Lines Matching +full:no +full:- +full:unaligned +full:- +full:direct +full:- +full:access
1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/fs/nfs/direct.c
7 * High-performance uncached I/O for the Linux NFS client
10 * depends on uncached access to file data. Database clusters
15 * from a local cache. A streaming video server, for instance, has no
21 * correct unaligned requests from applications. All requested bytes are
22 * held on permanent storage before a direct write system call returns to
27 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
33 * 18 Dec 2001 Initial implementation for 2.4 --cel
34 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
35 * 08 Jun 2003 Port to 2.5 APIs --cel
36 * 31 Mar 2004 Handle direct I/O without VFS support --cel
37 * 15 Sep 2004 Parallel async reads --cel
38 * 04 May 2005 support O_DIRECT with aio --cel
76 atomic_inc(&dreq->io_count); in get_dreq()
81 return atomic_dec_and_test(&dreq->io_count); in put_dreq()
89 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) || in nfs_direct_handle_truncated()
90 test_bit(NFS_IOHDR_EOF, &hdr->flags))) in nfs_direct_handle_truncated()
92 if (dreq->max_count >= dreq_len) { in nfs_direct_handle_truncated()
93 dreq->max_count = dreq_len; in nfs_direct_handle_truncated()
94 if (dreq->count > dreq_len) in nfs_direct_handle_truncated()
95 dreq->count = dreq_len; in nfs_direct_handle_truncated()
97 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) in nfs_direct_handle_truncated()
98 dreq->error = hdr->error; in nfs_direct_handle_truncated()
100 dreq->error = 0; in nfs_direct_handle_truncated()
108 loff_t hdr_end = hdr->io_start + hdr->good_bytes; in nfs_direct_count_bytes()
111 if (hdr_end > dreq->io_start) in nfs_direct_count_bytes()
112 dreq_len = hdr_end - dreq->io_start; in nfs_direct_count_bytes()
116 if (dreq_len > dreq->max_count) in nfs_direct_count_bytes()
117 dreq_len = dreq->max_count; in nfs_direct_count_bytes()
119 if (dreq->count < dreq_len) in nfs_direct_count_bytes()
120 dreq->count = dreq_len; in nfs_direct_count_bytes()
124 * nfs_swap_rw - NFS address space operation for swap I/O
128 * Perform IO to the swap-file. This is much like direct IO.
155 cinfo->inode = dreq->inode; in nfs_init_cinfo_from_dreq()
156 cinfo->mds = &dreq->mds_cinfo; in nfs_init_cinfo_from_dreq()
157 cinfo->ds = &dreq->ds_cinfo; in nfs_init_cinfo_from_dreq()
158 cinfo->dreq = dreq; in nfs_init_cinfo_from_dreq()
159 cinfo->completion_ops = &nfs_direct_commit_completion_ops; in nfs_init_cinfo_from_dreq()
170 kref_init(&dreq->kref); in nfs_direct_req_alloc()
171 kref_get(&dreq->kref); in nfs_direct_req_alloc()
172 init_completion(&dreq->completion); in nfs_direct_req_alloc()
173 INIT_LIST_HEAD(&dreq->mds_cinfo.list); in nfs_direct_req_alloc()
174 pnfs_init_ds_commit_info(&dreq->ds_cinfo); in nfs_direct_req_alloc()
175 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); in nfs_direct_req_alloc()
176 spin_lock_init(&dreq->lock); in nfs_direct_req_alloc()
185 pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode); in nfs_direct_req_free()
186 if (dreq->l_ctx != NULL) in nfs_direct_req_free()
187 nfs_put_lock_context(dreq->l_ctx); in nfs_direct_req_free()
188 if (dreq->ctx != NULL) in nfs_direct_req_free()
189 put_nfs_open_context(dreq->ctx); in nfs_direct_req_free()
195 kref_put(&dreq->kref, nfs_direct_req_free); in nfs_direct_req_release()
200 return dreq->bytes_left; in nfs_dreq_bytes_left()
205 * Collects and returns the final error value/byte-count.
