Merge tag 'xtensa-20190510' of git://github.com/jcmvbkbc/linux-xtensa
[sfrench/cifs-2.6.git] / fs / orangefs / file.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * (C) 2001 Clemson University and The University of Chicago
4  * Copyright 2018 Omnibond Systems, L.L.C.
5  *
6  * See COPYING in top-level directory.
7  */
8
9 /*
10  *  Linux VFS file operations.
11  */
12
13 #include "protocol.h"
14 #include "orangefs-kernel.h"
15 #include "orangefs-bufmap.h"
16 #include <linux/fs.h>
17 #include <linux/pagemap.h>
18
19 static int flush_racache(struct inode *inode)
20 {
21         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
22         struct orangefs_kernel_op_s *new_op;
23         int ret;
24
25         gossip_debug(GOSSIP_UTILS_DEBUG,
26             "%s: %pU: Handle is %pU | fs_id %d\n", __func__,
27             get_khandle_from_ino(inode), &orangefs_inode->refn.khandle,
28             orangefs_inode->refn.fs_id);
29
30         new_op = op_alloc(ORANGEFS_VFS_OP_RA_FLUSH);
31         if (!new_op)
32                 return -ENOMEM;
33         new_op->upcall.req.ra_cache_flush.refn = orangefs_inode->refn;
34
35         ret = service_operation(new_op, "orangefs_flush_racache",
36             get_interruptible_flag(inode));
37
38         gossip_debug(GOSSIP_UTILS_DEBUG, "%s: got return value of %d\n",
39             __func__, ret);
40
41         op_release(new_op);
42         return ret;
43 }
44
45 /*
46  * Post and wait for the I/O upcall to finish
47  */
48 ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
49     loff_t *offset, struct iov_iter *iter, size_t total_size,
50     loff_t readahead_size, struct orangefs_write_range *wr, int *index_return)
51 {
52         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
53         struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
54         struct orangefs_kernel_op_s *new_op = NULL;
55         int buffer_index = -1;
56         ssize_t ret;
57         size_t copy_amount;
58
59         new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
60         if (!new_op)
61                 return -ENOMEM;
62
63         /* synchronous I/O */
64         new_op->upcall.req.io.readahead_size = readahead_size;
65         new_op->upcall.req.io.io_type = type;
66         new_op->upcall.req.io.refn = orangefs_inode->refn;
67
68 populate_shared_memory:
69         /* get a shared buffer index */
70         buffer_index = orangefs_bufmap_get();
71         if (buffer_index < 0) {
72                 ret = buffer_index;
73                 gossip_debug(GOSSIP_FILE_DEBUG,
74                              "%s: orangefs_bufmap_get failure (%zd)\n",
75                              __func__, ret);
76                 goto out;
77         }
78         gossip_debug(GOSSIP_FILE_DEBUG,
79                      "%s(%pU): GET op %p -> buffer_index %d\n",
80                      __func__,
81                      handle,
82                      new_op,
83                      buffer_index);
84
85         new_op->uses_shared_memory = 1;
86         new_op->upcall.req.io.buf_index = buffer_index;
87         new_op->upcall.req.io.count = total_size;
88         new_op->upcall.req.io.offset = *offset;
89         if (type == ORANGEFS_IO_WRITE && wr) {
90                 new_op->upcall.uid = from_kuid(&init_user_ns, wr->uid);
91                 new_op->upcall.gid = from_kgid(&init_user_ns, wr->gid);
92         }
93
94         gossip_debug(GOSSIP_FILE_DEBUG,
95                      "%s(%pU): offset: %llu total_size: %zd\n",
96                      __func__,
97                      handle,
98                      llu(*offset),
99                      total_size);
100         /*
101          * Stage 1: copy the buffers into client-core's address space
102          */
103         if (type == ORANGEFS_IO_WRITE && total_size) {
104                 ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
105                     total_size);
106                 if (ret < 0) {
107                         gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
108                             __func__, (long)ret);
109                         goto out;
110                 }
111         }
112
113         gossip_debug(GOSSIP_FILE_DEBUG,
114                      "%s(%pU): Calling post_io_request with tag (%llu)\n",
115                      __func__,
116                      handle,
117                      llu(new_op->tag));
118
119         /* Stage 2: Service the I/O operation */
120         ret = service_operation(new_op,
121                                 type == ORANGEFS_IO_WRITE ?
