09.04.2024 14:02, Paolo Bonzini wrote:
On Tue, Apr 9, 2024 at 12:59 PM Zhao Liu <zhao1....@intel.com> wrote:
Hi Michael & Paolo,
I debugged it manually, and found the problem occurs in tlb_index() with
mmu_idx=5.
For v7.2, the maximum mmu index supported by i386 is 4 (since
NB_MMU_MODES = 5 defined in target/i386/cpu-param.h).
On Michael's 7.2-i386-mmu-idx tree, the commit 9fc3a7828d25 ("target/i386:
use separate MMU indexes for 32-bit accesses") introduced more indexes
without relaxing the NB_MMU_MODES for i386.
Oh well. I starred a few times at the new MMU_foo_IDXes in
v8.2.0-1876-g90f641531c78
"target/i386: use separate MMU indexes for 32-bit accesses" a few times, it
looked
somehow wrong, but it didn't occur to me what exactly was wrong there :)
Before this fix, probe_access_internal() just got the wrong mmu_idx as 4,
and it's not out of bounds. After this fix, the right mmu_idx=5 is truly
out of bounds.
On the master branch, there's no such issue since the commits ffd824f3f32d
("include/exec: Set default NB_MMU_MODES to 16") and 6787318a5d86
("target/i386: Remove NB_MMU_MODES define") relaxed upper limit of MMU
index for i386.
Thanks Zhao! Alternatively, it's enough to set NB_MMU_MODES to 8 in
commit 9fc3a7828d25.
Thank you, both of you, this is exactly the prob. I fixed it by increasing
NB_MMU_NODES in "use separate MMU indexes for 32-bit accesses", since this
is the change which introduces the new indexes.
Now 7.2 work again, and appears to be under control, which is an excellent
news! :)
/mjt