209 ssize_t result = -EIOCBQUEUED; in nfs_direct_wait()
212 if (dreq->iocb) in nfs_direct_wait()
215 result = wait_for_completion_killable(&dreq->completion); in nfs_direct_wait()
218 result = dreq->count; in nfs_direct_wait()
219 WARN_ON_ONCE(dreq->count < 0); in nfs_direct_wait()
222 result = dreq->error; in nfs_direct_wait()
229 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
234 struct inode *inode = dreq->inode; in nfs_direct_complete()
238 if (dreq->iocb) { in nfs_direct_complete()
239 long res = (long) dreq->error; in nfs_direct_complete()
240 if (dreq->count != 0) { in nfs_direct_complete()
241 res = (long) dreq->count; in nfs_direct_complete()
242 WARN_ON_ONCE(dreq->count < 0); in nfs_direct_complete()
244 dreq->iocb->ki_complete(dreq->iocb, res); in nfs_direct_complete()
247 complete(&dreq->completion); in nfs_direct_complete()
255 struct nfs_direct_req *dreq = hdr->dreq; in nfs_direct_read_completion()
257 spin_lock(&dreq->lock); in nfs_direct_read_completion()
258 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { in nfs_direct_read_completion()
259 spin_unlock(&dreq->lock); in nfs_direct_read_completion()
264 spin_unlock(&dreq->lock); in nfs_direct_read_completion()
266 while (!list_empty(&hdr->pages)) { in nfs_direct_read_completion()
267 struct nfs_page *req = nfs_list_entry(hdr->pages.next); in nfs_direct_read_completion()
268 struct page *page = req->wb_page; in nfs_direct_read_completion()
270 if (!PageCompound(page) && bytes < hdr->good_bytes && in nfs_direct_read_completion()
271 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY)) in nfs_direct_read_completion()
273 bytes += req->wb_bytes; in nfs_direct_read_completion()
280 hdr->release(hdr); in nfs_direct_read_completion()
288 req = nfs_list_entry(head->next); in nfs_read_sync_pgio_error()
296 get_dreq(hdr->dreq); in nfs_direct_pgio_init()
310 * no requests have been sent, just return an error.
318 struct inode *inode = dreq->inode; in nfs_direct_read_schedule_iovec()
319 ssize_t result = -EINVAL; in nfs_direct_read_schedule_iovec()
321 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE); in nfs_direct_read_schedule_iovec()
323 nfs_pageio_init_read(&desc, dreq->inode, false, in nfs_direct_read_schedule_iovec()
341 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; in nfs_direct_read_schedule_iovec()
344 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); in nfs_direct_read_schedule_iovec()
346 req = nfs_create_request(dreq->ctx, pagevec[i], in nfs_direct_read_schedule_iovec()
352 req->wb_index = pos >> PAGE_SHIFT; in nfs_direct_read_schedule_iovec()
353 req->wb_offset = pos & ~PAGE_MASK; in nfs_direct_read_schedule_iovec()
360 bytes -= req_len; in nfs_direct_read_schedule_iovec()
363 dreq->bytes_left -= req_len; in nfs_direct_read_schedule_iovec()
374 * If no bytes were started, return the error, and let the in nfs_direct_read_schedule_iovec()
380 return result < 0 ? result : -EIO; in nfs_direct_read_schedule_iovec()
389 * nfs_file_direct_read - file direct read operation for NFS files
394 * We use this function for direct reads instead of calling
397 * to work, we must generate a GETATTR before each direct read, and
411 struct file *file = iocb->ki_filp; in nfs_file_direct_read()
412 struct address_space *mapping = file->f_mapping; in nfs_file_direct_read()
413 struct inode *inode = mapping->host; in nfs_file_direct_read()
418 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); in nfs_file_direct_read()
420 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", in nfs_file_direct_read()
421 file, count, (long long) iocb->ki_pos); in nfs_file_direct_read()
429 result = -ENOMEM; in nfs_file_direct_read()
434 dreq->inode = inode; in nfs_file_direct_read()
435 dreq->bytes_left = dreq->max_count = count; in nfs_file_direct_read()
436 dreq->io_start = iocb->ki_pos; in nfs_file_direct_read()
437 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); in nfs_file_direct_read()
438 l_ctx = nfs_get_lock_context(dreq->ctx); in nfs_file_direct_read()
444 dreq->l_ctx = l_ctx; in nfs_file_direct_read()
446 dreq->iocb = iocb; in nfs_file_direct_read()
449 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY; in nfs_file_direct_read()