122                                         "file_write" :
123                                         "file_read",
124                                 get_interruptible_flag(inode));
125
126         /*
127          * If service_operation() returns -EAGAIN #and# the operation was
128          * purged from orangefs_request_list or htable_ops_in_progress, then
129          * we know that the client was restarted, causing the shared memory
130          * area to be wiped clean.  To restart a  write operation in this
131          * case, we must re-copy the data from the user's iovec to a NEW
132          * shared memory location. To restart a read operation, we must get
133          * a new shared memory location.
134          */
135         if (ret == -EAGAIN && op_state_purged(new_op)) {
136                 orangefs_bufmap_put(buffer_index);
137                 buffer_index = -1;
138                 if (type == ORANGEFS_IO_WRITE)
139                         iov_iter_revert(iter, total_size);
140                 gossip_debug(GOSSIP_FILE_DEBUG,
141                              "%s:going to repopulate_shared_memory.\n",
142                              __func__);
143                 goto populate_shared_memory;
144         }
145
146         if (ret < 0) {
147                 if (ret == -EINTR) {
148                         /*
149                          * We can't return EINTR if any data was written,
150                          * it's not POSIX. It is minimally acceptable
151                          * to give a partial write, the way NFS does.
152                          *
153                          * It would be optimal to return all or nothing,
154                          * but if a userspace write is bigger than
155                          * an IO buffer, and the interrupt occurs
156                          * between buffer writes, that would not be
157                          * possible.
158                          */
159                         switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
160                         /*
161                          * If the op was waiting when the interrupt
162                          * occurred, then the client-core did not
163                          * trigger the write.
164                          */
165                         case OP_VFS_STATE_WAITING:
166                                 if (*offset == 0)
167                                         ret = -EINTR;
168                                 else
169                                         ret = 0;
170                                 break;
171                         /*
172                          * If the op was in progress when the interrupt
173                          * occurred, then the client-core was able to
174                          * trigger the write.
175                          */
176                         case OP_VFS_STATE_INPROGR:
177                                 if (type == ORANGEFS_IO_READ)
178                                         ret = -EINTR;
179                                 else
180                                         ret = total_size;
181                                 break;
182                         default:
183                                 gossip_err("%s: unexpected op state :%d:.\n",
184                                            __func__,
185                                            new_op->op_state);
186                                 ret = 0;
187                                 break;
188                         }
189                         gossip_debug(GOSSIP_FILE_DEBUG,
190                                      "%s: got EINTR, state:%d: %p\n",
191                                      __func__,
192                                      new_op->op_state,
193                                      new_op);
194                 } else {
195                         gossip_err("%s: error in %s handle %pU, returning %zd\n",
196                                 __func__,
197                                 type == ORANGEFS_IO_READ ?
198                                         "read from" : "write to",
199                                 handle, ret);
200                 }
201                 if (orangefs_cancel_op_in_progress(new_op))
202                         return ret;
203
204                 goto out;
205         }
206
207         /*
208          * Stage 3: Post copy buffers from client-core's address space
209          */
210         if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
211                 /*
212                  * NOTE: the iovector can either contain addresses which
213                  *       can futher be kernel-space or user-space addresses.
214                  *       or it can pointers to struct page's
215                  */
216
217                 /*
218                  * When reading, readahead_size will only be zero when
219                  * we're doing O_DIRECT, otherwise we got here from
220                  * orangefs_readpage.
221                  *
222                  * If we got here from orangefs_readpage we want to
223                  * copy either a page or the whole file into the io
224                  * vector, whichever is smaller.
