Re: HPET regression in 2.6.26 versus 2.6.25 -- found another user with the same regression

From: David Witbrodt
Date: Wed Aug 20 2008 - 00:51:57 EST

> > $ dmesg | grep -i hpet
> > ACPI: HPET 77FE80C0, 0038 (r1 RS690 AWRDACPI 42302E31 AWRD 98)
> > ACPI: HPET id: 0x10b9a201 base: 0xfed00000
> > hpet clockevent registered
> > hpet0: at MMIO 0xfed00000, IRQs 2, 8, 0, 0
> > hpet0: 4 32-bit timers, 14318180 Hz
> > hpet_resources: 0xfed00000 is busy
> btw., you might also want to look into drivers/char/hpet.c and
> instrument that a bit. In particular the ioremap()s done there will show
> exactly how the hpet is mapped.

Well, I exhausted about 80% of the list of experiments I put together
on Saturday, but I still can't make a 2.6.2[67] kernel boot without
"hpet=disable" or reverting commits 3def3d6d... and 1e934dda...

I spent so many hours on this today... My head is spinning, and I'm
afraid springs and smoke will start emanating from my hard drive soon
from all the recompiling and rebooting!

I need to warn you all: I discovered today, for the first time, that
I am not the first user to report this bug. This guy got bit by it
back in May, at version 2.6.26-rc2:


He has different hardware from mine, so when 2.6.26 starts hitting the
distros you may see a flood of complaints -- and I came to LKML partly
with the purpose of providing a bug fix patch (or, less preferably, a
reversion patch) for Debian, my distro of choice.

I am not giving up. I _did_ look at drivers/char/hpet.c as requested,
but since this code did not change before and after 3def3d6d..., I
was not sure what to look for that would be harmful. The same turned out
to be true about the "connection" I found between HPET and the calls
of insert_resource(), though this could actually be affected if my latest
hypothesis pans out. (All of my ideas have failed so far, though, so it
will not surprise me if my new idea fails as well.)

I found another item in arch/x86/kernel/acpi/boot.c -- which I suspect
_is_ a bug -- but which had no effect on my lockup issue when I "fixed"
it. I will let a guru decide if I have found anything important:

===== BEGIN DIFF ==========================
diff --git a/arch/x86/kernel/acpi/boot.c b/arch/x86/kernel/acpi/boot.c
index 2cdc9de..d5a9d9d 100644
--- a/arch/x86/kernel/acpi/boot.c
+++ b/arch/x86/kernel/acpi/boot.c
@@ -669,7 +669,7 @@ static int __init acpi_parse_hpet(struct acpi_table_header *table)

memset(hpet_res, 0, sizeof(*hpet_res));
hpet_res->name = (void *)&hpet_res[1];
- hpet_res->flags = IORESOURCE_MEM;
snprintf((char *)hpet_res->name, HPET_RESOURCE_NAME_SIZE, "HPET %u",
===== END DIFF ==========================

The dynamically-allocated structure that hpet_res points to eventually gets
added to the resource tree by

static __init int hpet_insert_resource(void);

which calls insert_resource(). My thought is that we are supposed to be
marking the memory region as unavailable, so that nothing else will touch
it later, right?

Anyway, what happened in 3def3d6d to cause the regression? I have a new
hypothesis to test, but I'm too tired to continue right now -- so I'll hit
it again tomorrow before I go to work:

Up until now, I have focused on the fact that request_resource() was
replaced by insert_resource(). I did not pay attention to another aspect
of that commit -- the large change in the order of execution of where the
kernel memory regions are added to the resource list.

The original (2.6.25) approach, which works on my machine, is to identify
RAM resources as they are added to the resource list, then tack on the
kernel memory regions to the (proper) resource as it is added to the tree:

for (...) {
if ([i].type == E820_RAM) {
request_resource(res, code_resource);
request_resource(res, data_resource);
request_resource(res, bss_resource);
insert_resource(&iomem_resource, res);

The problem commit moves those 3 request_resource() calls out of
e820_reserve_resources() [arch/x86/kernel/e820{,_64}.c] and into
setup_arch() [arch/x86/kernel/setup{,_64}.c]. The original code
would have this happen when setup_arch() directly callse
820_reserve_resources(), but the commit moved those lines into
setup_arch() itself, and they run sooner now than before...
potentially affecting any resources added afterward.

I don't see what effect this reordering could possibly have, since
insert_resource() ignores the IORESOURCE_BUSY flag. But that
commit changed SOMETHING... and the two most obvious changes are
the {request,insert}_resource() switch, and the repositioning of
the request_resource() calls for {code,data,bss}_resource.

I did a LOT of testing of insert_resource() last weekend, and it
completely checked out: it was only called about a dozen times,
and it always inserted the resource without returning errors or
accessing code paths for special cases. That function is not
broken internally, though its proper functioning might have
unintended side effects I have not understood yet.

I had the idea of setting up a side-by-side test -- taking the
two versions of e820_reserve_resources() from before and after
3def3d6d, renaming them, and writing a tiny replacement of
e820_reserve_resources() which would call both versions... with
the idea that I could recurse the resulting trees and compare
their contents for differences.

While reading drivers/char/hpet.c and looking at the functions
used there, I discovered request_region(), and realized that it
would be difficult to compare the entire iomem_resource tree to
a dummy tree only containing resources added by insert_resource()
and request_resource(). It might be simpler to have my tiny
e820_reserve_resources() replacement add each resource to 3 trees
-- the real iomem_resource tree, and 2 dummy trees -- which could
then be compared for differences just before the kernel locks up.

Dave W.

/* head hits pillow */
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at