454 NFS_I(inode)->read_io += count; in nfs_file_direct_read()
455 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos); in nfs_file_direct_read()
463 requested -= result; in nfs_file_direct_read()
464 iocb->ki_pos += result; in nfs_file_direct_read()
483 if (req->wb_head != req || req->wb_this_page == req) in nfs_direct_join_group()
485 for (next = req->wb_this_page; in nfs_direct_join_group()
486 next != req->wb_head; in nfs_direct_join_group()
487 next = next->wb_this_page) { in nfs_direct_join_group()
500 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); in nfs_direct_write_scan_commit_list()
502 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0); in nfs_direct_write_scan_commit_list()
503 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); in nfs_direct_write_scan_commit_list()
515 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); in nfs_direct_write_reschedule()
517 nfs_direct_join_group(&reqs, dreq->inode); in nfs_direct_write_reschedule()
519 dreq->count = 0; in nfs_direct_write_reschedule()
520 dreq->max_count = 0; in nfs_direct_write_reschedule()
522 dreq->max_count += req->wb_bytes; in nfs_direct_write_reschedule()
523 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo); in nfs_direct_write_reschedule()
526 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, in nfs_direct_write_reschedule()
532 req->wb_nio++; in nfs_direct_write_reschedule()
535 spin_lock(&cinfo.inode->i_lock); in nfs_direct_write_reschedule()
536 dreq->flags = 0; in nfs_direct_write_reschedule()
538 dreq->error = desc.pg_error; in nfs_direct_write_reschedule()
540 dreq->error = -EIO; in nfs_direct_write_reschedule()
541 spin_unlock(&cinfo.inode->i_lock); in nfs_direct_write_reschedule()
559 const struct nfs_writeverf *verf = data->res.verf; in nfs_direct_commit_complete()
560 struct nfs_direct_req *dreq = data->dreq; in nfs_direct_commit_complete()
563 int status = data->task.tk_status; in nfs_direct_commit_complete()
569 dreq->error = status; in nfs_direct_commit_complete()
570 dreq->max_count = 0; in nfs_direct_commit_complete()
571 dreq->count = 0; in nfs_direct_commit_complete()
572 dreq->flags = NFS_ODIRECT_DONE; in nfs_direct_commit_complete()
574 status = dreq->error; in nfs_direct_commit_complete()
579 while (!list_empty(&data->pages)) { in nfs_direct_commit_complete()
580 req = nfs_list_entry(data->pages.next); in nfs_direct_commit_complete()
583 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; in nfs_direct_commit_complete()
588 req->wb_nio = 0; in nfs_direct_commit_complete()
602 struct nfs_direct_req *dreq = cinfo->dreq; in nfs_direct_resched_write()
606 spin_lock(&dreq->lock); in nfs_direct_resched_write()
607 if (dreq->flags != NFS_ODIRECT_DONE) in nfs_direct_resched_write()
608 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; in nfs_direct_resched_write()
609 spin_unlock(&dreq->lock); in nfs_direct_resched_write()
625 nfs_scan_commit(dreq->inode, &mds_list, &cinfo); in nfs_direct_commit_schedule()
626 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); in nfs_direct_commit_schedule()
627 if (res < 0) /* res == -ENOMEM */ in nfs_direct_commit_schedule()
638 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); in nfs_direct_write_clear_reqs()
651 int flags = dreq->flags; in nfs_direct_write_schedule_work()
653 dreq->flags = 0; in nfs_direct_write_schedule_work()
663 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping); in nfs_direct_write_schedule_work()
671 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */ in nfs_direct_write_complete()
676 struct nfs_direct_req *dreq = hdr->dreq; in nfs_direct_write_completion()
678 struct nfs_page *req = nfs_list_entry(hdr->pages.next); in nfs_direct_write_completion()
685 spin_lock(&dreq->lock); in nfs_direct_write_completion()
686 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { in nfs_direct_write_completion()
687 spin_unlock(&dreq->lock); in nfs_direct_write_completion()
692 if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags)) { in nfs_direct_write_completion()
693 if (!dreq->flags) in nfs_direct_write_completion()
694 dreq->flags = NFS_ODIRECT_DO_COMMIT; in nfs_direct_write_completion()
695 flags = dreq->flags; in nfs_direct_write_completion()
