amd64: stop using top of the thread' kernel stack for FPU user save area

[kib]

Summary

Instead do one more allocation at the thread creation time. This frees a lot of space on the stack.

Also do not use alloca() for temporal storage in signal delivery sendsig() function and signal return syscall sys_sigreturn(). This saves equal amount of space, again by the cost of one more allocation at the thread creation time.

amd64: eliminate td_md.md_fpu_scratch

For signal send, copyout from the user FPU save area directly.

For sigreturn, we are in sleepable context and can do temporal allocation of the transient save area. We cannot copying from userspace directly to user save area because XSAVE state needs to be validated, also partial copyins can corrupt it.

amd64: move signal handling and register structures manipulations into sig_machdep.c from machdep.c which is too large pile of unrelated things. Some ptrace functions are moved from machdep.c to ptrace_machdep.c.

linux32: add a hack to avoid redefining the type of the savefpu tag when compiling in amd64 kernel environment with -m32. This is a temporal workaround for some future proper (but unclear) fix.

Per-commit view is available at https://kib.kiev.ua/git/gitweb.cgi?p=deviant3.git;a=shortlog;h=refs/heads/alloca

[mhorne]

sys/amd64/amd64/ptrace_machdep.c

285 ↗

(On Diff #95085)

I think ptrace_machdep.c is the logical place for these functions to live, but it would be helpful to make the organization consistent across architectures, where possible.

The current situation is somewhat of a mess, but can you do the same move for at least i386?

[kib]

sys/amd64/amd64/ptrace_machdep.c

285 ↗

(On Diff #95085)

I can do something similar to i386. There is it even more sense to move signal support to separate file, because there are more compat stuff in machdep.c (and not in ia32/...). I was somewhat hesitant to touch i386 due to general hostility against the arch.

But I will do, just not in this patch series, but later.

[markj]

sys/amd64/amd64/machdep.c

1964

Why do we zero the entire save area for thread0 but not for other threads?

sys/amd64/amd64/sig_machdep.c

70 ↗

(On Diff #95085)

I'm fairly sure many of these are not needed, at least asan.h, csan.h, msan.h are not.

112 ↗

(On Diff #95085)

BTW, CS_SECURE and EFL_SECURE are duplicated in several places.

sys/amd64/amd64/vm_machdep.c

392 ↗

(On Diff #95085)

I would initialize these after td->td_frame, so the save area and related PCB initialization is done together. I also wonder if we really need a separate thread0_init_fpu_save().

[kib]

sys/amd64/amd64/machdep.c

1964

This is just historic. I removed thread0_init_fpu_state() so this question should become not important.

[kib]

Use cpu_thread_alloc() for thread0.
Reorder statements in cpu_thread_alloc().
Centralize CS_SECURE.
Reduce includes pollution in sig_machdep.c.

[markj]

sys/amd64/amd64/machdep.c

1910

Aren't we relying on the set_top_of_stack_td() call here?

[kib]

sys/amd64/amd64/machdep.c

1910

Yes, it is done at line 1503 above.

[jhb]

I'd be happy btw with fixing i386 to stop using alloca() rather than my GCC warning hack. I do wonder why we can't just use the normal save area though?

sys/amd64/amd64/sig_machdep.c

143 ↗

(On Diff #95155)

I've wondered if we could just save the state to the "normal" save area for sendsig? Similarly, sigreturn could just copy into the normal save area? We could assert that you aren't using the save area for kernel state (which should never be true for sendsig and sigreturn)?

[kib]

sys/amd64/amd64/sig_machdep.c

143 ↗

(On Diff #95155)

Might be, it would work. This will be a follow-up work, I do not want to do this in the current batch.

[kib]

amd64: eliminate td_md.md_fpu_scratch
Also remove alloca() from ia32_signal.c.

[kib]

One more pass over _SECURE macros, e.g. linuxolator has RFLAGS_SECURE etc, remove them as well.
Consistently use uprintf() to report weird situations in sigreturn.

[jrtc27]

sys/amd64/amd64/sig_machdep.c

346 ↗

(On Diff #95249)

Nothing below this point is about signals but about ptrace and exec so this feels like the wrong place for them (or the wrong name for the file)

[kib]

sys/amd64/amd64/sig_machdep.c

346 ↗

(On Diff #95249)

If you provide a sensitive and better name for the file, I am all for the change. Otherwise, we actually have a precedents like ia32_signal.c.

[jhb]

sys/amd64/amd64/sig_machdep.c

346 ↗

(On Diff #95249)

Nothing below this point is about signals but about ptrace and exec so this feels like the wrong place for them (or the wrong name for the file)

Hmm, it's hard to say for some things like get_mcontext as the ucontext is generally about signals (esp given that POSIX has deprecated all the swapcontext and related ilk). It's true that exec_setregs() isn't really signal specific. I think MIPS puts all this stuff in a 'pm_machdep.c' where the "pm" == "process management". Maybe 'proc_machdep.c' would be a better name?