Scott Wood [off-list ref] wrote on 16/11/2009 22:27:41:

On Mon, Nov 16, 2009 at 03:00:09PM -0600, Scott Wood wrote:

quoted

On Sun, Nov 15, 2009 at 06:09:27PM +0100, Joakim Tjernlund wrote:

quoted

[I used the wrong branch, this will hopefully not conflict]

This is hopfully the last iteration of the series.
Rex & Scott, please test and signoff.
Changes since last version:
 - Added mandatory pinning of iTLB
 - Added "DTLB Miss cleanup"

I still get the EFAULTs.

umm, this worked before this series with your own
version, right?

Did I screw up with the  "DTLB Miss cleanup"?

Does the self modifying code version work any better?

This snippet of strace is particularly interesting:

mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) =
0x30020000
read(0, 0x30020000, 1024)               = -1 EFAULT (Bad address)

I remember Rex saw something similar?

Or this:

brk(0x10036000)                         = 0x10036000
fstat64(1, {st_mode=S_IFCHR|0660, st_rdev=makedev(5, 1), ...}) = 0
open("/etc/mtab", O_RDONLY|O_LARGEFILE) = 3
fstat64(3, {st_mode=S_IFREG|0444, st_size=0, ...}) = 0
read(3, 0x10016000, 4096)               = -1 EFAULT (Bad address)

It seems that the failures are related to reading from /proc (/etc/mtab is
linked to /proc/mounts), though some /proc reads work
(/proc/self/oom_score).  I also see the uname syscall failing in some apps
(but not in the uname app itself, oddly), especially /proc-related apps such
as ps and top, though they're not actually accessing /proc with this call:

uname(0xffef7d0)                        = -1 EFAULT (Bad address)

-Scott