mlock & COW

[Matthew Dillon]

:>     vm_map_lookup() line 3161 and line 3297 (FreeBSD current).  Commenting
:>     those out will allow the COW on the user-wired page:
:>
:>         if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
:>             (entry->eflags & MAP_ENTRY_COW) &&
:>             (fault_type & VM_PROT_WRITE) &&
:>             (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
:>                 RETURN(KERN_PROTECTION_FAILURE);
:>         }

    Ok, never mind on commenting out that code in vm_map_lookup(),
    it won't work.  In fact, it will crash the machine even faster.

    It's messy both ways, but marginally less messy if you do it in the
    vm_fault() code.

    The problem with fixing it in the fault code is that the vm_fault()
    routine is called by the wiring and unwiring code and also by code which
    undoes failed wirings (where the entry flags do not reflect what the
    caller wants vm_fault() to do), so vm_fault() can't just check the entry
    flags and automatically user-wire.

    To do it without rewriting the whole mess (and rewriting is not a bad
    idea), the platform trap code needs to call vm_fault() with a new
    flag VM_FAULT_AUTOWIRE, then vm_map_lookup() needs to set the contents
    of &fs.wired whether it succeeds or fails, then vm_fault(), upon
    seeing result == KERN_PROTECTION_FAILURE, must check whether 
    VM_FAULT_AUTOWIRE was specified and if so check fs.wired to see if the
    mapping failed due to incompatible protections on a user wired 
    mapping, and THEN it can set VM_FAULT_USER_WIRE in fault_flags and
    continue normally.

    And on top of all of that I'm still not sure whether the originally
    underlying read-only paged that was COW'd will be properly unwired.

    As I said, messy.

					-Matt
					Matthew Dillon 
					<dillon at backplane.com>