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
 11struct slot_map {
 12	int c;
 13	wait_queue_head_t q;
 14	int count;
 15	unsigned long *map;
 16};
 17
 18static struct slot_map rw_map = {
 19	.c = -1,
 20	.q = __WAIT_QUEUE_HEAD_INITIALIZER(rw_map.q)
 21};
 22static struct slot_map readdir_map = {
 23	.c = -1,
 24	.q = __WAIT_QUEUE_HEAD_INITIALIZER(readdir_map.q)
 25};
 26
 27
 28static 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
 37static 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
 44static 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
 68static 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
 81static 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
124static 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 */
140struct orangefs_bufmap_desc {
141	void *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
147static 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
166static DEFINE_SPINLOCK(orangefs_bufmap_lock);
167
168static void
169orangefs_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
177static void
178orangefs_bufmap_free(struct orangefs_bufmap *bufmap)
179{
180	kfree(bufmap->page_array);
181	kfree(bufmap->desc_array);
182	kfree(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
191int 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
203int 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
215static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
216static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq);
217
218/*
219 * orangefs_get_bufmap_init
220 *
221 * If bufmap_init is 1, then the shared memory system, including the
222 * buffer_index_array, is available.  Otherwise, it is not.
223 *
224 * returns the value of bufmap_init
225 */
226int orangefs_get_bufmap_init(void)
227{
228	return __orangefs_bufmap ? 1 : 0;
229}
230
231
232static struct orangefs_bufmap *
233orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc)
234{
235	struct orangefs_bufmap *bufmap;
236
237	bufmap = kzalloc(sizeof(*bufmap), GFP_KERNEL);
238	if (!bufmap)
239		goto out;
240
241	bufmap->total_size = user_desc->total_size;
242	bufmap->desc_count = user_desc->count;
243	bufmap->desc_size = user_desc->size;
244	bufmap->desc_shift = ilog2(bufmap->desc_size);
245
246	bufmap->buffer_index_array =
247		kzalloc(DIV_ROUND_UP(bufmap->desc_count, BITS_PER_LONG), GFP_KERNEL);
248	if (!bufmap->buffer_index_array)
249		goto out_free_bufmap;
250
251	bufmap->desc_array =
252		kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
253			GFP_KERNEL);
254	if (!bufmap->desc_array)
255		goto out_free_index_array;
256
257	bufmap->page_count = bufmap->total_size / PAGE_SIZE;
258
259	/* allocate storage to track our page mappings */
260	bufmap->page_array =
261		kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
262	if (!bufmap->page_array)
263		goto out_free_desc_array;
264
265	return bufmap;
266
267out_free_desc_array:
268	kfree(bufmap->desc_array);
269out_free_index_array:
270	kfree(bufmap->buffer_index_array);
271out_free_bufmap:
272	kfree(bufmap);
273out:
274	return NULL;
275}
276
277static int
278orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
279		struct ORANGEFS_dev_map_desc *user_desc)
280{
281	int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
282	int offset = 0, ret, i;
283
284	/* map the pages */
285	ret = get_user_pages_fast((unsigned long)user_desc->ptr,
286			     bufmap->page_count, 1, bufmap->page_array);
287
288	if (ret < 0)
289		return ret;
290
291	if (ret != bufmap->page_count) {
292		gossip_err("orangefs error: asked for %d pages, only got %d.\n",
293				bufmap->page_count, ret);
294
295		for (i = 0; i < ret; i++) {
296			SetPageError(bufmap->page_array[i]);
297			put_page(bufmap->page_array[i]);
298		}
299		return -ENOMEM;
300	}
301
302	/*
303	 * ideally we want to get kernel space pointers for each page, but
304	 * we can't kmap that many pages at once if highmem is being used.
305	 * so instead, we just kmap/kunmap the page address each time the
306	 * kaddr is needed.
