e799bb4474f6c6c6921de103769b020d5a96c84f
[cascardo/linux.git] / fs / orangefs / file.c
1 /*
2  * (C) 2001 Clemson University and The University of Chicago
3  *
4  * See COPYING in top-level directory.
5  */
6
7 /*
8  *  Linux VFS file operations.
9  */
10
11 #include "protocol.h"
12 #include "orangefs-kernel.h"
13 #include "orangefs-bufmap.h"
14 #include <linux/fs.h>
15 #include <linux/pagemap.h>
16
17 /*
18  * Copy to client-core's address space from the buffers specified
19  * by the iovec upto total_size bytes.
20  * NOTE: the iovector can either contain addresses which
21  *       can futher be kernel-space or user-space addresses.
22  *       or it can pointers to struct page's
23  */
24 static int precopy_buffers(int buffer_index,
25                            struct iov_iter *iter,
26                            size_t total_size)
27 {
28         int ret = 0;
29         /*
30          * copy data from application/kernel by pulling it out
31          * of the iovec.
32          */
33
34
35         if (total_size) {
36                 ret = orangefs_bufmap_copy_from_iovec(iter,
37                                                       buffer_index,
38                                                       total_size);
39                 if (ret < 0)
40                 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
41                            __func__,
42                            (long)ret);
43         }
44
45         if (ret < 0)
46                 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
47                         __func__,
48                         (long)ret);
49         return ret;
50 }
51
52 /*
53  * Copy from client-core's address space to the buffers specified
54  * by the iovec upto total_size bytes.
55  * NOTE: the iovector can either contain addresses which
56  *       can futher be kernel-space or user-space addresses.
57  *       or it can pointers to struct page's
58  */
59 static int postcopy_buffers(int buffer_index,
60                             struct iov_iter *iter,
61                             size_t total_size)
62 {
63         int ret = 0;
64         /*
65          * copy data to application/kernel by pushing it out to
66          * the iovec. NOTE; target buffers can be addresses or
67          * struct page pointers.
68          */
69         if (total_size) {
70                 ret = orangefs_bufmap_copy_to_iovec(iter,
71                                                     buffer_index,
72                                                     total_size);
73                 if (ret < 0)
74                         gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
75                                 __func__,
76                                 (long)ret);
77         }
78         return ret;
79 }
80
81 /*
82  * Post and wait for the I/O upcall to finish
83  */
84 static ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
85                 loff_t *offset, struct iov_iter *iter,
86                 size_t total_size, loff_t readahead_size)
87 {
88         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
89         struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
90         struct orangefs_kernel_op_s *new_op = NULL;
91         struct iov_iter saved = *iter;
92         int buffer_index = -1;
93         ssize_t ret;
94
95         new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
96         if (!new_op)
97                 return -ENOMEM;
98
99         /* synchronous I/O */
100         new_op->upcall.req.io.readahead_size = readahead_size;
101         new_op->upcall.req.io.io_type = type;
102         new_op->upcall.req.io.refn = orangefs_inode->refn;
103
104 populate_shared_memory:
105         /* get a shared buffer index */
106         buffer_index = orangefs_bufmap_get();
107         if (buffer_index < 0) {
108                 ret = buffer_index;
109                 gossip_debug(GOSSIP_FILE_DEBUG,
110                              "%s: orangefs_bufmap_get failure (%zd)\n",
111                              __func__, ret);
112                 goto out;
113         }
114         gossip_debug(GOSSIP_FILE_DEBUG,
115                      "%s(%pU): GET op %p -> buffer_index %d\n",
116                      __func__,
117                      handle,
118                      new_op,
119                      buffer_index);
120
121         new_op->uses_shared_memory = 1;
122         new_op->upcall.req.io.buf_index = buffer_index;
123         new_op->upcall.req.io.count = total_size;
124         new_op->upcall.req.io.offset = *offset;
125
126         gossip_debug(GOSSIP_FILE_DEBUG,
127                      "%s(%pU): offset: %llu total_size: %zd\n",
128                      __func__,
129                      handle,
130                      llu(*offset),
131                      total_size);
132         /*
133          * Stage 1: copy the buffers into client-core's address space
134          * precopy_buffers only pertains to writes.
