On Sat, Jun 27, 2026 at 2:27 AM Qu Wenruo <[email protected]> wrote: > > > > 在 2026/6/27 00:48, Alexey Charkov 写道: > > Add support for generic directory iteration with opendir(), readdir() and > > closedir() in the btrfs filesystem driver. > > > > Signed-off-by: Alexey Charkov <[email protected]> > > --- > > fs/btrfs/btrfs.c | 98 > > +++++++++++++++++++++++++++++++++++++++++++++++++++++ > > fs/btrfs/ctree.h | 2 ++ > > fs/btrfs/dir-item.c | 73 +++++++++++++++++++++++++++++++++++++++ > > fs/fs.c | 4 ++- > > include/btrfs.h | 5 +++ > > 5 files changed, 181 insertions(+), 1 deletion(-) > > > > diff --git a/fs/btrfs/btrfs.c b/fs/btrfs/btrfs.c > > index f3087f690fa4..6f034861da90 100644 > > --- a/fs/btrfs/btrfs.c > > +++ b/fs/btrfs/btrfs.c > > @@ -8,7 +8,9 @@ > > #include <config.h> > > #include <malloc.h> > > #include <u-boot/uuid.h> > > +#include <linux/kernel.h> > > #include <linux/time.h> > > +#include <fs.h> > > #include "btrfs.h" > > #include "crypto/hash.h" > > #include "disk-io.h" > > @@ -159,6 +161,102 @@ int btrfs_ls(const char *path) > > return 0; > > } > > > > +/* > > + * The fs layer closes and re-probes btrfs between readdir() calls (see > > + * fs_readdir() in fs/fs.c), freeing and reallocating fs_info, so root > > cannot > > + * be stored directly. The subvolume id and inode number are stable > > though, so > > + * re-resolve the root from the current fs_info by subvolume id, which > > avoids > > + * a full path walk and is much faster. > > + */ > > +struct btrfs_dir_stream { > > + struct fs_dir_stream parent; > > + struct fs_dirent dirent; > > + u64 subvolid; > > + u64 ino; > > + u64 offset; > > +}; > > + > > +int btrfs_opendir(const char *dirname, struct fs_dir_stream **dirsp) > > +{ > > + struct btrfs_fs_info *fs_info = current_fs_info; > > + struct btrfs_dir_stream *dirs; > > + struct btrfs_root *root; > > + u64 ino; > > + u8 type; > > + int ret; > > + > > + *dirsp = NULL; > > + ASSERT(fs_info); > > + > > + ret = btrfs_lookup_path(fs_info->fs_root, BTRFS_FIRST_FREE_OBJECTID, > > + dirname, &root, &ino, &type, 40); > > + if (ret < 0) > > + return ret; > > + if (type != BTRFS_FT_DIR) > > + return -ENOTDIR; > > + > > + dirs = calloc(1, sizeof(*dirs)); > > + if (!dirs) > > + return -ENOMEM; > > + dirs->subvolid = root->root_key.objectid; > > + dirs->ino = ino; > > + > > + *dirsp = &dirs->parent; > > + return 0; > > +} > > + > > +static unsigned int btrfs_dirent_type_to_fs_type(u8 dirent_type) > > +{ > > + switch (dirent_type) { > > + case BTRFS_FT_DIR: > > + return FS_DT_DIR; > > + case BTRFS_FT_SYMLINK: > > + return FS_DT_LNK; > > + default: > > + return FS_DT_REG; > > + } > > +} > > + > > +int btrfs_readdir(struct fs_dir_stream *fs_dirs, struct fs_dirent **dentp) > > +{ > > + struct btrfs_dir_stream *dirs = container_of(fs_dirs, struct > > btrfs_dir_stream, parent); > > + struct btrfs_fs_info *fs_info = current_fs_info; > > + struct fs_dirent *dent = &dirs->dirent; > > + struct btrfs_root *root; > > + struct btrfs_key key; > > + u8 type; > > + int ret; > > + > > + *dentp = NULL; > > + ASSERT(fs_info); > > + > > + key.objectid = dirs->subvolid; > > + key.type = BTRFS_ROOT_ITEM_KEY; > > + key.offset = (u64)-1; > > + root = btrfs_read_fs_root(fs_info, &key); > > You can save the root pointer into btrfs_dir_stream, so we can avoid > subvolume root lookup for each dentry.
That was in fact the first thing I tried, but turns out that it doesn't work because a root pointer created with one fs_info structure stops working once fs_close at the end of the generic fs opendir caller closes the fs and deallocates the fs_info structure, due to it referencing trees allocated for that instance of fs_info. The result is a synchronous abort. Hence the comment above the btrfs_dir_stream definition and this second-best solution. Please let me know if I've missed a more elegant option! Closing and reopening the fs at every call which the FS layer does sounds like a huge overhead, but I must assume it was unavoidable when the interface was designed. > You do not need to bother the lifespan of subvolume roots either, they > are properly released during the close of the fs. > > Otherwise looks good to me, except an unrelated question just lines below. > > > + if (IS_ERR(root)) > > + return PTR_ERR(root); > > + > > + memset(dent, 0, sizeof(*dent)); > > + ret = btrfs_next_dir_entry(root, dirs->ino, &dirs->offset, dent->name, > > + sizeof(dent->name), &type); > > + if (ret < 0) > > + return ret; > > + if (ret > 0) > > + return -ENOENT; > > I'm not sure what is the proper/preferred/sane handling of end of directory. > > Ext4/fat returns -ENOENT, squashfs returned -1, erofs return 1, exfat > returns 0 but without populating @dentp. > > Personally I found the exfat behavior more sane, but considering the > caller fs_ls_generic() doesn't really bother the return value but only > cares if @dent is populated, it should be fine either way. > > But still, returning -ENOENT will populate errno, which may be confusing > for debugging. > > Anyway it's an unrelated nitpick. Happy to switch to exfat-like behavior. I was only looking at ext4 implementation as that is likely the most widely used and tested in various corner cases, but it sounds like either way should work. Thanks a lot, Alexey
