1 // SPDX-License-Identifier: GPL-2.0
3 * (C) 2001 Clemson University and The University of Chicago
5 * See COPYING in top-level directory.
8 #include "orangefs-kernel.h"
9 #include "orangefs-bufmap.h"
18 static struct slot_map rw_map = {
20 .q = __WAIT_QUEUE_HEAD_INITIALIZER(rw_map.q)
22 static struct slot_map readdir_map = {
24 .q = __WAIT_QUEUE_HEAD_INITIALIZER(readdir_map.q)
28 static void install(struct slot_map *m, int count, unsigned long *map)
30 spin_lock(&m->q.lock);
31 m->c = m->count = count;
33 wake_up_all_locked(&m->q);
34 spin_unlock(&m->q.lock);
37 static void mark_killed(struct slot_map *m)
39 spin_lock(&m->q.lock);
41 spin_unlock(&m->q.lock);
44 static void run_down(struct slot_map *m)
47 spin_lock(&m->q.lock);
50 if (likely(list_empty(&wait.entry)))
51 __add_wait_queue_entry_tail(&m->q, &wait);
52 set_current_state(TASK_UNINTERRUPTIBLE);
57 spin_unlock(&m->q.lock);
59 spin_lock(&m->q.lock);
61 __remove_wait_queue(&m->q, &wait);
62 __set_current_state(TASK_RUNNING);
65 spin_unlock(&m->q.lock);
68 static void put(struct slot_map *m, int slot)
71 spin_lock(&m->q.lock);
72 __clear_bit(slot, m->map);
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);
81 static int wait_for_free(struct slot_map *m)
83 long left = slot_timeout_secs * HZ;
88 if (likely(list_empty(&wait.entry)))
89 __add_wait_queue_entry_tail_exclusive(&m->q, &wait);
90 set_current_state(TASK_INTERRUPTIBLE);
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;
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)
107 left = t + (left - n);
108 if (signal_pending(current))
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);
118 if (likely(left > 0))
121 return left < 0 ? -EINTR : -ETIMEDOUT;
124 static int get(struct slot_map *m)
127 spin_lock(&m->q.lock);
128 if (unlikely(m->c <= 0))
129 res = wait_for_free(m);
132 res = find_first_zero_bit(m->map, m->count);
133 __set_bit(res, m->map);
135 spin_unlock(&m->q.lock);
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;
147 static struct orangefs_bufmap {
154 struct page **page_array;
155 struct orangefs_bufmap_desc *desc_array;
157 /* array to track usage of buffer descriptors */
158 unsigned long *buffer_index_array;
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];
164 } *__orangefs_bufmap;
166 static DEFINE_SPINLOCK(orangefs_bufmap_lock);
169 orangefs_bufmap_unmap(struct orangefs_bufmap *bufmap)
171 unpin_user_pages(bufmap->page_array, bufmap->page_count);
175 orangefs_bufmap_free(struct orangefs_bufmap *bufmap)
177 kfree(bufmap->page_array);
178 kfree(bufmap->desc_array);
179 kfree(bufmap->buffer_index_array);
184 * XXX: Can the size and shift change while the caller gives up the
185 * XXX: lock between calling this and doing something useful?
188 int orangefs_bufmap_size_query(void)
190 struct orangefs_bufmap *bufmap;
192 spin_lock(&orangefs_bufmap_lock);
193 bufmap = __orangefs_bufmap;
195 size = bufmap->desc_size;
196 spin_unlock(&orangefs_bufmap_lock);
200 int orangefs_bufmap_shift_query(void)
202 struct orangefs_bufmap *bufmap;
204 spin_lock(&orangefs_bufmap_lock);
205 bufmap = __orangefs_bufmap;
207 shift = bufmap->desc_shift;
208 spin_unlock(&orangefs_bufmap_lock);
212 static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
213 static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq);
215 static struct orangefs_bufmap *
216 orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc)
218 struct orangefs_bufmap *bufmap;
220 bufmap = kzalloc(sizeof(*bufmap), GFP_KERNEL);
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);
229 bufmap->buffer_index_array =
230 kzalloc(DIV_ROUND_UP(bufmap->desc_count, BITS_PER_LONG), GFP_KERNEL);
231 if (!bufmap->buffer_index_array)
232 goto out_free_bufmap;
235 kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
237 if (!bufmap->desc_array)
238 goto out_free_index_array;
240 bufmap->page_count = bufmap->total_size / PAGE_SIZE;
242 /* allocate storage to track our page mappings */
244 kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
245 if (!bufmap->page_array)
246 goto out_free_desc_array;
251 kfree(bufmap->desc_array);
252 out_free_index_array:
253 kfree(bufmap->buffer_index_array);
261 orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
262 struct ORANGEFS_dev_map_desc *user_desc)
264 int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
265 int offset = 0, ret, i;
268 ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
269 bufmap->page_count, FOLL_WRITE, bufmap->page_array);
274 if (ret != bufmap->page_count) {
275 gossip_err("orangefs error: asked for %d pages, only got %d.\n",
276 bufmap->page_count, ret);
278 for (i = 0; i < ret; i++) {
279 SetPageError(bufmap->page_array[i]);
280 unpin_user_page(bufmap->page_array[i]);
286 * ideally we want to get kernel space pointers for each page, but
287 * we can't kmap that many pages at once if highmem is being used.
288 * so instead, we just kmap/kunmap the page address each time the
291 for (i = 0; i < bufmap->page_count; i++)
292 flush_dcache_page(bufmap->page_array[i]);
294 /* build a list of available descriptors */
295 for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
296 bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
297 bufmap->desc_array[i].array_count = pages_per_desc;
298 bufmap->desc_array[i].uaddr =
299 (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
300 offset += pages_per_desc;
307 * orangefs_bufmap_initialize()
309 * initializes the mapped buffer interface
311 * returns 0 on success, -errno on failure
313 int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
315 struct orangefs_bufmap *bufmap;
318 gossip_debug(GOSSIP_BUFMAP_DEBUG,
319 "orangefs_bufmap_initialize: called (ptr ("
320 "%p) sz (%d) cnt(%d).\n",
325 if (user_desc->total_size < 0 ||
326 user_desc->size < 0 ||
327 user_desc->count < 0)
331 * sanity check alignment and size of buffer that caller wants to
334 if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
335 (unsigned long)user_desc->ptr) {
336 gossip_err("orangefs error: memory alignment (front). %p\n",
341 if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
342 != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
343 gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
345 user_desc->total_size);
349 if (user_desc->total_size != (user_desc->size * user_desc->count)) {
350 gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
351 user_desc->total_size,
357 if ((user_desc->size % PAGE_SIZE) != 0) {
358 gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
364 bufmap = orangefs_bufmap_alloc(user_desc);
368 ret = orangefs_bufmap_map(bufmap, user_desc);
370 goto out_free_bufmap;
373 spin_lock(&orangefs_bufmap_lock);
374 if (__orangefs_bufmap) {
375 spin_unlock(&orangefs_bufmap_lock);
376 gossip_err("orangefs: error: bufmap already initialized.\n");
378 goto out_unmap_bufmap;
380 __orangefs_bufmap = bufmap;
383 bufmap->buffer_index_array);
384 install(&readdir_map,
385 ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
386 bufmap->readdir_index_array);
387 spin_unlock(&orangefs_bufmap_lock);
389 gossip_debug(GOSSIP_BUFMAP_DEBUG,
390 "orangefs_bufmap_initialize: exiting normally\n");
394 orangefs_bufmap_unmap(bufmap);
396 orangefs_bufmap_free(bufmap);
402 * orangefs_bufmap_finalize()
404 * shuts down the mapped buffer interface and releases any resources
409 void orangefs_bufmap_finalize(void)
411 struct orangefs_bufmap *bufmap = __orangefs_bufmap;
414 gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
415 mark_killed(&rw_map);
416 mark_killed(&readdir_map);
417 gossip_debug(GOSSIP_BUFMAP_DEBUG,
418 "orangefs_bufmap_finalize: exiting normally\n");
421 void orangefs_bufmap_run_down(void)
423 struct orangefs_bufmap *bufmap = __orangefs_bufmap;
427 run_down(&readdir_map);
428 spin_lock(&orangefs_bufmap_lock);
429 __orangefs_bufmap = NULL;
430 spin_unlock(&orangefs_bufmap_lock);
431 orangefs_bufmap_unmap(bufmap);
432 orangefs_bufmap_free(bufmap);
436 * orangefs_bufmap_get()
438 * gets a free mapped buffer descriptor, will sleep until one becomes
439 * available if necessary
441 * returns slot on success, -errno on failure
443 int orangefs_bufmap_get(void)
449 * orangefs_bufmap_put()
451 * returns a mapped buffer descriptor to the collection
455 void orangefs_bufmap_put(int buffer_index)
457 put(&rw_map, buffer_index);
461 * orangefs_readdir_index_get()
463 * gets a free descriptor, will sleep until one becomes
464 * available if necessary.
465 * Although the readdir buffers are not mapped into kernel space
466 * we could do that at a later point of time. Regardless, these
467 * indices are used by the client-core.
469 * returns slot on success, -errno on failure
471 int orangefs_readdir_index_get(void)
473 return get(&readdir_map);
476 void orangefs_readdir_index_put(int buffer_index)
478 put(&readdir_map, buffer_index);
482 * we've been handed an iovec, we need to copy it to
483 * the shared memory descriptor at "buffer_index".
485 int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
489 struct orangefs_bufmap_desc *to;
492 gossip_debug(GOSSIP_BUFMAP_DEBUG,
493 "%s: buffer_index:%d: size:%zu:\n",
494 __func__, buffer_index, size);
496 to = &__orangefs_bufmap->desc_array[buffer_index];
497 for (i = 0; size; i++) {
498 struct page *page = to->page_array[i];
502 if (copy_page_from_iter(page, 0, n, iter) != n)
510 * we've been handed an iovec, we need to fill it from
511 * the shared memory descriptor at "buffer_index".
513 int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
517 struct orangefs_bufmap_desc *from;
520 from = &__orangefs_bufmap->desc_array[buffer_index];
521 gossip_debug(GOSSIP_BUFMAP_DEBUG,
522 "%s: buffer_index:%d: size:%zu:\n",
523 __func__, buffer_index, size);
526 for (i = 0; size; i++) {
527 struct page *page = from->page_array[i];
531 n = copy_page_to_iter(page, 0, n, iter);
539 void orangefs_bufmap_page_fill(void *page_to,
543 struct orangefs_bufmap_desc *from;
546 from = &__orangefs_bufmap->desc_array[buffer_index];
547 page_from = kmap_atomic(from->page_array[slot_index]);
548 memcpy(page_to, page_from, PAGE_SIZE);
549 kunmap_atomic(page_from);