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/pfn_t.h>
13 #include <linux/iomap.h>
14 #include <linux/interval_tree.h>
17 * Default memory range size. A power of 2 so it agrees with common FUSE_INIT
18 * map_alignment values 4KB and 64KB.
20 #define FUSE_DAX_SHIFT 21
21 #define FUSE_DAX_SZ (1 << FUSE_DAX_SHIFT)
22 #define FUSE_DAX_PAGES (FUSE_DAX_SZ / PAGE_SIZE)
24 /* Number of ranges reclaimer will try to free in one invocation */
25 #define FUSE_DAX_RECLAIM_CHUNK (10)
28 * Dax memory reclaim threshold in percetage of total ranges. When free
29 * number of free ranges drops below this threshold, reclaim can trigger
32 #define FUSE_DAX_RECLAIM_THRESHOLD (20)
34 /** Translation information for file offsets to DAX window offsets */
35 struct fuse_dax_mapping {
36 /* Pointer to inode where this memory range is mapped */
39 /* Will connect in fcd->free_ranges to keep track of free memory */
40 struct list_head list;
42 /* For interval tree in file/inode */
43 struct interval_tree_node itn;
45 /* Will connect in fc->busy_ranges to keep track busy memory */
46 struct list_head busy_list;
48 /** Position in DAX window */
51 /** Length of mapping, in bytes */
54 /* Is this mapping read-only or read-write */
57 /* reference count when the mapping is used by dax iomap. */
61 /* Per-inode dax map */
62 struct fuse_inode_dax {
63 /* Semaphore to protect modifications to the dmap tree */
64 struct rw_semaphore sem;
66 /* Sorted rb tree of struct fuse_dax_mapping elements */
67 struct rb_root_cached tree;
71 struct fuse_conn_dax {
73 struct dax_device *dev;
75 /* Lock protecting accessess to members of this structure */
78 /* List of memory ranges which are busy */
79 unsigned long nr_busy_ranges;
80 struct list_head busy_ranges;
82 /* Worker to free up memory ranges */
83 struct delayed_work free_work;
85 /* Wait queue for a dax range to become free */
86 wait_queue_head_t range_waitq;
88 /* DAX Window Free Ranges */
90 struct list_head free_ranges;
92 unsigned long nr_ranges;
95 static inline struct fuse_dax_mapping *
96 node_to_dmap(struct interval_tree_node *node)
101 return container_of(node, struct fuse_dax_mapping, itn);
104 static struct fuse_dax_mapping *
105 alloc_dax_mapping_reclaim(struct fuse_conn_dax *fcd, struct inode *inode);
108 __kick_dmap_free_worker(struct fuse_conn_dax *fcd, unsigned long delay_ms)
110 unsigned long free_threshold;
112 /* If number of free ranges are below threshold, start reclaim */
113 free_threshold = max_t(unsigned long, fcd->nr_ranges * FUSE_DAX_RECLAIM_THRESHOLD / 100,
115 if (fcd->nr_free_ranges < free_threshold)
116 queue_delayed_work(system_long_wq, &fcd->free_work,
117 msecs_to_jiffies(delay_ms));
120 static void kick_dmap_free_worker(struct fuse_conn_dax *fcd,
121 unsigned long delay_ms)
123 spin_lock(&fcd->lock);
124 __kick_dmap_free_worker(fcd, delay_ms);
125 spin_unlock(&fcd->lock);
128 static struct fuse_dax_mapping *alloc_dax_mapping(struct fuse_conn_dax *fcd)
130 struct fuse_dax_mapping *dmap;
132 spin_lock(&fcd->lock);
133 dmap = list_first_entry_or_null(&fcd->free_ranges,
134 struct fuse_dax_mapping, list);
136 list_del_init(&dmap->list);
137 WARN_ON(fcd->nr_free_ranges <= 0);
138 fcd->nr_free_ranges--;
140 spin_unlock(&fcd->lock);
142 kick_dmap_free_worker(fcd, 0);
146 /* This assumes fcd->lock is held */
147 static void __dmap_remove_busy_list(struct fuse_conn_dax *fcd,
148 struct fuse_dax_mapping *dmap)
150 list_del_init(&dmap->busy_list);
151 WARN_ON(fcd->nr_busy_ranges == 0);
152 fcd->nr_busy_ranges--;
155 static void dmap_remove_busy_list(struct fuse_conn_dax *fcd,
156 struct fuse_dax_mapping *dmap)
158 spin_lock(&fcd->lock);
