Provide the CloudABI vDSO to its executables.

[ed]

Summary

CloudABI executables already provide support for passing in vDSOs. This
functionality is used by the emulator for OS X to inject system call
handlers. On FreeBSD, we could use it to optimize calls to
gettimeofday(), etc.

Though I don't have any plans to optimize any system calls right now,
let's go ahead and already pass in a vDSO. This will allow us to
simplify the executables, as the traditional "syscall" shims can be
removed.

Test Plan

Seems to work both on x86-64 and aarch64, both compiled into the kernel
and built as a kernel module.

[ed]

Hi Kostik,

The reason why I'm putting you down as a reviewer: could you please take a look at the code in cloudabi_vdso.c to see whether I'm using the VM API correctly?

Thanks,
Ed

[kib]

sys/compat/cloudabi/cloudabi_vdso.c

56 ↗

(On Diff #19117)

Why do you check this, instead of assigning sv_shared_page_len to the needed value ?

70 ↗

(On Diff #19117)

vdso_length is size_t, i.e. unsigned. I suspect that this calculation does not do what you want, on the last partial page of the vdso, the first argument to MIN() is very large unsigned value..

[ed]

Simply update the vDSO length dynamically.

There is no need to have a fixed space hardcoded in the sysentvec.
cloudabi_vdso_init() can simply compute those values correctly for us.

[ed]

Thanks for the feedback. What do you think of this version?

sys/compat/cloudabi/cloudabi_vdso.c

56 ↗

(On Diff #19117)

That's a good question. I have to confess I inherited this construct from how we do the vDSO for Linux. I've just fixed it up to adjust these fields dynamically. Thanks!

70 ↗

(On Diff #19117)

i * PAGE_SIZE is guaranteed to not exceed vdso_length in this case, so it should never be negative. i is strictly smaller than the number of pages needed to hold the vDSO.