135          */
136         if (type == ORANGEFS_IO_WRITE) {
137                 ret = precopy_buffers(buffer_index,
138                                       iter,
139                                       total_size);
140                 if (ret < 0)
141                         goto out;
142         }
143
144         gossip_debug(GOSSIP_FILE_DEBUG,
145                      "%s(%pU): Calling post_io_request with tag (%llu)\n",
146                      __func__,
147                      handle,
148                      llu(new_op->tag));
149
150         /* Stage 2: Service the I/O operation */
151         ret = service_operation(new_op,
152                                 type == ORANGEFS_IO_WRITE ?
153                                         "file_write" :
154                                         "file_read",
155                                 get_interruptible_flag(inode));
156
157         /*
158          * If service_operation() returns -EAGAIN #and# the operation was
159          * purged from orangefs_request_list or htable_ops_in_progress, then
160          * we know that the client was restarted, causing the shared memory
161          * area to be wiped clean.  To restart a  write operation in this
162          * case, we must re-copy the data from the user's iovec to a NEW
163          * shared memory location. To restart a read operation, we must get
164          * a new shared memory location.
165          */
166         if (ret == -EAGAIN && op_state_purged(new_op)) {
167                 orangefs_bufmap_put(buffer_index);
168                 buffer_index = -1;
169                 if (type == ORANGEFS_IO_WRITE)
170                         *iter = saved;
171                 gossip_debug(GOSSIP_FILE_DEBUG,
172                              "%s:going to repopulate_shared_memory.\n",
173                              __func__);
174                 goto populate_shared_memory;
175         }
176
177         if (ret < 0) {
178                 if (ret == -EINTR) {
179                         /*
180                          * We can't return EINTR if any data was written,
181                          * it's not POSIX. It is minimally acceptable
182                          * to give a partial write, the way NFS does.
183                          *
184                          * It would be optimal to return all or nothing,
185                          * but if a userspace write is bigger than
186                          * an IO buffer, and the interrupt occurs
187                          * between buffer writes, that would not be
188                          * possible.
189                          */
190                         switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
191                         /*
192                          * If the op was waiting when the interrupt
193                          * occurred, then the client-core did not
194                          * trigger the write.
195                          */
196                         case OP_VFS_STATE_WAITING:
197                                 if (*offset == 0)
198                                         ret = -EINTR;
199                                 else
200                                         ret = 0;
201                                 break;
202                         /* 
203                          * If the op was in progress when the interrupt
204                          * occurred, then the client-core was able to
205                          * trigger the write.
206                          */
207                         case OP_VFS_STATE_INPROGR:
208                                 ret = total_size;
209                                 break;
210                         default:
211                                 gossip_err("%s: unexpected op state :%d:.\n",
212                                            __func__,
213                                            new_op->op_state);
214                                 ret = 0;
215                                 break;
216                         }
217                         gossip_debug(GOSSIP_FILE_DEBUG,
218                                      "%s: got EINTR, state:%d: %p\n",
219                                      __func__,
220                                      new_op->op_state,
221                                      new_op);
222                 } else {
223                         gossip_err("%s: error in %s handle %pU, returning %zd\n",
224                                 __func__,
225                                 type == ORANGEFS_IO_READ ?
226                                         "read from" : "write to",
227                                 handle, ret);
228                 }
229                 if (orangefs_cancel_op_in_progress(new_op))
230                         return ret;
231
232                 goto out;
233         }
234
235         /*
236          * Stage 3: Post copy buffers from client-core's address space
237          * postcopy_buffers only pertains to reads.
