sh: Move over sh64 switch_to and stack unwinder.

Signed-off-by: Paul Mundt <lethal@linux-sh.org>

3 files changed, 527 insertions, 1 deletions

diff --git a/arch/sh/kernel/cpu/sh5/Makefile b/arch/sh/kernel/cpu/sh5/Makefile
index 9778f9b..7ea1073 100644
--- a/arch/sh/kernel/cpu/sh5/Makefile
+++ b/arch/sh/kernel/cpu/sh5/Makefile
@@ -1 +1,3 @@
-obj-y := entry.o
+obj-y := entry.o switchto.o
+
+obj-$(CONFIG_KALLSYMS)		+= unwind.o
diff --git a/arch/sh/kernel/cpu/sh5/switchto.S b/arch/sh/kernel/cpu/sh5/switchto.S
new file mode 100644
index 0000000..45b2d90
--- /dev/null
+++ b/arch/sh/kernel/cpu/sh5/switchto.S
@@ -0,0 +1,198 @@
+/*
+ * arch/sh64/kernel/switchto.S
+ *
+ * sh64 context switch
+ *
+ * Copyright (C) 2004  Richard Curnow
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+*/
+
+	.section .text..SHmedia32,"ax"
+	.little
+
+	.balign 32
+
+	.type sh64_switch_to,@function
+	.global sh64_switch_to
+	.global __sh64_switch_to_end
+sh64_switch_to:
+
+/* Incoming args
+   r2 - prev
+   r3 - &prev->thread
+   r4 - next
+   r5 - &next->thread
+
+   Outgoing results
+   r2 - last (=prev) : this just stays in r2 throughout
+
+   Want to create a full (struct pt_regs) on the stack to allow backtracing
+   functions to work.  However, we only need to populate the callee-save
+   register slots in this structure; since we're a function our ancestors must
+   have themselves preserved all caller saved state in the stack.  This saves
+   some wasted effort since we won't need to look at the values.
+
+   In particular, all caller-save registers are immediately available for
+   scratch use.
+
+*/
+
+#define FRAME_SIZE (76*8 + 8)
+
+	movi	FRAME_SIZE, r0
+	sub.l	r15, r0, r15
+	! Do normal-style register save to support backtrace
+
+	st.l	r15,   0, r18	! save link reg
+	st.l	r15,   4, r14	! save fp
+	add.l	r15, r63, r14	! setup frame pointer
+
+	! hopefully this looks normal to the backtrace now.
+
+	addi.l	r15,   8, r1    ! base of pt_regs
+	addi.l	r1,   24, r0    ! base of pt_regs.regs
+	addi.l	r0, (63*8), r8	! base of pt_regs.trregs
+
+	/* Note : to be fixed?
+	   struct pt_regs is really designed for holding the state on entry
+	   to an exception, i.e. pc,sr,regs etc.  However, for the context
+	   switch state, some of this is not required.  But the unwinder takes
+	   struct pt_regs * as an arg so we have to build this structure
+	   to allow unwinding switched tasks in show_state() */
+
+	st.q	r0, ( 9*8), r9
+	st.q	r0, (10*8), r10
+	st.q	r0, (11*8), r11
+	st.q	r0, (12*8), r12
+	st.q	r0, (13*8), r13
+	st.q	r0, (14*8), r14 ! for unwind, want to look as though we took a trap at
+	! the point where the process is left in suspended animation, i.e. current
+	! fp here, not the saved one.
