map,
382 		ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
383 		bufmap->readdir_index_array);
384 	spin_unlock(&orangefs_bufmap_lock);
385 
386 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
387 		     "orangefs_bufmap_initialize: exiting normally\n");
388 	return 0;
389 
390 out_unmap_bufmap:
391 	orangefs_bufmap_unmap(bufmap);
392 out_free_bufmap:
393 	orangefs_bufmap_free(bufmap);
394 out:
395 	return ret;
396 }
397 
398 /*
399  * orangefs_bufmap_finalize()
400  *
401  * shuts down the mapped buffer interface and releases any resources
402  * associated with it
403  *
404  * no return value
405  */
406 void orangefs_bufmap_finalize(void)
407 {
408 	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
409 	if (!bufmap)
410 		return;
411 	gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
412 	mark_killed(&rw_map);
413 	mark_killed(&readdir_map);
414 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
415 		     "orangefs_bufmap_finalize: exiting normally\n");
416 }
417 
418 void orangefs_bufmap_run_down(void)
419 {
420 	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
421 	if (!bufmap)
422 		return;
423 	run_down(&rw_map);
424 	run_down(&readdir_map);
425 	spin_lock(&orangefs_bufmap_lock);
426 	__orangefs_bufmap = NULL;
427 	spin_unlock(&orangefs_bufmap_lock);
428 	orangefs_bufmap_unmap(bufmap);
429 	orangefs_bufmap_free(bufmap);
430 }
431 
432 /*
433  * orangefs_bufmap_get()
434  *
435  * gets a free mapped buffer descriptor, will sleep until one becomes
436  * available if necessary
437  *
438  * returns slot on success, -errno on failure
439  */
440 int orangefs_bufmap_get(void)
441 {
442 	return get(&rw_map);
443 }
444 
445 /*
446  * orangefs_bufmap_put()
447  *
448  * returns a mapped buffer descriptor to the collection
449  *
450  * no return value
451  */
452 void orangefs_bufmap_put(int buffer_index)
453 {
454 	put(&rw_map, buffer_index);
455 }
456 
457 /*
458  * orangefs_readdir_index_get()
459  *
460  * gets a free descriptor, will sleep until one becomes
461  * available if necessary.
462  * Although the readdir buffers are not mapped into kernel space
463  * we could do that at a later point of time. Regardless, these
464  * indices are used by the client-core.
465  *
466  * returns slot on success, -errno on failure
467  */
468 int orangefs_readdir_index_get(void)
469 {
470 	return get(&readdir_map);
471 }
472 
473 void orangefs_readdir_index_put(int buffer_index)
474 {
475 	put(&readdir_map, buffer_index);
476 }
477 
478 /*
479  * we've been handed an iovec, we need to copy it to
480  * the shared memory descriptor at "buffer_index".
481  */
482 int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
483 				int buffer_index,
484 				size_t size)
485 {
486 	struct orangefs_bufmap_desc *to;
487 	int i;
488 
489 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
490 		     "%s: buffer_index:%d: size:%zu:\n",
491 		     __func__, buffer_index, size);
492 
493 	to = &__orangefs_bufmap->desc_array[buffer_index];
494 	for (i = 0; size; i++) {
495 		struct page *page = to->page_array[i];
496 		size_t n = size;
497 		if (n > PAGE_SIZE)
498 			n = PAGE_SIZE;
499 		if (copy_page_from_iter(page, 0, n, iter) != n)
500 			return -EFAULT;
501 		size -= n;
502 	}
503 	return 0;
504 }
505 
506 /*
507  * we've been handed an iovec, we need to fill it from
508  * the shared memory descriptor at "buffer_index".
509  */
510 int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
511 				    int buffer_index,
512 				    size_t size)
513 {
514 	struct orangefs_bufmap_desc *from;
515 	int i;
516 
517 	from = &__orangefs_bufmap->desc_array[buffer_index];
518 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
519 		     "%s: buffer_index:%d: size:%zu:\n",
520 		     __func__, buffer_index, size);
521 
522 
523 	for (i = 0; size; i++) {
524 		struct page *page = from->page_array[i];
525 		size_t n = size;
526 		if (n > PAGE_SIZE)
527 			n = PAGE_SIZE;
528 		n = copy_page_to_iter(page, 0, n, iter);
529 		if (!n)
530 			return -EFAULT;
531 		size -= n;
532 	}
533 	return 0;
534 }
535 
536 void orangefs_bufmap_page_fill(void *page_to,
537 				int buffer_index,
538 				int slot_index)
539 {
540 	struct orangefs_bufmap_desc *from;
541 	void *page_from;
542 
543 	from = &__orangefs_bufmap->desc_array[buffer_index];
544 	page_from = kmap_atomic(from->page_array[slot_index]);
545 	memcpy(page_to, page_from, PAGE_SIZE);
546 	kunmap_atomic(page_from);
547 }
548