Skip to contentSkip to conversation

riscv: Discard vector state on syscalls

PATCH Superseded Björn Töpel 4 participants 6 messages Jun 22, 2023 – Jun 26, 2023 RISC-V

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 riscv: Discard vector state on syscalls

From: Björn Töpel <[email protected]>

The RISC-V vector specification states:
Executing a system call causes all caller-saved vector registers (v0-v31, vl, vtype) and vstart to become unspecified.

The vector registers are cleared, vill is set (invalid), and the vector status is set to Initial.

That way we can prevent userspace from accidentally relying on the stated save.

Rémi pointed out [1] that clearing the registers might be superfluous, and setting vill is sufficient.

---

I figured I'd sent out a proper patch. I like Andy's optimization patch, but TBH I think we should do that as a follow up.

As Rémi pointed out, the clearing might be opted out, but I left it in here.

Björn

---
 arch/riscv/include/asm/vector.h | 25 +++++++++++++++++++++++++
 arch/riscv/kernel/traps.c       |  2 ++
 2 files changed, 27 insertions(+)
arch/riscv/include/asm/vector.h +25 -0
index 04c0b07bf6cd..692ce55e4a69 100644
@@ -163,6 +163,30 @@ static inline void __switch_to_vector(struct task_struct *prev,
 void riscv_v_vstate_ctrl_init(struct task_struct *tsk);
 bool riscv_v_vstate_ctrl_user_allowed(void);
 
+static inline void riscv_v_vstate_discard(struct pt_regs *regs)
+{
+	unsigned long vl, vtype_inval = 1UL << (BITS_PER_LONG - 1);
+
+	if (!riscv_v_vstate_query(regs))
+		return;
+
+	riscv_v_enable();
+	asm volatile (
+		".option push\n\t"
+		".option arch, +v\n\t"
+		"vsetvli	%0, x0, e8, m8, ta, ma\n\t"
+		"vmv.v.i	v0, 0\n\t"
+		"vmv.v.i	v8, 0\n\t"
+		"vmv.v.i	v16, 0\n\t"
+		"vmv.v.i	v24, 0\n\t"
+		"vsetvl		%0, x0, %1\n\t"
+		".option pop\n\t"
+		: "=&r" (vl) : "r" (vtype_inval) : "memory");
+	riscv_v_disable();
+
+	riscv_v_vstate_on(regs);
+}
+
 #else /* ! CONFIG_RISCV_ISA_V  */
 
 struct pt_regs;
@@ -178,6 +202,7 @@ static inline bool riscv_v_vstate_ctrl_user_allowed(void) { return false; }
 #define __switch_to_vector(__prev, __next)	do {} while (0)
 #define riscv_v_vstate_off(regs)		do {} while (0)
 #define riscv_v_vstate_on(regs)			do {} while (0)
+#define riscv_v_vstate_discard(regs)		do {} while (0)
 
 #endif /* CONFIG_RISCV_ISA_V */
 
arch/riscv/kernel/traps.c +2 -0
index 05ffdcd1424e..00c68b57ff88 100644
@@ -295,6 +295,8 @@ asmlinkage __visible __trap_section void do_trap_ecall_u(struct pt_regs *regs)
 		regs->epc += 4;
 		regs->orig_a0 = regs->a0;
 
+		riscv_v_vstate_discard(regs);
+
 		syscall = syscall_enter_from_user_mode(regs, syscall);
 
 		if (syscall < NR_syscalls)

base-commit: 4681dacadeefa5ca6017e00736adc1d7dc963c6a

Signature · 1 line
2.39.2
maintainer
#1 lore raw
Quoted 19 lines · On Thu, 22 Jun 2023 10:36:13 PDT (-0700), [email protected] wrote:

From: Björn Töpel <[email protected]>

The RISC-V vector specification states:
Executing a system call causes all caller-saved vector registers (v0-v31, vl, vtype) and vstart to become unspecified.

The vector registers are cleared, vill is set (invalid), and the vector status is set to Initial.

That way we can prevent userspace from accidentally relying on the stated save.

Rémi pointed out [1] that clearing the registers might be superfluous, and setting vill is sufficient.

Quoted 7 lines

---

I figured I'd sent out a proper patch. I like Andy's optimization patch, but TBH I think we should do that as a follow up.

As Rémi pointed out, the clearing might be opted out, but I left it in here.

I think we're going to end up with a bunch of uarch-specific stuff here, but for now having the heavy hammer seems safest. There's no V hardware yet so we can't really tell how anything performs, at least this way we're definately not leaking anything.

So I'm OK with this. I'd also be fine with clearing to all-1s, I think it's kind of splitting hairs at this point: the 1s are nice because they're what the rest of V does, but setting vill should make everything trap anyway so maybe it doesn't matter -- and it's not clear if 1 or 0 will allow initial, so who knows.

Darius: I'm cool swapping over to the 1s if you feel strongly about it. Bjorn says Sweeden is on vacation, so just LMK and I'll re-spin it with the 1s.

Regardless I'd like to pick up something that blows away V state for this merge window, as it'll make sure the uABI is quite strictly enforced.

Quoted 64 lines

Björn

---
 arch/riscv/include/asm/vector.h | 25 +++++++++++++++++++++++++
 arch/riscv/kernel/traps.c       |  2 ++
 2 files changed, 27 insertions(+)
arch/riscv/include/asm/vector.h +25 -0
index 04c0b07bf6cd..692ce55e4a69 100644
@@ -163,6 +163,30 @@ static inline void __switch_to_vector(struct task_struct *prev,
 void riscv_v_vstate_ctrl_init(struct task_struct *tsk);
 bool riscv_v_vstate_ctrl_user_allowed(void);
 
+static inline void riscv_v_vstate_discard(struct pt_regs *regs)
+{
+	unsigned long vl, vtype_inval = 1UL << (BITS_PER_LONG - 1);
+
+	if (!riscv_v_vstate_query(regs))
+		return;
+
+	riscv_v_enable();
+	asm volatile (
+		".option push\n\t"
+		".option arch, +v\n\t"
+		"vsetvli	%0, x0, e8, m8, ta, ma\n\t"
+		"vmv.v.i	v0, 0\n\t"
+		"vmv.v.i	v8, 0\n\t"
+		"vmv.v.i	v16, 0\n\t"
+		"vmv.v.i	v24, 0\n\t"
+		"vsetvl		%0, x0, %1\n\t"
+		".option pop\n\t"
+		: "=&r" (vl) : "r" (vtype_inval) : "memory");
+	riscv_v_disable();
+
+	riscv_v_vstate_on(regs);
+}
+
 #else /* ! CONFIG_RISCV_ISA_V  */
 
 struct pt_regs;
@@ -178,6 +202,7 @@ static inline bool riscv_v_vstate_ctrl_user_allowed(void) { return false; }
 #define __switch_to_vector(__prev, __next)	do {} while (0)
 #define riscv_v_vstate_off(regs)		do {} while (0)
 #define riscv_v_vstate_on(regs)			do {} while (0)
+#define riscv_v_vstate_discard(regs)		do {} while (0)
 
 #endif /* CONFIG_RISCV_ISA_V */
 
arch/riscv/kernel/traps.c +2 -0
index 05ffdcd1424e..00c68b57ff88 100644
@@ -295,6 +295,8 @@ asmlinkage __visible __trap_section void do_trap_ecall_u(struct pt_regs *regs)
 		regs->epc += 4;
 		regs->orig_a0 = regs->a0;
 
+		riscv_v_vstate_discard(regs);
+
 		syscall = syscall_enter_from_user_mode(regs, syscall);
 
 		if (syscall < NR_syscalls)
Quoted 1 line

base-commit: 4681dacadeefa5ca6017e00736adc1d7dc963c6a

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

#2 lore raw
Quoted 39 lines · Palmer Dabbelt
Quoted 19 lines · On Thu, 22 Jun 2023 10:36:13 PDT (-0700), [email protected] wrote:

From: Björn Töpel <[email protected]>

The RISC-V vector specification states:
Executing a system call causes all caller-saved vector registers (v0-v31, vl, vtype) and vstart to become unspecified.

The vector registers are cleared, vill is set (invalid), and the vector status is set to Initial.

That way we can prevent userspace from accidentally relying on the stated save.

Rémi pointed out [1] that clearing the registers might be superfluous, and setting vill is sufficient.

