Re: [PATCH] i386: Omit pmap workaround on i486 or later.

[Masanori Ogino]

Hello,

2016-07-08 17:43 GMT+09:00 Justus Winter <justus@gnupg.org>:
> Hi :)
>
> Masanori Ogino <masanori.ogino@gmail.com> writes:
>> How about this patch?
>> Did I share this with you in a wrong way, or is it simply useless?
>>
>> 2016-06-18 22:04 GMT+09:00 Masanori Ogino <masanori.ogino@gmail.com>:
>>> As described in XXX comments, the workaround for memory mapping is
>>> implemented for 80386 and it is unnecessary on i486 or later.  Thus, it
>>> is safe to omit that if the kernel is built for the recent (1989~)
>>> processors.
>
> Sweet, thanks for the patch.  I included it in my builds, and it seems
> good.  However, I did not understand the change in detail, could you
> motivate it a little?  Or if others agree with the change, I'll merge
> it.

OK. Let me explain it.

According to the comment in pmap.c, line 1715 and below, i386 ignores
the read-write bit of page table entries (PTEs) in kernel (a.k.a.
supervisor or privileged) mode. This is why pmap_protect (pmap.c, line
1684~) and pmap_enter (pmap.c, line 1791~) treats the combination of
read-only mode and kernel mode as a special case. Moreover,
kernel_trap (trap.c, line 1521~) have to try read-write access to pmap
first when a page fault occurs.

(I couldn't find any page describing the i386 bug, though. Probably
there were certain revisions with the bug but the others worked fine.)

The patch is intended to remove the workaround to use hardware
write-protection properly in kernel mode on recent processors and it
might boost performance.

Anyway, I found a bug in the patch now; the vm_fault call in trap.c,
line 235 should check whether (code & T_PF_WRITE) is set or not on
i486 or later. In other words, `VM_PROT_READ|VM_PROT_WRITE` in line
237 should be `(code & T_PF_WRITE) ? VM_PROT_READ|VM_PROT_WRITE :
VM_PROT_READ` on recent processors.

I will share the v2 patch after checking the code again.

-- 
Masanori Ogino