xref: /openbmc/linux/fs/orangefs/orangefs-bufmap.c (revision ca481398)
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.entry)))
50 				__add_wait_queue_entry_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.entry)))
88 			__add_wait_queue_entry_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.entry))
112 		list_del(&wait.entry);
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 		goto out_free_bufmap;
249 
250 	bufmap->desc_array =
251 		kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
252 			GFP_KERNEL);
253 	if (!bufmap->desc_array)
254 		goto out_free_index_array;
255 
256 	bufmap->page_count = bufmap->total_size / PAGE_SIZE;
257 
258 	/* allocate storage to track our page mappings */
259 	bufmap->page_array =
260 		kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
261 	if (!bufmap->page_array)
262 		goto out_free_desc_array;
263 
264 	return bufmap;
265 
266 out_free_desc_array:
267 	kfree(bufmap->desc_array);
268 out_free_index_array:
269 	kfree(bufmap->buffer_index_array);
270 out_free_bufmap:
271 	kfree(bufmap);
272 out:
273 	return NULL;
274 }
275 
276 static int
277 orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
278 		struct ORANGEFS_dev_map_desc *user_desc)
279 {
280 	int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
281 	int offset = 0, ret, i;
282 
283 	/* map the pages */
284 	ret = get_user_pages_fast((unsigned long)user_desc->ptr,
285 			     bufmap->page_count, 1, bufmap->page_array);
286 
287 	if (ret < 0)
288 		return ret;
289 
290 	if (ret != bufmap->page_count) {
291 		gossip_err("orangefs error: asked for %d pages, only got %d.\n",
292 				bufmap->page_count, ret);
293 
294 		for (i = 0; i < ret; i++) {
295 			SetPageError(bufmap->page_array[i]);
296 			put_page(bufmap->page_array[i]);
297 		}
298 		return -ENOMEM;
299 	}
300 
301 	/*
302 	 * ideally we want to get kernel space pointers for each page, but
303 	 * we can't kmap that many pages at once if highmem is being used.
304 	 * so instead, we just kmap/kunmap the page address each time the
305 	 * kaddr is needed.
306 	 */
307 	for (i = 0; i < bufmap->page_count; i++)
308 		flush_dcache_page(bufmap->page_array[i]);
309 
310 	/* build a list of available descriptors */
311 	for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
312 		bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
313 		bufmap->desc_array[i].array_count = pages_per_desc;
314 		bufmap->desc_array[i].uaddr =
315 		    (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
316 		offset += pages_per_desc;
317 	}
318 
319 	return 0;
320 }
321 
322 /*
323  * orangefs_bufmap_initialize()
324  *
325  * initializes the mapped buffer interface
326  *
327  * returns 0 on success, -errno on failure
328  */
329 int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
330 {
331 	struct orangefs_bufmap *bufmap;
332 	int ret = -EINVAL;
333 
334 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
335 		     "orangefs_bufmap_initialize: called (ptr ("
336 		     "%p) sz (%d) cnt(%d).\n",
337 		     user_desc->ptr,
338 		     user_desc->size,
339 		     user_desc->count);
340 
341 	if (user_desc->total_size < 0 ||
342 	    user_desc->size < 0 ||
343 	    user_desc->count < 0)
344 		goto out;
345 
346 	/*
347 	 * sanity check alignment and size of buffer that caller wants to
348 	 * work with
349 	 */
350 	if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
351 	    (unsigned long)user_desc->ptr) {
352 		gossip_err("orangefs error: memory alignment (front). %p\n",
353 			   user_desc->ptr);
354 		goto out;
355 	}
356 
357 	if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
358 	    != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
359 		gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
360 			   user_desc->ptr,
361 			   user_desc->total_size);
362 		goto out;
363 	}
364 
365 	if (user_desc->total_size != (user_desc->size * user_desc->count)) {
366 		gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
367 			   user_desc->total_size,
368 			   user_desc->size,
369 			   user_desc->count);
370 		goto out;
371 	}
372 
373 	if ((user_desc->size % PAGE_SIZE) != 0) {
374 		gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
375 			   user_desc->size);
376 		goto out;
377 	}
378 
379 	ret = -ENOMEM;
380 	bufmap = orangefs_bufmap_alloc(user_desc);
381 	if (!bufmap)
382 		goto out;
383 
384 	ret = orangefs_bufmap_map(bufmap, user_desc);
385 	if (ret)
386 		goto out_free_bufmap;