307	 */
308	for (i = 0; i < bufmap->page_count; i++)
309		flush_dcache_page(bufmap->page_array[i]);
310
311	/* build a list of available descriptors */
312	for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
313		bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
314		bufmap->desc_array[i].array_count = pages_per_desc;
315		bufmap->desc_array[i].uaddr =
316		    (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
317		offset += pages_per_desc;
318	}
319
320	return 0;
321}
322
323/*
324 * orangefs_bufmap_initialize()
325 *
326 * initializes the mapped buffer interface
327 *
328 * returns 0 on success, -errno on failure
329 */
330int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
331{
332	struct orangefs_bufmap *bufmap;
333	int ret = -EINVAL;
334
335	gossip_debug(GOSSIP_BUFMAP_DEBUG,
336		     "orangefs_bufmap_initialize: called (ptr ("
337		     "%p) sz (%d) cnt(%d).\n",
338		     user_desc->ptr,
339		     user_desc->size,
340		     user_desc->count);
341
342	if (user_desc->total_size < 0 ||
343	    user_desc->size < 0 ||
344	    user_desc->count < 0)
345		goto out;
346
347	/*
348	 * sanity check alignment and size of buffer that caller wants to
349	 * work with
350	 */
351	if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
352	    (unsigned long)user_desc->ptr) {
353		gossip_err("orangefs error: memory alignment (front). %p\n",
354			   user_desc->ptr);
355		goto out;
356	}
357
358	if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
359	    != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
360		gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
361			   user_desc->ptr,
362			   user_desc->total_size);
363		goto out;
364	}
365
366	if (user_desc->total_size != (user_desc->size * user_desc->count)) {
367		gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
368			   user_desc->total_size,
369			   user_desc->size,
370			   user_desc->count);
371		goto out;
372	}
373
374	if ((user_desc->size % PAGE_SIZE) != 0) {
375		gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
376			   user_desc->size);
377		goto out;
378	}
379
380	ret = -ENOMEM;
381	bufmap = orangefs_bufmap_alloc(user_desc);
382	if (!bufmap)
383		goto out;
384
385	ret = orangefs_bufmap_map(bufmap, user_desc);
386	if (ret)
387		goto out_free_bufmap;
388
389
390	spin_lock(&orangefs_bufmap_lock);
391	if (__orangefs_bufmap) {
392		spin_unlock(&orangefs_bufmap_lock);
393		gossip_err("orangefs: error: bufmap already initialized.\n");
394		ret = -EINVAL;
395		goto out_unmap_bufmap;
396	}
397	__orangefs_bufmap = bufmap;
398	install(&rw_map,
399		bufmap->desc_count,
400		bufmap->buffer_index_array);
401	install(&readdir_map,
402		ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
403		bufmap->readdir_index_array);
404	spin_unlock(&orangefs_bufmap_lock);
405
406	gossip_debug(GOSSIP_BUFMAP_DEBUG,
407		     "orangefs_bufmap_initialize: exiting normally\n");
408	return 0;
409
410out_unmap_bufmap:
411	orangefs_bufmap_unmap(bufmap);
412out_free_bufmap:
413	orangefs_bufmap_free(bufmap);
414out:
415	return ret;
416}
417
418/*
419 * orangefs_bufmap_finalize()
420 *
421 * shuts down the mapped buffer interface and releases any resources
422 * associated with it
423 *
424 * no return value
425 */
426void orangefs_bufmap_finalize(void)
427{
428	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
429	if (!bufmap)
430		return;
431	gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
432	mark_killed(&rw_map);
433	mark_killed(&readdir_map);
434	gossip_debug(GOSSIP_BUFMAP_DEBUG,
435		     "orangefs_bufmap_finalize: exiting normally\n");
436}
437
438void orangefs_bufmap_run_down(void)
439{
440	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
441	if (!bufmap)
442		return;
443	run_down(&rw_map);
444	run_down(&readdir_map);
445	spin_lock(&orangefs_bufmap_lock);
446	__orangefs_bufmap = NULL;
447	spin_unlock(&orangefs_bufmap_lock);
448	orangefs_bufmap_unmap(bufmap);
449	orangefs_bufmap_free(bufmap);
450}
451
452/*
453 * orangefs_bufmap_get()
454 *
455 * gets a free mapped buffer descriptor, will sleep until one becomes
456 * available if necessary
457 *
458 * returns slot on success, -errno on failure
459 */
460int orangefs_bufmap_get(void)
461{
462	return get(&rw_map);
463}
464
465/*
466 * orangefs_bufmap_put()
467 *
468 * returns a mapped buffer descriptor to the collection
469 *
470 * no return value
471 */
472void orangefs_bufmap_put(int buffer_index)
473{
474	put(&rw_map, buffer_index);
475}
476
477/*
478 * orangefs_readdir_index_get()
479 *
480 * gets a free descriptor, will sleep until one becomes
481 * available if necessary.
482 * Although the readdir buffers are not mapped into kernel space
483 * we could do that at a later point of time. Regardless, these
484 * indices are used by the client-core.
485 *
486 * returns slot on success, -errno on failure
487 */
488int orangefs_readdir_index_get(void)
489{
490	return get(&readdir_map);
491}
492
493void orangefs_readdir_index_put(int buffer_index)
494{
495	put(&readdir_map, buffer_index);
496}
497
498/*
499 * we've been handed an iovec, we need to copy it to 
500 * the shared memory descriptor at "buffer_index".
501 */
502int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
503				int buffer_index,
504				size_t size)
505{
506	struct orangefs_bufmap_desc *to;
507	int i;
508
509	gossip_debug(GOSSIP_BUFMAP_DEBUG,
510		     "%s: buffer_index:%d: size:%zu:\n",
511		     __func__, buffer_index, size);
512
513	to = &__orangefs_bufmap->desc_array[buffer_index];
514	for (i = 0; size; i++) {
515		struct page *page = to->page_array[i];
516		size_t n = size;
517		if (n > PAGE_SIZE)
518			n = PAGE_SIZE;
519		if (copy_page_from_iter(page, 0, n, iter) != n)
520			return -EFAULT;
521		size -= n;
522	}
523	return 0;
524}
525
526/*
527 * we've been handed an iovec, we need to fill it from
528 * the shared memory descriptor at "buffer_index".
529 */
530int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
531				    int buffer_index,
532				    size_t size)
533{
534	struct orangefs_bufmap_desc *from;
535	int i;
536
537	from = &__orangefs_bufmap->desc_array[buffer_index];
538	gossip_debug(GOSSIP_BUFMAP_DEBUG,
539		     "%s: buffer_index:%d: size:%zu:\n",
540		     __func__, buffer_index, size);
541
542
543	for (i = 0; size; i++) {
544		struct page *page = from->page_array[i];
545		size_t n = size;
546		if (n > PAGE_SIZE)
547			n = PAGE_SIZE;
548		n = copy_page_to_iter(page, 0, n, iter);
549		if (!n)
550			return -EFAULT;
551		size -= n;
552	}
553	return 0;
554}