697 spin_unlock(&dreq->lock); in nfs_direct_write_completion()
699 while (!list_empty(&hdr->pages)) { in nfs_direct_write_completion()
701 req = nfs_list_entry(hdr->pages.next); in nfs_direct_write_completion()
704 kref_get(&req->wb_kref); in nfs_direct_write_completion()
705 memcpy(&req->wb_verf, &hdr->verf.verifier, in nfs_direct_write_completion()
706 sizeof(req->wb_verf)); in nfs_direct_write_completion()
707 nfs_mark_request_commit(req, hdr->lseg, &cinfo, in nfs_direct_write_completion()
708 hdr->ds_commit_idx); in nfs_direct_write_completion()
710 kref_get(&req->wb_kref); in nfs_direct_write_completion()
719 hdr->release(hdr); in nfs_direct_write_completion()
727 req = nfs_list_entry(head->next); in nfs_write_sync_pgio_error()
735 struct nfs_direct_req *dreq = hdr->dreq; in nfs_direct_write_reschedule_io()
739 spin_lock(&dreq->lock); in nfs_direct_write_reschedule_io()
740 if (dreq->error == 0) { in nfs_direct_write_reschedule_io()
741 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; in nfs_direct_write_reschedule_io()
743 hdr->verf.committed = NFS_UNSTABLE; in nfs_direct_write_reschedule_io()
744 hdr->good_bytes = hdr->args.offset + hdr->args.count - in nfs_direct_write_reschedule_io()
745 hdr->io_start; in nfs_direct_write_reschedule_io()
747 spin_unlock(&dreq->lock); in nfs_direct_write_reschedule_io()
767 * no requests have been sent, just return an error.
774 struct inode *inode = dreq->inode; in nfs_direct_write_schedule_iovec()
777 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); in nfs_direct_write_schedule_iovec()
787 NFS_I(inode)->write_io += iov_iter_count(iter); in nfs_direct_write_schedule_iovec()
800 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; in nfs_direct_write_schedule_iovec()
803 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); in nfs_direct_write_schedule_iovec()
805 req = nfs_create_request(dreq->ctx, pagevec[i], in nfs_direct_write_schedule_iovec()
819 req->wb_index = pos >> PAGE_SHIFT; in nfs_direct_write_schedule_iovec()
820 req->wb_offset = pos & ~PAGE_MASK; in nfs_direct_write_schedule_iovec()
827 bytes -= req_len; in nfs_direct_write_schedule_iovec()
830 dreq->bytes_left -= req_len; in nfs_direct_write_schedule_iovec()
840 * If no bytes were started, return the error, and let the in nfs_direct_write_schedule_iovec()
846 return result < 0 ? result : -EIO; in nfs_direct_write_schedule_iovec()
855 * nfs_file_direct_write - file direct write operation for NFS files
860 * We use this function for direct writes instead of calling
867 * We eliminate local atime updates, see direct read above.
872 * Note that O_APPEND is not supported for NFS direct writes, as there
873 * is no atomic O_APPEND write facility in the NFS protocol.
880 struct file *file = iocb->ki_filp; in nfs_file_direct_write()
881 struct address_space *mapping = file->f_mapping; in nfs_file_direct_write()
882 struct inode *inode = mapping->host; in nfs_file_direct_write()
887 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", in nfs_file_direct_write()
888 file, iov_iter_count(iter), (long long) iocb->ki_pos); in nfs_file_direct_write()
898 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); in nfs_file_direct_write()
900 pos = iocb->ki_pos; in nfs_file_direct_write()
901 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT; in nfs_file_direct_write()
905 result = -ENOMEM; in nfs_file_direct_write()
910 dreq->inode = inode; in nfs_file_direct_write()
911 dreq->bytes_left = dreq->max_count = count; in nfs_file_direct_write()
912 dreq->io_start = pos; in nfs_file_direct_write()
913 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); in nfs_file_direct_write()
914 l_ctx = nfs_get_lock_context(dreq->ctx); in nfs_file_direct_write()
920 dreq->l_ctx = l_ctx; in nfs_file_direct_write()
922 dreq->iocb = iocb; in nfs_file_direct_write()
923 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode); in nfs_file_direct_write()
934 if (mapping->nrpages) { in nfs_file_direct_write()
945 requested -= result; in nfs_file_direct_write()
946 iocb->ki_pos = pos + result; in nfs_file_direct_write()
962 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
973 return -ENOMEM; in nfs_init_directcache()
979 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures