fs/orangefs/orangefs-bufmap.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * (C) 2001 Clemson University and The University of Chicago
0004  *
0005  * See COPYING in top-level directory.
0006  */
0007 #include "protocol.h"
0008 #include "orangefs-kernel.h"
0009 #include "orangefs-bufmap.h"
0010
0011 struct slot_map {
0012     int c;
0013     wait_queue_head_t q;
0014     int count;
0015     unsigned long *map;
0016 };
0017
0018 static struct slot_map rw_map = {
0019     .c = -1,
0020     .q = __WAIT_QUEUE_HEAD_INITIALIZER(rw_map.q)
0021 };
0022 static struct slot_map readdir_map = {
0023     .c = -1,
0024     .q = __WAIT_QUEUE_HEAD_INITIALIZER(readdir_map.q)
0025 };
0026
0027
0028 static void install(struct slot_map *m, int count, unsigned long *map)
0029 {
0030     spin_lock(&m->q.lock);
0031     m->c = m->count = count;
0032     m->map = map;
0033     wake_up_all_locked(&m->q);
0034     spin_unlock(&m->q.lock);
0035 }
0036
0037 static void mark_killed(struct slot_map *m)
0038 {
0039     spin_lock(&m->q.lock);
0040     m->c -= m->count + 1;
0041     spin_unlock(&m->q.lock);
0042 }
0043
0044 static void run_down(struct slot_map *m)
0045 {
0046     DEFINE_WAIT(wait);
0047     spin_lock(&m->q.lock);
0048     if (m->c != -1) {
0049         for (;;) {
0050             if (likely(list_empty(&wait.entry)))
0051                 __add_wait_queue_entry_tail(&m->q, &wait);
0052             set_current_state(TASK_UNINTERRUPTIBLE);
0053
0054             if (m->c == -1)
0055                 break;
0056
0057             spin_unlock(&m->q.lock);
0058             schedule();
0059             spin_lock(&m->q.lock);
0060         }
0061         __remove_wait_queue(&m->q, &wait);
0062         __set_current_state(TASK_RUNNING);
0063     }
0064     m->map = NULL;
0065     spin_unlock(&m->q.lock);
0066 }
0067
0068 static void put(struct slot_map *m, int slot)
0069 {
0070     int v;
0071     spin_lock(&m->q.lock);
0072     __clear_bit(slot, m->map);
0073     v = ++m->c;
0074     if (v > 0)
0075         wake_up_locked(&m->q);
0076     if (unlikely(v == -1))     /* finished dying */
0077         wake_up_all_locked(&m->q);
0078     spin_unlock(&m->q.lock);
0079 }
0080
0081 static int wait_for_free(struct slot_map *m)
0082 {
0083     long left = slot_timeout_secs * HZ;
0084     DEFINE_WAIT(wait);
0085
0086     do {
0087         long n = left, t;
0088         if (likely(list_empty(&wait.entry)))
0089             __add_wait_queue_entry_tail_exclusive(&m->q, &wait);
0090         set_current_state(TASK_INTERRUPTIBLE);
0091
0092         if (m->c > 0)
0093             break;
0094
0095         if (m->c < 0) {
0096             /* we are waiting for map to be installed */
0097             /* it would better be there soon, or we go away */
0098             if (n > ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ)
0099                 n = ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ;
0100         }
0101         spin_unlock(&m->q.lock);
0102         t = schedule_timeout(n);
0103         spin_lock(&m->q.lock);
0104         if (unlikely(!t) && n != left && m->c < 0)
0105             left = t;
0106         else
0107             left = t + (left - n);
0108         if (signal_pending(current))
0109             left = -EINTR;
0110     } while (left > 0);
0111
0112     if (!list_empty(&wait.entry))
0113         list_del(&wait.entry);
0114     else if (left <= 0 && waitqueue_active(&m->q))
0115         __wake_up_locked_key(&m->q, TASK_INTERRUPTIBLE, NULL);
0116     __set_current_state(TASK_RUNNING);
0117
0118     if (likely(left > 0))
0119         return 0;
0120
0121     return left < 0 ? -EINTR : -ETIMEDOUT;
0122 }
0123
0124 static int get(struct slot_map *m)
0125 {
0126     int res = 0;
0127     spin_lock(&m->q.lock);
0128     if (unlikely(m->c <= 0))
0129         res = wait_for_free(m);
0130     if (likely(!res)) {
0131         m->c--;
0132         res = find_first_zero_bit(m->map, m->count);
0133         __set_bit(res, m->map);
0134     }
0135     spin_unlock(&m->q.lock);
0136     return res;
0137 }
0138
0139 /* used to describe mapped buffers */
0140 struct orangefs_bufmap_desc {
0141     void __user *uaddr;     /* user space address pointer */
0142     struct page **page_array;   /* array of mapped pages */
0143     int array_count;        /* size of above arrays */
0144     struct list_head list_link;
0145 };
0146
0147 static struct orangefs_bufmap {
0148     int desc_size;
0149     int desc_shift;
0150     int desc_count;
0151     int total_size;
0152     int page_count;
0153
0154     struct page **page_array;
0155     struct orangefs_bufmap_desc *desc_array;
0156
0157     /* array to track usage of buffer descriptors */
0158     unsigned long *buffer_index_array;
0159
0160     /* array to track usage of buffer descriptors for readdir */
0161 #define N DIV_ROUND_UP(ORANGEFS_READDIR_DEFAULT_DESC_COUNT, BITS_PER_LONG)
0162     unsigned long readdir_index_array[N];
0163 #undef N
0164 } *__orangefs_bufmap;
0165
0166 static DEFINE_SPINLOCK(orangefs_bufmap_lock);
0167
0168 static void
0169 orangefs_bufmap_unmap(struct orangefs_bufmap *bufmap)
0170 {
0171     unpin_user_pages(bufmap->page_array, bufmap->page_count);
0172 }
0173
0174 static void
0175 orangefs_bufmap_free(struct orangefs_bufmap *bufmap)
0176 {
0177     kfree(bufmap->page_array);
0178     kfree(bufmap->desc_array);
0179     bitmap_free(bufmap->buffer_index_array);
0180     kfree(bufmap);
0181 }
0182
0183 /*
0184  * XXX: Can the size and shift change while the caller gives up the
0185  * XXX: lock between calling this and doing something useful?
0186  */
0187
0188 int orangefs_bufmap_size_query(void)
0189 {
0190     struct orangefs_bufmap *bufmap;
0191     int size = 0;
0192     spin_lock(&orangefs_bufmap_lock);
0193     bufmap = __orangefs_bufmap;
0194     if (bufmap)
0195         size = bufmap->desc_size;
0196     spin_unlock(&orangefs_bufmap_lock);
0197     return size;
0198 }
0199
0200 int orangefs_bufmap_shift_query(void)
0201 {
0202     struct orangefs_bufmap *bufmap;
0203     int shift = 0;
0204     spin_lock(&orangefs_bufmap_lock);
0205     bufmap = __orangefs_bufmap;
0206     if (bufmap)
0207         shift = bufmap->desc_shift;
0208     spin_unlock(&orangefs_bufmap_lock);
0209     return shift;
0210 }
0211
0212 static DECLARE_WAIT_QUEUE_HEAD(bufmap_waitq);
0213 static DECLARE_WAIT_QUEUE_HEAD(readdir_waitq);
0214
0215 static struct orangefs_bufmap *
0216 orangefs_bufmap_alloc(struct ORANGEFS_dev_map_desc *user_desc)
0217 {
0218     struct orangefs_bufmap *bufmap;
0219
0220     bufmap = kzalloc(sizeof(*bufmap), GFP_KERNEL);
0221     if (!bufmap)
0222         goto out;
0223
0224     bufmap->total_size = user_desc->total_size;
0225     bufmap->desc_count = user_desc->count;
0226     bufmap->desc_size = user_desc->size;
0227     bufmap->desc_shift = ilog2(bufmap->desc_size);
0228
0229     bufmap->buffer_index_array = bitmap_zalloc(bufmap->desc_count, GFP_KERNEL);
0230     if (!bufmap->buffer_index_array)
0231         goto out_free_bufmap;
0232
0233     bufmap->desc_array =
0234         kcalloc(bufmap->desc_count, sizeof(struct orangefs_bufmap_desc),
0235             GFP_KERNEL);
0236     if (!bufmap->desc_array)
0237         goto out_free_index_array;
0238
0239     bufmap->page_count = bufmap->total_size / PAGE_SIZE;
0240
0241     /* allocate storage to track our page mappings */
0242     bufmap->page_array =
0243         kcalloc(bufmap->page_count, sizeof(struct page *), GFP_KERNEL);
0244     if (!bufmap->page_array)
0245         goto out_free_desc_array;
0246
0247     return bufmap;
0248
0249 out_free_desc_array:
0250     kfree(bufmap->desc_array);
0251 out_free_index_array:
0252     bitmap_free(bufmap->buffer_index_array);
0253 out_free_bufmap:
0254     kfree(bufmap);
0255 out:
0256     return NULL;
0257 }
0258
0259 static int
0260 orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
0261         struct ORANGEFS_dev_map_desc *user_desc)
0262 {
0263     int pages_per_desc = bufmap->desc_size / PAGE_SIZE;
0264     int offset = 0, ret, i;
0265
0266     /* map the pages */
0267     ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
0268                  bufmap->page_count, FOLL_WRITE, bufmap->page_array);
0269
0270     if (ret < 0)
0271         return ret;
0272
0273     if (ret != bufmap->page_count) {
0274         gossip_err("orangefs error: asked for %d pages, only got %d.\n",
0275                 bufmap->page_count, ret);
0276
0277         for (i = 0; i < ret; i++) {
0278             SetPageError(bufmap->page_array[i]);
0279             unpin_user_page(bufmap->page_array[i]);
0280         }
0281         return -ENOMEM;
0282     }
0283
0284     /*
0285      * ideally we want to get kernel space pointers for each page, but
0286      * we can't kmap that many pages at once if highmem is being used.
0287      * so instead, we just kmap/kunmap the page address each time the
0288      * kaddr is needed.
0289      */
0290     for (i = 0; i < bufmap->page_count; i++)
0291         flush_dcache_page(bufmap->page_array[i]);
0292
0293     /* build a list of available descriptors */
0294     for (offset = 0, i = 0; i < bufmap->desc_count; i++) {
0295         bufmap->desc_array[i].page_array = &bufmap->page_array[offset];
0296         bufmap->desc_array[i].array_count = pages_per_desc;
0297         bufmap->desc_array[i].uaddr =
0298             (user_desc->ptr + (i * pages_per_desc * PAGE_SIZE));
0299         offset += pages_per_desc;
0300     }
0301
0302     return 0;
0303 }
0304
0305 /*
0306  * orangefs_bufmap_initialize()
0307  *
0308  * initializes the mapped buffer interface
0309  *
0310  * returns 0 on success, -errno on failure
0311  */
0312 int orangefs_bufmap_initialize(struct ORANGEFS_dev_map_desc *user_desc)
0313 {
0314     struct orangefs_bufmap *bufmap;
0315     int ret = -EINVAL;
0316
0317     gossip_debug(GOSSIP_BUFMAP_DEBUG,
0318              "orangefs_bufmap_initialize: called (ptr ("
0319              "%p) sz (%d) cnt(%d).\n",
0320              user_desc->ptr,
0321              user_desc->size,
0322              user_desc->count);
0323
0324     if (user_desc->total_size < 0 ||
0325         user_desc->size < 0 ||
0326         user_desc->count < 0)
0327         goto out;
0328
0329     /*
0330      * sanity check alignment and size of buffer that caller wants to
0331      * work with
0332      */
0333     if (PAGE_ALIGN((unsigned long)user_desc->ptr) !=
0334         (unsigned long)user_desc->ptr) {
0335         gossip_err("orangefs error: memory alignment (front). %p\n",
0336                user_desc->ptr);
0337         goto out;
0338     }
0339
0340     if (PAGE_ALIGN(((unsigned long)user_desc->ptr + user_desc->total_size))
0341         != (unsigned long)(user_desc->ptr + user_desc->total_size)) {
