orangefs-bufmap.c

  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 (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 __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
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	unpin_user_pages(bufmap->page_array, bufmap->page_count);
172}
173
174static void
175orangefs_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
188int 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
200static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
201static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq);
202
203static struct orangefs_bufmap *
204orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc)
205{
206	struct orangefs_bufmap *bufmap;
207
208	bufmap = kzalloc(sizeof(*bufmap), GFP_KERNEL);
209	if (!bufmap)
210		goto out;
211
212	bufmap->total_size = user_desc->total_size;
213	bufmap->desc_count = user_desc->count;
214	bufmap->desc_size = user_desc->size;
215	bufmap->desc_shift = ilog2(bufmap->desc_size);
216
217	bufmap->buffer_index_array = bitmap_zalloc(bufmap->desc_count, GFP_KERNEL);
218	if (!bufmap->buffer_index_array)
219		goto out_free_bufmap;
220
221	bufmap->desc_array =
222		kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
223			GFP_KERNEL);
224	if (!bufmap->desc_array)
225		goto out_free_index_array;
226
227	bufmap->page_count = bufmap->total_size / PAGE_SIZE;
228
229	/* allocate storage to track our page mappings */
230	bufmap->page_array =
231		kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
232	if (!bufmap->page_array)
233		goto out_free_desc_array;
234
235	return bufmap;
236
237out_free_desc_array:
238	kfree(bufmap->desc_array);
239out_free_index_array:
240	bitmap_free(bufmap->buffer_index_array);
241out_free_bufmap:
242	kfree(bufmap);
243out:
244	return NULL;
245}
246
247static int
248orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
249		struct ORANGEFS_dev_map_desc *user_desc)
250{
251	int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
252	int offset = 0, ret, i;
253
254	/* map the pages */
255	ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
256			     bufmap->page_count, FOLL_WRITE, bufmap->page_array);
257
258	if (ret < 0)
259		return ret;
260
261	if (ret != bufmap->page_count) {
262		gossip_err("orangefs error: asked for %d pages, only got %d.\n",
263				bufmap->page_count, ret);
264
265		for (i = 0; i < ret; i++)
266			unpin_user_page(bufmap->page_array[i]);
267		return -ENOMEM;
268	}
269
270	/*
271	 * ideally we want to get kernel space pointers for each page, but
272	 * we can't kmap that many pages at once if highmem is being used.
273	 * so instead, we just kmap/kunmap the page address each time the
274	 * kaddr is needed.
275	 */
276	for (i = 0; i < bufmap->page_count; i++)
277		flush_dcache_page(bufmap->page_array[i]);
278
279	/* build a list of available descriptors */
280	for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
281		bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
282		bufmap->desc_array[i].array_count = pages_per_desc;
283		bufmap->desc_array[i].uaddr =
284		    (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
285		offset += pages_per_desc;
286	}
287
288	return 0;
289}
290
291/*
292 * orangefs_bufmap_initialize()
293 *
294 * initializes the mapped buffer interface
295 *
296 * returns 0 on success, -errno on failure
297 */
298int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
299{
300	struct orangefs_bufmap *bufmap;
301	int ret = -EINVAL;
302
303	gossip_debug(GOSSIP_BUFMAP_DEBUG,
304		     "orangefs_bufmap_initialize: called (ptr ("
305		     "%p) sz (%d) cnt(%d).\n",
306		     user_desc->ptr,
307		     user_desc->size,
308		     user_desc->count);
309
310	if (user_desc->total_size < 0 ||
311	    user_desc->size < 0 ||
312	    user_desc->count < 0)
313		goto out;
314
315	/*
316	 * sanity check alignment and size of buffer that caller wants to
317	 * work with
318	 */
319	if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
320	    (unsigned long)user_desc->ptr) {
321		gossip_err("orangefs error: memory alignment (front). %p\n",
322			   user_desc->ptr);
323		goto out;
324	}
325
326	if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
327	    != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
328		gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
329			   user_desc->ptr,
330			   user_desc->total_size);
331		goto out;
332	}
333
334	if (user_desc->total_size != (user_desc->size * user_desc->count)) {
335		gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
336			   user_desc->total_size,
337			   user_desc->size,
338			   user_desc->count);
