[PATCH 5/5] target/riscv: enforce SSE precedence for ssp CSR access

[PATCH 5/5] target/riscv: enforce SSE precedence for ssp CSR access

[wangyang25]

When menvcfg.SSE is clear, ssp access below M mode raises an
illegal-instruction exception before virtual-instruction classification.
Apply the henvcfg/senvcfg checks for VS/VU and preserve the user-only
path.
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4226

Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>
---
 target/riscv/tcg/csr.c | 29 ++++++++++++++++++++++++-----
 1 file changed, 24 insertions(+), 5 deletions(-)

diff --git a/target/riscv/tcg/csr.c b/target/riscv/tcg/csr.c
index fd00538..17a6fc1 100644
--- a/target/riscv/tcg/csr.c
+++ b/target/riscv/tcg/csr.c
@@ -201,15 +201,34 @@ static RISCVException cfi_ss(CPURISCVState *env, int csrno)
         return RISCV_EXCP_NONE;
     }
 
-    /* if bcfi not active for current env, access to csr is illegal */
-    if (!cpu_get_bcfien(env)) {
 #if !defined(CONFIG_USER_ONLY)
-        if (env->debugger) {
-            return RISCV_EXCP_NONE;
-        } else if (env->virt_enabled) {
+    if (env->debugger) {
+        return RISCV_EXCP_NONE;
+    }
+
+    /* priv < M and menvcfg.SSE = 0 must raise illegal-instruction */
+    if (!(env->menvcfg & MENVCFG_SSE)) {
+        return RISCV_EXCP_ILLEGAL_INST;
+    }
+
+    if (env->virt_enabled) {
+        if (env->priv == PRV_S) {
+            if (env->henvcfg & HENVCFG_SSE) {
+                return RISCV_EXCP_NONE;
+            }
+            return RISCV_EXCP_VIRT_INSTRUCTION_FAULT;
+        }
+        if (env->priv == PRV_U) {
+            if ((env->henvcfg & HENVCFG_SSE) &&
+                (env->senvcfg & SENVCFG_SSE)) {
+                return RISCV_EXCP_NONE;
+            }
             return RISCV_EXCP_VIRT_INSTRUCTION_FAULT;
         }
+    }
 #endif
+
+    if (!cpu_get_bcfien(env)) {
         return RISCV_EXCP_ILLEGAL_INST;
     }
 
-- 
2.55.0.windows.2

[PATCH 3/5] target/riscv: mask inactive senvcfg.SSE on read

[wangyang25]

The CFI rules make senvcfg.SSE read as zero when menvcfg.SSE is clear.
In VU mode it also reads as zero when henvcfg.SSE is clear. Mask the
readback to match those rules.
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4224

Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>
---
 target/riscv/tcg/csr.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/target/riscv/tcg/csr.c b/target/riscv/tcg/csr.c
index 5ffd858..7078736 100644
--- a/target/riscv/tcg/csr.c
+++ b/target/riscv/tcg/csr.c
@@ -3305,6 +3305,14 @@ static RISCVException read_senvcfg(CPURISCVState *env, int csrno,
     }
 
     *val = env->senvcfg;
+
+    if (env_archcpu(env)->cfg.ext_zicfiss) {
+        if (!(env->menvcfg & MENVCFG_SSE) ||
+            (env->virt_enabled && !(env->henvcfg & HENVCFG_SSE))) {
+            *val &= ~SENVCFG_SSE;
+        }
+    }
+
     return RISCV_EXCP_NONE;
 }
 
-- 
2.55.0.windows.2

[PATCH 0/5] target/riscv: fix CFI and envcfg CSR handling

[wangyang25]

This series fixes five independent RISC-V CSR and trap-handling defects.

The patches cover:

 - making mnstatus.MNPELP writable when Zicfilp is implemented;
 - saving ELP in mstatus.MPELP for M-mode RNMI exceptions;
 - masking inactive senvcfg.SSE on readback;
 - gating envcfg CBO fields on Zicbom and Zicboz;
 - enforcing menvcfg.SSE precedence for ssp CSR access.

Each patch includes a corresponding GitLab Work Item and a Signed-off-by
line. The reports were found with an automated differential/emulation tool
and manually triaged against QEMU master eea8fe61b8be.

The witness families were rerun on the Linux x86_64 execution plane. The
patches apply cleanly and pass git diff --check and checkpatch.pl with zero
errors and zero warnings.

wangyang (5):
  target/riscv: make mnstatus.MNPELP writable with Zicfilp
  target/riscv: save ELP in MPELP for RNMI exceptions
  target/riscv: mask inactive senvcfg.SSE on read
  target/riscv: gate envcfg CBO fields on extensions
  target/riscv: enforce SSE precedence for ssp CSR access

 target/riscv/tcg/cpu_helper.c |  7 +---
 target/riscv/tcg/csr.c        | 79 +++++++++++++++++++++++++++++++----
 2 files changed, 71 insertions(+), 15 deletions(-)

-- 
2.55.0.windows.2

[PATCH 1/5] target/riscv: make mnstatus.MNPELP writable with Zicfilp

[wangyang25]

Zicfilp adds MNPELP to mnstatus, and MNRET consumes it when restoring ELP.
Include the field in write_mnstatus() when Zicfilp is implemented.
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4222

Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>
---
 target/riscv/tcg/csr.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/target/riscv/tcg/csr.c b/target/riscv/tcg/csr.c
index 36f2004..5ffd858 100644
--- a/target/riscv/tcg/csr.c
+++ b/target/riscv/tcg/csr.c
@@ -5563,6 +5563,10 @@ static RISCVException write_mnstatus(CPURISCVState *env, int csrno,
         mask |= MNSTATUS_MNPV;
     }
 
+    if (env_archcpu(env)->cfg.ext_zicfilp) {
+        mask |= MNSTATUS_MNPELP;
+    }
+
     /* mnstatus.mnie can only be cleared by hardware. */
     env->mnstatus = (env->mnstatus & MNSTATUS_NMIE) | (val & mask);
     return RISCV_EXCP_NONE;
-- 
2.55.0.windows.2

[PATCH 2/5] target/riscv: save ELP in MPELP for RNMI exceptions

[wangyang25]

An M-mode exception with mnstatus.NMIE clear uses the RNMI vector, but
Smrnmi otherwise specifies the normal M-mode state updates. MRET restores
ELP from mstatus.MPELP, so save it there rather than in MNPELP.
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4223

Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>
---
 target/riscv/tcg/cpu_helper.c | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/target/riscv/tcg/cpu_helper.c b/target/riscv/tcg/cpu_helper.c
index 07d9222..6337e91 100644
--- a/target/riscv/tcg/cpu_helper.c
+++ b/target/riscv/tcg/cpu_helper.c
@@ -2275,12 +2275,7 @@ void riscv_cpu_do_interrupt(CPUState *cs)
         /* handle the trap in M-mode */
         /* save elp status */
         if (cpu_get_fcfien(env)) {
-            if (nnmi_excep) {
-                env->mnstatus = set_field(env->mnstatus, MNSTATUS_MNPELP,
-                                          env->elp);
-            } else {
-                env->mstatus = set_field(env->mstatus, MSTATUS_MPELP, env->elp);
-            }
+            env->mstatus = set_field(env->mstatus, MSTATUS_MPELP, env->elp);
         }
 
         if (riscv_has_ext(env, RVH)) {
-- 
2.55.0.windows.2

[PATCH 4/5] target/riscv: gate envcfg CBO fields on extensions

[wangyang25]

The envcfg CBO fields are provided by Zicbom and Zicboz. Keep them
read-only zero when the corresponding extension is absent in all three
envcfg CSRs.
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4225

Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>
---
 target/riscv/tcg/csr.c | 38 ++++++++++++++++++++++++++++++++++----
 1 file changed, 34 insertions(+), 4 deletions(-)

diff --git a/target/riscv/tcg/csr.c b/target/riscv/tcg/csr.c
index 7078736..fd00538 100644
--- a/target/riscv/tcg/csr.c
+++ b/target/riscv/tcg/csr.c
@@ -3212,10 +3212,19 @@ static RISCVException write_menvcfg(CPURISCVState *env, int csrno,
                                     target_ulong val, uintptr_t ra)
 {
     const RISCVCPUConfig *cfg = riscv_cpu_cfg(env);
-    uint64_t mask = MENVCFG_FIOM | MENVCFG_CBIE | MENVCFG_CBCFE |
-                    MENVCFG_CBZE;
+    uint64_t mask = MENVCFG_FIOM;
     bool stce_changed = false;
 
+    /* CBIE/CBCFE are read-only zero when Zicbom is not implemented. */
+    if (cfg->ext_zicbom) {
+        mask |= MENVCFG_CBIE | MENVCFG_CBCFE;
+    }
+
+    /* CBZE is read-only zero when Zicboz is not implemented. */
+    if (cfg->ext_zicboz) {
+        mask |= MENVCFG_CBZE;
+    }
+
     if (riscv_cpu_mxl(env) == MXL_RV64) {
         mask |= (cfg->ext_svpbmt ? MENVCFG_PBMTE : 0) |
                 (cfg->ext_sstc ? MENVCFG_STCE : 0) |
@@ -3319,8 +3328,19 @@ static RISCVException read_senvcfg(CPURISCVState *env, int csrno,
 static RISCVException write_senvcfg(CPURISCVState *env, int csrno,
                                     target_ulong val, uintptr_t ra)
 {
-    uint64_t mask = SENVCFG_FIOM | SENVCFG_CBIE | SENVCFG_CBCFE | SENVCFG_CBZE;
+    uint64_t mask = SENVCFG_FIOM;
     RISCVException ret;
+
+    /* CBIE/CBCFE are read-only zero when Zicbom is not implemented. */
+    if (env_archcpu(env)->cfg.ext_zicbom) {
+        mask |= SENVCFG_CBIE | SENVCFG_CBCFE;
+    }
+
+    /* CBZE is read-only zero when Zicboz is not implemented. */
+    if (env_archcpu(env)->cfg.ext_zicboz) {
+        mask |= SENVCFG_CBZE;
+    }
+
     /* Update PMM field only if the value is valid according to Zjpm v1.0 */
     if (env_archcpu(env)->cfg.ext_ssnpm &&
         riscv_cpu_mxl(env) == MXL_RV64 &&
@@ -3377,7 +3397,7 @@ static RISCVException write_henvcfg(CPURISCVState *env, int csrno,
                                     target_ulong val, uintptr_t ra)
 {
     const RISCVCPUConfig *cfg = riscv_cpu_cfg(env);
-    uint64_t mask = HENVCFG_FIOM | HENVCFG_CBIE | HENVCFG_CBCFE | HENVCFG_CBZE;
+    uint64_t mask = HENVCFG_FIOM;
     RISCVException ret;
     bool stce_changed = false;
 
@@ -3386,6 +3406,16 @@ static RISCVException write_henvcfg(CPURISCVState *env, int csrno,
         return ret;
     }
 
+    /* CBIE/CBCFE are read-only zero when Zicbom is not implemented. */
+    if (cfg->ext_zicbom) {
+        mask |= HENVCFG_CBIE | HENVCFG_CBCFE;
+    }
+
+    /* CBZE is read-only zero when Zicboz is not implemented. */
+    if (cfg->ext_zicboz) {
+        mask |= HENVCFG_CBZE;
+    }
+
     if (riscv_cpu_mxl(env) == MXL_RV64) {
         mask |= env->menvcfg & (HENVCFG_PBMTE | HENVCFG_STCE | HENVCFG_ADUE |
                                 HENVCFG_DTE);
-- 
2.55.0.windows.2

Re: [PATCH 4/5] target/riscv: gate envcfg CBO fields on extensions

[Alistair Francis]

On Sun, 2026-08-23 at 21:14 +0800, wangyang25@otcaix.iscas.ac.cn wrote:
> The envcfg CBO fields are provided by Zicbom and Zicboz. Keep them
> read-only zero when the corresponding extension is absent in all
> three
> envcfg CSRs.
> Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4225
> 
> Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>

Reviewed-by: Alistair Francis <alistair.francis@wdc.com>

Alistair

> ---
>  target/riscv/tcg/csr.c | 38 ++++++++++++++++++++++++++++++++++----
>  1 file changed, 34 insertions(+), 4 deletions(-)
> 
> diff --git a/target/riscv/tcg/csr.c b/target/riscv/tcg/csr.c
> index 7078736..fd00538 100644
> --- a/target/riscv/tcg/csr.c
> +++ b/target/riscv/tcg/csr.c
> @@ -3212,10 +3212,19 @@ static RISCVException
> write_menvcfg(CPURISCVState *env, int csrno,
>                                      target_ulong val, uintptr_t ra)
>  {
>      const RISCVCPUConfig *cfg = riscv_cpu_cfg(env);
> -    uint64_t mask = MENVCFG_FIOM | MENVCFG_CBIE | MENVCFG_CBCFE |
> -                    MENVCFG_CBZE;
> +    uint64_t mask = MENVCFG_FIOM;
>      bool stce_changed = false;
>  
> +    /* CBIE/CBCFE are read-only zero when Zicbom is not implemented.
> */
> +    if (cfg->ext_zicbom) {
> +        mask |= MENVCFG_CBIE | MENVCFG_CBCFE;
> +    }
> +
> +    /* CBZE is read-only zero when Zicboz is not implemented. */
> +    if (cfg->ext_zicboz) {
> +        mask |= MENVCFG_CBZE;
> +    }
> +
>      if (riscv_cpu_mxl(env) == MXL_RV64) {
>          mask |= (cfg->ext_svpbmt ? MENVCFG_PBMTE : 0) |
>                  (cfg->ext_sstc ? MENVCFG_STCE : 0) |
> @@ -3319,8 +3328,19 @@ static RISCVException
> read_senvcfg(CPURISCVState *env, int csrno,
>  static RISCVException write_senvcfg(CPURISCVState *env, int csrno,
>                                      target_ulong val, uintptr_t ra)
>  {
> -    uint64_t mask = SENVCFG_FIOM | SENVCFG_CBIE | SENVCFG_CBCFE |
> SENVCFG_CBZE;
> +    uint64_t mask = SENVCFG_FIOM;
>      RISCVException ret;
> +
> +    /* CBIE/CBCFE are read-only zero when Zicbom is not implemented.
> */
> +    if (env_archcpu(env)->cfg.ext_zicbom) {
> +        mask |= SENVCFG_CBIE | SENVCFG_CBCFE;
> +    }
> +
> +    /* CBZE is read-only zero when Zicboz is not implemented. */
> +    if (env_archcpu(env)->cfg.ext_zicboz) {
> +        mask |= SENVCFG_CBZE;
> +    }
> +
>      /* Update PMM field only if the value is valid according to Zjpm
> v1.0 */
>      if (env_archcpu(env)->cfg.ext_ssnpm &&
>          riscv_cpu_mxl(env) == MXL_RV64 &&
> @@ -3377,7 +3397,7 @@ static RISCVException
> write_henvcfg(CPURISCVState *env, int csrno,
>                                      target_ulong val, uintptr_t ra)
>  {
>      const RISCVCPUConfig *cfg = riscv_cpu_cfg(env);
> -    uint64_t mask = HENVCFG_FIOM | HENVCFG_CBIE | HENVCFG_CBCFE |
> HENVCFG_CBZE;
> +    uint64_t mask = HENVCFG_FIOM;
>      RISCVException ret;
>      bool stce_changed = false;
>  
> @@ -3386,6 +3406,16 @@ static RISCVException
> write_henvcfg(CPURISCVState *env, int csrno,
>          return ret;
>      }
>  
> +    /* CBIE/CBCFE are read-only zero when Zicbom is not implemented.
> */
> +    if (cfg->ext_zicbom) {
> +        mask |= HENVCFG_CBIE | HENVCFG_CBCFE;
> +    }
> +
> +    /* CBZE is read-only zero when Zicboz is not implemented. */
> +    if (cfg->ext_zicboz) {
> +        mask |= HENVCFG_CBZE;
> +    }
> +
>      if (riscv_cpu_mxl(env) == MXL_RV64) {
>          mask |= env->menvcfg & (HENVCFG_PBMTE | HENVCFG_STCE |
> HENVCFG_ADUE |
>                                  HENVCFG_DTE);

Re: [PATCH 1/5] target/riscv: make mnstatus.MNPELP writable with Zicfilp

[Alistair Francis]

On Sun, 2026-08-23 at 21:14 +0800, wangyang25@otcaix.iscas.ac.cn wrote:
> Zicfilp adds MNPELP to mnstatus, and MNRET consumes it when restoring
> ELP.
> Include the field in write_mnstatus() when Zicfilp is implemented.
> Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4222
> 
> Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>

Reviewed-by: Alistair Francis <alistair.francis@wdc.com>

Alistair

> ---
>  target/riscv/tcg/csr.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/target/riscv/tcg/csr.c b/target/riscv/tcg/csr.c
> index 36f2004..5ffd858 100644
> --- a/target/riscv/tcg/csr.c
> +++ b/target/riscv/tcg/csr.c
> @@ -5563,6 +5563,10 @@ static RISCVException
> write_mnstatus(CPURISCVState *env, int csrno,
>          mask |= MNSTATUS_MNPV;
>      }
>  
> +    if (env_archcpu(env)->cfg.ext_zicfilp) {
> +        mask |= MNSTATUS_MNPELP;
> +    }
> +
>      /* mnstatus.mnie can only be cleared by hardware. */
>      env->mnstatus = (env->mnstatus & MNSTATUS_NMIE) | (val & mask);
>      return RISCV_EXCP_NONE;

Re: [PATCH 2/5] target/riscv: save ELP in MPELP for RNMI exceptions

[Alistair Francis]

On Sun, 2026-08-23 at 21:14 +0800, wangyang25@otcaix.iscas.ac.cn wrote:
> An M-mode exception with mnstatus.NMIE clear uses the RNMI vector,
> but
> Smrnmi otherwise specifies the normal M-mode state updates. MRET
> restores
> ELP from mstatus.MPELP, so save it there rather than in MNPELP.

This commit message isn't really clear. What are you fixing?

Alistair

> Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4223
> 
> Signed-off-by: wangyang <wangyang25@otcaix.iscas.ac.cn>
> ---
>  target/riscv/tcg/cpu_helper.c | 7 +------
>  1 file changed, 1 insertion(+), 6 deletions(-)
> 
> diff --git a/target/riscv/tcg/cpu_helper.c
> b/target/riscv/tcg/cpu_helper.c
> index 07d9222..6337e91 100644
> --- a/target/riscv/tcg/cpu_helper.c
> +++ b/target/riscv/tcg/cpu_helper.c
> @@ -2275,12 +2275,7 @@ void riscv_cpu_do_interrupt(CPUState *cs)
>          /* handle the trap in M-mode */
>          /* save elp status */
>          if (cpu_get_fcfien(env)) {
> -            if (nnmi_excep) {
> -                env->mnstatus = set_field(env->mnstatus,
> MNSTATUS_MNPELP,
> -                                          env->elp);
> -            } else {
> -                env->mstatus = set_field(env->mstatus,
> MSTATUS_MPELP, env->elp);
> -            }
> +            env->mstatus = set_field(env->mstatus, MSTATUS_MPELP,
> env->elp);
>          }
>  
>          if (riscv_has_ext(env, RVH)) {

Re: [PATCH 2/5] target/riscv: save ELP in MPELP for RNMI exceptions

[Wang Yang]

Hi Alistair,

Thank you for the feedback, and sorry that the commit message was unclear.

This patch fixes the Smrnmi and Zicfilp case where an M-mode exception occurs with mnstatus.NMIE=0. QEMU currently saves ELP in mnstatus.MNPELP when redirecting to the RNMI exception vector, but this path returns with MRET, which restores ELP from mstatus.MPELP. As a result, ELP is lost and a subsequent indirect branch requiring a landing pad may avoid the expected software-check exception. The patch saves ELP in mstatus.MPELP for this path; the true RNMI interrupt path using MNPELP and MNRET is unchanged.

I also found that Zephyr Li has submitted an equivalent fix, including a TCG regression test: https://www.mail-archive.com/qemu-devel@nongnu.org/msg1222084.html. Since that patch fixes the same issue, I will not submit a duplicate revision. Please let me know if any further clarification is needed.

Best regards,
Wang Yang