225                  */
226                 if (readahead_size)
227                         copy_amount =
228                                 min(new_op->downcall.resp.io.amt_complete,
229                                         (__s64)PAGE_SIZE);
230                 else
231                         copy_amount = new_op->downcall.resp.io.amt_complete;
232
233                 ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
234                         copy_amount);
235                 if (ret < 0) {
236                         gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
237                             __func__, (long)ret);
238                         goto out;
239                 }
240         }
241         gossip_debug(GOSSIP_FILE_DEBUG,
242             "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
243             __func__,
244             handle,
245             type == ORANGEFS_IO_READ ?  "read" : "written",
246             (int)new_op->downcall.resp.io.amt_complete);
247
248         ret = new_op->downcall.resp.io.amt_complete;
249
250 out:
251         if (buffer_index >= 0) {
252                 if ((readahead_size) && (type == ORANGEFS_IO_READ)) {
253                         /* readpage */
254                         *index_return = buffer_index;
255                         gossip_debug(GOSSIP_FILE_DEBUG,
256                                 "%s: hold on to buffer_index :%d:\n",
257                                 __func__, buffer_index);
258                 } else {
259                         /* O_DIRECT */
260                         orangefs_bufmap_put(buffer_index);
261                         gossip_debug(GOSSIP_FILE_DEBUG,
262                                 "%s(%pU): PUT buffer_index %d\n",
263                                 __func__, handle, buffer_index);
264                 }
265                 buffer_index = -1;
266         }
267         op_release(new_op);
268         return ret;
269 }
270
271 int orangefs_revalidate_mapping(struct inode *inode)
272 {
273         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
274         struct address_space *mapping = inode->i_mapping;
275         unsigned long *bitlock = &orangefs_inode->bitlock;
276         int ret;
277
278         while (1) {
279                 ret = wait_on_bit(bitlock, 1, TASK_KILLABLE);
280                 if (ret)
281                         return ret;
282                 spin_lock(&inode->i_lock);
283                 if (test_bit(1, bitlock)) {
284                         spin_unlock(&inode->i_lock);
285                         continue;
286                 }
287                 if (!time_before(jiffies, orangefs_inode->mapping_time))
288                         break;
289                 spin_unlock(&inode->i_lock);
290                 return 0;
291         }
292
293         set_bit(1, bitlock);
294         smp_wmb();
295         spin_unlock(&inode->i_lock);
296
297         unmap_mapping_range(mapping, 0, 0, 0);
298         ret = filemap_write_and_wait(mapping);
299         if (!ret)
300                 ret = invalidate_inode_pages2(mapping);
301
302         orangefs_inode->mapping_time = jiffies +
303             orangefs_cache_timeout_msecs*HZ/1000;
304
305         clear_bit(1, bitlock);
306         smp_mb__after_atomic();
307         wake_up_bit(bitlock, 1);
308
309         return ret;
310 }
311
312 static ssize_t orangefs_file_read_iter(struct kiocb *iocb,
313     struct iov_iter *iter)
314 {
315         int ret;
316         struct orangefs_read_options *ro;
317
318         orangefs_stats.reads++;
319
320         /*
321          * Remember how they set "count" in read(2) or pread(2) or whatever -
322          * users can use count as a knob to control orangefs io size and later
323          * we can try to help them fill as many pages as possible in readpage.
324          */
325         if (!iocb->ki_filp->private_data) {
326                 iocb->ki_filp->private_data = kmalloc(sizeof *ro, GFP_KERNEL);
327                 if (!iocb->ki_filp->private_data)
328                         return(ENOMEM);
329                 ro = iocb->ki_filp->private_data;
330                 ro->blksiz = iter->count;
331         }
332
333         down_read(&file_inode(iocb->ki_filp)->i_rwsem);
334         ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
335         if (ret)
336                 goto out;
337
338         ret = generic_file_read_iter(iocb, iter);
339 out:
340         up_read(&file_inode(iocb->ki_filp)->i_rwsem);
341         return ret;
342 }
343
344 static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
345     struct iov_iter *iter)
346 {
347         int ret;
348         orangefs_stats.writes++;
349
350         if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
351                 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
352                 if (ret)
353                         return ret;
354         }
355
356         ret = generic_file_write_iter(iocb, iter);
357         return ret;
358 }
359
360 /*
361  * Perform a miscellaneous operation on a file.
