RFC because it *still* doesn't work, just like v1. However this time the changes are extremely minimal, and now I really *really* don't understand why they don't work, because as far as I can tell the new locking is *identical* with the current i/o path.
In particular, the actual access happens between _prepare and _complete, just like before. Indeed, since this doesn't work as-is, then I can't believe the atomic path works either. I have included the TLB_BSWAP stuff just so that the patch set is self-contained. I did experiment with qemu_ram_addr_from_host_nofail(), but an assertion for equality with addr + iotlbentry->addr does not trigger. As expected, really, since that's what the i/o path uses. One more cleanup for the atomics path... Ideas? r~ Richard Henderson (5): cputlb: Disable __always_inline__ without optimization cputlb: Replace switches in load/store_helper with callback cputlb: Introduce TLB_BSWAP exec: Adjust notdirty tracing cputlb: Move NOTDIRTY handling from I/O path to TLB path include/exec/cpu-all.h | 2 + include/exec/cpu-common.h | 1 - include/qemu/compiler.h | 11 ++ accel/tcg/cputlb.c | 243 +++++++++++++++++++++----------------- exec.c | 57 +-------- memory.c | 20 ---- trace-events | 4 +- 7 files changed, 157 insertions(+), 181 deletions(-) -- 2.17.1