Implement the ksyms mmap interface using d_mmap_single.

[markj]

Summary

This attempts to fix mmap(/dev/ksyms). In particular, we now
preallocate a VM object for the pages backing the symbol table and allow
it to be shared among processes that share the device file handle.

While here, fix some unrelated bugs:

• don't call the dtor if ksyms_open() fails, since devfs_open() will do that for us
• check for errors from copyout()
• use an sx lock instead of a mutex to allow M_WAITOK allocations

Test Plan

Only basic functional testing of the mmap and ioctl interfaces so far.

[kib]

sys/dev/ksyms/ksyms.c

396

Why is this limitation useful ?

560

Can the addition overflow ?

581

As was noted in the presentation, close might happen in the context of the process which is different from the process where open occured.

[markj]

sys/dev/ksyms/ksyms.c

396

I can't see a reason.

560

I think so.

581

Hm, right. I think this behaviour of mapping on open is not really supportable - there is no reliable way to remove the mapping on close.

[markj]

Further cleanups:

• Remove the ksyms ioctl interface. lockstat was modified to work without it, and the interface isn't present in Solaris/illumos.
• Don't map the object into the opening process. We have no good way of removing the mapping when the device is closed.
• Use uiomove_object() to copy the ELF file into the object, and in ksyms_read().

[markj]

sys/dev/ksyms/ksyms.c

396

I think I'll remove this, but would prefer to do so in a separate revision.

581

I addressed this by simply removing the ioctl interface, which effectively removes the need to manually map the image into the process' address space.

[kib]

sys/dev/ksyms/ksyms.c

571

You may remove empty ksyms_close(). kern_conf.c automatically substitutes NULL d_close with the same stub.