362  */
363 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
364 {
365         int ret = -ENOTTY;
366         __u64 val = 0;
367         unsigned long uval;
368
369         gossip_debug(GOSSIP_FILE_DEBUG,
370                      "orangefs_ioctl: called with cmd %d\n",
371                      cmd);
372
373         /*
374          * we understand some general ioctls on files, such as the immutable
375          * and append flags
376          */
377         if (cmd == FS_IOC_GETFLAGS) {
378                 val = 0;
379                 ret = orangefs_inode_getxattr(file_inode(file),
380                                               "user.pvfs2.meta_hint",
381                                               &val, sizeof(val));
382                 if (ret < 0 && ret != -ENODATA)
383                         return ret;
384                 else if (ret == -ENODATA)
385                         val = 0;
386                 uval = val;
387                 gossip_debug(GOSSIP_FILE_DEBUG,
388                              "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
389                              (unsigned long long)uval);
390                 return put_user(uval, (int __user *)arg);
391         } else if (cmd == FS_IOC_SETFLAGS) {
392                 ret = 0;
393                 if (get_user(uval, (int __user *)arg))
394                         return -EFAULT;
395                 /*
396                  * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
397                  * is turned on for a file. The user is not allowed to turn
398                  * on this bit, but the bit is present if the user first gets
399                  * the flags and then updates the flags with some new
400                  * settings. So, we ignore it in the following edit. bligon.
401                  */
402                 if ((uval & ~ORANGEFS_MIRROR_FL) &
403                     (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) {
404                         gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
405                         return -EINVAL;
406                 }
407                 val = uval;
408                 gossip_debug(GOSSIP_FILE_DEBUG,
409                              "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
410                              (unsigned long long)val);
411                 ret = orangefs_inode_setxattr(file_inode(file),
412                                               "user.pvfs2.meta_hint",
413                                               &val, sizeof(val), 0);
414         }
415
416         return ret;
417 }
418
419 static vm_fault_t orangefs_fault(struct vm_fault *vmf)
420 {
421         struct file *file = vmf->vma->vm_file;
422         int ret;
423         ret = orangefs_inode_getattr(file->f_mapping->host,
424             ORANGEFS_GETATTR_SIZE);
425         if (ret == -ESTALE)
426                 ret = -EIO;
427         if (ret) {
428                 gossip_err("%s: orangefs_inode_getattr failed, "
429                     "ret:%d:.\n", __func__, ret);
430                 return VM_FAULT_SIGBUS;
431         }
432         return filemap_fault(vmf);
433 }
434
435 static const struct vm_operations_struct orangefs_file_vm_ops = {
436         .fault = orangefs_fault,
437         .map_pages = filemap_map_pages,
438         .page_mkwrite = orangefs_page_mkwrite,
439 };
440
441 /*
442  * Memory map a region of a file.
443  */
444 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
445 {
446         int ret;
447
448         ret = orangefs_revalidate_mapping(file_inode(file));
449         if (ret)
450                 return ret;
451
452         gossip_debug(GOSSIP_FILE_DEBUG,
453                      "orangefs_file_mmap: called on %s\n",
454                      (file ?
455                         (char *)file->f_path.dentry->d_name.name :
456                         (char *)"Unknown"));
457
458         /* set the sequential readahead hint */
459         vma->vm_flags |= VM_SEQ_READ;
460         vma->vm_flags &= ~VM_RAND_READ;
461
462         file_accessed(file);
463         vma->vm_ops = &orangefs_file_vm_ops;
464         return 0;
465 }
466
467 #define mapping_nrpages(idata) ((idata)->nrpages)
468
469 /*
470  * Called to notify the module that there are no more references to
471  * this file (i.e. no processes have it open).
472  *
473  * \note Not called when each file is closed.
474  */
475 static int orangefs_file_release(struct inode *inode, struct file *file)
476 {
477         gossip_debug(GOSSIP_FILE_DEBUG,
478                      "orangefs_file_release: called on %pD\n",
479                      file);
480
481         /*
482          * remove all associated inode pages from the page cache and
483          * readahead cache (if any); this forces an expensive refresh of
484          * data for the next caller of mmap (or 'get_block' accesses)
485          */
486         if (file_inode(file) &&
487             file_inode(file)->i_mapping &&
488             mapping_nrpages(&file_inode(file)->i_data)) {
489                 if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
490                         gossip_debug(GOSSIP_INODE_DEBUG,
491                             "calling flush_racache on %pU\n",
492                             get_khandle_from_ino(inode));
493                         flush_racache(inode);
494                         gossip_debug(GOSSIP_INODE_DEBUG,
495                             "flush_racache finished\n");
496                 }
497
498         }
499         return 0;
500 }
501
502 /*
503  * Push all data for a specific file onto permanent storage.
