I thought I had 4 GiB of memory, but just to be sure, let's ask the BIOS about that:

?: sudo dmidecode --type 20
# dmidecode 2.12
SMBIOS 2.6 present.

Handle 0x000B, DMI type 20, 19 bytes
Memory Device Mapped Address
    Starting Address: 0x00000000000
    Ending Address: 0x0007FFFFFFF
    Range Size: 2 GB
    Physical Device Handle: 0x000A
    Memory Array Mapped Address Handle: 0x000E
    Partition Row Position: Unknown
    Interleave Position: Unknown
    Interleaved Data Depth: Unknown

Handle 0x000D, DMI type 20, 19 bytes
Memory Device Mapped Address
    Starting Address: 0x00080000000
    Ending Address: 0x000FFFFFFFF
    Range Size: 2 GB
    Physical Device Handle: 0x000C
    Memory Array Mapped Address Handle: 0x000E
    Partition Row Position: Unknown
    Interleave Position: Unknown
    Interleaved Data Depth: Unknown

Alright, 4 GiB it is. But I can't use all of it:

?: cat /proc/meminfo | head -n 1
MemTotal:        3913452 kB

Somehow, somewhere, I lost 274 MiB. Where did 6% of my memory go?

Now I know the address ranges in DMI are incorrect, because the ACPI memory map reports usable ranges well beyond the ending address of the second memory module:

?: dmesg | grep -E "BIOS-e820: .* usable"
[    0.000000] BIOS-e820: [mem 0x0000000000000000-0x000000000009e7ff] usable
[    0.000000] BIOS-e820: [mem 0x0000000000100000-0x00000000dee7bfff] usable
[    0.000000] BIOS-e820: [mem 0x0000000100000000-0x0000000117ffffff] usable

I get pretty much the same info from /proc/iomem (except for the 4 kiB hole 0x000-0xFFF), which also shows that the kernel only accounts for less than 8 MiB.

I guess 0x00000000-0x7FFFFFFF is indeed mapped to the first memory module, and 0x80000000-0xDFFFFFFF to part of the second memory module (a bunch of ACPI NVS things live between 0xDEE7C000 and 0xDEF30FFF, and the remaining 16-something MiB of that range are just 'reserved'). I guess the highest 0x18000000 bytes of the second memory module are mapped above the 4 GiB mark.

But even then, there are two problems:

• 128 MiB (0x08000000 bytes, living somewhere between 0xE0000000 and 0xFFFFFFFF) are still completely unaccounted for. To note, my graphics card is on PCI-Express and (allegedly) has 1 GiB dedicated memory, so that shouldn't be the culprit. Did the BIOS screw up in moving the memory, leaving it partially shadowed by MMIO?
• Even with this mediocre explanation, I only 'found' 128 MiB. But /proc/meminfo is reporting a much larger deficit; where's the other 146 MiB? How does Linux count MemTotal?