bcachefs: mkwrite() now only dirties one page

Don't dirty the whole folio - fixes write amplification with
applications doing mmaped writes.

https://www.reddit.com/r/bcachefs/comments/1klzcg1/incredible_amounts_of_write_amplification_when/

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

diff --git a/fs/bcachefs/fs-io-pagecache.c b/fs/bcachefs/fs-io-pagecache.c
index e072900e6a5b9eb11ca21b62ac95766181401242..fbae9c1de746a11338875daf5d9713c35ac46acc 100644 (file)
--- a/fs/bcachefs/fs-io-pagecache.c
+++ b/fs/bcachefs/fs-io-pagecache.c
@@ -605,10 +605,14 @@ vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
        struct address_space *mapping = file->f_mapping;
        struct bch_fs *c = inode->v.i_sb->s_fs_info;
        struct bch2_folio_reservation res;
-       unsigned len;
-       loff_t isize;
        vm_fault_t ret;
 
+       loff_t file_offset = round_down(vmf->pgoff << PAGE_SHIFT, block_bytes(c));
+       unsigned offset = file_offset - folio_pos(folio);
+       unsigned len = max(PAGE_SIZE, block_bytes(c));
+
+       BUG_ON(offset + len > folio_size(folio));
+
        bch2_folio_reservation_init(c, inode, &res);
 
        sb_start_pagefault(inode->v.i_sb);
@@ -623,24 +627,24 @@ vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
        bch2_pagecache_add_get(inode);
 
        folio_lock(folio);
-       isize = i_size_read(&inode->v);
+       u64 isize = i_size_read(&inode->v);
 
-       if (folio->mapping != mapping || folio_pos(folio) >= isize) {
+       if (folio->mapping != mapping || file_offset >= isize) {
                folio_unlock(folio);
                ret = VM_FAULT_NOPAGE;
                goto out;
        }
 
-       len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio));
+       len = min_t(unsigned, len, isize - file_offset);
 
        if (bch2_folio_set(c, inode_inum(inode), &folio, 1) ?:
-           bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) {
+           bch2_folio_reservation_get(c, inode, folio, &res, offset, len)) {
                folio_unlock(folio);
                ret = VM_FAULT_SIGBUS;
                goto out;
        }
 
-       bch2_set_folio_dirty(c, inode, folio, &res, 0, len);
+       bch2_set_folio_dirty(c, inode, folio, &res, offset, len);
        bch2_folio_reservation_put(c, inode, &res);
 
        folio_wait_stable(folio);