504  */
505 static int orangefs_fsync(struct file *file,
506                        loff_t start,
507                        loff_t end,
508                        int datasync)
509 {
510         int ret;
511         struct orangefs_inode_s *orangefs_inode =
512                 ORANGEFS_I(file_inode(file));
513         struct orangefs_kernel_op_s *new_op = NULL;
514
515         ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
516             start, end);
517         if (ret < 0)
518                 return ret;
519
520         new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
521         if (!new_op)
522                 return -ENOMEM;
523         new_op->upcall.req.fsync.refn = orangefs_inode->refn;
524
525         ret = service_operation(new_op,
526                         "orangefs_fsync",
527                         get_interruptible_flag(file_inode(file)));
528
529         gossip_debug(GOSSIP_FILE_DEBUG,
530                      "orangefs_fsync got return value of %d\n",
531                      ret);
532
533         op_release(new_op);
534         return ret;
535 }
536
537 /*
538  * Change the file pointer position for an instance of an open file.
539  *
540  * \note If .llseek is overriden, we must acquire lock as described in
541  *       Documentation/filesystems/Locking.
542  *
543  * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
544  * require much changes to the FS
545  */
546 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
547 {
548         int ret = -EINVAL;
549         struct inode *inode = file_inode(file);
550
551         if (origin == SEEK_END) {
552                 /*
553                  * revalidate the inode's file size.
554                  * NOTE: We are only interested in file size here,
555                  * so we set mask accordingly.
556                  */
557                 ret = orangefs_inode_getattr(file->f_mapping->host,
558                     ORANGEFS_GETATTR_SIZE);
559                 if (ret == -ESTALE)
560                         ret = -EIO;
561                 if (ret) {
562                         gossip_debug(GOSSIP_FILE_DEBUG,
563                                      "%s:%s:%d calling make bad inode\n",
564                                      __FILE__,
565                                      __func__,
566                                      __LINE__);
567                         return ret;
568                 }
569         }
570
571         gossip_debug(GOSSIP_FILE_DEBUG,
572                      "orangefs_file_llseek: offset is %ld | origin is %d"
573                      " | inode size is %lu\n",
574                      (long)offset,
575                      origin,
576                      (unsigned long)i_size_read(inode));
577
578         return generic_file_llseek(file, offset, origin);
579 }
580
581 /*
582  * Support local locks (locks that only this kernel knows about)
583  * if Orangefs was mounted -o local_lock.
584  */
585 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
586 {
587         int rc = -EINVAL;
588
589         if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
590                 if (cmd == F_GETLK) {
591                         rc = 0;
592                         posix_test_lock(filp, fl);
593                 } else {
594                         rc = posix_lock_file(filp, fl, NULL);
595                 }
596         }
597
598         return rc;
599 }
600
601 static int orangefs_file_open(struct inode * inode, struct file *file)
602 {
603         file->private_data = NULL;
604         return generic_file_open(inode, file);
605 }
606
607 static int orangefs_flush(struct file *file, fl_owner_t id)
608 {
609         /*
610          * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
611          * service_operation in orangefs_fsync.
612          *
613          * Do not send fsync to OrangeFS server on a close.  Do send fsync
614          * on an explicit fsync call.  This duplicates historical OrangeFS
615          * behavior.
616          */
617         struct inode *inode = file->f_mapping->host;
618         int r;
619
620         kfree(file->private_data);
621         file->private_data = NULL;
622
623         if (inode->i_state & I_DIRTY_TIME) {
624                 spin_lock(&inode->i_lock);
625                 inode->i_state &= ~I_DIRTY_TIME;
626                 spin_unlock(&inode->i_lock);
627                 mark_inode_dirty_sync(inode);
628         }
629
630         r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
631         if (r > 0)
632                 return 0;
633         else
634                 return r;
635 }
636
637 /** ORANGEFS implementation of VFS file operations */
638 const struct file_operations orangefs_file_operations = {
639         .llseek         = orangefs_file_llseek,
640         .read_iter      = orangefs_file_read_iter,
641         .write_iter     = orangefs_file_write_iter,
642         .lock           = orangefs_lock,
643         .unlocked_ioctl = orangefs_ioctl,
644         .mmap           = orangefs_file_mmap,
645         .open           = orangefs_file_open,
646         .flush          = orangefs_flush,
647         .release        = orangefs_file_release,
648         .fsync          = orangefs_fsync,
649 };