159 __dmap_remove_busy_list(fcd, dmap);
160 spin_unlock(&fcd->lock);
163 /* This assumes fcd->lock is held */
164 static void __dmap_add_to_free_pool(struct fuse_conn_dax *fcd,
165 struct fuse_dax_mapping *dmap)
167 list_add_tail(&dmap->list, &fcd->free_ranges);
168 fcd->nr_free_ranges++;
169 wake_up(&fcd->range_waitq);
172 static void dmap_add_to_free_pool(struct fuse_conn_dax *fcd,
173 struct fuse_dax_mapping *dmap)
175 /* Return fuse_dax_mapping to free list */
176 spin_lock(&fcd->lock);
177 __dmap_add_to_free_pool(fcd, dmap);
178 spin_unlock(&fcd->lock);
181 static int fuse_setup_one_mapping(struct inode *inode, unsigned long start_idx,
182 struct fuse_dax_mapping *dmap, bool writable,
185 struct fuse_mount *fm = get_fuse_mount(inode);
186 struct fuse_conn_dax *fcd = fm->fc->dax;
187 struct fuse_inode *fi = get_fuse_inode(inode);
188 struct fuse_setupmapping_in inarg;
189 loff_t offset = start_idx << FUSE_DAX_SHIFT;
193 WARN_ON(fcd->nr_free_ranges < 0);
195 /* Ask fuse daemon to setup mapping */
196 memset(&inarg, 0, sizeof(inarg));
197 inarg.foffset = offset;
199 inarg.moffset = dmap->window_offset;
200 inarg.len = FUSE_DAX_SZ;
201 inarg.flags |= FUSE_SETUPMAPPING_FLAG_READ;
203 inarg.flags |= FUSE_SETUPMAPPING_FLAG_WRITE;
204 args.opcode = FUSE_SETUPMAPPING;
205 args.nodeid = fi->nodeid;
207 args.in_args[0].size = sizeof(inarg);
208 args.in_args[0].value = &inarg;
209 err = fuse_simple_request(fm, &args);
212 dmap->writable = writable;
215 * We don't take a refernce on inode. inode is valid right now
216 * and when inode is going away, cleanup logic should first
217 * cleanup dmap entries.
220 dmap->itn.start = dmap->itn.last = start_idx;
221 /* Protected by fi->dax->sem */
222 interval_tree_insert(&dmap->itn, &fi->dax->tree);
224 spin_lock(&fcd->lock);
225 list_add_tail(&dmap->busy_list, &fcd->busy_ranges);
226 fcd->nr_busy_ranges++;
227 spin_unlock(&fcd->lock);
232 static int fuse_send_removemapping(struct inode *inode,
233 struct fuse_removemapping_in *inargp,
234 struct fuse_removemapping_one *remove_one)
236 struct fuse_inode *fi = get_fuse_inode(inode);
237 struct fuse_mount *fm = get_fuse_mount(inode);
240 args.opcode = FUSE_REMOVEMAPPING;
241 args.nodeid = fi->nodeid;
243 args.in_args[0].size = sizeof(*inargp);
244 args.in_args[0].value = inargp;
245 args.in_args[1].size = inargp->count * sizeof(*remove_one);
246 args.in_args[1].value = remove_one;
247 return fuse_simple_request(fm, &args);
250 static int dmap_removemapping_list(struct inode *inode, unsigned int num,
251 struct list_head *to_remove)
253 struct fuse_removemapping_one *remove_one, *ptr;
254 struct fuse_removemapping_in inarg;
255 struct fuse_dax_mapping *dmap;
256 int ret, i = 0, nr_alloc;
258 nr_alloc = min_t(unsigned int, num, FUSE_REMOVEMAPPING_MAX_ENTRY);
259 remove_one = kmalloc_array(nr_alloc, sizeof(*remove_one), GFP_NOFS);
264 list_for_each_entry(dmap, to_remove, list) {
265 ptr->moffset = dmap->window_offset;
266 ptr->len = dmap->length;
270 if (i >= nr_alloc || num == 0) {
271 memset(&inarg, 0, sizeof(inarg));
273 ret = fuse_send_removemapping(inode, &inarg,
287 * Cleanup dmap entry and add back to free list. This should be called with
290 static void dmap_reinit_add_to_free_pool(struct fuse_conn_dax *fcd,
291 struct fuse_dax_mapping *dmap)
293 pr_debug("fuse: freeing memory range start_idx=0x%lx end_idx=0x%lx window_offset=0x%llx length=0x%llx\n",
294 dmap->itn.start, dmap->itn.last, dmap->window_offset,
296 __dmap_remove_busy_list(fcd, dmap);
298 dmap->itn.start = dmap->itn.last = 0;
299 __dmap_add_to_free_pool(fcd, dmap);
303 * Free inode dmap entries whose range falls inside [start, end].
