daxfs: Read holes as zeros instead of returning short

[congwang-mk]

Problem

A file extended by truncate has no overlay pages behind the new range and no base image data either. daxfs_read_iter() treated that as end of file and stopped, so read() reported EOF in the middle of a file that stat() says is larger:

: > f; truncate -s 16384 f
wc -c < f    ->  16384
dd if=f      ->  0 bytes

POSIX requires a hole to read as zeros. The mmap path already does this, since daxfs_copy_page() zero-fills when there is no source, so read() and mmap() disagreed about the same range. This is also why the fio benchmarking in #15 had to lay files out with dd rather than truncating them into place.

Fix

Zero-fill the hole a page at a time and continue the loop, so a read spanning a hole still finds the data after it rather than stopping at the gap.

The part that needed care: the caller could not tell a hole from a failure. daxfs_base_file_data() returned NULL both when nothing backed the range and when daxfs_pcache_get_page() failed with -EIO, -EFBIG or -ENOENT. Blindly zero-filling would have converted a pcache read error into silent zeros, which is a worse bug than the short read it replaces.

So the function now returns ERR_PTR for failures and keeps NULL for holes, and the contract is documented at the definition. daxfs_read_iter() propagates the error (or returns the short count if it already copied something) and zero-fills only for a genuine NULL.

The other five callers are updated for the new contract, which is not optional: they previously passed the NULL to daxfs_copy_page(), which zero-fills, and with an ERR_PTR that would have memcpy'd from an error pointer.

• daxfs_write_prealloc() leaves the page unpublished so the per-page path retries and surfaces the error
• daxfs_write_iter() fails the write rather than COW zeros over unreadable data
• daxfs_dax_fault() and daxfs_dax_pfn_mkwrite() fall through to VM_FAULT_SIGBUS, taking care to do it via the existing !data check after sb_end_pagefault() rather than returning from inside the critical section

Testing

New regression test in tests/test_overlay.sh: a truncate-extended file must report the full size from read(), the range must be zeros, and a read spanning the hole must still find data written after it.

Verified it catches the bug rather than passing vacuously:

without the fix:   FAIL: Hole reads
                   Tests run: 21, passed: 20, failed: 1

with the fix:      PASS: Holes read as zeros, not short
                   Tests run: 21, passed: 21, failed: 0

Also checked, since this touches the hot read loop:

• 256 MiB random payload round-trips with a matching md5
• psync 4k random read unchanged at 4912 MiB/s (1 job) and 21576 MiB/s (8 jobs) on a 40 CPU host, in line with the post-daxfs: Skip the hardened usercopy check on region copies #15 baseline
• no new dmesg warnings

Note

#16 is open against main and also adds a test after test_overlay_truncate in tests/test_overlay.sh. Whichever merges second will need a trivial rebase for the test registration list; the C changes do not overlap.