1 // SPDX-License-Identifier: GPL-2.0
3 * dax: direct host memory access
4 * Copyright (C) 2020 Red Hat, Inc.
9 #include <linux/delay.h>
10 #include <linux/dax.h>
11 #include <linux/uio.h>
12 #include <linux/pagemap.h>
13 #include <linux/pfn_t.h>
14 #include <linux/iomap.h>
15 #include <linux/interval_tree.h>
18 * Default memory range size. A power of 2 so it agrees with common FUSE_INIT
19 * map_alignment values 4KB and 64KB.
21 #define FUSE_DAX_SHIFT 21
22 #define FUSE_DAX_SZ (1 << FUSE_DAX_SHIFT)
23 #define FUSE_DAX_PAGES (FUSE_DAX_SZ / PAGE_SIZE)
25 /* Number of ranges reclaimer will try to free in one invocation */
26 #define FUSE_DAX_RECLAIM_CHUNK (10)
29 * Dax memory reclaim threshold in percetage of total ranges. When free
30 * number of free ranges drops below this threshold, reclaim can trigger
33 #define FUSE_DAX_RECLAIM_THRESHOLD (20)
35 /** Translation information for file offsets to DAX window offsets */
36 struct fuse_dax_mapping {
37 /* Pointer to inode where this memory range is mapped */
40 /* Will connect in fcd->free_ranges to keep track of free memory */
41 struct list_head list;
43 /* For interval tree in file/inode */
44 struct interval_tree_node itn;
46 /* Will connect in fc->busy_ranges to keep track busy memory */
47 struct list_head busy_list;
49 /** Position in DAX window */
52 /** Length of mapping, in bytes */
55 /* Is this mapping read-only or read-write */
58 /* reference count when the mapping is used by dax iomap. */
62 /* Per-inode dax map */
63 struct fuse_inode_dax {
64 /* Semaphore to protect modifications to the dmap tree */
65 struct rw_semaphore sem;
67 /* Sorted rb tree of struct fuse_dax_mapping elements */
68 struct rb_root_cached tree;
72 struct fuse_conn_dax {
74 struct dax_device *dev;
76 /* Lock protecting accessess to members of this structure */
79 /* List of memory ranges which are busy */
80 unsigned long nr_busy_ranges;
81 struct list_head busy_ranges;
83 /* Worker to free up memory ranges */
84 struct delayed_work free_work;
86 /* Wait queue for a dax range to become free */
87 wait_queue_head_t range_waitq;
89 /* DAX Window Free Ranges */
91 struct list_head free_ranges;
93 unsigned long nr_ranges;
96 static inline struct fuse_dax_mapping *
97 node_to_dmap(struct interval_tree_node *node)
102 return container_of(node, struct fuse_dax_mapping, itn);
105 static struct fuse_dax_mapping *
106 alloc_dax_mapping_reclaim(struct fuse_conn_dax *fcd, struct inode *inode);
109 __kick_dmap_free_worker(struct fuse_conn_dax *fcd, unsigned long delay_ms)
111 unsigned long free_threshold;
113 /* If number of free ranges are below threshold, start reclaim */
114 free_threshold = max_t(unsigned long, fcd->nr_ranges * FUSE_DAX_RECLAIM_THRESHOLD / 100,
116 if (fcd->nr_free_ranges < free_threshold)
117 queue_delayed_work(system_long_wq, &fcd->free_work,
118 msecs_to_jiffies(delay_ms));
121 static void kick_dmap_free_worker(struct fuse_conn_dax *fcd,
122 unsigned long delay_ms)
124 spin_lock(&fcd->lock);
125 __kick_dmap_free_worker(fcd, delay_ms);
126 spin_unlock(&fcd->lock);
129 static struct fuse_dax_mapping *alloc_dax_mapping(struct fuse_conn_dax *fcd)
131 struct fuse_dax_mapping *dmap;
133 spin_lock(&fcd->lock);
134 dmap = list_first_entry_or_null(&fcd->free_ranges,
135 struct fuse_dax_mapping, list);
137 list_del_init(&dmap->list);
138 WARN_ON(fcd->nr_free_ranges <= 0);
139 fcd->nr_free_ranges--;
141 __kick_dmap_free_worker(fcd, 0);
142 spin_unlock(&fcd->lock);
147 /* This assumes fcd->lock is held */
148 static void __dmap_remove_busy_list(struct fuse_conn_dax *fcd,
149 struct fuse_dax_mapping *dmap)
151 list_del_init(&dmap->busy_list);
152 WARN_ON(fcd->nr_busy_ranges == 0);
153 fcd->nr_busy_ranges--;
156 static void dmap_remove_busy_list(struct fuse_conn_dax *fcd,
157 struct fuse_dax_mapping *dmap)
159 spin_lock(&fcd->lock);
160 __dmap_remove_busy_list(fcd, dmap);
161 spin_unlock(&fcd->lock);
164 /* This assumes fcd->lock is held */
165 static void __dmap_add_to_free_pool(struct fuse_conn_dax *fcd,
166 struct fuse_dax_mapping *dmap)
168 list_add_tail(&dmap->list, &fcd->free_ranges);
169 fcd->nr_free_ranges++;
170 wake_up(&fcd->range_waitq);
173 static void dmap_add_to_free_pool(struct fuse_conn_dax *fcd,
174 struct fuse_dax_mapping *dmap)
176 /* Return fuse_dax_mapping to free list */
177 spin_lock(&fcd->lock);
178 __dmap_add_to_free_pool(fcd, dmap);
179 spin_unlock(&fcd->lock);
182 static int fuse_setup_one_mapping(struct inode *inode, unsigned long start_idx,
183 struct fuse_dax_mapping *dmap, bool writable,
186 struct fuse_mount *fm = get_fuse_mount(inode);
187 struct fuse_conn_dax *fcd = fm->fc->dax;
188 struct fuse_inode *fi = get_fuse_inode(inode);
189 struct fuse_setupmapping_in inarg;
190 loff_t offset = start_idx << FUSE_DAX_SHIFT;
194 WARN_ON(fcd->nr_free_ranges < 0);
196 /* Ask fuse daemon to setup mapping */
197 memset(&inarg, 0, sizeof(inarg));
198 inarg.foffset = offset;
200 inarg.moffset = dmap->window_offset;
201 inarg.len = FUSE_DAX_SZ;
202 inarg.flags |= FUSE_SETUPMAPPING_FLAG_READ;
204 inarg.flags |= FUSE_SETUPMAPPING_FLAG_WRITE;
205 args.opcode = FUSE_SETUPMAPPING;
206 args.nodeid = fi->nodeid;
208 args.in_args[0].size = sizeof(inarg);
209 args.in_args[0].value = &inarg;
210 err = fuse_simple_request(fm, &args);
213 dmap->writable = writable;
216 * We don't take a reference on inode. inode is valid right now
217 * and when inode is going away, cleanup logic should first
218 * cleanup dmap entries.
221 dmap->itn.start = dmap->itn.last = start_idx;
222 /* Protected by fi->dax->sem */
223 interval_tree_insert(&dmap->itn, &fi->dax->tree);
225 spin_lock(&fcd->lock);
226 list_add_tail(&dmap->busy_list, &fcd->busy_ranges);
227 fcd->nr_busy_ranges++;
228 spin_unlock(&fcd->lock);
233 static int fuse_send_removemapping(struct inode *inode,
234 struct fuse_removemapping_in *inargp,
235 struct fuse_removemapping_one *remove_one)
237 struct fuse_inode *fi = get_fuse_inode(inode);
238 struct fuse_mount *fm = get_fuse_mount(inode);
241 args.opcode = FUSE_REMOVEMAPPING;
242 args.nodeid = fi->nodeid;
244 args.in_args[0].size = sizeof(*inargp);
245 args.in_args[0].value = inargp;
246 args.in_args[1].size = inargp->count * sizeof(*remove_one);
247 args.in_args[1].value = remove_one;
248 return fuse_simple_request(fm, &args);
251 static int dmap_removemapping_list(struct inode *inode, unsigned int num,
252 struct list_head *to_remove)
254 struct fuse_removemapping_one *remove_one, *ptr;
255 struct fuse_removemapping_in inarg;
256 struct fuse_dax_mapping *dmap;
257 int ret, i = 0, nr_alloc;
259 nr_alloc = min_t(unsigned int, num, FUSE_REMOVEMAPPING_MAX_ENTRY);
260 remove_one = kmalloc_array(nr_alloc, sizeof(*remove_one), GFP_NOFS);
265 list_for_each_entry(dmap, to_remove, list) {
266 ptr->moffset = dmap->window_offset;
267 ptr->len = dmap->length;
271 if (i >= nr_alloc || num == 0) {
272 memset(&inarg, 0, sizeof(inarg));
274 ret = fuse_send_removemapping(inode, &inarg,
288 * Cleanup dmap entry and add back to free list. This should be called with
291 static void dmap_reinit_add_to_free_pool(struct fuse_conn_dax *fcd,
292 struct fuse_dax_mapping *dmap)
294 pr_debug("fuse: freeing memory range start_idx=0x%lx end_idx=0x%lx window_offset=0x%llx length=0x%llx\n",
295 dmap->itn.start, dmap->itn.last, dmap->window_offset,
297 __dmap_remove_busy_list(fcd, dmap);
299 dmap->itn.start = dmap->itn.last = 0;
300 __dmap_add_to_free_pool(fcd, dmap);
304 * Free inode dmap entries whose range falls inside [start, end].