238          */
239         if (type == ORANGEFS_IO_READ) {
240                 ret = postcopy_buffers(buffer_index,
241                                        iter,
242                                        new_op->downcall.resp.io.amt_complete);
243                 if (ret < 0)
244                         goto out;
245         }
246         gossip_debug(GOSSIP_FILE_DEBUG,
247             "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
248             __func__,
249             handle,
250             type == ORANGEFS_IO_READ ?  "read" : "written",
251             (int)new_op->downcall.resp.io.amt_complete);
252
253         ret = new_op->downcall.resp.io.amt_complete;
254
255 out:
256         if (buffer_index >= 0) {
257                 orangefs_bufmap_put(buffer_index);
258                 gossip_debug(GOSSIP_FILE_DEBUG,
259                              "%s(%pU): PUT buffer_index %d\n",
260                              __func__, handle, buffer_index);
261                 buffer_index = -1;
262         }
263         op_release(new_op);
264         return ret;
265 }
266
267 /*
268  * Common entry point for read/write/readv/writev
269  * This function will dispatch it to either the direct I/O
270  * or buffered I/O path depending on the mount options and/or
271  * augmented/extended metadata attached to the file.
272  * Note: File extended attributes override any mount options.
273  */
274 static ssize_t do_readv_writev(enum ORANGEFS_io_type type, struct file *file,
275                 loff_t *offset, struct iov_iter *iter)
276 {
277         struct inode *inode = file->f_mapping->host;
278         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
279         struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
280         size_t count = iov_iter_count(iter);
281         ssize_t total_count = 0;
282         ssize_t ret = -EINVAL;
283
284         gossip_debug(GOSSIP_FILE_DEBUG,
285                 "%s-BEGIN(%pU): count(%d) after estimate_max_iovecs.\n",
286                 __func__,
287                 handle,
288                 (int)count);
289
290         if (type == ORANGEFS_IO_WRITE) {
291                 gossip_debug(GOSSIP_FILE_DEBUG,
292                              "%s(%pU): proceeding with offset : %llu, "
293                              "size %d\n",
294                              __func__,
295                              handle,
296                              llu(*offset),
297                              (int)count);
298         }
299
300         if (count == 0) {
301                 ret = 0;
302                 goto out;
303         }
304
305         while (iov_iter_count(iter)) {
306                 size_t each_count = iov_iter_count(iter);
307                 size_t amt_complete;
308
309                 /* how much to transfer in this loop iteration */
310                 if (each_count > orangefs_bufmap_size_query())
311                         each_count = orangefs_bufmap_size_query();
312
313                 gossip_debug(GOSSIP_FILE_DEBUG,
314                              "%s(%pU): size of each_count(%d)\n",
315                              __func__,
316                              handle,
317                              (int)each_count);
318                 gossip_debug(GOSSIP_FILE_DEBUG,
319                              "%s(%pU): BEFORE wait_for_io: offset is %d\n",
320                              __func__,
321                              handle,
322                              (int)*offset);
323
324                 ret = wait_for_direct_io(type, inode, offset, iter,
325                                 each_count, 0);
326                 gossip_debug(GOSSIP_FILE_DEBUG,
327                              "%s(%pU): return from wait_for_io:%d\n",
328                              __func__,
329                              handle,
330                              (int)ret);
331
332                 if (ret < 0)
333                         goto out;
334
335                 *offset += ret;
336                 total_count += ret;
337                 amt_complete = ret;
338
339                 gossip_debug(GOSSIP_FILE_DEBUG,
340                              "%s(%pU): AFTER wait_for_io: offset is %d\n",
341                              __func__,
342                              handle,
343                              (int)*offset);
344
345                 /*
346                  * if we got a short I/O operations,
347                  * fall out and return what we got so far
348                  */
349                 if (amt_complete < each_count)
350                         break;
351         } /*end while */
352
353 out:
354         if (total_count > 0)
355                 ret = total_count;
356         if (ret > 0) {
357                 if (type == ORANGEFS_IO_READ) {
358                         file_accessed(file);
359                 } else {
360                         SetMtimeFlag(orangefs_inode);
361                         inode->i_mtime = current_time(inode);
362                         mark_inode_dirty_sync(inode);
363                 }
364         }
365
366         gossip_debug(GOSSIP_FILE_DEBUG,
367                      "%s(%pU): Value(%d) returned.\n",
368                      __func__,
369                      handle,
370                      (int)ret);
371
372         return ret;
373 }
374
375 /*
376  * Read data from a specified offset in a file (referenced by inode).
377  * Data may be placed either in a user or kernel buffer.
378  */
379 ssize_t orangefs_inode_read(struct inode *inode,
380                             struct iov_iter *iter,
381                             loff_t *offset,
382                             loff_t readahead_size)
383 {
384         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
385         size_t count = iov_iter_count(iter);
386         size_t bufmap_size;
387         ssize_t ret = -EINVAL;
388
389         g_orangefs_stats.reads++;
390
391         bufmap_size = orangefs_bufmap_size_query();
392         if (count > bufmap_size) {
393                 gossip_debug(GOSSIP_FILE_DEBUG,
394                              "%s: count is too large (%zd/%zd)!\n",
395                              __func__, count, bufmap_size);
396                 return -EINVAL;
397         }
398
399         gossip_debug(GOSSIP_FILE_DEBUG,
400                      "%s(%pU) %zd@%llu\n",
401                      __func__,
402                      &orangefs_inode->refn.khandle,
403                      count,
404                      llu(*offset));
405
406         ret = wait_for_direct_io(ORANGEFS_IO_READ, inode, offset, iter,
407                         count, readahead_size);
408         if (ret > 0)
409                 *offset += ret;
410
411         gossip_debug(GOSSIP_FILE_DEBUG,
412                      "%s(%pU): Value(%zd) returned.\n",
413                      __func__,
414                      &orangefs_inode->refn.khandle,
415                      ret);
416
417         return ret;
418 }
419
420 static ssize_t orangefs_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
421 {
422         struct file *file = iocb->ki_filp;
423         loff_t pos = *(&iocb->ki_pos);
424         ssize_t rc = 0;
425
426         BUG_ON(iocb->private);
427
428         gossip_debug(GOSSIP_FILE_DEBUG, "orangefs_file_read_iter\n");
429
430         g_orangefs_stats.reads++;
431
432         rc = do_readv_writev(ORANGEFS_IO_READ, file, &pos, iter);
433         iocb->ki_pos = pos;
434
435         return rc;
436 }
437
438 static ssize_t orangefs_file_write_iter(struct kiocb *iocb, struct iov_iter *iter)
439 {
440         struct file *file = iocb->ki_filp;
441         loff_t pos;
442         ssize_t rc;
443
444         BUG_ON(iocb->private);
445
446         gossip_debug(GOSSIP_FILE_DEBUG, "orangefs_file_write_iter\n");
447
448         inode_lock(file->f_mapping->host);
449
450         /* Make sure generic_write_checks sees an up to date inode size. */
451         if (file->f_flags & O_APPEND) {
452                 rc = orangefs_inode_getattr(file->f_mapping->host, 0, 1);
453                 if (rc == -ESTALE)
454                         rc = -EIO;
455                 if (rc) {
456                         gossip_err("%s: orangefs_inode_getattr failed, "
457                             "rc:%zd:.\n", __func__, rc);
458                         goto out;
459                 }
460         }
461
462         if (file->f_pos > i_size_read(file->f_mapping->host))
463                 orangefs_i_size_write(file->f_mapping->host, file->f_pos);
464
465         rc = generic_write_checks(iocb, iter);
466
467         if (rc <= 0) {
468                 gossip_err("%s: generic_write_checks failed, rc:%zd:.\n",
469                            __func__, rc);
470                 goto out;
471         }
472
473         /*
474          * if we are appending, generic_write_checks would have updated
475          * pos to the end of the file, so we will wait till now to set
476          * pos...
477          */
478         pos = *(&iocb->ki_pos);
479
480         rc = do_readv_writev(ORANGEFS_IO_WRITE,
481                              file,
482                              &pos,
483                              iter);
484         if (rc < 0) {
485                 gossip_err("%s: do_readv_writev failed, rc:%zd:.\n",
486                            __func__, rc);
487                 goto out;
488         }
489
490         iocb->ki_pos = pos;
491         g_orangefs_stats.writes++;
492
493 out:
494
495         inode_unlock(file->f_mapping->host);
496         return rc;
497 }
498
499 /*
500  * Perform a miscellaneous operation on a file.
501  */
502 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
503 {
504         int ret = -ENOTTY;
505         __u64 val = 0;
506         unsigned long uval;
507
508         gossip_debug(GOSSIP_FILE_DEBUG,
509                      "orangefs_ioctl: called with cmd %d\n",
510                      cmd);
511
512         /*
513          * we understand some general ioctls on files, such as the immutable
514          * and append flags
515          */
516         if (cmd == FS_IOC_GETFLAGS) {
517                 val = 0;
518                 ret = orangefs_inode_getxattr(file_inode(file),
519                                               "user.pvfs2.meta_hint",
520                                               &val, sizeof(val));
521                 if (ret < 0 && ret != -ENODATA)
522                         return ret;
523                 else if (ret == -ENODATA)
524                         val = 0;
525                 uval = val;
526                 gossip_debug(GOSSIP_FILE_DEBUG,
527                              "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
528                              (unsigned long long)uval);
529                 return put_user(uval, (int __user *)arg);
530         } else if (cmd == FS_IOC_SETFLAGS) {
531                 ret = 0;
532                 if (get_user(uval, (int __user *)arg))
533                         return -EFAULT;
534                 /*
535                  * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
536                  * is turned on for a file. The user is not allowed to turn
537                  * on this bit, but the bit is present if the user first gets
538                  * the flags and then updates the flags with some new
539                  * settings. So, we ignore it in the following edit. bligon.
540                  */
541                 if ((uval & ~ORANGEFS_MIRROR_FL) &
542                     (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) {
543                         gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
544                         return -EINVAL;
545                 }
546                 val = uval;
547                 gossip_debug(GOSSIP_FILE_DEBUG,
548                              "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
549                              (unsigned long long)val);
550                 ret = orangefs_inode_setxattr(file_inode(file),
551                                               "user.pvfs2.meta_hint",
552                                               &val, sizeof(val), 0);
553         }
554
555         return ret;
556 }
557
558 /*
559  * Memory map a region of a file.
560  */
561 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
562 {
563         gossip_debug(GOSSIP_FILE_DEBUG,
564                      "orangefs_file_mmap: called on %s\n",
565                      (file ?
566                         (char *)file->f_path.dentry->d_name.name :
567                         (char *)"Unknown"));
568
569         /* set the sequential readahead hint */
570         vma->vm_flags |= VM_SEQ_READ;
571         vma->vm_flags &= ~VM_RAND_READ;
572
573         /* Use readonly mmap since we cannot support writable maps. */
574         return generic_file_readonly_mmap(file, vma);
575 }
576
577 #define mapping_nrpages(idata) ((idata)->nrpages)
578
579 /*
580  * Called to notify the module that there are no more references to
581  * this file (i.e. no processes have it open).
582  *
583  * \note Not called when each file is closed.
584  */
585 static int orangefs_file_release(struct inode *inode, struct file *file)
586 {
587         gossip_debug(GOSSIP_FILE_DEBUG,
588                      "orangefs_file_release: called on %s\n",
589                      file->f_path.dentry->d_name.name);
590
591         orangefs_flush_inode(inode);
592
593         /*
594          * remove all associated inode pages from the page cache and mmap
595          * readahead cache (if any); this forces an expensive refresh of
596          * data for the next caller of mmap (or 'get_block' accesses)
597          */
598         if (file->f_path.dentry->d_inode &&
599             file->f_path.dentry->d_inode->i_mapping &&
600             mapping_nrpages(&file->f_path.dentry->d_inode->i_data))
601                 truncate_inode_pages(file->f_path.dentry->d_inode->i_mapping,
602                                      0);
603         return 0;
604 }
605
606 /*
607  * Push all data for a specific file onto permanent storage.
608  */
609 static int orangefs_fsync(struct file *file,
610                        loff_t start,
611                        loff_t end,
612                        int datasync)
613 {
614         int ret = -EINVAL;
615         struct orangefs_inode_s *orangefs_inode =
616                 ORANGEFS_I(file->f_path.dentry->d_inode);
617         struct orangefs_kernel_op_s *new_op = NULL;
618
619         /* required call */
620         filemap_write_and_wait_range(file->f_mapping, start, end);
621
622         new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
623         if (!new_op)
624                 return -ENOMEM;
625         new_op->upcall.req.fsync.refn = orangefs_inode->refn;
626
627         ret = service_operation(new_op,
628                         "orangefs_fsync",
629                         get_interruptible_flag(file->f_path.dentry->d_inode));
630
631         gossip_debug(GOSSIP_FILE_DEBUG,
632                      "orangefs_fsync got return value of %d\n",
633                      ret);
634
635         op_release(new_op);
636
637         orangefs_flush_inode(file->f_path.dentry->d_inode);
638         return ret;
639 }
640
641 /*
642  * Change the file pointer position for an instance of an open file.
643  *
644  * \note If .llseek is overriden, we must acquire lock as described in
645  *       Documentation/filesystems/Locking.
646  *
647  * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
648  * require much changes to the FS
649  */
650 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
651 {
652         int ret = -EINVAL;
653         struct inode *inode = file_inode(file);
654
655         if (origin == SEEK_END) {
656                 /*
657                  * revalidate the inode's file size.
658                  * NOTE: We are only interested in file size here,
659                  * so we set mask accordingly.
660                  */
661                 ret = orangefs_inode_getattr(file->f_mapping->host, 0, 1);
662                 if (ret == -ESTALE)
663                         ret = -EIO;
664                 if (ret) {
665                         gossip_debug(GOSSIP_FILE_DEBUG,
666                                      "%s:%s:%d calling make bad inode\n",
667                                      __FILE__,
668                                      __func__,
669                                      __LINE__);
670                         return ret;
671                 }
672         }
673
674         gossip_debug(GOSSIP_FILE_DEBUG,
675                      "orangefs_file_llseek: offset is %ld | origin is %d"
676                      " | inode size is %lu\n",
677                      (long)offset,
678                      origin,
679                      (unsigned long)i_size_read(inode));
680
681         return generic_file_llseek(file, offset, origin);
682 }
683
684 /*
685  * Support local locks (locks that only this kernel knows about)
686  * if Orangefs was mounted -o local_lock.
687  */
688 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
689 {
690         int rc = -EINVAL;
691
692         if (ORANGEFS_SB(filp->f_inode->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
693                 if (cmd == F_GETLK) {
694                         rc = 0;
695                         posix_test_lock(filp, fl);
696                 } else {
697                         rc = posix_lock_file(filp, fl, NULL);
698                 }
699         }
700
701         return rc;
702 }
703
704 /** ORANGEFS implementation of VFS file operations */
705 const struct file_operations orangefs_file_operations = {
706         .llseek         = orangefs_file_llseek,
707         .read_iter      = orangefs_file_read_iter,
708         .write_iter     = orangefs_file_write_iter,
709         .lock           = orangefs_lock,
710         .unlocked_ioctl = orangefs_ioctl,
711         .mmap           = orangefs_file_mmap,
712         .open           = generic_file_open,
713         .release        = orangefs_file_release,
714         .fsync          = orangefs_fsync,
715 };