Re: [Crash-utility] increase __PHYSICAL_MASK_SHIFT_XEN?
Hi Jiri,
I'm not an expert on Xen internals, but there's quite likely no better
person to answer. So, here we go...
On Thu, 21 Jan 2021 17:18:29 +0100
Jiri Bohac <jbohac(a)suse.cz> wrote:
Hi,
I've just come across a situation where crash failed to open a
dump generated by a 4.12 XEN PV dom0 kernel, terminating with
this message:
crash: read error: physical address: ffffffffffffffff type: "pud page"
The problem is a failed machine-to-physical translation.
xen_m2p() returns an error (-1UL) and x86_64_pud_offset() than
uses that value as a physical address.
I debugged the problem by running crash inside gdb. The backtrace was:
#0 xen_m2p (machine=973135872) at kernel.c:9714
#1 0x000000000053ed4a in x86_64_pud_offset (pgd_pte=3299508019303,
vaddr=18446744072642223552, verbose=0, IS_XEN=1)
at x86_64.c:1889
#2 0x0000000000540c8a in x86_64_kvtop_xen_wpt (tc=0x0, kvaddr=18446744072642223552,
paddr=0x7fffffffdb30, verbose=0)
at x86_64.c:2523
#3 0x00000000005407d0 in x86_64_kvtop (tc=0x0, kvaddr=18446744072642223552,
paddr=0x7fffffffdb30, verbose=0)
at x86_64.c:2413
#4 0x0000000000491a97 in kvtop (tc=0x0, kvaddr=18446744072642223552,
paddr=0x7fffffffdb30, verbose=0) at memory.c:3062
#5 0x000000000048f3f0 in readmem (addr=18446744072642223552, memtype=1, buffer=0xba63a0
<shared_bufs>, size=832,
type=0x92e1f2 "module struct", error_handle=6) at memory.c:2314
#6 0x00000000005071e2 in module_init () at kernel.c:3699
....
I tracked the problem to a wrong value of
__PHYSICAL_MASK_SHIFT_XEN. The current value of 40 does not
correspond to the current kernel value of 52 since kernel commit
6f0e8bf16730a36ff6773802d8c8df56d10e60cd (xen: support 52 bit
physical addresses in pv guests).
The result is visible in the above backtrace:
x86_64_pud_offset() is called with pgd_pte=0x3003a00e067 and that
value is wrongly masked by "pud_paddr = pgd_pte &
PHYSICAL_PAGE_MASK" and passed to xen_m2p() as 0x3a00e000 instead
of 0x3003a00e000, causing the m2p translation to fail.
Thank you for this detailed analysis! AFAICS it is totally correct.
Setting __PHYSICAL_MASK_SHIFT_XEN to 52 fixes the problem with
this dump.
But I am not confident it's a safe change. My understanding is
that it should be safe, as all the unused bits of the physical
address inside the PTEs should be 0 and thus having the mask wider
than necessary should not hurt. But I am suspicious if my
reasoning is correct. Why does crash go into such trouble
differentiating between different kernels and sets
machdep->machspec->physical_mask_shift dynamically to one of
__PHYSICAL_MASK_SHIFT_XEN (40), __PHYSICAL_MASK_SHIFT_2_6 (46),
and __PHYSICAL_MASK_SHIFT_5LEVEL (52)? Would something break if
it were always set to 52? The commit adding the logic is
307e7f35.
I have asked myself the same question when writing libkdumpfile. ;-)
FWIW libkdumpfile does not go into this trouble, using 52 for all
kernels, and it has been tested on many current and historical crash
dumps with no issues. So, I believe it _is_ in fact safe to use 52
unconditionally as an architectural limit.
IIUC the only reason for having a physical mask at all is that all page
table bits beyond the current architectural limit are marked as
reserved by Intel, so they could theoretically be used for any purpose
in a future revision of the architecture. Intel has always recommended
to initialize reserved bits to zero, but I think that crash utility
developers are quite conservative and afraid of introducing regressions
for cases that are known to work, so whenever they introduce a new
feature (such as 5-level paging), they try to preserve the existing
working cases with absolutely no change.
My two cents,
Petr T
Attachments:
- attachment.sig (application/pgp-signature — 488 bytes)