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
 10struct slot_map {
 11	int c;
 12	wait_queue_head_t q;
 13	int count;
 14	unsigned long *map;
 15};
 16
 17static struct slot_map rw_map = {
 18	.c = -1,
 19	.q = __WAIT_QUEUE_HEAD_INITIALIZER(rw_map.q)
 20};
 21static struct slot_map readdir_map = {
 22	.c = -1,
 23	.q = __WAIT_QUEUE_HEAD_INITIALIZER(readdir_map.q)
 24};
 25
 26
 27static 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
 36static 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
 43static 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.task_list)))
 50				__add_wait_queue_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
 67static 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
 80static 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.task_list)))
 88			__add_wait_queue_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.task_list))
112		list_del(&wait.task_list);
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
123static 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 */
139struct 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
146static 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
165static DEFINE_SPINLOCK(orangefs_bufmap_lock);
166
167static void
168orangefs_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
176static void
177orangefs_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
190int 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
202int 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
214static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
215static 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 */
225int orangefs_get_bufmap_init(void)
226{
227	return __orangefs_bufmap ? 1 : 0;
228}
229
230
231static struct orangefs_bufmap *
232orangefs_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		gossip_err("orangefs: could not allocate %d buffer indices\n",
249				bufmap->desc_count);
250		goto out_free_bufmap;
251	}
252
253	bufmap->desc_array =
254		kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
255			GFP_KERNEL);
256	if (!bufmap->desc_array) {
257		gossip_err("orangefs: could not allocate %d descriptors\n",
258				bufmap->desc_count);
259		goto out_free_index_array;
260	}
261
262	bufmap->page_count = bufmap->total_size / PAGE_SIZE;
263
264	/* allocate storage to track our page mappings */
265	bufmap->page_array =
266		kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
267	if (!bufmap->page_array)
268		goto out_free_desc_array;
269
270	return bufmap;
271
272out_free_desc_array:
273	kfree(bufmap->desc_array);
274out_free_index_array:
275	kfree(bufmap->buffer_index_array);
276out_free_bufmap:
277	kfree(bufmap);
278out:
279	return NULL;
280}
281
282static int
283orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
284		struct ORANGEFS_dev_map_desc *user_desc)
285{
286	int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
287	int offset = 0, ret, i;
288
289	/* map the pages */
290	ret = get_user_pages_fast((unsigned long)user_desc->ptr,
291			     bufmap->page_count, 1, bufmap->page_array);
292
293	if (ret < 0)
294		return ret;
295
296	if (ret != bufmap->page_count) {
297		gossip_err("orangefs error: asked for %d pages, only got %d.\n",
298				bufmap->page_count, ret);
299
300		for (i = 0; i < ret; i++) {
301			SetPageError(bufmap->page_array[i]);
302			put_page(bufmap->page_array[i]);
303		}
304		return -ENOMEM;
305	}
306
307	/*
308	 * ideally we want to get kernel space pointers for each page, but
309	 * we can't kmap that many pages at once if highmem is being used.
310	 * so instead, we just kmap/kunmap the page address each time the
311	 * kaddr is needed.
312	 */
313	for (i = 0; i < bufmap->page_count; i++)
314		flush_dcache_page(bufmap->page_array[i]);
315
316	/* build a list of available descriptors */
317	for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
318		bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
319		bufmap->desc_array[i].array_count = pages_per_desc;
320		bufmap->desc_array[i].uaddr =
321		    (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
322		offset += pages_per_desc;
323	}
324
325	return 0;
326}
327
328/*
329 * orangefs_bufmap_initialize()
330 *
331 * initializes the mapped buffer interface
332 *
333 * returns 0 on success, -errno on failure
334 */
335int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
336{
337	struct orangefs_bufmap *bufmap;
338	int ret = -EINVAL;
339
340	gossip_debug(GOSSIP_BUFMAP_DEBUG,
341		     "orangefs_bufmap_initialize: called (ptr ("
342		     "%p) sz (%d) cnt(%d).\n",
343		     user_desc->ptr,
344		     user_desc->size,
345		     user_desc->count);
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		n = copy_page_from_iter(page, 0, n, iter);
520		if (!n)
521			return -EFAULT;
522		size -= n;
523	}
524	return 0;
525
526}
527
528/*
529 * we've been handed an iovec, we need to fill it from
530 * the shared memory descriptor at "buffer_index".
531 */
532int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
533				    int buffer_index,
534				    size_t size)
535{
536	struct orangefs_bufmap_desc *from;
537	int i;
538
539	from = &__orangefs_bufmap->desc_array[buffer_index];
540	gossip_debug(GOSSIP_BUFMAP_DEBUG,
541		     "%s: buffer_index:%d: size:%zu:\n",
542		     __func__, buffer_index, size);
543
544
545	for (i = 0; size; i++) {
546		struct page *page = from->page_array[i];
547		size_t n = size;
548		if (n > PAGE_SIZE)
549			n = PAGE_SIZE;
550		n = copy_page_to_iter(page, 0, n, iter);
551		if (!n)
552			return -EFAULT;
553		size -= n;
554	}
555	return 0;
556}