339		goto out;
340	}
341
342	if ((user_desc->size % PAGE_SIZE) != 0) {
343		gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
344			   user_desc->size);
345		goto out;
346	}
347
348	ret = -ENOMEM;
349	bufmap = orangefs_bufmap_alloc(user_desc);
350	if (!bufmap)
351		goto out;
352
353	ret = orangefs_bufmap_map(bufmap, user_desc);
354	if (ret)
355		goto out_free_bufmap;
356
357
358	spin_lock(&orangefs_bufmap_lock);
359	if (__orangefs_bufmap) {
360		spin_unlock(&orangefs_bufmap_lock);
361		gossip_err("orangefs: error: bufmap already initialized.\n");
362		ret = -EINVAL;
363		goto out_unmap_bufmap;
364	}
365	__orangefs_bufmap = bufmap;
366	install(&rw_map,
367		bufmap->desc_count,
368		bufmap->buffer_index_array);
369	install(&readdir_map,
370		ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
371		bufmap->readdir_index_array);
372	spin_unlock(&orangefs_bufmap_lock);
373
374	gossip_debug(GOSSIP_BUFMAP_DEBUG,
375		     "orangefs_bufmap_initialize: exiting normally\n");
376	return 0;
377
378out_unmap_bufmap:
379	orangefs_bufmap_unmap(bufmap);
380out_free_bufmap:
381	orangefs_bufmap_free(bufmap);
382out:
383	return ret;
384}
385
386/*
387 * orangefs_bufmap_finalize()
388 *
389 * shuts down the mapped buffer interface and releases any resources
390 * associated with it
391 *
392 * no return value
393 */
394void orangefs_bufmap_finalize(void)
395{
396	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
397	if (!bufmap)
398		return;
399	gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
400	mark_killed(&rw_map);
401	mark_killed(&readdir_map);
402	gossip_debug(GOSSIP_BUFMAP_DEBUG,
403		     "orangefs_bufmap_finalize: exiting normally\n");
404}
405
406void orangefs_bufmap_run_down(void)
407{
408	struct orangefs_bufmap *bufmap = __orangefs_bufmap;
409	if (!bufmap)
410		return;
411	run_down(&rw_map);
412	run_down(&readdir_map);
413	spin_lock(&orangefs_bufmap_lock);
414	__orangefs_bufmap = NULL;
415	spin_unlock(&orangefs_bufmap_lock);
416	orangefs_bufmap_unmap(bufmap);
417	orangefs_bufmap_free(bufmap);
418}
419
420/*
421 * orangefs_bufmap_get()
422 *
423 * gets a free mapped buffer descriptor, will sleep until one becomes
424 * available if necessary
425 *
426 * returns slot on success, -errno on failure
427 */
428int orangefs_bufmap_get(void)
429{
430	return get(&rw_map);
431}
432
433/*
434 * orangefs_bufmap_put()
435 *
436 * returns a mapped buffer descriptor to the collection
437 *
438 * no return value
439 */
440void orangefs_bufmap_put(int buffer_index)
441{
442	put(&rw_map, buffer_index);
443}
444
445/*
446 * orangefs_readdir_index_get()
447 *
448 * gets a free descriptor, will sleep until one becomes
449 * available if necessary.
450 * Although the readdir buffers are not mapped into kernel space
451 * we could do that at a later point of time. Regardless, these
452 * indices are used by the client-core.
453 *
454 * returns slot on success, -errno on failure
455 */
456int orangefs_readdir_index_get(void)
457{
458	return get(&readdir_map);
459}
460
461void orangefs_readdir_index_put(int buffer_index)
462{
463	put(&readdir_map, buffer_index);
464}
465
466/*
467 * we've been handed an iovec, we need to copy it to
468 * the shared memory descriptor at "buffer_index".
469 */
470int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
471				int buffer_index,
472				size_t size)
473{
474	struct orangefs_bufmap_desc *to;
475	int i;
476
477	gossip_debug(GOSSIP_BUFMAP_DEBUG,
478		     "%s: buffer_index:%d: size:%zu:\n",
479		     __func__, buffer_index, size);
480
481	to = &__orangefs_bufmap->desc_array[buffer_index];
482	for (i = 0; size; i++) {
483		struct page *page = to->page_array[i];
484		size_t n = size;
485		if (n > PAGE_SIZE)
486			n = PAGE_SIZE;
487		if (copy_page_from_iter(page, 0, n, iter) != n)
488			return -EFAULT;
489		size -= n;
490	}
491	return 0;
492}
493
494/*
495 * we've been handed an iovec, we need to fill it from
496 * the shared memory descriptor at "buffer_index".
497 */
498int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
499				    int buffer_index,
500				    size_t size)
501{
502	struct orangefs_bufmap_desc *from;
503	int i;
504
505	from = &__orangefs_bufmap->desc_array[buffer_index];
506	gossip_debug(GOSSIP_BUFMAP_DEBUG,
507		     "%s: buffer_index:%d: size:%zu:\n",
508		     __func__, buffer_index, size);
509
510
511	for (i = 0; size; i++) {
512		struct page *page = from->page_array[i];
513		size_t n = size;
514		if (n > PAGE_SIZE)
515			n = PAGE_SIZE;
516		n = copy_page_to_iter(page, 0, n, iter);
517		if (!n)
518			return -EFAULT;
519		size -= n;
520	}
521	return 0;
522}