On Thu, 23 Feb 2023 16:10:59 PST (-0800), Richard Henderson wrote:
On 2/23/23 13:44, Daniel Henrique Barboza wrote:
probe_access_flags() as it is today uses probe_access_full(), which in
turn uses probe_access_internal() with size = 0. probe_access_internal()
then uses the size to call the tlb_fill() callback for the given CPU.
This size param ('fault_size' as probe_access_internal() calls it) is
ignored by most existing .tlb_fill callback implementations, e.g.
arm_cpu_tlb_fill(), ppc_cpu_tlb_fill(), x86_cpu_tlb_fill() and
mips_cpu_tlb_fill() to name a few.
But RISC-V riscv_cpu_tlb_fill() actually uses it. The 'size' parameter
is used to check for PMP (Physical Memory Protection) access. This is
necessary because PMP does not make any guarantees about all the bytes
of the same page having the same permissions, i.e. the same page can
have different PMP properties, so we're forced to make sub-page range
checks. To allow RISC-V emulation to do a probe_acess_flags() that
covers PMP, we need to either add a 'size' param to the existing
probe_acess_flags() or create a new interface (e.g.
probe_access_range_flags).
There are quite a few probe_* APIs already, so let's add a 'size' param
to probe_access_flags() and re-use this API. This is done by open coding
what probe_access_full() does inside probe_acess_flags() and passing the
'size' param to probe_acess_internal(). Existing probe_access_flags()
callers use size = 0 to not change their current API usage. 'size' is
asserted to enforce single page access like probe_access() already does.
No behavioral changes intended.
Signed-off-by: Daniel Henrique Barboza<dbarb...@ventanamicro.com>
---
accel/stubs/tcg-stub.c | 2 +-
accel/tcg/cputlb.c | 17 ++++++++++++++---
accel/tcg/user-exec.c | 5 +++--
include/exec/exec-all.h | 3 ++-
semihosting/uaccess.c | 2 +-
target/arm/ptw.c | 2 +-
target/arm/sve_helper.c | 2 +-
target/s390x/tcg/mem_helper.c | 6 +++---
8 files changed, 26 insertions(+), 13 deletions(-)
Queueing to tcg-next.
Unless I'm missing something, that's not in Peter's tree yet? I Ack'd
the cover, it's fine with me if you want to take these via the TCG tree.
Doing some sort of shared tag for the first one works for me too, I've
also got some other stuff in the RISC-V queue.
Thanks!
