Fork of memtest86+ 5.01. Improves its correctness, code health, and code readability while doing no harm to performance and coverage.
NON-goals of this fork:
- Support UEFI.
- Build a 64-bit binary.
- Stabilize SMP support. SMP remains experimental.
- Hardware detection: reporting SPD and DMI metadata, or any controller- specific metadata is orthogonal to actually testing the memory. Such functionality can only break or go out-of-date.
-
Apply the rest of the ubuntu/debian patches. Get in touch with distro maintainers. what's the next step?
-
Attempt to update reloc.c from a recent glibc that knows how to relocate more ELF section types. I suspect we aren't completely relocating everything we need to; see comment on the STATIC macro in test.c.
-
Update the version number to 6.0
-
Specify units on major APIs, data structures, and shared constants. Never again wonder whether a given integer is a VA, PA, byte address, dword index, 4k page index, and so on.
-
Add an ASSERT() macro for validating assumptions. Prints a message in yellow to distinguish asserts from memory errors (printed in red.)
-
Add a unit test for the core memtest86+ routines in test.c. The "self_test" allows running these routines in gdb or adding debug prints. Prod code can use the DEBUGF macro to print messages.
The self_test does not yet support SMP or error injection. It does cover all the major test routines. You need a 32-bit libc to build the self_test (sorry.)
-
Clean up the duplicative and brittle loop tests in test.c. There were address-dependent bugs here, eg. they would only manifest if a segment ended at exactly 4G or had a length that was a near multiple of SPINSZ.
Some bugs were silent, for example test 7 would stop early (without testing all memory) if the range from calculate_chunk() maps to the exact top of 4GB address space. Other tests may have the same bug.
To deny bugs their habitat, all tests share a new 'foreach_segment()' routine providing the segment-loop logic in a single central place. The new logic avoids overflow, not by being excruciatingly careful, but by computing loop bounds in terms of dword indices instead of byte addresses. Thus all quantities are far from the 4G overflow line.
-
Avoid 1- and 2-letter variable names. This is a work in progress. Favor names that are meaningful, or at least searchable.
-
Indent consistently. Replace tab characters with spaces so the code will render the same for everyone (the jwz standard.)
-
Rewrite test 1 (addr_tst1) to fix a bug. Before, it could possibly write to stray addresses that aren't mapped to DRAM -- and could be mapped to hardware devices, uh oh! The simplified test combines the unbanked and banked modes, since the banked mode pretty much covered the unbanked mode. It uses the foreach_segment() routine and dword-indexed-based loop tests to avoid overflows.