Quoted 7 lines

---

I figured I'd sent out a proper patch. I like Andy's optimization patch, but TBH I think we should do that as a follow up.

As Rémi pointed out, the clearing might be opted out, but I left it in here.

I think we're going to end up with a bunch of uarch-specific stuff here, but for now having the heavy hammer seems safest. There's no V hardware yet so we can't really tell how anything performs, at least this way we're definately not leaking anything.

So I'm OK with this. I'd also be fine with clearing to all-1s, I think it's kind of splitting hairs at this point: the 1s are nice because they're what the rest of V does, but setting vill should make everything trap anyway so maybe it doesn't matter -- and it's not clear if 1 or 0 will allow initial, so who knows.

Quoted 3 lines

Darius: I'm cool swapping over to the 1s if you feel strongly about it. Bjorn says Sweeden is on vacation, so just LMK and I'll re-spin it with the 1s.

I think all 1s would be preferred, but I don't think it's particularly critical (splitting hairs, like you said), so I'll let you make the call.

Quoted 69 lines

Regardless I'd like to pick up something that blows away V state for this merge window, as it'll make sure the uABI is quite strictly enforced.

Quoted 65 lines

Björn

---
 arch/riscv/include/asm/vector.h | 25 +++++++++++++++++++++++++
 arch/riscv/kernel/traps.c       |  2 ++
 2 files changed, 27 insertions(+)
arch/riscv/include/asm/vector.h +25 -0
index 04c0b07bf6cd..692ce55e4a69 100644
@@ -163,6 +163,30 @@ static inline void __switch_to_vector(struct task_struct *prev,
 void riscv_v_vstate_ctrl_init(struct task_struct *tsk);
 bool riscv_v_vstate_ctrl_user_allowed(void);
 
+static inline void riscv_v_vstate_discard(struct pt_regs *regs)
+{
+	unsigned long vl, vtype_inval = 1UL << (BITS_PER_LONG - 1);
+
+	if (!riscv_v_vstate_query(regs))
+		return;
+
+	riscv_v_enable();
+	asm volatile (
+		".option push\n\t"
+		".option arch, +v\n\t"
+		"vsetvli	%0, x0, e8, m8, ta, ma\n\t"
+		"vmv.v.i	v0, 0\n\t"
+		"vmv.v.i	v8, 0\n\t"
+		"vmv.v.i	v16, 0\n\t"
+		"vmv.v.i	v24, 0\n\t"
+		"vsetvl		%0, x0, %1\n\t"
+		".option pop\n\t"
+		: "=&r" (vl) : "r" (vtype_inval) : "memory");
+	riscv_v_disable();
+
+	riscv_v_vstate_on(regs);
+}
+
 #else /* ! CONFIG_RISCV_ISA_V  */
 
 struct pt_regs;
@@ -178,6 +202,7 @@ static inline bool riscv_v_vstate_ctrl_user_allowed(void) { return false; }
 #define __switch_to_vector(__prev, __next)	do {} while (0)
 #define riscv_v_vstate_off(regs)		do {} while (0)
 #define riscv_v_vstate_on(regs)			do {} while (0)
+#define riscv_v_vstate_discard(regs)		do {} while (0)
 
 #endif /* CONFIG_RISCV_ISA_V */
 
arch/riscv/kernel/traps.c +2 -0
index 05ffdcd1424e..00c68b57ff88 100644
@@ -295,6 +295,8 @@ asmlinkage __visible __trap_section void do_trap_ecall_u(struct pt_regs *regs)
 		regs->epc += 4;
 		regs->orig_a0 = regs->a0;
 
+		riscv_v_vstate_discard(regs);
+
 		syscall = syscall_enter_from_user_mode(regs, syscall);
 
 		if (syscall < NR_syscalls)

base-commit: 4681dacadeefa5ca6017e00736adc1d7dc963c6a

maintainer
#3 lore raw
Quoted 40 lines · Darius Rad
Quoted 39 lines · Palmer Dabbelt
Quoted 19 lines · On Thu, 22 Jun 2023 10:36:13 PDT (-0700), [email protected] wrote:

From: Björn Töpel <[email protected]>

The RISC-V vector specification states:
Executing a system call causes all caller-saved vector registers (v0-v31, vl, vtype) and vstart to become unspecified.

The vector registers are cleared, vill is set (invalid), and the vector status is set to Initial.

That way we can prevent userspace from accidentally relying on the stated save.

Rémi pointed out [1] that clearing the registers might be superfluous, and setting vill is sufficient.

Quoted 7 lines

---

I figured I'd sent out a proper patch. I like Andy's optimization patch, but TBH I think we should do that as a follow up.

As Rémi pointed out, the clearing might be opted out, but I left it in here.

I think we're going to end up with a bunch of uarch-specific stuff here, but for now having the heavy hammer seems safest. There's no V hardware yet so we can't really tell how anything performs, at least this way we're definately not leaking anything.

So I'm OK with this. I'd also be fine with clearing to all-1s, I think it's kind of splitting hairs at this point: the 1s are nice because they're what the rest of V does, but setting vill should make everything trap anyway so maybe it doesn't matter -- and it's not clear if 1 or 0 will allow initial, so who knows.

Quoted 6 lines
Quoted 3 lines

Darius: I'm cool swapping over to the 1s if you feel strongly about it. Bjorn says Sweeden is on vacation, so just LMK and I'll re-spin it with the 1s.

I think all 1s would be preferred, but I don't think it's particularly critical (splitting hairs, like you said), so I'll let you make the call.

OK, I'm just going to take this then as it's the less work option ;)

Quoted 76 lines
Quoted 70 lines

Regardless I'd like to pick up something that blows away V state for this merge window, as it'll make sure the uABI is quite strictly enforced.

Quoted 65 lines

Björn

---
 arch/riscv/include/asm/vector.h | 25 +++++++++++++++++++++++++
 arch/riscv/kernel/traps.c       |  2 ++
 2 files changed, 27 insertions(+)
arch/riscv/include/asm/vector.h +25 -0
index 04c0b07bf6cd..692ce55e4a69 100644
@@ -163,6 +163,30 @@ static inline void __switch_to_vector(struct task_struct *prev,
 void riscv_v_vstate_ctrl_init(struct task_struct *tsk);
 bool riscv_v_vstate_ctrl_user_allowed(void);
 
+static inline void riscv_v_vstate_discard(struct pt_regs *regs)
+{
+	unsigned long vl, vtype_inval = 1UL << (BITS_PER_LONG - 1);
+
+	if (!riscv_v_vstate_query(regs))
+		return;
+
+	riscv_v_enable();
+	asm volatile (
+		".option push\n\t"
+		".option arch, +v\n\t"
+		"vsetvli	%0, x0, e8, m8, ta, ma\n\t"
+		"vmv.v.i	v0, 0\n\t"
+		"vmv.v.i	v8, 0\n\t"
+		"vmv.v.i	v16, 0\n\t"
+		"vmv.v.i	v24, 0\n\t"
+		"vsetvl		%0, x0, %1\n\t"
+		".option pop\n\t"
+		: "=&r" (vl) : "r" (vtype_inval) : "memory");
+	riscv_v_disable();
+
+	riscv_v_vstate_on(regs);
+}
+
 #else /* ! CONFIG_RISCV_ISA_V  */
 
 struct pt_regs;
@@ -178,6 +202,7 @@ static inline bool riscv_v_vstate_ctrl_user_allowed(void) { return false; }
 #define __switch_to_vector(__prev, __next)	do {} while (0)
 #define riscv_v_vstate_off(regs)		do {} while (0)
 #define riscv_v_vstate_on(regs)			do {} while (0)
+#define riscv_v_vstate_discard(regs)		do {} while (0)
 
 #endif /* CONFIG_RISCV_ISA_V */
 
arch/riscv/kernel/traps.c +2 -0
index 05ffdcd1424e..00c68b57ff88 100644
@@ -295,6 +295,8 @@ asmlinkage __visible __trap_section void do_trap_ecall_u(struct pt_regs *regs)
 		regs->epc += 4;
 		regs->orig_a0 = regs->a0;
 
+		riscv_v_vstate_discard(regs);
+
 		syscall = syscall_enter_from_user_mode(regs, syscall);
 
 		if (syscall < NR_syscalls)

base-commit: 4681dacadeefa5ca6017e00736adc1d7dc963c6a

Signature · 4 lines
You received this message because you are subscribed to the Google Groups "linux" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
To view this discussion on the web visit https://groups.google.com/a/rivosinc.com/d/msgid/linux/ZJSLKrB/4xaYB75d%40bruce.bluespec.com.
For more options, visit https://groups.google.com/a/rivosinc.com/d/optout.
#4 lore raw
In reply to Björn Töpel
Quoted 5 lines · Björn Töpel

From: Björn Töpel <[email protected]>

The RISC-V vector specification states:
Executing a system call causes all caller-saved vector registers (v0-v31, vl, vtype) and vstart to become unspecified.

So the only use case of saving/restoring V regs in __switch_to is an interrupt induced task switch ?

Quoted 88 lines

The vector registers are cleared, vill is set (invalid), and the vector status is set to Initial.

That way we can prevent userspace from accidentally relying on the stated save.

Rémi pointed out [1] that clearing the registers might be superfluous, and setting vill is sufficient.

---

I figured I'd sent out a proper patch. I like Andy's optimization patch, but TBH I think we should do that as a follow up.

As Rémi pointed out, the clearing might be opted out, but I left it in here.

Björn

---
  arch/riscv/include/asm/vector.h | 25 +++++++++++++++++++++++++
  arch/riscv/kernel/traps.c       |  2 ++
  2 files changed, 27 insertions(+)
arch/riscv/include/asm/vector.h +25 -0
index 04c0b07bf6cd..692ce55e4a69 100644
@@ -163,6 +163,30 @@ static inline void __switch_to_vector(struct task_struct *prev,
  void riscv_v_vstate_ctrl_init(struct task_struct *tsk);
  bool riscv_v_vstate_ctrl_user_allowed(void);
 
+static inline void riscv_v_vstate_discard(struct pt_regs *regs)
+{
+	unsigned long vl, vtype_inval = 1UL << (BITS_PER_LONG - 1);
+
+	if (!riscv_v_vstate_query(regs))
+		return;
+
+	riscv_v_enable();
+	asm volatile (
+		".option push\n\t"
+		".option arch, +v\n\t"
+		"vsetvli	%0, x0, e8, m8, ta, ma\n\t"
+		"vmv.v.i	v0, 0\n\t"
+		"vmv.v.i	v8, 0\n\t"
+		"vmv.v.i	v16, 0\n\t"
+		"vmv.v.i	v24, 0\n\t"
+		"vsetvl		%0, x0, %1\n\t"
+		".option pop\n\t"
+		: "=&r" (vl) : "r" (vtype_inval) : "memory");
+	riscv_v_disable();
+
+	riscv_v_vstate_on(regs);
+}
+
  #else /* ! CONFIG_RISCV_ISA_V  */
 
  struct pt_regs;
@@ -178,6 +202,7 @@ static inline bool riscv_v_vstate_ctrl_user_allowed(void) { return false; }
  #define __switch_to_vector(__prev, __next)	do {} while (0)
  #define riscv_v_vstate_off(regs)		do {} while (0)
  #define riscv_v_vstate_on(regs)			do {} while (0)
+#define riscv_v_vstate_discard(regs)		do {} while (0)
 
  #endif /* CONFIG_RISCV_ISA_V */
 
arch/riscv/kernel/traps.c +2 -0
index 05ffdcd1424e..00c68b57ff88 100644
@@ -295,6 +295,8 @@ asmlinkage __visible __trap_section void do_trap_ecall_u(struct pt_regs *regs)
  		regs->epc += 4;
  		regs->orig_a0 = regs->a0;
 
+		riscv_v_vstate_discard(regs);
+
  		syscall = syscall_enter_from_user_mode(regs, syscall);
 
  		if (syscall < NR_syscalls)

base-commit: 4681dacadeefa5ca6017e00736adc1d7dc963c6a

#5 lore raw
Quoted 9 lines · Vineet Gupta
Quoted 5 lines · Björn Töpel

From: Björn Töpel <[email protected]>

The RISC-V vector specification states:
Executing a system call causes all caller-saved vector registers (v0-v31, vl, vtype) and vstart to become unspecified.

So the only use case of saving/restoring V regs in __switch_to is an interrupt induced task switch ?

Yes!

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv