OpenE2K
/
qemu-e2k
13
4
Fork 0
Code Issues 4 Pull Requests Projects 2 Releases Activity

target/arm: Move mte check for store-exclusive 

Push the mte check behind the exclusive_addr check.
Document the several ways that we are still out of spec
with this implementation.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20230530191438.411344-18-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
This commit is contained in:
Richard Henderson 2023-06-06 10:19:39 +01:00 committed by Peter Maydell
parent c1a1f80518
commit 5096ec5b32
1 changed files with 36 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
target/arm/tcg/translate-a64.c
View File

@ -2512,17 +2512,47 @@ static void gen_store_exclusive(DisasContext *s, int rd, int rt, int rt2,
*/
TCGLabel *fail_label = gen_new_label();
TCGLabel *done_label = gen_new_label();
TCGv_i64 tmp, dirty_addr, clean_addr;
TCGv_i64 tmp, clean_addr;
MemOp memop;
memop = (size + is_pair) | MO_ALIGN;
memop = finalize_memop(s, memop);
dirty_addr = cpu_reg_sp(s, rn);
clean_addr = gen_mte_check1(s, dirty_addr, true, rn != 31, memop);
/*
* FIXME: We are out of spec here. We have recorded only the address
* from load_exclusive, not the entire range, and we assume that the
* size of the access on both sides match. The architecture allows the
* store to be smaller than the load, so long as the stored bytes are
* within the range recorded by the load.
*/
/* See AArch64.ExclusiveMonitorsPass() and AArch64.IsExclusiveVA(). */
clean_addr = clean_data_tbi(s, cpu_reg_sp(s, rn));
tcg_gen_brcond_i64(TCG_COND_NE, clean_addr, cpu_exclusive_addr, fail_label);
/*
* The write, and any associated faults, only happen if the virtual
* and physical addresses pass the exclusive monitor check. These
* faults are exceedingly unlikely, because normally the guest uses
* the exact same address register for the load_exclusive, and we
* would have recognized these faults there.
*
* It is possible to trigger an alignment fault pre-LSE2, e.g. with an
* unaligned 4-byte write within the range of an aligned 8-byte load.
* With LSE2, the store would need to cross a 16-byte boundary when the
* load did not, which would mean the store is outside the range
* recorded for the monitor, which would have failed a corrected monitor
* check above. For now, we assume no size change and retain the
* MO_ALIGN to let tcg know what we checked in the load_exclusive.
*
* It is possible to trigger an MTE fault, by performing the load with
* a virtual address with a valid tag and performing the store with the
* same virtual address and a different invalid tag.
*/
memop = size + is_pair;
if (memop == MO_128 || !dc_isar_feature(aa64_lse2, s)) {
memop |= MO_ALIGN;
}
memop = finalize_memop(s, memop);
gen_mte_check1(s, cpu_reg_sp(s, rn), true, rn != 31, memop);
tmp = tcg_temp_new_i64();
if (is_pair) {
if (size == 2) {