it is necessary to tag these symbols as functions otherwise the
linker gets confused; this previously manifested as the internal
functions (pre-rename) like __getcontext leaking into the symbol
table of things linked against libucontext that used the ucontext
POSIX API through the weak aliases
it also had another bad effect and that is if you tried to use
libucontext's API (post-rename), the linker would warn you during
compile time that the type is unknown, and the resulting program
would crash at runtime
after properly tagging everything, I no longer notice any leakage,
i.e. there don't seem to be any references to the aliased symbols
in the resulting symbol table when using the aliases, and using
the libucontext prefixed symbols directly also works
getcontext cannot be correctly implemented in C.
If this calls another function, as it does to call syscall, it needs to
first spill its return address to the stack. If, after getcontext returns,
its caller then calls other functions, this saved return address can be
clobbered. When the context saved by getcontext is later restored, the
(now clobbered) return address will be reloaded from the stack, and the
second return from getcontext will return to the wrong location.
Because the powerpc swapcontext syscall allows either the old context or
new context pointers to be null, it is usable for implementing all of
get/set/swapcontext.
We therefore rewrite swapcontext in assembly, and get/setcontext as simple
assembly function wrappers around swapcontext.
The one piece we keep in C is the code to check the return value of the
system call and to set errno. This code was actually unnecessary before --
libc does this within syscall. However, now that the system call is made
directly in assembly, bypassing libc, it is truly necessary. Because errno
is thread-local and the details of how to set it can vary by libc, this
code remains written in C.
The ELFv2 ABI used on PPC64 differs from the ELFv1 ABI used on PPC32 here.
On PPC64, once there are any parameters that need to be passed on the
stack, space needs to be reserved on the stack to pass all parameters.
Parameters 0-7 are still only passed by register, but if the callee needs
to spill them, it can use the stack space reserved for the corresponding
parameter to do so.
The switch statement is simpler as an if/else, and removing the argp
variable makes the code more symmetric between the register and stack
parameter cases.
This was previously stored either in the CR (ppc64) or LR (ppc32) save
area of the stack, or to one of the parameter save slots.
In either case, the saved value was unused. This value is also passed
to __start_context via r31, so there's no need to pass it on the stack.
The ABI states that sp[0] should point to the previous stack frame, or be
zero if there is no previous stack frame. makecontext previously set this
slot to point to the __start_context function, rather than to a valid
stack frame.
Daniel Kolesa
Ariadne Conill