[PATCH] riscv: vector: treat VS_INITIAL as discard

[Andy Chiu]

The purpose of riscv_v_vstate_discard() is to invalidate v context at
entries of syscalls. So users happen to use v after a syscall without
re-configuring would see a failure. It was achieved by setting vector
registers and CSRs to -1 and marking the context busy. However, this
results in redundant saving of v-context if the process is scheduled out
in a syscall. Moreover, restoring the invalidated context from memory is
a costly operation. In fact, all can be prevented if we can delay
vstate_discard before returning back to the user space. To be more
specific, the kernel can mark v-context as INITIAL and set the restore
flag at syscall entries. This is the indication for the vstate_restore,
so it awares that the vstate has to be invalidated before returning back
to the user space.

After applying this patch, the context switch performance has improved
6.78% on vector enabled lmbench running on a FPGA with VLEN=512. The
result was obtained by averaging the output from the following command.

$ lat_ctx 2
Before the patch: 599.8357692
After the patch: 559.1748148

Signed-off-by: Andy Chiu <andy.chiu at sifive.com>
---
 arch/riscv/include/asm/vector.h | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/arch/riscv/include/asm/vector.h b/arch/riscv/include/asm/vector.h
index be7d309cca8a..1221f8b8f564 100644
--- a/arch/riscv/include/asm/vector.h
+++ b/arch/riscv/include/asm/vector.h
@@ -170,8 +170,8 @@ static inline void riscv_v_vstate_discard(struct pt_regs *regs)
 	if ((regs->status & SR_VS) == SR_VS_OFF)
 		return;
 
-	__riscv_v_vstate_discard();
-	__riscv_v_vstate_dirty(regs);
+	set_tsk_thread_flag(current, TIF_RISCV_V_DEFER_RESTORE);
+	riscv_v_vstate_on(regs);
 }
 
 static inline void riscv_v_vstate_save(struct __riscv_v_ext_state *vstate,
@@ -186,7 +186,9 @@ static inline void riscv_v_vstate_save(struct __riscv_v_ext_state *vstate,
 static inline void riscv_v_vstate_restore(struct __riscv_v_ext_state *vstate,
 					  struct pt_regs *regs)
 {
-	if ((regs->status & SR_VS) != SR_VS_OFF) {
+	if ((regs->status & SR_VS) == SR_VS_INITIAL) {
+		__riscv_v_vstate_discard();
+	} else if ((regs->status & SR_VS) != SR_VS_OFF) {
 		__riscv_v_vstate_restore(vstate, vstate->datap);
 		__riscv_v_vstate_clean(regs);
 	}
@@ -197,7 +199,7 @@ static inline void riscv_v_vstate_set_restore(struct task_struct *task,
 {
 	if ((regs->status & SR_VS) != SR_VS_OFF) {
 		set_tsk_thread_flag(task, TIF_RISCV_V_DEFER_RESTORE);
-		riscv_v_vstate_on(regs);
+		__riscv_v_vstate_clean(regs);
 	}
 }
 

---
base-commit: 60a6707f582ebbdfb6b378f45d7bf929106a1cd5
change-id: 20240628-dev-vstate_discard-5784652968a3

Best regards,
-- 
Andy Chiu <andy.chiu at sifive.com>