305 * Does not take any locks. At this point of time it should only be
306 * called from evict_inode() path where we know all dmap entries can be
309 static void inode_reclaim_dmap_range(struct fuse_conn_dax *fcd,
311 loff_t start, loff_t end)
313 struct fuse_inode *fi = get_fuse_inode(inode);
314 struct fuse_dax_mapping *dmap, *n;
316 LIST_HEAD(to_remove);
317 unsigned long start_idx = start >> FUSE_DAX_SHIFT;
318 unsigned long end_idx = end >> FUSE_DAX_SHIFT;
319 struct interval_tree_node *node;
322 node = interval_tree_iter_first(&fi->dax->tree, start_idx,
326 dmap = node_to_dmap(node);
327 /* inode is going away. There should not be any users of dmap */
328 WARN_ON(refcount_read(&dmap->refcnt) > 1);
329 interval_tree_remove(&dmap->itn, &fi->dax->tree);
331 list_add(&dmap->list, &to_remove);
334 /* Nothing to remove */
335 if (list_empty(&to_remove))
338 WARN_ON(fi->dax->nr < num);
340 err = dmap_removemapping_list(inode, num, &to_remove);
341 if (err && err != -ENOTCONN) {
342 pr_warn("Failed to removemappings. start=0x%llx end=0x%llx\n",
345 spin_lock(&fcd->lock);
346 list_for_each_entry_safe(dmap, n, &to_remove, list) {
347 list_del_init(&dmap->list);
348 dmap_reinit_add_to_free_pool(fcd, dmap);
350 spin_unlock(&fcd->lock);
353 static int dmap_removemapping_one(struct inode *inode,
354 struct fuse_dax_mapping *dmap)
356 struct fuse_removemapping_one forget_one;
357 struct fuse_removemapping_in inarg;
359 memset(&inarg, 0, sizeof(inarg));
361 memset(&forget_one, 0, sizeof(forget_one));
362 forget_one.moffset = dmap->window_offset;
363 forget_one.len = dmap->length;
365 return fuse_send_removemapping(inode, &inarg, &forget_one);
369 * It is called from evict_inode() and by that time inode is going away. So
370 * this function does not take any locks like fi->dax->sem for traversing
371 * that fuse inode interval tree. If that lock is taken then lock validator
372 * complains of deadlock situation w.r.t fs_reclaim lock.
374 void fuse_dax_inode_cleanup(struct inode *inode)
376 struct fuse_conn *fc = get_fuse_conn(inode);
377 struct fuse_inode *fi = get_fuse_inode(inode);
380 * fuse_evict_inode() has already called truncate_inode_pages_final()
381 * before we arrive here. So we should not have to worry about any
382 * pages/exception entries still associated with inode.
384 inode_reclaim_dmap_range(fc->dax, inode, 0, -1);
385 WARN_ON(fi->dax->nr);
388 static void fuse_fill_iomap_hole(struct iomap *iomap, loff_t length)
390 iomap->addr = IOMAP_NULL_ADDR;
391 iomap->length = length;
392 iomap->type = IOMAP_HOLE;
395 static void fuse_fill_iomap(struct inode *inode, loff_t pos, loff_t length,
396 struct iomap *iomap, struct fuse_dax_mapping *dmap,
400 loff_t i_size = i_size_read(inode);
402 offset = pos - (dmap->itn.start << FUSE_DAX_SHIFT);
403 len = min(length, dmap->length - offset);
405 /* If length is beyond end of file, truncate further */
406 if (pos + len > i_size)
410 iomap->addr = dmap->window_offset + offset;
412 if (flags & IOMAP_FAULT)
413 iomap->length = ALIGN(len, PAGE_SIZE);
414 iomap->type = IOMAP_MAPPED;
416 * increace refcnt so that reclaim code knows this dmap is in
417 * use. This assumes fi->dax->sem mutex is held either
420 refcount_inc(&dmap->refcnt);
422 /* iomap->private should be NULL */
423 WARN_ON_ONCE(iomap->private);
424 iomap->private = dmap;
426 /* Mapping beyond end of file is hole */
427 fuse_fill_iomap_hole(iomap, length);
431 static int fuse_setup_new_dax_mapping(struct inode *inode, loff_t pos,
432 loff_t length, unsigned int flags,
435 struct fuse_inode *fi = get_fuse_inode(inode);
436 struct fuse_conn *fc = get_fuse_conn(inode);
437 struct fuse_conn_dax *fcd = fc->dax;
438 struct fuse_dax_mapping *dmap, *alloc_dmap = NULL;
440 bool writable = flags & IOMAP_WRITE;
441 unsigned long start_idx = pos >> FUSE_DAX_SHIFT;
442 struct interval_tree_node *node;
445 * Can't do inline reclaim in fault path. We call
446 * dax_layout_busy_page() before we free a range. And
447 * fuse_wait_dax_page() drops mapping->invalidate_lock and requires it.
448 * In fault path we enter with mapping->invalidate_lock held and can't
449 * drop it. Also in fault path we hold mapping->invalidate_lock shared
450 * and not exclusive, so that creates further issues with
451 * fuse_wait_dax_page(). Hence return -EAGAIN and fuse_dax_fault()
452 * will wait for a memory range to become free and retry.