0342         gossip_err("orangefs error: memory alignment (back).(%p + %d)\n",
0343                user_desc->ptr,
0344                user_desc->total_size);
0345         goto out;
0346     }
0347
0348     if (user_desc->total_size != (user_desc->size * user_desc->count)) {
0349         gossip_err("orangefs error: user provided an oddly sized buffer: (%d, %d, %d)\n",
0350                user_desc->total_size,
0351                user_desc->size,
0352                user_desc->count);
0353         goto out;
0354     }
0355
0356     if ((user_desc->size % PAGE_SIZE) != 0) {
0357         gossip_err("orangefs error: bufmap size not page size divisible (%d).\n",
0358                user_desc->size);
0359         goto out;
0360     }
0361
0362     ret = -ENOMEM;
0363     bufmap = orangefs_bufmap_alloc(user_desc);
0364     if (!bufmap)
0365         goto out;
0366
0367     ret = orangefs_bufmap_map(bufmap, user_desc);
0368     if (ret)
0369         goto out_free_bufmap;
0370
0371
0372     spin_lock(&orangefs_bufmap_lock);
0373     if (__orangefs_bufmap) {
0374         spin_unlock(&orangefs_bufmap_lock);
0375         gossip_err("orangefs: error: bufmap already initialized.\n");
0376         ret = -EINVAL;
0377         goto out_unmap_bufmap;
0378     }
0379     __orangefs_bufmap = bufmap;
0380     install(&rw_map,
0381         bufmap->desc_count,
0382         bufmap->buffer_index_array);
0383     install(&readdir_map,
0384         ORANGEFS_READDIR_DEFAULT_DESC_COUNT,
0385         bufmap->readdir_index_array);
0386     spin_unlock(&orangefs_bufmap_lock);
0387
0388     gossip_debug(GOSSIP_BUFMAP_DEBUG,
0389              "orangefs_bufmap_initialize: exiting normally\n");
0390     return 0;
0391
0392 out_unmap_bufmap:
0393     orangefs_bufmap_unmap(bufmap);
0394 out_free_bufmap:
0395     orangefs_bufmap_free(bufmap);
0396 out:
0397     return ret;
0398 }
0399
0400 /*
0401  * orangefs_bufmap_finalize()
0402  *
0403  * shuts down the mapped buffer interface and releases any resources
0404  * associated with it
0405  *
0406  * no return value
0407  */
0408 void orangefs_bufmap_finalize(void)
0409 {
0410     struct orangefs_bufmap *bufmap = __orangefs_bufmap;
0411     if (!bufmap)
0412         return;
0413     gossip_debug(GOSSIP_BUFMAP_DEBUG, "orangefs_bufmap_finalize: called\n");
0414     mark_killed(&rw_map);
0415     mark_killed(&readdir_map);
0416     gossip_debug(GOSSIP_BUFMAP_DEBUG,
0417              "orangefs_bufmap_finalize: exiting normally\n");
0418 }
0419
0420 void orangefs_bufmap_run_down(void)
0421 {
0422     struct orangefs_bufmap *bufmap = __orangefs_bufmap;
0423     if (!bufmap)
0424         return;
0425     run_down(&rw_map);
0426     run_down(&readdir_map);
0427     spin_lock(&orangefs_bufmap_lock);
0428     __orangefs_bufmap = NULL;
0429     spin_unlock(&orangefs_bufmap_lock);
0430     orangefs_bufmap_unmap(bufmap);
0431     orangefs_bufmap_free(bufmap);
0432 }
0433
0434 /*
0435  * orangefs_bufmap_get()
0436  *
0437  * gets a free mapped buffer descriptor, will sleep until one becomes
0438  * available if necessary
0439  *
0440  * returns slot on success, -errno on failure
0441  */
0442 int orangefs_bufmap_get(void)
0443 {
0444     return get(&rw_map);
0445 }
0446
0447 /*
0448  * orangefs_bufmap_put()
0449  *
0450  * returns a mapped buffer descriptor to the collection
0451  *
0452  * no return value
0453  */
0454 void orangefs_bufmap_put(int buffer_index)
0455 {
0456     put(&rw_map, buffer_index);
0457 }
0458
0459 /*
0460  * orangefs_readdir_index_get()
0461  *
0462  * gets a free descriptor, will sleep until one becomes
0463  * available if necessary.
0464  * Although the readdir buffers are not mapped into kernel space
0465  * we could do that at a later point of time. Regardless, these
0466  * indices are used by the client-core.
0467  *
0468  * returns slot on success, -errno on failure
0469  */
0470 int orangefs_readdir_index_get(void)
0471 {
0472     return get(&readdir_map);
0473 }
0474
0475 void orangefs_readdir_index_put(int buffer_index)
0476 {
0477     put(&readdir_map, buffer_index);
0478 }
0479
0480 /*
0481  * we've been handed an iovec, we need to copy it to
0482  * the shared memory descriptor at "buffer_index".
0483  */
0484 int orangefs_bufmap_copy_from_iovec(struct iov_iter *iter,
0485                 int buffer_index,
0486                 size_t size)
0487 {
0488     struct orangefs_bufmap_desc *to;
0489     int i;
0490
0491     gossip_debug(GOSSIP_BUFMAP_DEBUG,
0492              "%s: buffer_index:%d: size:%zu:\n",
0493              __func__, buffer_index, size);
0494
0495     to = &__orangefs_bufmap->desc_array[buffer_index];
0496     for (i = 0; size; i++) {
0497         struct page *page = to->page_array[i];
0498         size_t n = size;
0499         if (n > PAGE_SIZE)
0500             n = PAGE_SIZE;
0501         if (copy_page_from_iter(page, 0, n, iter) != n)
0502             return -EFAULT;
0503         size -= n;
0504     }
0505     return 0;
0506 }
0507
0508 /*
0509  * we've been handed an iovec, we need to fill it from
0510  * the shared memory descriptor at "buffer_index".
0511  */
0512 int orangefs_bufmap_copy_to_iovec(struct iov_iter *iter,
0513                     int buffer_index,
0514                     size_t size)
0515 {
0516     struct orangefs_bufmap_desc *from;
0517     int i;
0518
0519     from = &__orangefs_bufmap->desc_array[buffer_index];
0520     gossip_debug(GOSSIP_BUFMAP_DEBUG,
0521              "%s: buffer_index:%d: size:%zu:\n",
0522              __func__, buffer_index, size);
0523
0524
0525     for (i = 0; size; i++) {
0526         struct page *page = from->page_array[i];
0527         size_t n = size;
0528         if (n > PAGE_SIZE)
0529             n = PAGE_SIZE;
0530         n = copy_page_to_iter(page, 0, n, iter);
0531         if (!n)
0532             return -EFAULT;
0533         size -= n;
0534     }
0535     return 0;
0536 }
0537
0538 void orangefs_bufmap_page_fill(void *page_to,
0539                 int buffer_index,
0540                 int slot_index)
0541 {
0542     struct orangefs_bufmap_desc *from;
0543     void *page_from;
0544
0545     from = &__orangefs_bufmap->desc_array[buffer_index];
0546     page_from = kmap_atomic(from->page_array[slot_index]);
0547     memcpy(page_to, page_from, PAGE_SIZE);
0548     kunmap_atomic(page_from);
0549 }