On 10/27/25 18:00, Danila Zhebryakov wrote:
These changes make powerpc booke206 mmu pass TLB_BSWAP flag for pages marked as
LE, and also fixes all the issues this causes.
- added TLB_BSWAP to non-MMIO flags
- inserted additional check for LE page when bswap-ing instructions
- removed assert for cross-page TLB_BSWAP accesses
- added gdbstub fix to account for TLB_BSWAP of the currently running code
BE- and LE- atomic operations (used by load-and-reserve and store-conditional
instructions) were unified to support TLB_BSWAP
Some of the fixes are based on what is done in
https://github.com/YetAnotherMod/qemu
V2:
- Removed usage of PAGE_USER bit for LE (comment on previous version)
- Refactored MMU code so directly setting TLB_BSWAP is possible
- moved translation LE page detect to occur 1 time per translation (comment
on previous version)
- Removed redundant assert in atomic_mmu_lookup (comment on previous version)
- Added assert on cross-endian accesses (ideally, there should be an
exception)
- Changed atomic need_bswap handling (comment on previous version)
- Rebased and fixed atomics
V3
- unfix assumed typo (if vs iff). Everything else unchanged
V4
- fix leftovers from BE/LE atomic op tables
- fix uninitialized CPUTLBEntryFull full in mmu_common
According to RM, unaligned accesses between LE and BE pages should cause an
exception on e500.
However, (as I understand it) supporting this would require adding a new
callback to TCGCPUOps, likely just for E500 or maybe a couple other CPUs doing
this.
Such accesses are clearly not normal and usually don't occur in real
applications, so I think just placing assert is OK here.
Danila Zhebryakov (4):
accel/tcg: Unify big- and little- endian atomic ops
target/ppc: refactor MMU helpers
target/ppc: Add support for LE pages on PowerPC booke206 mmu
target/ppc: fix GDB stub to work correctly with LE pages
accel/tcg/atomic_common.c.inc | 36 +--
accel/tcg/atomic_template.h | 326 ++++++++++------------------
accel/tcg/cputlb.c | 29 ++-
accel/tcg/tcg-runtime.h | 48 +---
accel/tcg/user-exec.c | 2 +-
include/accel/tcg/cpu-ldst-common.h | 51 ++---
target/m68k/op_helper.c | 4 +-
target/ppc/gdbstub.c | 17 +-
target/ppc/internal.h | 3 +-
target/ppc/mmu-booke.c | 67 +++---
target/ppc/mmu-booke.h | 2 +-
target/ppc/mmu-hash32.c | 30 +--
target/ppc/mmu-hash32.h | 3 +-
target/ppc/mmu-hash64.c | 14 +-
target/ppc/mmu-hash64.h | 2 +-
target/ppc/mmu-radix64.c | 20 +-
target/ppc/mmu-radix64.h | 2 +-
target/ppc/mmu_common.c | 75 ++++---
target/ppc/mmu_helper.c | 15 +-
target/ppc/translate.c | 42 +++-
target/s390x/tcg/mem_helper.c | 6 +-
tcg/tcg-op-ldst.c | 51 ++---
22 files changed, 374 insertions(+), 471 deletions(-)
Hey Danila,
This series is failing on QEMU_CI. You can check it here :
https://gitlab.com/rathc/qemu/-/pipelines/2131344109
<https://gitlab.com/rathc/qemu/-/pipelines/2131344109>
Here are some logs for quick reference :
|configuring with: '../configure' '--enable-werror' '--disable-docs'
'--cross-prefix=powerpc64le-linux-gnu-' '--disable-tools' '--enable-kvm'
'--disable-tcg' '--without-default-devices'
...
|||
|[1690/1838] Linking target qemu-system-ppc64|
|
||
FAILED: qemu-system-ppc64
||
powerpc64le-linux-gnu-gcc -m64 -mlittle-endian @qemu-system-ppc64.rsp
||
/usr/lib/gcc-cross/powerpc64le-linux-gnu/14/../../../../powerpc64le-linux-gnu/bin/ld:
libqemu-ppc64-softmmu.a.p/target_ppc_gdbstub.c.o: in function
`ppc_maybe_bswap_register':
||
/builds/rathc/qemu/build/../target/ppc/gdbstub.c:96:(.text+0xd0):
undefined reference to `probe_access_full_mmu'
||
collect2: error: ld returned 1 exit status
||
[1691/1838] Linking target qemu-system-ppc
||
FAILED: qemu-system-ppc
||
powerpc64le-linux-gnu-gcc -m64 -mlittle-endian @qemu-system-ppc.rsp
||
/usr/lib/gcc-cross/powerpc64le-linux-gnu/14/../../../../powerpc64le-linux-gnu/bin/ld:
libqemu-ppc-softmmu.a.p/target_ppc_gdbstub.c.o: in function
`ppc_maybe_bswap_register':
||
/builds/rathc/qemu/build/../target/ppc/gdbstub.c:96:(.text+0xd0):
undefined reference to `probe_access_full_mmu'
||
collect2: error: ld returned 1 exit status
||
[1692/1838] Generating pc-bios/edk2-arm-code.fd with a custom command
(wrapped by meson to capture output)
||
ninja: build stopped: subcommand failed.
||make: *** [Makefile:168: run-ninja] Error 1
|
Thanks and regards,
Chinmay