454 if (flags & IOMAP_FAULT) {
455 alloc_dmap = alloc_dax_mapping(fcd);
459 alloc_dmap = alloc_dax_mapping_reclaim(fcd, inode);
460 if (IS_ERR(alloc_dmap))
461 return PTR_ERR(alloc_dmap);
464 /* If we are here, we should have memory allocated */
465 if (WARN_ON(!alloc_dmap))
469 * Take write lock so that only one caller can try to setup mapping
472 down_write(&fi->dax->sem);
474 * We dropped lock. Check again if somebody else setup
477 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
479 dmap = node_to_dmap(node);
480 fuse_fill_iomap(inode, pos, length, iomap, dmap, flags);
481 dmap_add_to_free_pool(fcd, alloc_dmap);
482 up_write(&fi->dax->sem);
486 /* Setup one mapping */
487 ret = fuse_setup_one_mapping(inode, pos >> FUSE_DAX_SHIFT, alloc_dmap,
490 dmap_add_to_free_pool(fcd, alloc_dmap);
491 up_write(&fi->dax->sem);
494 fuse_fill_iomap(inode, pos, length, iomap, alloc_dmap, flags);
495 up_write(&fi->dax->sem);
499 static int fuse_upgrade_dax_mapping(struct inode *inode, loff_t pos,
500 loff_t length, unsigned int flags,
503 struct fuse_inode *fi = get_fuse_inode(inode);
504 struct fuse_dax_mapping *dmap;
506 unsigned long idx = pos >> FUSE_DAX_SHIFT;
507 struct interval_tree_node *node;
510 * Take exclusive lock so that only one caller can try to setup
511 * mapping and others wait.
513 down_write(&fi->dax->sem);
514 node = interval_tree_iter_first(&fi->dax->tree, idx, idx);
516 /* We are holding either inode lock or invalidate_lock, and that should
517 * ensure that dmap can't be truncated. We are holding a reference
518 * on dmap and that should make sure it can't be reclaimed. So dmap
519 * should still be there in tree despite the fact we dropped and
520 * re-acquired the fi->dax->sem lock.
526 dmap = node_to_dmap(node);
528 /* We took an extra reference on dmap to make sure its not reclaimd.
529 * Now we hold fi->dax->sem lock and that reference is not needed
532 if (refcount_dec_and_test(&dmap->refcnt)) {
533 /* refcount should not hit 0. This object only goes
534 * away when fuse connection goes away
539 /* Maybe another thread already upgraded mapping while we were not
542 if (dmap->writable) {
547 ret = fuse_setup_one_mapping(inode, pos >> FUSE_DAX_SHIFT, dmap, true,
552 fuse_fill_iomap(inode, pos, length, iomap, dmap, flags);
554 up_write(&fi->dax->sem);
558 /* This is just for DAX and the mapping is ephemeral, do not use it for other
559 * purposes since there is no block device with a permanent mapping.
561 static int fuse_iomap_begin(struct inode *inode, loff_t pos, loff_t length,
562 unsigned int flags, struct iomap *iomap,
563 struct iomap *srcmap)
565 struct fuse_inode *fi = get_fuse_inode(inode);
566 struct fuse_conn *fc = get_fuse_conn(inode);
567 struct fuse_dax_mapping *dmap;
568 bool writable = flags & IOMAP_WRITE;
569 unsigned long start_idx = pos >> FUSE_DAX_SHIFT;
570 struct interval_tree_node *node;
572 /* We don't support FIEMAP */
573 if (WARN_ON(flags & IOMAP_REPORT))
579 iomap->dax_dev = fc->dax->dev;
582 * Both read/write and mmap path can race here. So we need something
583 * to make sure if we are setting up mapping, then other path waits
585 * For now, use a semaphore for this. It probably needs to be
588 down_read(&fi->dax->sem);
589 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
591 dmap = node_to_dmap(node);
592 if (writable && !dmap->writable) {
593 /* Upgrade read-only mapping to read-write. This will
594 * require exclusive fi->dax->sem lock as we don't want
595 * two threads to be trying to this simultaneously
596 * for same dmap. So drop shared lock and acquire
599 * Before dropping fi->dax->sem lock, take reference
600 * on dmap so that its not freed by range reclaim.
602 refcount_inc(&dmap->refcnt);
603 up_read(&fi->dax->sem);
604 pr_debug("%s: Upgrading mapping at offset 0x%llx length 0x%llx\n",
605 __func__, pos, length);
606 return fuse_upgrade_dax_mapping(inode, pos, length,
609 fuse_fill_iomap(inode, pos, length, iomap, dmap, flags);
610 up_read(&fi->dax->sem);
614 up_read(&fi->dax->sem);
615 pr_debug("%s: no mapping at offset 0x%llx length 0x%llx\n",
616 __func__, pos, length);
617 if (pos >= i_size_read(inode))
620 return fuse_setup_new_dax_mapping(inode, pos, length, flags,
625 * If read beyond end of file happens, fs code seems to return
629 fuse_fill_iomap_hole(iomap, length);
630 pr_debug("%s returning hole mapping. pos=0x%llx length_asked=0x%llx length_returned=0x%llx\n",
631 __func__, pos, length, iomap->length);
635 static int fuse_iomap_end(struct inode *inode, loff_t pos, loff_t length,
636 ssize_t written, unsigned int flags,
639 struct fuse_dax_mapping *dmap = iomap->private;
642 if (refcount_dec_and_test(&dmap->refcnt)) {
643 /* refcount should not hit 0. This object only goes
644 * away when fuse connection goes away
650 /* DAX writes beyond end-of-file aren't handled using iomap, so the
651 * file size is unchanged and there is nothing to do here.
656 static const struct iomap_ops fuse_iomap_ops = {
657 .iomap_begin = fuse_iomap_begin,
658 .iomap_end = fuse_iomap_end,
661 static void fuse_wait_dax_page(struct inode *inode)
663 filemap_invalidate_unlock(inode->i_mapping);
665 filemap_invalidate_lock(inode->i_mapping);
668 /* Should be called with mapping->invalidate_lock held exclusively */
669 static int __fuse_dax_break_layouts(struct inode *inode, bool *retry,
670 loff_t start, loff_t end)
674 page = dax_layout_busy_page_range(inode->i_mapping, start, end);
679 return ___wait_var_event(&page->_refcount,
680 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
681 0, 0, fuse_wait_dax_page(inode));
684 /* dmap_end == 0 leads to unmapping of whole file */
685 int fuse_dax_break_layouts(struct inode *inode, u64 dmap_start,
693 ret = __fuse_dax_break_layouts(inode, &retry, dmap_start,
695 } while (ret == 0 && retry);
700 ssize_t fuse_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
702 struct inode *inode = file_inode(iocb->ki_filp);
705 if (iocb->ki_flags & IOCB_NOWAIT) {
706 if (!inode_trylock_shared(inode))
709 inode_lock_shared(inode);
712 ret = dax_iomap_rw(iocb, to, &fuse_iomap_ops);
713 inode_unlock_shared(inode);
715 /* TODO file_accessed(iocb->f_filp) */
719 static bool file_extending_write(struct kiocb *iocb, struct iov_iter *from)
721 struct inode *inode = file_inode(iocb->ki_filp);
723 return (iov_iter_rw(from) == WRITE &&
724 ((iocb->ki_pos) >= i_size_read(inode) ||
725 (iocb->ki_pos + iov_iter_count(from) > i_size_read(inode))));
728 static ssize_t fuse_dax_direct_write(struct kiocb *iocb, struct iov_iter *from)
730 struct inode *inode = file_inode(iocb->ki_filp);
731 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
734 ret = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
736 fuse_write_update_attr(inode, iocb->ki_pos, ret);
740 ssize_t fuse_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
742 struct inode *inode = file_inode(iocb->ki_filp);
745 if (iocb->ki_flags & IOCB_NOWAIT) {
746 if (!inode_trylock(inode))
752 ret = generic_write_checks(iocb, from);
756 ret = file_remove_privs(iocb->ki_filp);
759 /* TODO file_update_time() but we don't want metadata I/O */
761 /* Do not use dax for file extending writes as write and on
762 * disk i_size increase are not atomic otherwise.
764 if (file_extending_write(iocb, from))
765 ret = fuse_dax_direct_write(iocb, from);
767 ret = dax_iomap_rw(iocb, from, &fuse_iomap_ops);
773 ret = generic_write_sync(iocb, ret);
777 static int fuse_dax_writepages(struct address_space *mapping,
778 struct writeback_control *wbc)
781 struct inode *inode = mapping->host;
782 struct fuse_conn *fc = get_fuse_conn(inode);
784 return dax_writeback_mapping_range(mapping, fc->dax->dev, wbc);
787 static vm_fault_t __fuse_dax_fault(struct vm_fault *vmf,
788 enum page_entry_size pe_size, bool write)
791 struct inode *inode = file_inode(vmf->vma->vm_file);
792 struct super_block *sb = inode->i_sb;
795 struct fuse_conn *fc = get_fuse_conn(inode);
796 struct fuse_conn_dax *fcd = fc->dax;
800 sb_start_pagefault(sb);
802 if (retry && !(fcd->nr_free_ranges > 0))
803 wait_event(fcd->range_waitq, (fcd->nr_free_ranges > 0));
806 * We need to serialize against not only truncate but also against
807 * fuse dax memory range reclaim. While a range is being reclaimed,
808 * we do not want any read/write/mmap to make progress and try
809 * to populate page cache or access memory we are trying to free.
811 filemap_invalidate_lock_shared(inode->i_mapping);
812 ret = dax_iomap_fault(vmf, pe_size, &pfn, &error, &fuse_iomap_ops);
813 if ((ret & VM_FAULT_ERROR) && error == -EAGAIN) {
816 filemap_invalidate_unlock_shared(inode->i_mapping);
820 if (ret & VM_FAULT_NEEDDSYNC)
821 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
822 filemap_invalidate_unlock_shared(inode->i_mapping);
825 sb_end_pagefault(sb);
830 static vm_fault_t fuse_dax_fault(struct vm_fault *vmf)
832 return __fuse_dax_fault(vmf, PE_SIZE_PTE,
833 vmf->flags & FAULT_FLAG_WRITE);
836 static vm_fault_t fuse_dax_huge_fault(struct vm_fault *vmf,
837 enum page_entry_size pe_size)
839 return __fuse_dax_fault(vmf, pe_size, vmf->flags & FAULT_FLAG_WRITE);
842 static vm_fault_t fuse_dax_page_mkwrite(struct vm_fault *vmf)
844 return __fuse_dax_fault(vmf, PE_SIZE_PTE, true);
847 static vm_fault_t fuse_dax_pfn_mkwrite(struct vm_fault *vmf)
849 return __fuse_dax_fault(vmf, PE_SIZE_PTE, true);
852 static const struct vm_operations_struct fuse_dax_vm_ops = {
853 .fault = fuse_dax_fault,
854 .huge_fault = fuse_dax_huge_fault,
855 .page_mkwrite = fuse_dax_page_mkwrite,
856 .pfn_mkwrite = fuse_dax_pfn_mkwrite,
859 int fuse_dax_mmap(struct file *file, struct vm_area_struct *vma)
862 vma->vm_ops = &fuse_dax_vm_ops;
863 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
867 static int dmap_writeback_invalidate(struct inode *inode,
868 struct fuse_dax_mapping *dmap)
871 loff_t start_pos = dmap->itn.start << FUSE_DAX_SHIFT;
872 loff_t end_pos = (start_pos + FUSE_DAX_SZ - 1);
874 ret = filemap_fdatawrite_range(inode->i_mapping, start_pos, end_pos);
876 pr_debug("fuse: filemap_fdatawrite_range() failed. err=%d start_pos=0x%llx, end_pos=0x%llx\n",
877 ret, start_pos, end_pos);
881 ret = invalidate_inode_pages2_range(inode->i_mapping,
882 start_pos >> PAGE_SHIFT,
883 end_pos >> PAGE_SHIFT);
885 pr_debug("fuse: invalidate_inode_pages2_range() failed err=%d\n",
891 static int reclaim_one_dmap_locked(struct inode *inode,
892 struct fuse_dax_mapping *dmap)
895 struct fuse_inode *fi = get_fuse_inode(inode);
898 * igrab() was done to make sure inode won't go under us, and this
899 * further avoids the race with evict().
901 ret = dmap_writeback_invalidate(inode, dmap);
905 /* Remove dax mapping from inode interval tree now */
906 interval_tree_remove(&dmap->itn, &fi->dax->tree);
909 /* It is possible that umount/shutdown has killed the fuse connection
910 * and worker thread is trying to reclaim memory in parallel. Don't
913 ret = dmap_removemapping_one(inode, dmap);
914 if (ret && ret != -ENOTCONN) {
915 pr_warn("Failed to remove mapping. offset=0x%llx len=0x%llx ret=%d\n",
916 dmap->window_offset, dmap->length, ret);
921 /* Find first mapped dmap for an inode and return file offset. Caller needs
922 * to hold fi->dax->sem lock either shared or exclusive.
924 static struct fuse_dax_mapping *inode_lookup_first_dmap(struct inode *inode)
926 struct fuse_inode *fi = get_fuse_inode(inode);
927 struct fuse_dax_mapping *dmap;
928 struct interval_tree_node *node;
930 for (node = interval_tree_iter_first(&fi->dax->tree, 0, -1); node;
931 node = interval_tree_iter_next(node, 0, -1)) {
932 dmap = node_to_dmap(node);
934 if (refcount_read(&dmap->refcnt) > 1)
944 * Find first mapping in the tree and free it and return it. Do not add
945 * it back to free pool.
947 static struct fuse_dax_mapping *
948 inode_inline_reclaim_one_dmap(struct fuse_conn_dax *fcd, struct inode *inode,
951 struct fuse_inode *fi = get_fuse_inode(inode);
952 struct fuse_dax_mapping *dmap;
953 u64 dmap_start, dmap_end;
954 unsigned long start_idx;
956 struct interval_tree_node *node;
958 filemap_invalidate_lock(inode->i_mapping);
960 /* Lookup a dmap and corresponding file offset to reclaim. */
961 down_read(&fi->dax->sem);
962 dmap = inode_lookup_first_dmap(inode);
964 start_idx = dmap->itn.start;
965 dmap_start = start_idx << FUSE_DAX_SHIFT;
966 dmap_end = dmap_start + FUSE_DAX_SZ - 1;
968 up_read(&fi->dax->sem);
973 * Make sure there are no references to inode pages using
976 ret = fuse_dax_break_layouts(inode, dmap_start, dmap_end);
978 pr_debug("fuse: fuse_dax_break_layouts() failed. err=%d\n",
984 down_write(&fi->dax->sem);
985 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
986 /* Range already got reclaimed by somebody else */
990 goto out_write_dmap_sem;
993 dmap = node_to_dmap(node);
995 if (refcount_read(&dmap->refcnt) > 1) {
999 goto out_write_dmap_sem;
1002 ret = reclaim_one_dmap_locked(inode, dmap);
1004 dmap = ERR_PTR(ret);
1005 goto out_write_dmap_sem;
1008 /* Clean up dmap. Do not add back to free list */
1009 dmap_remove_busy_list(fcd, dmap);
1011 dmap->itn.start = dmap->itn.last = 0;
1013 pr_debug("fuse: %s: inline reclaimed memory range. inode=%p, window_offset=0x%llx, length=0x%llx\n",
1014 __func__, inode, dmap->window_offset, dmap->length);
1017 up_write(&fi->dax->sem);
1019 filemap_invalidate_unlock(inode->i_mapping);
1023 static struct fuse_dax_mapping *
1024 alloc_dax_mapping_reclaim(struct fuse_conn_dax *fcd, struct inode *inode)
1026 struct fuse_dax_mapping *dmap;
1027 struct fuse_inode *fi = get_fuse_inode(inode);
1032 dmap = alloc_dax_mapping(fcd);
1036 dmap = inode_inline_reclaim_one_dmap(fcd, inode, &retry);
1038 * Either we got a mapping or it is an error, return in both
1044 /* If we could not reclaim a mapping because it
1045 * had a reference or some other temporary failure,
1046 * Try again. We want to give up inline reclaim only
1047 * if there is no range assigned to this node. Otherwise
1048 * if a deadlock is possible if we sleep with
1049 * mapping->invalidate_lock held and worker to free memory
1050 * can't make progress due to unavailability of
1051 * mapping->invalidate_lock. So sleep only if fi->dax->nr=0
1056 * There are no mappings which can be reclaimed. Wait for one.
