~700k ipc requests

+ Eliminated need for save/load_regs, now the register save area is
  behind a pointer which _slightly_ increases overall syscall delay, but
  speeds up ipc requests by a fair bit.

6 files changed, 39 insertions, 28 deletions

diff --git a/arch/riscv64/asm/asm-offsets.c b/arch/riscv64/asm/asm-offsets.c
index 3619765..1239489 100644
--- a/arch/riscv64/asm/asm-offsets.c
+++ b/arch/riscv64/asm/asm-offsets.c
@@ -83,4 +83,9 @@ void asm_offsets()
 	OFFSETOF(ar3, struct sys_ret);
 	OFFSETOF(ar4, struct sys_ret);
 	SIZEOF(sys_ret, struct sys_ret);
+
+	OFFSETOF(exec, struct tcb);
+	OFFSETOF(regs, struct tcb);
+	OFFSETOF(tcbd, struct tcb);
+	SIZEOF(tcb, struct tcb);
 }
diff --git a/arch/riscv64/include/tcb.h b/arch/riscv64/include/tcb.h
index ed2fd71..829ce6a 100644
--- a/arch/riscv64/include/tcb.h
+++ b/arch/riscv64/include/tcb.h
@@ -16,6 +16,7 @@
  * saving of something
  */
 struct arch_tcbd {
+	long scratch;
 };
 
 #endif /* ARCH_RISCV_TCB_H */
diff --git a/arch/riscv64/kernel/entry.S b/arch/riscv64/kernel/entry.S
index 355cd3d..f099065 100644
--- a/arch/riscv64/kernel/entry.S
+++ b/arch/riscv64/kernel/entry.S
@@ -23,13 +23,9 @@ handle_irq:
 	 * now just spin in place */
 	1: j 1b
 _save_context:
-	/* store user stack pointer and replace it with kernel stack */
-	addi	tp, tp,-sizeof_registers
-	sr	sp, offsetof_sp(tp)
-	mv	sp, tp
-
-	/* move thread pointer back */
-	addi	tp, tp, sizeof_registers
+	sr	sp, offsetof_tcbd(tp)
+	lr	sp, offsetof_regs(tp)
+	addi	sp, sp, -sizeof_registers
 	csrrw	tp, CSR_SSCRATCH, tp
 
 	/* save registers */
@@ -67,6 +63,9 @@ _save_context:
 	/* get current tcb into tp and set scratch to 0 so we can figure out if
 	 * exception occured in kernel or userspace */
 	csrrw tp, CSR_SSCRATCH, x0
+	lr	s0, offsetof_tcbd(tp)
+	sr	s0, offsetof_sp(sp)
+	addi	sp, tp, -sizeof_registers
 
 	/* load supervisor cause */
 	csrr s4, CSR_SCAUSE
@@ -85,7 +84,7 @@ handle_exception:
 handle_dispatch:
 	/* store execution continuation point */
 	csrr    s0, CSR_SEPC
-	sr	s0, 0(tp)
+	sr	s0, offsetof_exec(tp)
 	/* allocate space for sys_ret structure on stack and shift argument
 	 * registers down one to make room for the pointer to this structure */
 	addi	sp, sp, -sizeof_sys_ret
@@ -109,10 +108,10 @@ handle_dispatch:
 	/* if we had a thread switch, load kernel stack of current thread and
 	 * restore its context */
 	/* get associated kernel stack */
-	mv	sp, tp
+	lr	sp, offsetof_regs(tp)
 	addi	sp, sp, -sizeof_registers
 	/* set execution continuation */
-	lr	s0, 0(tp)
+	lr	s0, offsetof_exec(tp)
 	csrw	CSR_SEPC, s0
 	/* restore system call result */
 	j	restore_noreturn
diff --git a/arch/riscv64/kernel/proc.c b/arch/riscv64/kernel/proc.c
index b8f6c5d..4180bf8 100644
--- a/arch/riscv64/kernel/proc.c
+++ b/arch/riscv64/kernel/proc.c
@@ -30,7 +30,7 @@ void run_init(struct tcb *t, void *fdt)
 
 void set_args(struct tcb *t, struct sys_ret a)
 {
-	struct riscv_regs *r = (struct riscv_regs *)(t) - 1;
+	struct riscv_regs *r = (struct riscv_regs *)(t->regs) - 1;
 	r->a0 = a.s;
 	r->a1 = a.ar0;
 	r->a2 = a.ar1;
@@ -41,7 +41,7 @@ void set_args(struct tcb *t, struct sys_ret a)
 
 struct sys_ret get_args(struct tcb *t)
 {
-	struct riscv_regs *r = (struct riscv_regs *)(t) - 1;
+	struct riscv_regs *r = (struct riscv_regs *)(t->regs) - 1;
 	return SYS_RET6(r->a0, r->a1, r->a2, r->a3, r->a4, r->a5);
 }
 
@@ -49,7 +49,7 @@ void set_thread(struct tcb *t)
 {
 	/* get location of registers in memory */
 	/** \todo check alignment, should be fine but just to be sure */
-	struct riscv_regs *r = (struct riscv_regs *)(t) - 1;
+	struct riscv_regs *r = (struct riscv_regs *)(t->regs) - 1;
 
 	/* insert important values into register slots */
 	r->sp = (long)t->thread_stack_top;
@@ -58,28 +58,28 @@ void set_thread(struct tcb *t)
 
 vm_t get_stack(struct tcb *t)
 {
-	struct riscv_regs *r = (struct riscv_regs *)(t) - 1;
+	struct riscv_regs *r = (struct riscv_regs *)(t->regs) - 1;
 	return r->sp;
 }
 
 void save_regs(struct tcb *t, void *p)
 {
-	struct riscv_regs *r = (struct riscv_regs *)(t) - 1;
-	struct riscv_regs *rp = (struct riscv_regs *)(p);
+	struct riscv_regs *r = (struct riscv_regs *)(t->regs) - 1;
+	struct riscv_regs *rp = (struct riscv_regs *)(p) - 1;
 	*rp = *r;
 }
 
 void load_regs(void *p, struct tcb *t)
 {
-	struct riscv_regs *r = (struct riscv_regs *)(t) - 1;
-	struct riscv_regs *rp = (struct riscv_regs *)(p);
+	struct riscv_regs *r = (struct riscv_regs *)(t->regs) - 1;
+	struct riscv_regs *rp = (struct riscv_regs *)(p) - 1;
 	*r = *rp;
 }
 
 void clone_regs(struct tcb *d, struct tcb *s)
 {
-	struct riscv_regs *rd = (struct riscv_regs *)(d) - 1;
-	struct riscv_regs *rs = (struct riscv_regs *)(s) - 1;
+	struct riscv_regs *rd = (struct riscv_regs *)(d->regs) - 1;
+	struct riscv_regs *rs = (struct riscv_regs *)(s->regs) - 1;
 	*rd = *rs;
 }
 
diff --git a/common/tcb.c b/common/tcb.c
index 478bd05..f774fbc 100644
--- a/common/tcb.c
+++ b/common/tcb.c
@@ -163,6 +163,8 @@ struct tcb *create_thread(struct tcb *p)
 	t->rpc.vmem = create_vmem();
 	__setup_rpc_stack(t, __call_stack_size);
 
+	t->regs = (vm_t)t;
+
 	set_canary(t);
 	return t;
 }
@@ -180,7 +182,6 @@ static stat_t __copy_proc(struct tcb *p, struct tcb *n)
 	/** @todo I think keeping track of userspace stack stuff is unnecessary,
 	 * unless we want unlimited stack size but that sounds dumb. Anycase, we
 	 * need to duplicate stack info, whatever we do. */
-	/** @todo should there be in-kernel child tracking? */
 	n->exec = p->exec;
 	n->callback = p->callback;
 	n->thread_stack = p->thread_stack;
@@ -409,6 +410,7 @@ static void mark_rpc_accessible(struct tcb *t, vm_t start, vm_t end)
 
 struct call_ctx {
 	vm_t exec;
+	vm_t regs;
 	vm_t rpc_stack;
 	id_t eid, pid;
 };
@@ -420,27 +422,29 @@ void save_context(struct tcb *t)
 		/** @todo what if user uses their own stack? */
 		rpc_stack = align_down(get_stack(t), BASE_PAGE_SIZE);
 
-	rpc_stack -= BASE_PAGE_SIZE;
 
-	struct call_ctx *ctx = (struct call_ctx *)rpc_stack;
+	struct call_ctx *ctx = (struct call_ctx *)(rpc_stack) - 1;
 	ctx->exec = t->exec;
 	ctx->pid = t->pid;
 	ctx->eid = t->eid;
-	ctx->rpc_stack = rpc_stack + BASE_PAGE_SIZE;
-	save_regs(t, ctx + 1);
+	ctx->regs = t->regs;
+	ctx->rpc_stack = rpc_stack;
+
+	rpc_stack -= BASE_PAGE_SIZE;
 
 	mark_rpc_inaccessible(t, rpc_stack, t->rpc_stack);
 	t->rpc_stack = rpc_stack;
+	t->regs = (vm_t)ctx;
 }
 
 void load_context(struct tcb *t)
 {
-	vm_t rpc_stack = t->rpc_stack;
-	struct call_ctx *ctx = (struct call_ctx *)rpc_stack;
+	vm_t rpc_stack = t->rpc_stack + BASE_PAGE_SIZE;
+	struct call_ctx *ctx = (struct call_ctx *)(rpc_stack) - 1;
 	set_return(t, ctx->exec);
 	mark_rpc_accessible(t, t->rpc_stack, ctx->rpc_stack);
-	load_regs(ctx + 1, t);
 	t->rpc_stack = ctx->rpc_stack;
 	t->pid = ctx->pid;
 	t->eid = ctx->eid;
+	t->regs = ctx->regs;
 }
diff --git a/include/apos/tcb.h b/include/apos/tcb.h
index 9ed916f..23e6451 100644
--- a/include/apos/tcb.h
+++ b/include/apos/tcb.h
@@ -78,6 +78,8 @@ struct tcb {
 	/** Execution continuation point. Important that it is first. */
 	vm_t exec;
 
+	vm_t regs;
+
 	/** Arch-specific data. */
 	struct arch_tcbd tcbd;