8:52 a.m.
----- "Bob Montgomery" <bob.montgomery(a)hp.com> wrote:
I have a dump from a 2.6.31-based x86_64 system where the number of
"possible" cpus equals the system's NR_CPUS (32).
On that system, the __per_cpu_offset table in the kernel consists of 32
valid offset pointers.
When crash loads this table into its __per_cpu_offset[NR_CPUS=4096]
array in struct kernel_table, it knows the length of the kernel's array
(32*sizeof(long)), and copies the 32 pointers, leaving the rest of its
(much longer) array full of 0x0s.
(This happens in kernel.c)
193 if (symbol_exists("__per_cpu_offset")) {
194 if (LKCD_KERNTYPES())
195 i = get_cpus_possible();
196 else
197 i = get_array_length("__per_cpu_offset", NULL, 0);
198 get_symbol_data("__per_cpu_offset",
199 sizeof(long)*((i && (i <= NR_CPUS)) ? i : NR_CPUS),
200 &kt->__per_cpu_offset[0]);
201 kt->flags |= PER_CPU_OFF;
202 }
Later, in a couple of places, crash checks for the maximum valid
__per_cpu_offset by reading the cpu_number value out of each per_cpu
area and comparing it to the expected number until the comparison fails.
(Remember NR_CPUS in crash is much larger then the kernel's NR_CPUS, and
that's OK).
>From x86_64.c:
4201 for (i = cpus = 0; i < NR_CPUS; i++) {
4202 readmem(symbol_value("per_cpu__cpu_number") +
4203 kt->__per_cpu_offset[i], KVADDR,
4204 &cpunumber, sizeof(int),
4205 "cpu number (per_cpu)", FAULT_ON_ERROR);
4206 if (cpunumber != cpus)
4207 break;
4208 cpus++;
4209 }
This works well when the kernel's array has fewer real per_cpu_offsets
than its own NR_CPUS, since the kernel preloads its array with a pointer
(BOOT_PERCPU_OFFSET) and when this loop runs past the real
per_cpu_offset pointers and tries to use the BOOT_PERCPU_OFFSET, it
reads a bogus value for cpunumber and terminates.
But when the kernel's table is full of valid per_cpu_offset pointers,
this loop continues off the end of that into the part of crash's
__per_cpu_offset array that has the 0x0 initial values, and dies with:
crash: invalid kernel virtual address: cc08 type: "cpu number (per_cpu)"
The cc08 comes from the symbol_value of per_cpu__cpu_number:
000000000000cc08 D per_cpu__cpu_number
Bottom line: Crash is assuming an insufficient array termination for
the kernel's __per_cpu_offset array (a pointer that points to an invalid
cpu_number).
The included patch adds an additional loop termination so that crash
doesn't run off the end of what it loaded from the dump. It just checks
for a NULL 0x0 value in kt->__per_cpu_offset[i].
Bob Montgomery,
Working at HP
I have a similar-but-different fix queued for this, but instead of
checking for a NULL kt->__per_cpu_offset[i] entry, it changes the
readmem() call to RETURN_ON_ERROR|QUIET instead of FAULT_ON_ERROR
like this:
if (!readmem(symbol_value("per_cpu__cpu_number") +
kt->__per_cpu_offset[i],
KVADDR, &cpunumber, sizeof(int),
"cpu number (per_cpu)", QUIET|RETURN_ON_ERROR))
break;
That should prevent the failure you're seeing.
But another question is in the (extremely) rare circumstance of a
non-CONFIG_SMP kernel. In that case, the kt->__per_cpu_offset[] array
would be all NULL, and the symbol_value("per_cpu__cpu_number")
call would return the qualified unity-mapped address. So the
virtual address calculation should work in x86_64_per_cpu_init(),
and the loop wouldn't even be entered in x86_64_get_smp_cpus()
That being said, I don't think I've seen a recent x86_64 kernel
that was not compiled CONFIG_SMP, so I can't confirm that it's
ever been tested.
So for sanity's sake, maybe your patch should also be applied,
but should also check if the "i" index is non-zero?
Thanks,
Dave
12:24 p.m.
New subject: invalid kernel virtual address: cc08 type: "cpu number (per_cpu)"
On Wed, 2009-11-11 at 14:52 +0000, Dave Anderson wrote:
----- "Bob Montgomery" <bob.montgomery(a)hp.com>
wrote:
> I have a dump from a 2.6.31-based x86_64 system where the number of
> "possible" cpus equals the system's NR_CPUS (32).
> On that system, the __per_cpu_offset table in the kernel consists of 32
> valid offset pointers.
I have a similar-but-different fix queued for this, but instead of
checking for a NULL kt->__per_cpu_offset[i] entry, it changes the
readmem() call to RETURN_ON_ERROR|QUIET instead of FAULT_ON_ERROR
like this:
if (!readmem(symbol_value("per_cpu__cpu_number") +
kt->__per_cpu_offset[i],
KVADDR, &cpunumber, sizeof(int),
"cpu number (per_cpu)", QUIET|RETURN_ON_ERROR))
break;
That should prevent the failure you're seeing.
I did that first, and thought it was sort of cheating :-)
But another question is in the (extremely) rare circumstance of a
non-CONFIG_SMP kernel. In that case, the kt->__per_cpu_offset[] array
would be all NULL, and the symbol_value("per_cpu__cpu_number")
call would return the qualified unity-mapped address. So the
virtual address calculation should work in x86_64_per_cpu_init(),
and the loop wouldn't even be entered in x86_64_get_smp_cpus()
That being said, I don't think I've seen a recent x86_64 kernel
that was not compiled CONFIG_SMP, so I can't confirm that it's
ever been tested.
So for sanity's sake, maybe your patch should also be applied,
but should also check if the "i" index is non-zero?
So like this?
+ if (i && (kt->__per_cpu_offset[i] == NULL))
+ break;
So it's always ok to try the readmem on the first element of
the array. And the RETURN_ON_ERROR would deal with something going
wrong with that, although that case would presumably be a real problem
with the dump, right? (cpus == 0)
Thanks,
Bob M.
5494
days inactive
5494
days old
devel@lists.crash-utility.osci.io
1 comments
2 participants
participants (2)
-
Bob Montgomery -
Dave Anderson