387 
388 
389 	spin_lock(&orangefs_bufmap_lock);
390 	if (__orangefs_bufmap) {
391 		spin_unlock(&orangefs_bufmap_lock);
392 		gossip_err("orangefs: error: bufmap already initialized.\n");
393 		ret = -EINVAL;
394 		goto out_unmap_bufmap;
395 	}
396 	__orangefs_bufmap = bufmap;
397 	install(&rw_map,
398 		bufmap->desc_count,
399 		bufmap->buffer_index_array);
400 	install(&readdir_map,
401 		ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
402 		bufmap->readdir_index_array);
403 	spin_unlock(&orangefs_bufmap_lock);
404 
405 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
406 		     "orangefs_bufmap_initialize: exiting normally\n");
407 	return 0;
408 
409 out_unmap_bufmap:
410 	orangefs_bufmap_unmap(bufmap);
411 out_free_bufmap:
412 	orangefs_bufmap_free(bufmap);
413 out:
414 	return ret;
415 }
416 
417 /*
418  * orangefs_bufmap_finalize()
419  *
420  * shuts down the mapped buffer interface and releases any resources
421  * associated with it
422  *
423  * no return value
424  */
425 void orangefs_bufmap_finalize(void)
426 {
427 	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
428 	if (!bufmap)
429 		return;
430 	gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
431 	mark_killed(&rw_map);
432 	mark_killed(&readdir_map);
433 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
434 		     "orangefs_bufmap_finalize: exiting normally\n");
435 }
436 
437 void orangefs_bufmap_run_down(void)
438 {
439 	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
440 	if (!bufmap)
441 		return;
442 	run_down(&rw_map);
443 	run_down(&readdir_map);
444 	spin_lock(&orangefs_bufmap_lock);
445 	__orangefs_bufmap = NULL;
446 	spin_unlock(&orangefs_bufmap_lock);
447 	orangefs_bufmap_unmap(bufmap);
448 	orangefs_bufmap_free(bufmap);
449 }
450 
451 /*
452  * orangefs_bufmap_get()
453  *
454  * gets a free mapped buffer descriptor, will sleep until one becomes
455  * available if necessary
456  *
457  * returns slot on success, -errno on failure
458  */
459 int orangefs_bufmap_get(void)
460 {
461 	return get(&rw_map);
462 }
463 
464 /*
465  * orangefs_bufmap_put()
466  *
467  * returns a mapped buffer descriptor to the collection
468  *
469  * no return value
470  */
471 void orangefs_bufmap_put(int buffer_index)
472 {
473 	put(&rw_map, buffer_index);
474 }
475 
476 /*
477  * orangefs_readdir_index_get()
478  *
479  * gets a free descriptor, will sleep until one becomes
480  * available if necessary.
481  * Although the readdir buffers are not mapped into kernel space
482  * we could do that at a later point of time. Regardless, these
483  * indices are used by the client-core.
484  *
485  * returns slot on success, -errno on failure
486  */
487 int orangefs_readdir_index_get(void)
488 {
489 	return get(&readdir_map);
490 }
491 
492 void orangefs_readdir_index_put(int buffer_index)
493 {
494 	put(&readdir_map, buffer_index);
495 }
496 
497 /*
498  * we've been handed an iovec, we need to copy it to
499  * the shared memory descriptor at "buffer_index".
500  */
501 int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
502 				int buffer_index,
503 				size_t size)
504 {
505 	struct orangefs_bufmap_desc *to;
506 	int i;
507 
508 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
509 		     "%s: buffer_index:%d: size:%zu:\n",
510 		     __func__, buffer_index, size);
511 
512 	to = &__orangefs_bufmap->desc_array[buffer_index];
513 	for (i = 0; size; i++) {
514 		struct page *page = to->page_array[i];
515 		size_t n = size;
516 		if (n > PAGE_SIZE)
517 			n = PAGE_SIZE;
518 		if (copy_page_from_iter(page, 0, n, iter) != n)
519 			return -EFAULT;
520 		size -= n;
521 	}
522 	return 0;
523 }
524 
525 /*
526  * we've been handed an iovec, we need to fill it from
527  * the shared memory descriptor at "buffer_index".
528  */
529 int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
530 				    int buffer_index,
531 				    size_t size)
532 {
533 	struct orangefs_bufmap_desc *from;
534 	int i;
535 
536 	from = &__orangefs_bufmap->desc_array[buffer_index];
537 	gossip_debug(GOSSIP_BUFMAP_DEBUG,
538 		     "%s: buffer_index:%d: size:%zu:\n",
539 		     __func__, buffer_index, size);
540 
541 
542 	for (i = 0; size; i++) {
543 		struct page *page = from->page_array[i];
544 		size_t n = size;
545 		if (n > PAGE_SIZE)
546 			n = PAGE_SIZE;
547 		n = copy_page_to_iter(page, 0, n, iter);
548 		if (!n)
549 			return -EFAULT;
550 		size -= n;
551 	}
552 	return 0;
553 }
554