+	st.q	r0, (16*8), r16
+
+	st.q	r0, (24*8), r24
+	st.q	r0, (25*8), r25
+	st.q	r0, (26*8), r26
+	st.q	r0, (27*8), r27
+	st.q	r0, (28*8), r28
+	st.q	r0, (29*8), r29
+	st.q	r0, (30*8), r30
+	st.q	r0, (31*8), r31
+	st.q	r0, (32*8), r32
+	st.q	r0, (33*8), r33
+	st.q	r0, (34*8), r34
+	st.q	r0, (35*8), r35
+
+	st.q	r0, (44*8), r44
+	st.q	r0, (45*8), r45
+	st.q	r0, (46*8), r46
+	st.q	r0, (47*8), r47
+	st.q	r0, (48*8), r48
+	st.q	r0, (49*8), r49
+	st.q	r0, (50*8), r50
+	st.q	r0, (51*8), r51
+	st.q	r0, (52*8), r52
+	st.q	r0, (53*8), r53
+	st.q	r0, (54*8), r54
+	st.q	r0, (55*8), r55
+	st.q	r0, (56*8), r56
+	st.q	r0, (57*8), r57
+	st.q	r0, (58*8), r58
+	st.q	r0, (59*8), r59
+
+	! do this early as pta->gettr has no pipeline forwarding (=> 5 cycle latency)
+	! Use a local label to avoid creating a symbol that will confuse the !
+	! backtrace
+	pta	.Lsave_pc, tr0
+
+	gettr	tr5, r45
+	gettr	tr6, r46
+	gettr	tr7, r47
+	st.q	r8, (5*8), r45
+	st.q	r8, (6*8), r46
+	st.q	r8, (7*8), r47
+
+	! Now switch context
+	gettr	tr0, r9
+	st.l	r3, 0, r15	! prev->thread.sp
+	st.l	r3, 8, r1	! prev->thread.kregs
+	st.l	r3, 4, r9	! prev->thread.pc
+	st.q	r1, 0, r9	! save prev->thread.pc into pt_regs->pc
+
+	! Load PC for next task (init value or save_pc later)
+	ld.l	r5, 4, r18	! next->thread.pc
+	! Switch stacks
+	ld.l	r5, 0, r15	! next->thread.sp
+	ptabs	r18, tr0
+
+	! Update current
+	ld.l	r4, 4, r9	! next->thread_info (2nd element of next task_struct)
+	putcon	r9, kcr0	! current = next->thread_info
+
+	! go to save_pc for a reschedule, or the initial thread.pc for a new process
+	blink	tr0, r63
+
+	! Restore (when we come back to a previously saved task)
+.Lsave_pc:
+	addi.l	r15, 32, r0	! r0 = next's regs
+	addi.l	r0, (63*8), r8	! r8 = next's tr_regs
+
+	ld.q	r8, (5*8), r45
+	ld.q	r8, (6*8), r46
+	ld.q	r8, (7*8), r47
+	ptabs	r45, tr5
+	ptabs	r46, tr6
+	ptabs	r47, tr7
+
+	ld.q	r0, ( 9*8), r9
+	ld.q	r0, (10*8), r10
+	ld.q	r0, (11*8), r11
+	ld.q	r0, (12*8), r12
+	ld.q	r0, (13*8), r13
+	ld.q	r0, (14*8), r14
+	ld.q	r0, (16*8), r16
+
+	ld.q	r0, (24*8), r24
+	ld.q	r0, (25*8), r25
+	ld.q	r0, (26*8), r26
+	ld.q	r0, (27*8), r27
+	ld.q	r0, (28*8), r28
+	ld.q	r0, (29*8), r29
+	ld.q	r0, (30*8), r30
+	ld.q	r0, (31*8), r31
+	ld.q	r0, (32*8), r32
+	ld.q	r0, (33*8), r33
+	ld.q	r0, (34*8), r34
+	ld.q	r0, (35*8), r35
+
+	ld.q	r0, (44*8), r44
+	ld.q	r0, (45*8), r45
+	ld.q	r0, (46*8), r46
+	ld.q	r0, (47*8), r47
+	ld.q	r0, (48*8), r48
+	ld.q	r0, (49*8), r49
+	ld.q	r0, (50*8), r50
+	ld.q	r0, (51*8), r51
+	ld.q	r0, (52*8), r52
+	ld.q	r0, (53*8), r53
+	ld.q	r0, (54*8), r54
+	ld.q	r0, (55*8), r55
+	ld.q	r0, (56*8), r56
+	ld.q	r0, (57*8), r57
+	ld.q	r0, (58*8), r58
+	ld.q	r0, (59*8), r59
+
+	! epilogue
+	ld.l	r15, 0, r18
+	ld.l	r15, 4, r14
+	ptabs	r18, tr0
+	movi	FRAME_SIZE, r0
+	add	r15, r0, r15
+	blink	tr0, r63
+__sh64_switch_to_end:
+.LFE1:
+	.size	sh64_switch_to,.LFE1-sh64_switch_to
+
diff --git a/arch/sh/kernel/cpu/sh5/unwind.c b/arch/sh/kernel/cpu/sh5/unwind.c
new file mode 100644
index 0000000..1214c78
--- /dev/null
+++ b/arch/sh/kernel/cpu/sh5/unwind.c
@@ -0,0 +1,326 @@
+/*
+ * arch/sh64/kernel/unwind.c
+ *
+ * Copyright (C) 2004  Paul Mundt
+ * Copyright (C) 2004  Richard Curnow
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ */
+#include <linux/kallsyms.h>
+#include <linux/kernel.h>
+#include <linux/types.h>
+#include <linux/errno.h>
+#include <asm/page.h>
+#include <asm/ptrace.h>
+#include <asm/processor.h>
+#include <asm/io.h>
+
+static u8 regcache[63];
+
+/*
+ * Finding the previous stack frame isn't horribly straightforward as it is
+ * on some other platforms. In the sh64 case, we don't have "linked" stack
+ * frames, so we need to do a bit of work to determine the previous frame,
+ * and in turn, the previous r14/r18 pair.
+ *
+ * There are generally a few cases which determine where we can find out
+ * the r14/r18 values. In the general case, this can be determined by poking
+ * around the prologue of the symbol PC is in (note that we absolutely must
+ * have frame pointer support as well as the kernel symbol table mapped,
+ * otherwise we can't even get this far).
+ *
+ * In other cases, such as the interrupt/exception path, we can poke around
+ * the sp/fp.
+ *
+ * Notably, this entire approach is somewhat error prone, and in the event
+ * that the previous frame cannot be determined, that's all we can do.
+ * Either way, this still leaves us with a more correct backtrace then what
+ * we would be able to come up with by walking the stack (which is garbage
+ * for anything beyond the first frame).
+ *						-- PFM.
+ */
+static int lookup_prev_stack_frame(unsigned long fp, unsigned long pc,
+		      unsigned long *pprev_fp, unsigned long *pprev_pc,
+		      struct pt_regs *regs)
+{
+	const char *sym;
+	char namebuf[128];
+	unsigned long offset;
+	unsigned long prologue = 0;
+	unsigned long fp_displacement = 0;
+	unsigned long fp_prev = 0;
+	unsigned long offset_r14 = 0, offset_r18 = 0;
+	int i, found_prologue_end = 0;
+
+	sym = kallsyms_lookup(pc, NULL, &offset, NULL, namebuf);
+	if (!sym)
+		return -EINVAL;
+
+	prologue = pc - offset;
+	if (!prologue)
+		return -EINVAL;
+
+	/* Validate fp, to avoid risk of dereferencing a bad pointer later.
+	   Assume 128Mb since that's the amount of RAM on a Cayman.  Modify
+	   when there is an SH-5 board with more. */
+	if ((fp < (unsigned long) phys_to_virt(__MEMORY_START)) ||
+	    (fp >= (unsigned long)(phys_to_virt(__MEMORY_START)) + 128*1024*1024) ||
+	    ((fp & 7) != 0)) {
+		return -EINVAL;
+	}
+
+	/*
+	 * Depth to walk, depth is completely arbitrary.
+	 */
+	for (i = 0; i < 100; i++, prologue += sizeof(unsigned long)) {
+		unsigned long op;
+		u8 major, minor;
+		u8 src, dest, disp;
+
+		op = *(unsigned long *)prologue;
+
+		major = (op >> 26) & 0x3f;
+		src   = (op >> 20) & 0x3f;
+		minor = (op >> 16) & 0xf;
+		disp  = (op >> 10) & 0x3f;
+		dest  = (op >>  4) & 0x3f;
+
+		/*
+		 * Stack frame creation happens in a number of ways.. in the
+		 * general case when the stack frame is less than 511 bytes,
+		 * it's generally created by an addi or addi.l:
+		 *
+		 *	addi/addi.l r15, -FRAME_SIZE, r15
+		 *
+		 * in the event that the frame size is bigger than this, it's
+		 * typically created using a movi/sub pair as follows:
+		 *
+		 *	movi	FRAME_SIZE, rX
+		 *	sub	r15, rX, r15
+		 */
+
+		switch (major) {
+		case (0x00 >> 2):
+			switch (minor) {
+			case 0x8: /* add.l */
+			case 0x9: /* add */
+				/* Look for r15, r63, r14 */
+				if (src == 15 && disp == 63 && dest == 14)
+					found_prologue_end = 1;
+
+				break;
+			case 0xa: /* sub.l */
+			case 0xb: /* sub */
+				if (src != 15 || dest != 15)
+					continue;
+
+				fp_displacement -= regcache[disp];
+				fp_prev = fp - fp_displacement;
+				break;
+			}
+			break;
+		case (0xa8 >> 2): /* st.l */
+			if (src != 15)
+				continue;
+
+			switch (dest) {
+			case 14:
+				if (offset_r14 || fp_displacement == 0)
+					continue;
+
+				offset_r14 = (u64)(((((s64)op >> 10) & 0x3ff) << 54) >> 54);
+				offset_r14 *= sizeof(unsigned long);
+				offset_r14 += fp_displacement;
+				break;
+			case 18:
+				if (offset_r18 || fp_displacement == 0)
+					continue;
+
+				offset_r18 = (u64)(((((s64)op >> 10) & 0x3ff) << 54) >> 54);
+				offset_r18 *= sizeof(unsigned long);
+				offset_r18 += fp_displacement;
+				break;
+			}
+
+			break;
+		case (0xcc >> 2): /* movi */
+			if (dest >= 63) {
+				printk(KERN_NOTICE "%s: Invalid dest reg %d "
+				       "specified in movi handler. Failed "
+				       "opcode was 0x%lx: ", __FUNCTION__,
+				       dest, op);
+
+				continue;
+			}
+
+			/* Sign extend */
+			regcache[dest] =
+				((((s64)(u64)op >> 10) & 0xffff) << 54) >> 54;
+			break;
+		case (0xd0 >> 2): /* addi */
+		case (0xd4 >> 2): /* addi.l */
+			/* Look for r15, -FRAME_SIZE, r15 */
+			if (src != 15 || dest != 15)
+				continue;
+
+			/* Sign extended frame size.. */
+			fp_displacement +=
+				(u64)(((((s64)op >> 10) & 0x3ff) << 54) >> 54);
+			fp_prev = fp - fp_displacement;
+			break;
+		}
+
+		if (found_prologue_end && offset_r14 && (offset_r18 || *pprev_pc) && fp_prev)
+			break;
+	}
+
+	if (offset_r14 == 0 || fp_prev == 0) {
+		if (!offset_r14)
+			pr_debug("Unable to find r14 offset\n");
+		if (!fp_prev)
+			pr_debug("Unable to find previous fp\n");
+
+		return -EINVAL;
+	}
+
+	/* For innermost leaf function, there might not be a offset_r18 */
+	if (!*pprev_pc && (offset_r18 == 0))
+		return -EINVAL;
+
+	*pprev_fp = *(unsigned long *)(fp_prev + offset_r14);
+
+	if (offset_r18)
+		*pprev_pc = *(unsigned long *)(fp_prev + offset_r18);
+
+	*pprev_pc &= ~1;
+
+	return 0;
+}
+
+/* Don't put this on the stack since we'll want to call sh64_unwind
+ * when we're close to underflowing the stack anyway. */
+static struct pt_regs here_regs;
+
+extern const char syscall_ret;
+extern const char ret_from_syscall;
+extern const char ret_from_exception;
+extern const char ret_from_irq;
+
+static void sh64_unwind_inner(struct pt_regs *regs);
+
+static void unwind_nested (unsigned long pc, unsigned long fp)
+{
+	if ((fp >= __MEMORY_START) &&
+	    ((fp & 7) == 0)) {
+		sh64_unwind_inner((struct pt_regs *) fp);
+	}
+}
+
+static void sh64_unwind_inner(struct pt_regs *regs)
+{
+	unsigned long pc, fp;
+	int ofs = 0;
+	int first_pass;
+
+	pc = regs->pc & ~1;
+	fp = regs->regs[14];
+
+	first_pass = 1;
+	for (;;) {
+		int cond;
+		unsigned long next_fp, next_pc;
+
+		if (pc == ((unsigned long) &syscall_ret & ~1)) {
+			printk("SYSCALL\n");
+			unwind_nested(pc,fp);
+			return;
+		}
+
+		if (pc == ((unsigned long) &ret_from_syscall & ~1)) {
+			printk("SYSCALL (PREEMPTED)\n");
+			unwind_nested(pc,fp);
+			return;
+		}
+
+		/* In this case, the PC is discovered by lookup_prev_stack_frame but
+		   it has 4 taken off it to look like the 'caller' */
+		if (pc == ((unsigned long) &ret_from_exception & ~1)) {
+			printk("EXCEPTION\n");
+			unwind_nested(pc,fp);
+			return;
+		}
+
+		if (pc == ((unsigned long) &ret_from_irq & ~1)) {
+			printk("IRQ\n");
+			unwind_nested(pc,fp);
+			return;
+		}
+
+		cond = ((pc >= __MEMORY_START) && (fp >= __MEMORY_START) &&
+			((pc & 3) == 0) && ((fp & 7) == 0));
+
+		pc -= ofs;
+
+		printk("[<%08lx>] ", pc);
+		print_symbol("%s\n", pc);
+
+		if (first_pass) {
+			/* If the innermost frame is a leaf function, it's
+			 * possible that r18 is never saved out to the stack.
+			 */
+			next_pc = regs->regs[18];
+		} else {
+			next_pc = 0;
+		}
+
+		if (lookup_prev_stack_frame(fp, pc, &next_fp, &next_pc, regs) == 0) {
+			ofs = sizeof(unsigned long);
+			pc = next_pc & ~1;
+			fp = next_fp;
+		} else {
+			printk("Unable to lookup previous stack frame\n");
+			break;
+		}
+		first_pass = 0;
+	}
+
+	printk("\n");
+
+}
+
+void sh64_unwind(struct pt_regs *regs)
+{
+	if (!regs) {
+		/*
+		 * Fetch current regs if we have no other saved state to back
+		 * trace from.
+		 */
+		regs = &here_regs;
+
+		__asm__ __volatile__ ("ori r14, 0, %0" : "=r" (regs->regs[14]));
+		__asm__ __volatile__ ("ori r15, 0, %0" : "=r" (regs->regs[15]));
+		__asm__ __volatile__ ("ori r18, 0, %0" : "=r" (regs->regs[18]));
+
+		__asm__ __volatile__ ("gettr tr0, %0" : "=r" (regs->tregs[0]));
+		__asm__ __volatile__ ("gettr tr1, %0" : "=r" (regs->tregs[1]));
+		__asm__ __volatile__ ("gettr tr2, %0" : "=r" (regs->tregs[2]));
+		__asm__ __volatile__ ("gettr tr3, %0" : "=r" (regs->tregs[3]));
+		__asm__ __volatile__ ("gettr tr4, %0" : "=r" (regs->tregs[4]));
+		__asm__ __volatile__ ("gettr tr5, %0" : "=r" (regs->tregs[5]));
+		__asm__ __volatile__ ("gettr tr6, %0" : "=r" (regs->tregs[6]));
+		__asm__ __volatile__ ("gettr tr7, %0" : "=r" (regs->tregs[7]));
+
+		__asm__ __volatile__ (
+			"pta 0f, tr0\n\t"
+			"blink tr0, %0\n\t"
+			"0: nop"
+			: "=r" (regs->pc)
+		);
+	}
+
+	printk("\nCall Trace:\n");
+	sh64_unwind_inner(regs);
+}
+