1057 * We are not holding fi->dax->sem. So it is possible
1058 * that range gets added now. But as we are not holding
1059 * mapping->invalidate_lock, worker should still be able to
1060 * free up a range and wake us up.
1062 if (!fi->dax->nr && !(fcd->nr_free_ranges > 0)) {
1063 if (wait_event_killable_exclusive(fcd->range_waitq,
1064 (fcd->nr_free_ranges > 0))) {
1065 return ERR_PTR(-EINTR);
1071 static int lookup_and_reclaim_dmap_locked(struct fuse_conn_dax *fcd,
1072 struct inode *inode,
1073 unsigned long start_idx)
1076 struct fuse_inode *fi = get_fuse_inode(inode);
1077 struct fuse_dax_mapping *dmap;
1078 struct interval_tree_node *node;
1080 /* Find fuse dax mapping at file offset inode. */
1081 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
1083 /* Range already got cleaned up by somebody else */
1086 dmap = node_to_dmap(node);
1089 if (refcount_read(&dmap->refcnt) > 1)
1092 ret = reclaim_one_dmap_locked(inode, dmap);
1096 /* Cleanup dmap entry and add back to free list */
1097 spin_lock(&fcd->lock);
1098 dmap_reinit_add_to_free_pool(fcd, dmap);
1099 spin_unlock(&fcd->lock);
1104 * Free a range of memory.
1106 * 1. Take mapping->invalidate_lock to block dax faults.
1107 * 2. Take fi->dax->sem to protect interval tree and also to make sure
1108 * read/write can not reuse a dmap which we might be freeing.
1110 static int lookup_and_reclaim_dmap(struct fuse_conn_dax *fcd,
1111 struct inode *inode,
1112 unsigned long start_idx,
1113 unsigned long end_idx)
1116 struct fuse_inode *fi = get_fuse_inode(inode);
1117 loff_t dmap_start = start_idx << FUSE_DAX_SHIFT;
1118 loff_t dmap_end = (dmap_start + FUSE_DAX_SZ) - 1;
1120 filemap_invalidate_lock(inode->i_mapping);
1121 ret = fuse_dax_break_layouts(inode, dmap_start, dmap_end);
1123 pr_debug("virtio_fs: fuse_dax_break_layouts() failed. err=%d\n",
1128 down_write(&fi->dax->sem);
1129 ret = lookup_and_reclaim_dmap_locked(fcd, inode, start_idx);
1130 up_write(&fi->dax->sem);
1132 filemap_invalidate_unlock(inode->i_mapping);
1136 static int try_to_free_dmap_chunks(struct fuse_conn_dax *fcd,
1137 unsigned long nr_to_free)
1139 struct fuse_dax_mapping *dmap, *pos, *temp;
1140 int ret, nr_freed = 0;
1141 unsigned long start_idx = 0, end_idx = 0;
1142 struct inode *inode = NULL;
1144 /* Pick first busy range and free it for now*/
1146 if (nr_freed >= nr_to_free)
1150 spin_lock(&fcd->lock);
1152 if (!fcd->nr_busy_ranges) {
1153 spin_unlock(&fcd->lock);
1157 list_for_each_entry_safe(pos, temp, &fcd->busy_ranges,
1159 /* skip this range if it's in use. */
1160 if (refcount_read(&pos->refcnt) > 1)
1163 inode = igrab(pos->inode);
1165 * This inode is going away. That will free
1166 * up all the ranges anyway, continue to
1172 * Take this element off list and add it tail. If
1173 * this element can't be freed, it will help with
1174 * selecting new element in next iteration of loop.