304 * Does not take any locks. At this point of time it should only be
305 * called from evict_inode() path where we know all dmap entries can be
308 static void inode_reclaim_dmap_range(struct fuse_conn_dax *fcd,
310 loff_t start, loff_t end)
312 struct fuse_inode *fi = get_fuse_inode(inode);
313 struct fuse_dax_mapping *dmap, *n;
315 LIST_HEAD(to_remove);
316 unsigned long start_idx = start >> FUSE_DAX_SHIFT;
317 unsigned long end_idx = end >> FUSE_DAX_SHIFT;
318 struct interval_tree_node *node;
321 node = interval_tree_iter_first(&fi->dax->tree, start_idx,
325 dmap = node_to_dmap(node);
326 /* inode is going away. There should not be any users of dmap */
327 WARN_ON(refcount_read(&dmap->refcnt) > 1);
328 interval_tree_remove(&dmap->itn, &fi->dax->tree);
330 list_add(&dmap->list, &to_remove);
333 /* Nothing to remove */
334 if (list_empty(&to_remove))
337 WARN_ON(fi->dax->nr < num);
339 err = dmap_removemapping_list(inode, num, &to_remove);
340 if (err && err != -ENOTCONN) {
341 pr_warn("Failed to removemappings. start=0x%llx end=0x%llx\n",
344 spin_lock(&fcd->lock);
345 list_for_each_entry_safe(dmap, n, &to_remove, list) {
346 list_del_init(&dmap->list);
347 dmap_reinit_add_to_free_pool(fcd, dmap);
349 spin_unlock(&fcd->lock);
352 static int dmap_removemapping_one(struct inode *inode,
353 struct fuse_dax_mapping *dmap)
355 struct fuse_removemapping_one forget_one;
356 struct fuse_removemapping_in inarg;
358 memset(&inarg, 0, sizeof(inarg));
360 memset(&forget_one, 0, sizeof(forget_one));
361 forget_one.moffset = dmap->window_offset;
362 forget_one.len = dmap->length;
364 return fuse_send_removemapping(inode, &inarg, &forget_one);
368 * It is called from evict_inode() and by that time inode is going away. So
369 * this function does not take any locks like fi->dax->sem for traversing
370 * that fuse inode interval tree. If that lock is taken then lock validator
371 * complains of deadlock situation w.r.t fs_reclaim lock.
373 void fuse_dax_inode_cleanup(struct inode *inode)
375 struct fuse_conn *fc = get_fuse_conn(inode);
376 struct fuse_inode *fi = get_fuse_inode(inode);
379 * fuse_evict_inode() has already called truncate_inode_pages_final()
380 * before we arrive here. So we should not have to worry about any
381 * pages/exception entries still associated with inode.
383 inode_reclaim_dmap_range(fc->dax, inode, 0, -1);
384 WARN_ON(fi->dax->nr);
387 static void fuse_fill_iomap_hole(struct iomap *iomap, loff_t length)
389 iomap->addr = IOMAP_NULL_ADDR;
390 iomap->length = length;
391 iomap->type = IOMAP_HOLE;
394 static void fuse_fill_iomap(struct inode *inode, loff_t pos, loff_t length,
395 struct iomap *iomap, struct fuse_dax_mapping *dmap,
399 loff_t i_size = i_size_read(inode);
401 offset = pos - (dmap->itn.start << FUSE_DAX_SHIFT);
402 len = min(length, dmap->length - offset);
404 /* If length is beyond end of file, truncate further */
405 if (pos + len > i_size)
409 iomap->addr = dmap->window_offset + offset;
411 if (flags & IOMAP_FAULT)
412 iomap->length = ALIGN(len, PAGE_SIZE);
413 iomap->type = IOMAP_MAPPED;
415 * increace refcnt so that reclaim code knows this dmap is in
416 * use. This assumes fi->dax->sem mutex is held either
419 refcount_inc(&dmap->refcnt);
421 /* iomap->private should be NULL */
422 WARN_ON_ONCE(iomap->private);
423 iomap->private = dmap;
425 /* Mapping beyond end of file is hole */
426 fuse_fill_iomap_hole(iomap, length);
430 static int fuse_setup_new_dax_mapping(struct inode *inode, loff_t pos,
431 loff_t length, unsigned int flags,
434 struct fuse_inode *fi = get_fuse_inode(inode);
435 struct fuse_conn *fc = get_fuse_conn(inode);
436 struct fuse_conn_dax *fcd = fc->dax;
437 struct fuse_dax_mapping *dmap, *alloc_dmap = NULL;
439 bool writable = flags & IOMAP_WRITE;
440 unsigned long start_idx = pos >> FUSE_DAX_SHIFT;
441 struct interval_tree_node *node;
444 * Can't do inline reclaim in fault path. We call
445 * dax_layout_busy_page() before we free a range. And
446 * fuse_wait_dax_page() drops fi->i_mmap_sem lock and requires it.
447 * In fault path we enter with fi->i_mmap_sem held and can't drop
448 * it. Also in fault path we hold fi->i_mmap_sem shared and not
449 * exclusive, so that creates further issues with fuse_wait_dax_page().
450 * Hence return -EAGAIN and fuse_dax_fault() will wait for a memory
451 * range to become free and retry.
