i386/77710: Linux page fault sigcontext information is wrong

[Peter S. Housel]

>Number:         77710
>Category:       i386
>Synopsis:       Linux page fault sigcontext information is wrong
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    freebsd-i386
>State:          open
>Quarter:        
>Keywords:       
>Date-Required:
>Class:          sw-bug
>Submitter-Id:   current-users
>Arrival-Date:   Sat Feb 19 04:30:27 GMT 2005
>Closed-Date:
>Last-Modified:
>Originator:     Peter S. Housel
>Release:        6.0-CURRENT
>Organization:
>Environment:
FreeBSD housel.dyndns.org 6.0-CURRENT FreeBSD 6.0-CURRENT #4: Thu Feb 10 19:49:38 PST 2005     housel at housel.dyndns.org:/usr/obj/usr/src/sys/HOUSEL  i386
>Description:
      (Originally posted to emulation@ and hackers@ in May 2004 with no response.)

I'm running some Linux code (a garbage collector) that needs to trap page faults. There are a couple of problems with the i386 code that builds linux signal context:

/* ... */
 frame.sf_sc.sc_err    = regs->tf_err;
 frame.sf_sc.sc_trapno = bsd_to_linux_trapcode(code);

The first problem is that the sc_cr2 field of the linux sigcontext is not initialized with the faulting address (obtained at page fault time from the cr2 register).  The second problem is that sc_err is being initialized incorrectly, because in trap_pfault the original tf_err field provided by the processor has already been overwritten:

 /* kludge to pass faulting virtual address to sendsig */
 frame->tf_err = eva;

The trap handler has worked this way since trap.c revision 1.25 (now past its tenth anniversary).  Surely there is some better way to pass eva to sendsig so thtat the 3-bit err field is not lost.  (It would be nice if said field were available in the FreeBSD sigcontext, too, but that would likely involve an ABI change.)

>How-To-Repeat:
      
>Fix:
      
>Release-Note:
>Audit-Trail:
>Unformatted: