The code still allows for up to 192 bytes, should that
be useful. Simply by re-assembling upxentry.asm then
re-running exeflat, the changes will be picked up.
This format has several advantages:
* The CONFIG block need not be moved.
* The entire compressed image (depacker and payload) need
not be moved another time before the UPX depacker's own
operation.
* The CONFIG block always lives at 00602h, and the kernel
need not be aware whether it was compressed for detecting
which CONFIG block to use.
* Support for compressed images beyond 64 KiB for free.
(The assembly define TEST_FILL_INIT_TEXT can be passed in
NASMENV to test this support with 32 KiB of LFSR output.)
* A subsequent commit will shorten the stub to 64 bytes,
compared to the prior 32 + 45 = 77 bytes, with no loss
of features. (The packed payload is a bit shorter too.)
* The new stub also sets ds and es to the segment value
that would point to the DOS/EXE process's PSP. This is
apparently not used by the UPX depacker but could be in
a future or past version, or if another packer is used.
Some assembly tricks:
* SMC instead of checking the XMS driver address
in the DOS DS stub,
* SMC so that the address goes right into a
`call far immediate` instruction,
* use `repe cmpsw` to compare multiple words (saves
space over the individual word compares),
* near calls to far functions use push cs to build
a far-call stack frame,
* segments 0 and FFFFh generated by segment arithmetic
instead of loading from memory,
* common case (A20 already enabled) made to be the case
where the conditional branch just falls through, which
may be slightly better.
A crash was introduced by the prior commit, ee255d2. I
tested that the kernel still did build with gcc, but I
failed to test whether it also still runs. Turns out that
apparently changing the pointers to static made it so gcc
put the structure into the init data segment and
implicitly called memcpy to put it on the stack. That's
wrong, however. We depend on #define to convert all calls
in init to the init text copies of memcpy etc. The
implicit call by gcc however attempts to call the HMA text
copy of memcpy which happens to be at another offset.
The fix is to make the entire array of structs static so
that memory for it is allocated only in the init data
segment, not on the stack. This seems to work for both
gcc and OpenWatcom (no warnings, it builds, it runs).
CONFIG takes precedence and is by default initialised
to the empty string, that is, skipped. ALTCONFIG is
used to specify the 'alternative' filename and defaults
to fdconfig.sys. Finally, OLDCONFIG defaults to the
config.sys filename.
The protocol for this extension is based on what lDebug
already does given the PROTOCOL=FREEDOS with the switch
CMDLINE=1 (which was originally defined for the lDOS and
RxDOS load protocols). Therefore the stack frame mirrors
that of the lDOS protocol.
Also adds a CheckDebugger byte to the CONFIG block and
support for reading and writing this setting in the
SYS CONFIG program. Default is 0 (no check assume absent).
Here we depend on the signal instruction that is patched
by exeflat to indicate whether we were compressed. This
replaces the prior check that detected the UPX help data
solely by checking the CONFIG signature at 05E2h.
This commit does not change the behaviour of the code
as it already did the right thing. However, it notes
that that should happen. (This is DR DOS compatible.
For example, SHSUCDX depends on DR DOS setting this
segment so that it can keep track of the DOS DS
across it being relocated. We do not yet relocate it
but keeping a reference here is still useful.)