1177 list_move_tail(&dmap->busy_list, &fcd->busy_ranges);
1178 start_idx = end_idx = dmap->itn.start;
1181 spin_unlock(&fcd->lock);
1185 ret = lookup_and_reclaim_dmap(fcd, inode, start_idx, end_idx);
1194 static void fuse_dax_free_mem_worker(struct work_struct *work)
1197 struct fuse_conn_dax *fcd = container_of(work, struct fuse_conn_dax,
1199 ret = try_to_free_dmap_chunks(fcd, FUSE_DAX_RECLAIM_CHUNK);
1201 pr_debug("fuse: try_to_free_dmap_chunks() failed with err=%d\n",
1205 /* If number of free ranges are still below threshold, requeue */
1206 kick_dmap_free_worker(fcd, 1);
1209 static void fuse_free_dax_mem_ranges(struct list_head *mem_list)
1211 struct fuse_dax_mapping *range, *temp;
1213 /* Free All allocated elements */
1214 list_for_each_entry_safe(range, temp, mem_list, list) {
1215 list_del(&range->list);
1216 if (!list_empty(&range->busy_list))
1217 list_del(&range->busy_list);
1222 void fuse_dax_conn_free(struct fuse_conn *fc)
1225 fuse_free_dax_mem_ranges(&fc->dax->free_ranges);
1231 static int fuse_dax_mem_range_init(struct fuse_conn_dax *fcd)
1233 long nr_pages, nr_ranges;
1234 struct fuse_dax_mapping *range;
1236 size_t dax_size = -1;
1239 init_waitqueue_head(&fcd->range_waitq);
1240 INIT_LIST_HEAD(&fcd->free_ranges);
1241 INIT_LIST_HEAD(&fcd->busy_ranges);
1242 INIT_DELAYED_WORK(&fcd->free_work, fuse_dax_free_mem_worker);
1244 id = dax_read_lock();
1245 nr_pages = dax_direct_access(fcd->dev, 0, PHYS_PFN(dax_size),
1246 DAX_ACCESS, NULL, NULL);
1247 dax_read_unlock(id);
1249 pr_debug("dax_direct_access() returned %ld\n", nr_pages);
1253 nr_ranges = nr_pages/FUSE_DAX_PAGES;
1254 pr_debug("%s: dax mapped %ld pages. nr_ranges=%ld\n",
1255 __func__, nr_pages, nr_ranges);
1257 for (i = 0; i < nr_ranges; i++) {
1258 range = kzalloc(sizeof(struct fuse_dax_mapping), GFP_KERNEL);
1263 /* TODO: This offset only works if virtio-fs driver is not
1264 * having some memory hidden at the beginning. This needs
1267 range->window_offset = i * FUSE_DAX_SZ;
1268 range->length = FUSE_DAX_SZ;
1269 INIT_LIST_HEAD(&range->busy_list);
1270 refcount_set(&range->refcnt, 1);
1271 list_add_tail(&range->list, &fcd->free_ranges);
1274 fcd->nr_free_ranges = nr_ranges;
1275 fcd->nr_ranges = nr_ranges;
1278 /* Free All allocated elements */
1279 fuse_free_dax_mem_ranges(&fcd->free_ranges);
1283 int fuse_dax_conn_alloc(struct fuse_conn *fc, enum fuse_dax_mode dax_mode,
1284 struct dax_device *dax_dev)
1286 struct fuse_conn_dax *fcd;
1289 fc->dax_mode = dax_mode;
1294 fcd = kzalloc(sizeof(*fcd), GFP_KERNEL);
1298 spin_lock_init(&fcd->lock);
1300 err = fuse_dax_mem_range_init(fcd);
1310 bool fuse_dax_inode_alloc(struct super_block *sb, struct fuse_inode *fi)
1312 struct fuse_conn *fc = get_fuse_conn_super(sb);
1316 fi->dax = kzalloc(sizeof(*fi->dax), GFP_KERNEL_ACCOUNT);
1320 init_rwsem(&fi->dax->sem);
1321 fi->dax->tree = RB_ROOT_CACHED;
1327 static const struct address_space_operations fuse_dax_file_aops = {
1328 .writepages = fuse_dax_writepages,
1329 .direct_IO = noop_direct_IO,
1330 .dirty_folio = noop_dirty_folio,
1333 static bool fuse_should_enable_dax(struct inode *inode, unsigned int flags)
1335 struct fuse_conn *fc = get_fuse_conn(inode);
1336 enum fuse_dax_mode dax_mode = fc->dax_mode;
1338 if (dax_mode == FUSE_DAX_NEVER)
1342 * fc->dax may be NULL in 'inode' mode when filesystem device doesn't
1343 * support DAX, in which case it will silently fallback to 'never' mode.
1348 if (dax_mode == FUSE_DAX_ALWAYS)
1351 /* dax_mode is FUSE_DAX_INODE* */
1352 return fc->inode_dax && (flags & FUSE_ATTR_DAX);
1355 void fuse_dax_inode_init(struct inode *inode, unsigned int flags)
1357 if (!fuse_should_enable_dax(inode, flags))
1360 inode->i_flags |= S_DAX;
1361 inode->i_data.a_ops = &fuse_dax_file_aops;
1364 void fuse_dax_dontcache(struct inode *inode, unsigned int flags)
1366 struct fuse_conn *fc = get_fuse_conn(inode);
1368 if (fuse_is_inode_dax_mode(fc->dax_mode) &&
1369 ((bool) IS_DAX(inode) != (bool) (flags & FUSE_ATTR_DAX)))
1370 d_mark_dontcache(inode);
1373 bool fuse_dax_check_alignment(struct fuse_conn *fc, unsigned int map_alignment)
1375 if (fc->dax && (map_alignment > FUSE_DAX_SHIFT)) {
1376 pr_warn("FUSE: map_alignment %u incompatible with dax mem range size %u\n",
1377 map_alignment, FUSE_DAX_SZ);
1383 void fuse_dax_cancel_work(struct fuse_conn *fc)
1385 struct fuse_conn_dax *fcd = fc->dax;
1388 cancel_delayed_work_sync(&fcd->free_work);
1391 EXPORT_SYMBOL_GPL(fuse_dax_cancel_work);