[PATCH] erofs: support SEEK_HOLE/SEEK_DATA in inode_share mode

[PATCH] erofs: support SEEK_HOLE/SEEK_DATA in inode_share mode

[Jingbo Xu]

When inode_share is enabled, erofs_ishare_fops.llseek falls back to
generic_file_llseek, which treats the whole file as data and always
returns i_size for SEEK_HOLE, hiding real holes in sparse files.

Switch it to erofs_file_llseek instead.  For user files f_mapping->host
is always the real erofs inode, so SEEK_HOLE/SEEK_DATA resolve the
per-file on-disk layout via iomap_seek_hole()/iomap_seek_data().

Reported-by: Yuanhe Shu <xiangzao@linux.alibaba.com>
Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
---
 fs/erofs/data.c     | 2 +-
 fs/erofs/internal.h | 2 ++
 fs/erofs/ishare.c   | 2 +-
 3 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/fs/erofs/data.c b/fs/erofs/data.c
index d2f01245ee79..09d668deb86d 100644
--- a/fs/erofs/data.c
+++ b/fs/erofs/data.c
@@ -509,7 +509,7 @@ static int erofs_file_mmap_prepare(struct vm_area_desc *desc)
 #define erofs_file_mmap_prepare	generic_file_readonly_mmap_prepare
 #endif
 
-static loff_t erofs_file_llseek(struct file *file, loff_t offset, int whence)
+loff_t erofs_file_llseek(struct file *file, loff_t offset, int whence)
 {
 	struct inode *inode = file->f_mapping->host;
 	const struct iomap_ops *ops = &erofs_iomap_ops;
diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
index 57bd21859c65..3263c11d714c 100644
--- a/fs/erofs/internal.h
+++ b/fs/erofs/internal.h
@@ -528,6 +528,8 @@ static inline struct inode *erofs_real_inode(struct inode *inode, bool *need_ipu
 }
 #endif
 
+loff_t erofs_file_llseek(struct file *file, loff_t offset, int whence);
+
 long erofs_ioctl(struct file *filp, unsigned int cmd, unsigned long arg);
 long erofs_compat_ioctl(struct file *filp, unsigned int cmd,
 			unsigned long arg);
diff --git a/fs/erofs/ishare.c b/fs/erofs/ishare.c
index fa7d4112dec5..4c7be60565cb 100644
--- a/fs/erofs/ishare.c
+++ b/fs/erofs/ishare.c
@@ -156,7 +156,7 @@ static int erofs_ishare_fadvise(struct file *file, loff_t offset,
 
 const struct file_operations erofs_ishare_fops = {
 	.open		= erofs_ishare_file_open,
-	.llseek		= generic_file_llseek,
+	.llseek		= erofs_file_llseek,
 	.read_iter	= erofs_ishare_file_read_iter,
 	.mmap		= erofs_ishare_mmap,
 	.release	= erofs_ishare_file_release,
-- 
2.19.1.6.gb485710b

Re: [PATCH] erofs: support SEEK_HOLE/SEEK_DATA in inode_share mode

[Zhan Xusheng]

On Tue, 18 Aug 2026 19:37:13 +0800, Jingbo Xu wrote:
> - .llseek   = generic_file_llseek,
> + .llseek   = erofs_file_llseek,

No objection to the change.  While in that table, .splice_read is the entry
left that touches the page cache without going through ->private_data:

  .read_iter   kiocb_clone() onto private_data, then filemap_read()
  .mmap        vma_set_file(vma, realfile)
  .fadvise     vfs_fadvise(file->private_data, ...)
  .splice_read filemap_splice_read

filemap_splice_read() does init_sync_kiocb(&iocb, in), and
filemap_get_pages() then takes iocb->ki_filp->f_mapping (mm/filemap.c:2686),
so it works on the user file's mapping, which after your patch is confirmed
to be the real inode's.  erofs_fill_inode() sets that mapping's a_ops
unconditionally, so the data splice() and sendfile() return is correct, but
they populate the per-file page cache from disk rather than the shared one
read() and mmap() use, caching the same content twice.

Thanks,
Zhan Xusheng

Re: [PATCH] erofs: support SEEK_HOLE/SEEK_DATA in inode_share mode

[Jingbo Xu]

On 8/20/26 11:10 AM, Zhan Xusheng wrote:
> On Tue, 18 Aug 2026 19:37:13 +0800, Jingbo Xu wrote:
>> - .llseek   = generic_file_llseek,
>> + .llseek   = erofs_file_llseek,
> 
> No objection to the change.  While in that table, .splice_read is the entry
> left that touches the page cache without going through ->private_data:
> 
>   .read_iter   kiocb_clone() onto private_data, then filemap_read()
>   .mmap        vma_set_file(vma, realfile)
>   .fadvise     vfs_fadvise(file->private_data, ...)
>   .splice_read filemap_splice_read
> 
> filemap_splice_read() does init_sync_kiocb(&iocb, in), and
> filemap_get_pages() then takes iocb->ki_filp->f_mapping (mm/filemap.c:2686),
> so it works on the user file's mapping, which after your patch is confirmed
> to be the real inode's.  erofs_fill_inode() sets that mapping's a_ops
> unconditionally, so the data splice() and sendfile() return is correct, but
> they populate the per-file page cache from disk rather than the shared one
> read() and mmap() use, caching the same content twice.

Right.  I think that would be another patch to fix that.

-- 
Thanks,
Jingbo