Comment 35 for bug 2056461

Created attachment 306444
Fix option 1

At last I've found the problem (which is, as I thought a week ago, due to 'stat'), but quite subtle.

With extra debugging enabled on the mount (debug=0x0f rather than debug=0x08) it works, so with debug=0x08, after adding additional debug messages in v9fs_stat2inode_dotl(), I can see that in the failing case there are 2 calls to that function immediately before the v9fs_file_read_iter() starts. In those calls the netfs->remote_i_size is updated from stat->st_size and I see st_size has different (increasing) values but smaller than the actual file size.

So, effectively, all the data hasn't arrived before the read and once the read starts the size is no longer updated.

When debugging the additional cumulative delays of writing extreme amounts of debug info to the kernel log gives enough time for the data to arrive so that when v9fs_stat2inode_dotl() is called the final size is correct.

This possible fix updates stat if a read returns 0 bytes.

Tested with 6.8.12 and 6.10-rc2 repeatedly (10+ iterations so far) and no failures with debug logging completely disabled.