Since the FAR_ELx fix at 38d931687fa1, it is reported that page granularity flushing is broken.
This makes sense, since TCG will record the entire virtual address in its TLB, not simply the 56 significant bits. With no other TCG support, the ARM backend should require 256 different page flushes to clear the virtual address of any possible tag. So I added a new tcg interface that allows passing the size of the virtual address. I thought a simple bit-count was a cleaner interface than passing in a mask, since it means that we couldn't be passed nonsensical masks like 0xdeadbeef. It also makes it easy to re-direct special cases. I don't have a test case that triggers the bug. All I can say is that (1) this still boots a normal kernel and (2) the code paths are triggered since the kernel enables tbi for EL0. Jordan, can you test this please? r~ Richard Henderson (2): accel/tcg: Add tlb_flush_page_bits_by_mmuidx* target/arm: Use tlb_flush_page_bits_by_mmuidx* include/exec/exec-all.h | 36 ++++++ accel/tcg/cputlb.c | 259 ++++++++++++++++++++++++++++++++++++++-- target/arm/helper.c | 46 +++++-- 3 files changed, 325 insertions(+), 16 deletions(-) -- 2.25.1
