target/riscv: preserve vl on element-zero fault
PATCH v3 1/1 target/riscv: preserve vl on element-zero fault
A unit-stride fault-only-first load must leave vl unchanged when element zero raises a synchronous exception. Keep the shortened value in a local bound until all loads complete, and only then update the architectural vl.
Add a bare-metal RVV regression test which faults vle32ff.v on element zero and checks that vl remains unchanged in the trap handler.
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3544
---
Changes in v3:
- Rebase onto current master.
- Regenerate the patch to fix the malformed v2 submission.
Changes in v2:
- Fix commit message formatting.
- Enable vector state in the bare-metal test.
- Use an aligned unmapped address and compare against the vsetvli result.
- Express the mstatus.VS setting as a field shift.
target/riscv/tcg/vector_helper.c | 23 +++++---- tests/tcg/riscv64/Makefile.softmmu-target | 5 ++ tests/tcg/riscv64/test-vle32ff.S | 57 +++++++++++++++++++++++ 3 files changed, 76 insertions(+), 9 deletions(-) create mode 100644 tests/tcg/riscv64/test-vle32ff.S
target/riscv/tcg/vector_helper.c +14 -9
@@ -686,7 +686,7 @@ vext_ldff(void *vd, void *v0, target_ulong base, CPURISCVState *env, uint32_t desc, vext_ldst_elem_fn_tlb *ldst_tlb, vext_ldst_elem_fn_host *ldst_host, uint32_t log2_esz, uintptr_t ra) { - uint32_t i, k, vl = 0; + uint32_t i, k, vl = 0, load_vl;uint32_t nf =vext_nf (desc);uint32_t vm =vext_vm (desc);uint32_t max_elems =vext_max_elems (desc, log2_esz);
@@ -752,22 +752,24 @@ vext_ldff(void *vd, void *v0, target_ulong base, CPURISCVState *env, } } ProbeSuccess: -/* load bytes from guest memory */ -if (vl !=0 ) { - env->vl = vl; - } +/* +* Keep a shortened vl local until all loads complete. In particular, +* an exception from element zero must leave the architectural vl alone. +*/ + load_vl = vl ? vl : env->vl; -if (env->vstart < env->vl) { +if (env->vstart < load_vl) {if (vm) {/* Load/store elements in the first page */ if (likely (elems)) { + elems =MIN (elems, load_vl - env->vstart);vext_page_ldst_us (env, vd, addr, elems, nf, max_elems, log2_esz, true, mmu_index, ldst_tlb, ldst_host, ra); }/* Load/store elements in the second page */ -if (unlikely (env->vstart < env->vl)) { +if (unlikely (env->vstart < load_vl)) {/* Cross page element */ if (unlikely (page_split % msize)) {for (k =0 ; k < nf; k++) {
@@ -780,7 +782,7 @@ ProbeSuccess: addr = base + ((env->vstart * nf) << log2_esz);/* Get number of elements of second page */ - elems = env->vl - env->vstart; + elems = load_vl - env->vstart;/* Load/store elements in the second page */ vext_page_ldst_us(env, vd, addr, elems, nf, max_elems,
@@ -788,7 +790,7 @@ ProbeSuccess: ldst_host, ra); } }else { -for (i = env->vstart; i < env->vl; i++) { +for (i = env->vstart; i < load_vl; i++) { k =0 ;while (k < nf) {if (!vext_elem_mask (v0, i)) {
@@ -806,6 +808,9 @@ ProbeSuccess: } } } +if (vl !=0 ) { + env->vl = vl; + } env->vstart =0 ;vext_set_tail_elems_1s (env->vl, vd, desc, nf, esz, max_elems);
tests/tcg/riscv64/Makefile.softmmu-target +5 -0
@@ -41,5 +41,10 @@ comma:= , run-test-crc32: test-crc32 $(call run-test, $<, $(QEMU) -cpu rv64$(comma)xlrbr=true $(QEMU_OPTS)$<) +EXTRA_RUNS += run-test-vle32ff +run-test-vle32ff: test-vle32ff + $(call run-test, $<, $(QEMU) -cpu rv64$(comma)v=true $(QEMU_OPTS)$<) +test-vle32ff: CFLAGS += -march=rv64gcv + # We don't currently support the multiarch system tests undefine MULTIARCH_TESTS
tests/tcg/riscv64/test-vle32ff.S new file +57 -0
57 lines not on this page · raw message
Signature · 1 line
2.43.0
Quoted 4 lines · Zephyr Li
A unit-stride fault-only-first load must leave vl unchanged when element zero raises a synchronous exception. Keep the shortened value in a local bound until all loads complete, and only then update the architectural vl.
The failing case is a corner case. When the element-0 faulting probe in vext_ldff() faults, it longjmps out of the helper, ProbeSuccess is never reached and env->vl is never touched -- that path was already correct.
The broken path is element 0 in a mapped MMIO page: probe_access() succeeds there, but QEMU cannot know a device transaction will fail without issuing it that broken the assumption of the pre probe_access() checking.
Quoted 2 lines
tests/tcg/riscv64/test-vle32ff.S new file +0
-0
...
Quoted 3 lines
tests/tcg/riscv64/test-vle32ff.S new file +2
-0+ vsetvli t2, t1, e32, m1, ta, ma + li t0, 0x18000000 /* Unmapped gap in the virt memory map.*/
Actually it is not unmapped -- it is mapped as unassigned MMIO, and that is exactly why the bug reproduces: an unmapped address would make the element-0 probe fault and the test would pass even with the bug present.
I think the fixed vext_ldff implementation is valid according to the RISC-V specification, considering the VL reduction rule and the non-idempotent constraint. Only the test case wording and the root cause explanation may need to be updated.
As for the patch:
Thanks Max, that makes sense.
I'll update the root-cause description and the test comment accordingly. The implementation itself will remain unchanged.
Thanks,
Zephyr
Quoted 36 lines · Max Chou
Quoted 3 lines · Zephyr Li
A unit-stride fault-only-first load must leave vl unchanged when element zero raises a synchronous exception. Keep the shortened value in a local bound until all loads complete, and only then update the architectural
vl.
Quoted 1 line
The failing case is a corner case. When the element-0 faulting probe in
vext_ldff()faults, it longjmps out of the helper, ProbeSuccess is never reached andenv->vlis never touched -- that path was already correct.The broken path is element 0 in a mapped MMIO page:
probe_access()succeeds there, but QEMU cannot know a device transaction will fail without issuing it that broken the assumption of the preprobe_access()checking.Quoted 2 lines
tests/tcg/riscv64/test-vle32ff.S new file +0
-0...
Quoted 3 lines
tests/tcg/riscv64/test-vle32ff.S new file +2
-0+ vsetvli t2, t1, e32, m1, ta, ma + li t0, 0x18000000 /* Unmapped gap in the virt memory map.*/
Actually it is not unmapped -- it is mapped as unassigned MMIO, and that is exactly why the bug reproduces: an unmapped address would make the element-0 probe fault and the test would pass even with the bug present.
I think the fixed
vext_ldffimplementation is valid according to the RISC-V specification, considering the VL reduction rule and the non-idempotent constraint. Only the test case wording and the root cause explanation may need to be updated.As for the patch:
Reviewed-by Max Chou <[email protected]>