The memory accounting in Linux is probably the issue. Both times I ran
the test were from a freshly booted VM. The test parses the output of
$(free -k) to determine the amount of free memory it should allocate
and write/read from, with a given stride of pages to skip before
touching the next page.
We grab the third column of numbers from the Mem output that looks like this
total used free shared buff/cache available
Mem: 65856604 4128688 48558952 11208 13168964 60942928
Swap: 1953788 118124 1835664
So my workstation has 48558952 free bytes. We take that, give it to
memtouch to allocate that much anonymous memory rounded down to the
nearest MB with mmap and randomly read/write the buffer.
For an 8GB machine, the UEFI will have the initial 0-0xA000 memory and
0x10_0000 to 0xC00_0000 (beginning of mmio hole) prevalidated. The
next 5GB is classified as the UEFI v2.9 memory type
EFI_RESOURCE_MEMORY_UNACCEPTED, 0x1_4000_000 to 0x2_0000_0000.
The Linux e820 map should see that range as unaccepted rather than
EFI_CONVENTIONAL_MEMORY (i.e., EDK2's EFI_RESOURCE_SYSTEM_MEMORY), but
I think it needs to be accounted as free conventional memory.
AFAIU the unaccepted memory also stays in buddy (first via slow path)
and should be accounted automatically in free?
So when I see 2044MB free vs 7089MB free in my VMs, the two are
roughly 5GB different.
Is it possible all memory got allocated with memblock? Maybe some
variable tests to validate with '/proc/meminfo | grep UnacceptedMem'
would give you more clue.
Thanks,
Pankaj