453 if (flags & IOMAP_FAULT) {
454 alloc_dmap = alloc_dax_mapping(fcd);
458 alloc_dmap = alloc_dax_mapping_reclaim(fcd, inode);
459 if (IS_ERR(alloc_dmap))
460 return PTR_ERR(alloc_dmap);
463 /* If we are here, we should have memory allocated */
464 if (WARN_ON(!alloc_dmap))
468 * Take write lock so that only one caller can try to setup mapping
471 down_write(&fi->dax->sem);
473 * We dropped lock. Check again if somebody else setup
476 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
478 dmap = node_to_dmap(node);
479 fuse_fill_iomap(inode, pos, length, iomap, dmap, flags);
480 dmap_add_to_free_pool(fcd, alloc_dmap);
481 up_write(&fi->dax->sem);
485 /* Setup one mapping */
486 ret = fuse_setup_one_mapping(inode, pos >> FUSE_DAX_SHIFT, alloc_dmap,
489 dmap_add_to_free_pool(fcd, alloc_dmap);
490 up_write(&fi->dax->sem);
493 fuse_fill_iomap(inode, pos, length, iomap, alloc_dmap, flags);
494 up_write(&fi->dax->sem);
498 static int fuse_upgrade_dax_mapping(struct inode *inode, loff_t pos,
499 loff_t length, unsigned int flags,
502 struct fuse_inode *fi = get_fuse_inode(inode);
503 struct fuse_dax_mapping *dmap;
505 unsigned long idx = pos >> FUSE_DAX_SHIFT;
506 struct interval_tree_node *node;
509 * Take exclusive lock so that only one caller can try to setup
510 * mapping and others wait.
512 down_write(&fi->dax->sem);
513 node = interval_tree_iter_first(&fi->dax->tree, idx, idx);
515 /* We are holding either inode lock or i_mmap_sem, and that should
516 * ensure that dmap can't be truncated. We are holding a reference
517 * on dmap and that should make sure it can't be reclaimed. So dmap
518 * should still be there in tree despite the fact we dropped and
519 * re-acquired the fi->dax->sem lock.
525 dmap = node_to_dmap(node);
527 /* We took an extra reference on dmap to make sure its not reclaimd.
528 * Now we hold fi->dax->sem lock and that reference is not needed
531 if (refcount_dec_and_test(&dmap->refcnt)) {
532 /* refcount should not hit 0. This object only goes
533 * away when fuse connection goes away
538 /* Maybe another thread already upgraded mapping while we were not
541 if (dmap->writable) {
546 ret = fuse_setup_one_mapping(inode, pos >> FUSE_DAX_SHIFT, dmap, true,
551 fuse_fill_iomap(inode, pos, length, iomap, dmap, flags);
553 up_write(&fi->dax->sem);
557 /* This is just for DAX and the mapping is ephemeral, do not use it for other
558 * purposes since there is no block device with a permanent mapping.
560 static int fuse_iomap_begin(struct inode *inode, loff_t pos, loff_t length,
561 unsigned int flags, struct iomap *iomap,
562 struct iomap *srcmap)
564 struct fuse_inode *fi = get_fuse_inode(inode);
565 struct fuse_conn *fc = get_fuse_conn(inode);
566 struct fuse_dax_mapping *dmap;
567 bool writable = flags & IOMAP_WRITE;
568 unsigned long start_idx = pos >> FUSE_DAX_SHIFT;
569 struct interval_tree_node *node;
571 /* We don't support FIEMAP */
572 if (WARN_ON(flags & IOMAP_REPORT))
578 iomap->dax_dev = fc->dax->dev;
581 * Both read/write and mmap path can race here. So we need something
582 * to make sure if we are setting up mapping, then other path waits
584 * For now, use a semaphore for this. It probably needs to be
587 down_read(&fi->dax->sem);
588 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
590 dmap = node_to_dmap(node);
591 if (writable && !dmap->writable) {
592 /* Upgrade read-only mapping to read-write. This will
593 * require exclusive fi->dax->sem lock as we don't want
594 * two threads to be trying to this simultaneously
595 * for same dmap. So drop shared lock and acquire
598 * Before dropping fi->dax->sem lock, take reference
599 * on dmap so that its not freed by range reclaim.
601 refcount_inc(&dmap->refcnt);
602 up_read(&fi->dax->sem);
603 pr_debug("%s: Upgrading mapping at offset 0x%llx length 0x%llx\n",
604 __func__, pos, length);
605 return fuse_upgrade_dax_mapping(inode, pos, length,
608 fuse_fill_iomap(inode, pos, length, iomap, dmap, flags);
609 up_read(&fi->dax->sem);
613 up_read(&fi->dax->sem);
614 pr_debug("%s: no mapping at offset 0x%llx length 0x%llx\n",
615 __func__, pos, length);
616 if (pos >= i_size_read(inode))
619 return fuse_setup_new_dax_mapping(inode, pos, length, flags,
624 * If read beyond end of file happnes, fs code seems to return
628 fuse_fill_iomap_hole(iomap, length);
629 pr_debug("%s returning hole mapping. pos=0x%llx length_asked=0x%llx length_returned=0x%llx\n",
630 __func__, pos, length, iomap->length);
634 static int fuse_iomap_end(struct inode *inode, loff_t pos, loff_t length,
635 ssize_t written, unsigned int flags,
638 struct fuse_dax_mapping *dmap = iomap->private;
641 if (refcount_dec_and_test(&dmap->refcnt)) {
642 /* refcount should not hit 0. This object only goes
643 * away when fuse connection goes away
649 /* DAX writes beyond end-of-file aren't handled using iomap, so the
650 * file size is unchanged and there is nothing to do here.
655 static const struct iomap_ops fuse_iomap_ops = {
656 .iomap_begin = fuse_iomap_begin,
657 .iomap_end = fuse_iomap_end,
660 static void fuse_wait_dax_page(struct inode *inode)
662 struct fuse_inode *fi = get_fuse_inode(inode);
664 up_write(&fi->i_mmap_sem);
666 down_write(&fi->i_mmap_sem);
669 /* Should be called with fi->i_mmap_sem lock held exclusively */
670 static int __fuse_dax_break_layouts(struct inode *inode, bool *retry,
671 loff_t start, loff_t end)
675 page = dax_layout_busy_page_range(inode->i_mapping, start, end);
680 return ___wait_var_event(&page->_refcount,
681 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
682 0, 0, fuse_wait_dax_page(inode));
685 /* dmap_end == 0 leads to unmapping of whole file */
686 int fuse_dax_break_layouts(struct inode *inode, u64 dmap_start,
694 ret = __fuse_dax_break_layouts(inode, &retry, dmap_start,
696 } while (ret == 0 && retry);
701 ssize_t fuse_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
703 struct inode *inode = file_inode(iocb->ki_filp);
706 if (iocb->ki_flags & IOCB_NOWAIT) {
707 if (!inode_trylock_shared(inode))
710 inode_lock_shared(inode);
713 ret = dax_iomap_rw(iocb, to, &fuse_iomap_ops);
714 inode_unlock_shared(inode);
716 /* TODO file_accessed(iocb->f_filp) */
720 static bool file_extending_write(struct kiocb *iocb, struct iov_iter *from)
722 struct inode *inode = file_inode(iocb->ki_filp);
724 return (iov_iter_rw(from) == WRITE &&
725 ((iocb->ki_pos) >= i_size_read(inode) ||
726 (iocb->ki_pos + iov_iter_count(from) > i_size_read(inode))));
729 static ssize_t fuse_dax_direct_write(struct kiocb *iocb, struct iov_iter *from)
731 struct inode *inode = file_inode(iocb->ki_filp);
732 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
735 ret = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
739 fuse_invalidate_attr(inode);
740 fuse_write_update_size(inode, iocb->ki_pos);
744 ssize_t fuse_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
746 struct inode *inode = file_inode(iocb->ki_filp);
749 if (iocb->ki_flags & IOCB_NOWAIT) {
750 if (!inode_trylock(inode))
756 ret = generic_write_checks(iocb, from);
760 ret = file_remove_privs(iocb->ki_filp);
763 /* TODO file_update_time() but we don't want metadata I/O */
765 /* Do not use dax for file extending writes as write and on
766 * disk i_size increase are not atomic otherwise.
768 if (file_extending_write(iocb, from))
769 ret = fuse_dax_direct_write(iocb, from);
771 ret = dax_iomap_rw(iocb, from, &fuse_iomap_ops);
777 ret = generic_write_sync(iocb, ret);
781 static int fuse_dax_writepages(struct address_space *mapping,
782 struct writeback_control *wbc)
785 struct inode *inode = mapping->host;
786 struct fuse_conn *fc = get_fuse_conn(inode);
788 return dax_writeback_mapping_range(mapping, fc->dax->dev, wbc);
791 static vm_fault_t __fuse_dax_fault(struct vm_fault *vmf,
792 enum page_entry_size pe_size, bool write)
795 struct inode *inode = file_inode(vmf->vma->vm_file);
796 struct super_block *sb = inode->i_sb;
799 struct fuse_conn *fc = get_fuse_conn(inode);
800 struct fuse_conn_dax *fcd = fc->dax;
804 sb_start_pagefault(sb);
806 if (retry && !(fcd->nr_free_ranges > 0))
807 wait_event(fcd->range_waitq, (fcd->nr_free_ranges > 0));
810 * We need to serialize against not only truncate but also against
811 * fuse dax memory range reclaim. While a range is being reclaimed,
812 * we do not want any read/write/mmap to make progress and try
813 * to populate page cache or access memory we are trying to free.
815 down_read(&get_fuse_inode(inode)->i_mmap_sem);
816 ret = dax_iomap_fault(vmf, pe_size, &pfn, &error, &fuse_iomap_ops);
817 if ((ret & VM_FAULT_ERROR) && error == -EAGAIN) {
820 up_read(&get_fuse_inode(inode)->i_mmap_sem);
824 if (ret & VM_FAULT_NEEDDSYNC)
825 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
826 up_read(&get_fuse_inode(inode)->i_mmap_sem);
829 sb_end_pagefault(sb);
834 static vm_fault_t fuse_dax_fault(struct vm_fault *vmf)
836 return __fuse_dax_fault(vmf, PE_SIZE_PTE,
837 vmf->flags & FAULT_FLAG_WRITE);
840 static vm_fault_t fuse_dax_huge_fault(struct vm_fault *vmf,
841 enum page_entry_size pe_size)
843 return __fuse_dax_fault(vmf, pe_size, vmf->flags & FAULT_FLAG_WRITE);
846 static vm_fault_t fuse_dax_page_mkwrite(struct vm_fault *vmf)
848 return __fuse_dax_fault(vmf, PE_SIZE_PTE, true);
851 static vm_fault_t fuse_dax_pfn_mkwrite(struct vm_fault *vmf)
853 return __fuse_dax_fault(vmf, PE_SIZE_PTE, true);
856 static const struct vm_operations_struct fuse_dax_vm_ops = {
857 .fault = fuse_dax_fault,
858 .huge_fault = fuse_dax_huge_fault,
859 .page_mkwrite = fuse_dax_page_mkwrite,
860 .pfn_mkwrite = fuse_dax_pfn_mkwrite,
863 int fuse_dax_mmap(struct file *file, struct vm_area_struct *vma)
866 vma->vm_ops = &fuse_dax_vm_ops;
867 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
871 static int dmap_writeback_invalidate(struct inode *inode,
872 struct fuse_dax_mapping *dmap)
875 loff_t start_pos = dmap->itn.start << FUSE_DAX_SHIFT;
876 loff_t end_pos = (start_pos + FUSE_DAX_SZ - 1);
878 ret = filemap_fdatawrite_range(inode->i_mapping, start_pos, end_pos);
880 pr_debug("fuse: filemap_fdatawrite_range() failed. err=%d start_pos=0x%llx, end_pos=0x%llx\n",
881 ret, start_pos, end_pos);
885 ret = invalidate_inode_pages2_range(inode->i_mapping,
886 start_pos >> PAGE_SHIFT,
887 end_pos >> PAGE_SHIFT);
889 pr_debug("fuse: invalidate_inode_pages2_range() failed err=%d\n",
895 static int reclaim_one_dmap_locked(struct inode *inode,
896 struct fuse_dax_mapping *dmap)
899 struct fuse_inode *fi = get_fuse_inode(inode);
902 * igrab() was done to make sure inode won't go under us, and this
903 * further avoids the race with evict().
905 ret = dmap_writeback_invalidate(inode, dmap);
909 /* Remove dax mapping from inode interval tree now */
910 interval_tree_remove(&dmap->itn, &fi->dax->tree);
913 /* It is possible that umount/shutdown has killed the fuse connection
914 * and worker thread is trying to reclaim memory in parallel. Don't
917 ret = dmap_removemapping_one(inode, dmap);
918 if (ret && ret != -ENOTCONN) {
919 pr_warn("Failed to remove mapping. offset=0x%llx len=0x%llx ret=%d\n",
920 dmap->window_offset, dmap->length, ret);
925 /* Find first mapped dmap for an inode and return file offset. Caller needs
926 * to hold fi->dax->sem lock either shared or exclusive.
928 static struct fuse_dax_mapping *inode_lookup_first_dmap(struct inode *inode)
930 struct fuse_inode *fi = get_fuse_inode(inode);
931 struct fuse_dax_mapping *dmap;
932 struct interval_tree_node *node;
934 for (node = interval_tree_iter_first(&fi->dax->tree, 0, -1); node;
935 node = interval_tree_iter_next(node, 0, -1)) {
936 dmap = node_to_dmap(node);
938 if (refcount_read(&dmap->refcnt) > 1)
948 * Find first mapping in the tree and free it and return it. Do not add
949 * it back to free pool.
951 static struct fuse_dax_mapping *
952 inode_inline_reclaim_one_dmap(struct fuse_conn_dax *fcd, struct inode *inode,
955 struct fuse_inode *fi = get_fuse_inode(inode);
956 struct fuse_dax_mapping *dmap;
957 u64 dmap_start, dmap_end;
958 unsigned long start_idx;
960 struct interval_tree_node *node;
962 down_write(&fi->i_mmap_sem);
964 /* Lookup a dmap and corresponding file offset to reclaim. */
965 down_read(&fi->dax->sem);
966 dmap = inode_lookup_first_dmap(inode);
968 start_idx = dmap->itn.start;
969 dmap_start = start_idx << FUSE_DAX_SHIFT;
970 dmap_end = dmap_start + FUSE_DAX_SZ - 1;
972 up_read(&fi->dax->sem);
977 * Make sure there are no references to inode pages using
980 ret = fuse_dax_break_layouts(inode, dmap_start, dmap_end);
982 pr_debug("fuse: fuse_dax_break_layouts() failed. err=%d\n",
988 down_write(&fi->dax->sem);
989 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
990 /* Range already got reclaimed by somebody else */
994 goto out_write_dmap_sem;
997 dmap = node_to_dmap(node);
999 if (refcount_read(&dmap->refcnt) > 1) {
1003 goto out_write_dmap_sem;
1006 ret = reclaim_one_dmap_locked(inode, dmap);
1008 dmap = ERR_PTR(ret);
1009 goto out_write_dmap_sem;
1012 /* Clean up dmap. Do not add back to free list */
1013 dmap_remove_busy_list(fcd, dmap);
1015 dmap->itn.start = dmap->itn.last = 0;
1017 pr_debug("fuse: %s: inline reclaimed memory range. inode=%p, window_offset=0x%llx, length=0x%llx\n",
1018 __func__, inode, dmap->window_offset, dmap->length);
1021 up_write(&fi->dax->sem);
1023 up_write(&fi->i_mmap_sem);
1027 static struct fuse_dax_mapping *
1028 alloc_dax_mapping_reclaim(struct fuse_conn_dax *fcd, struct inode *inode)
1030 struct fuse_dax_mapping *dmap;
1031 struct fuse_inode *fi = get_fuse_inode(inode);
1036 dmap = alloc_dax_mapping(fcd);
1040 dmap = inode_inline_reclaim_one_dmap(fcd, inode, &retry);
1042 * Either we got a mapping or it is an error, return in both
1048 /* If we could not reclaim a mapping because it
1049 * had a reference or some other temporary failure,
1050 * Try again. We want to give up inline reclaim only
1051 * if there is no range assigned to this node. Otherwise
1052 * if a deadlock is possible if we sleep with fi->i_mmap_sem
1053 * held and worker to free memory can't make progress due
1054 * to unavailability of fi->i_mmap_sem lock. So sleep
1055 * only if fi->dax->nr=0
1060 * There are no mappings which can be reclaimed. Wait for one.
1061 * We are not holding fi->dax->sem. So it is possible
1062 * that range gets added now. But as we are not holding
1063 * fi->i_mmap_sem, worker should still be able to free up
1064 * a range and wake us up.
1066 if (!fi->dax->nr && !(fcd->nr_free_ranges > 0)) {
1067 if (wait_event_killable_exclusive(fcd->range_waitq,
1068 (fcd->nr_free_ranges > 0))) {
1069 return ERR_PTR(-EINTR);
1075 static int lookup_and_reclaim_dmap_locked(struct fuse_conn_dax *fcd,
1076 struct inode *inode,
1077 unsigned long start_idx)
1080 struct fuse_inode *fi = get_fuse_inode(inode);
1081 struct fuse_dax_mapping *dmap;
1082 struct interval_tree_node *node;
1084 /* Find fuse dax mapping at file offset inode. */
1085 node = interval_tree_iter_first(&fi->dax->tree, start_idx, start_idx);
1087 /* Range already got cleaned up by somebody else */
1090 dmap = node_to_dmap(node);
1093 if (refcount_read(&dmap->refcnt) > 1)
1096 ret = reclaim_one_dmap_locked(inode, dmap);
1100 /* Cleanup dmap entry and add back to free list */
1101 spin_lock(&fcd->lock);
1102 dmap_reinit_add_to_free_pool(fcd, dmap);
1103 spin_unlock(&fcd->lock);
1108 * Free a range of memory.
1110 * 1. Take fi->i_mmap_sem to block dax faults.
1111 * 2. Take fi->dax->sem to protect interval tree and also to make sure
1112 * read/write can not reuse a dmap which we might be freeing.
1114 static int lookup_and_reclaim_dmap(struct fuse_conn_dax *fcd,
1115 struct inode *inode,
1116 unsigned long start_idx,
1117 unsigned long end_idx)
1120 struct fuse_inode *fi = get_fuse_inode(inode);
1121 loff_t dmap_start = start_idx << FUSE_DAX_SHIFT;
1122 loff_t dmap_end = (dmap_start + FUSE_DAX_SZ) - 1;
1124 down_write(&fi->i_mmap_sem);
1125 ret = fuse_dax_break_layouts(inode, dmap_start, dmap_end);
1127 pr_debug("virtio_fs: fuse_dax_break_layouts() failed. err=%d\n",
1132 down_write(&fi->dax->sem);
1133 ret = lookup_and_reclaim_dmap_locked(fcd, inode, start_idx);
1134 up_write(&fi->dax->sem);
1136 up_write(&fi->i_mmap_sem);
1140 static int try_to_free_dmap_chunks(struct fuse_conn_dax *fcd,
1141 unsigned long nr_to_free)
1143 struct fuse_dax_mapping *dmap, *pos, *temp;
1144 int ret, nr_freed = 0;
1145 unsigned long start_idx = 0, end_idx = 0;
1146 struct inode *inode = NULL;
1148 /* Pick first busy range and free it for now*/
1150 if (nr_freed >= nr_to_free)
1154 spin_lock(&fcd->lock);
1156 if (!fcd->nr_busy_ranges) {
1157 spin_unlock(&fcd->lock);
1161 list_for_each_entry_safe(pos, temp, &fcd->busy_ranges,
1163 /* skip this range if it's in use. */
1164 if (refcount_read(&pos->refcnt) > 1)
1167 inode = igrab(pos->inode);
1169 * This inode is going away. That will free
1170 * up all the ranges anyway, continue to
1176 * Take this element off list and add it tail. If
1177 * this element can't be freed, it will help with
1178 * selecting new element in next iteration of loop.
1181 list_move_tail(&dmap->busy_list, &fcd->busy_ranges);
1182 start_idx = end_idx = dmap->itn.start;
1185 spin_unlock(&fcd->lock);
1189 ret = lookup_and_reclaim_dmap(fcd, inode, start_idx, end_idx);
1198 static void fuse_dax_free_mem_worker(struct work_struct *work)
1201 struct fuse_conn_dax *fcd = container_of(work, struct fuse_conn_dax,
1203 ret = try_to_free_dmap_chunks(fcd, FUSE_DAX_RECLAIM_CHUNK);
1205 pr_debug("fuse: try_to_free_dmap_chunks() failed with err=%d\n",
1209 /* If number of free ranges are still below threhold, requeue */
1210 kick_dmap_free_worker(fcd, 1);
1213 static void fuse_free_dax_mem_ranges(struct list_head *mem_list)
1215 struct fuse_dax_mapping *range, *temp;
1217 /* Free All allocated elements */
1218 list_for_each_entry_safe(range, temp, mem_list, list) {
1219 list_del(&range->list);
1220 if (!list_empty(&range->busy_list))
1221 list_del(&range->busy_list);
1226 void fuse_dax_conn_free(struct fuse_conn *fc)
1229 fuse_free_dax_mem_ranges(&fc->dax->free_ranges);
1234 static int fuse_dax_mem_range_init(struct fuse_conn_dax *fcd)
1236 long nr_pages, nr_ranges;
1239 struct fuse_dax_mapping *range;
1241 size_t dax_size = -1;
1244 init_waitqueue_head(&fcd->range_waitq);
1245 INIT_LIST_HEAD(&fcd->free_ranges);
1246 INIT_LIST_HEAD(&fcd->busy_ranges);
1247 INIT_DELAYED_WORK(&fcd->free_work, fuse_dax_free_mem_worker);
1249 id = dax_read_lock();
1250 nr_pages = dax_direct_access(fcd->dev, 0, PHYS_PFN(dax_size), &kaddr,
1252 dax_read_unlock(id);
1254 pr_debug("dax_direct_access() returned %ld\n", nr_pages);
1258 nr_ranges = nr_pages/FUSE_DAX_PAGES;
1259 pr_debug("%s: dax mapped %ld pages. nr_ranges=%ld\n",
1260 __func__, nr_pages, nr_ranges);
1262 for (i = 0; i < nr_ranges; i++) {
1263 range = kzalloc(sizeof(struct fuse_dax_mapping), GFP_KERNEL);
1268 /* TODO: This offset only works if virtio-fs driver is not
1269 * having some memory hidden at the beginning. This needs
1272 range->window_offset = i * FUSE_DAX_SZ;
1273 range->length = FUSE_DAX_SZ;
1274 INIT_LIST_HEAD(&range->busy_list);
1275 refcount_set(&range->refcnt, 1);
1276 list_add_tail(&range->list, &fcd->free_ranges);
1279 fcd->nr_free_ranges = nr_ranges;
1280 fcd->nr_ranges = nr_ranges;
1283 /* Free All allocated elements */
1284 fuse_free_dax_mem_ranges(&fcd->free_ranges);
1288 int fuse_dax_conn_alloc(struct fuse_conn *fc, struct dax_device *dax_dev)
1290 struct fuse_conn_dax *fcd;
1296 fcd = kzalloc(sizeof(*fcd), GFP_KERNEL);
1300 spin_lock_init(&fcd->lock);
1302 err = fuse_dax_mem_range_init(fcd);
1312 bool fuse_dax_inode_alloc(struct super_block *sb, struct fuse_inode *fi)
1314 struct fuse_conn *fc = get_fuse_conn_super(sb);
1318 fi->dax = kzalloc(sizeof(*fi->dax), GFP_KERNEL_ACCOUNT);
1322 init_rwsem(&fi->dax->sem);
1323 fi->dax->tree = RB_ROOT_CACHED;
1329 static const struct address_space_operations fuse_dax_file_aops = {
1330 .writepages = fuse_dax_writepages,
1331 .direct_IO = noop_direct_IO,
1332 .set_page_dirty = noop_set_page_dirty,
1333 .invalidatepage = noop_invalidatepage,
1336 void fuse_dax_inode_init(struct inode *inode)
1338 struct fuse_conn *fc = get_fuse_conn(inode);
1343 inode->i_flags |= S_DAX;
1344 inode->i_data.a_ops = &fuse_dax_file_aops;
1347 bool fuse_dax_check_alignment(struct fuse_conn *fc, unsigned int map_alignment)
1349 if (fc->dax && (map_alignment > FUSE_DAX_SHIFT)) {
1350 pr_warn("FUSE: map_alignment %u incompatible with dax mem range size %u\n",
1351 map_alignment, FUSE_DAX_SZ);
1357 void fuse_dax_cancel_work(struct fuse_conn *fc)
1359 struct fuse_conn_dax *fcd = fc->dax;
1362 cancel_delayed_work_sync(&fcd->free_work);
1365 EXPORT_SYMBOL_GPL(fuse_dax_cancel_work);