tile: fast-path unaligned memory access for tilegx

This change enables unaligned userspace memory access via a kernel
fast path on tilegx.  The kernel tracks user PC/instruction pairs
per-thread using a direct-mapped cache in userspace.  The cache
maps those PC/instruction pairs to JIT'ed instruction sequences that
load or store using byte-wide load store intructions and then
synthesize 2-, 4- or 8-byte load or store results.  Once an
instruction has been seen to generate an unaligned access once,
subsequent hits on that instruction typically require overhead
of only around 50 cycles if cache and TLB is hot.

We support the prctl() PR_GET_UNALIGN / PR_SET_UNALIGN sys call to
enable or disable unaligned fixups on a per-process basis.

To do this we pull some of the tilepro unaligned support out of the
single_step.c file; tilepro uses instruction disassembly for both
single-step and unaligned access support.  Since tilegx actually has
hardware singlestep support, though, it's cleaner to keep the tilegx
unaligned access code in a separate file.  While we're at it,
properly rename the tilepro-specific types, etc., to have tilepro
suffixes instead of generic tile suffixes.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

1 files changed, 63 insertions, 53 deletions

diff --git a/arch/tile/kernel/single_step.c b/arch/tile/kernel/single_step.c
index 27742e87e255..5ef2e9eae5c5 100644
--- a/arch/tile/kernel/single_step.c
+++ b/arch/tile/kernel/single_step.c
@@ -12,41 +12,30 @@
  *   more details.
  *
  * A code-rewriter that enables instruction single-stepping.
- * Derived from iLib's single-stepping code.
  */
 
-#ifndef __tilegx__   /* Hardware support for single step unavailable. */
-
-/* These functions are only used on the TILE platform */
+#include <linux/smp.h>
+#include <linux/ptrace.h>
 #include <linux/slab.h>
 #include <linux/thread_info.h>
 #include <linux/uaccess.h>
 #include <linux/mman.h>
 #include <linux/types.h>
 #include <linux/err.h>
+#include <linux/prctl.h>
 #include <asm/cacheflush.h>
+#include <asm/traps.h>
+#include <asm/uaccess.h>
 #include <asm/unaligned.h>
 #include <arch/abi.h>
+#include <arch/spr_def.h>
 #include <arch/opcode.h>
 
-#define signExtend17(val) sign_extend((val), 17)
-#define TILE_X1_MASK (0xffffffffULL << 31)
-
-int unaligned_printk;
 
-static int __init setup_unaligned_printk(char *str)
-{
-	long val;
-	if (strict_strtol(str, 0, &val) != 0)
-		return 0;
-	unaligned_printk = val;
-	pr_info("Printk for each unaligned data accesses is %s\n",
-		unaligned_printk ? "enabled" : "disabled");
-	return 1;
-}
-__setup("unaligned_printk=", setup_unaligned_printk);
+#ifndef __tilegx__   /* Hardware support for single step unavailable. */
 
-unsigned int unaligned_fixup_count;
+#define signExtend17(val) sign_extend((val), 17)
+#define TILE_X1_MASK (0xffffffffULL << 31)
 
 enum mem_op {
 	MEMOP_NONE,
@@ -56,12 +45,13 @@ enum mem_op {
 	MEMOP_STORE_POSTINCR
 };
 
-static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, s32 offset)
+static inline tilepro_bundle_bits set_BrOff_X1(tilepro_bundle_bits n,
+	s32 offset)
 {
-	tile_bundle_bits result;
+	tilepro_bundle_bits result;
 
 	/* mask out the old offset */
-	tile_bundle_bits mask = create_BrOff_X1(-1);
+	tilepro_bundle_bits mask = create_BrOff_X1(-1);
 	result = n & (~mask);
 
 	/* or in the new offset */
@@ -70,10 +60,11 @@ static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, s32 offset)
 	return result;
 }
 
-static inline tile_bundle_bits move_X1(tile_bundle_bits n, int dest, int src)
+static inline tilepro_bundle_bits move_X1(tilepro_bundle_bits n, int dest,
+	int src)
 {
-	tile_bundle_bits result;
-	tile_bundle_bits op;
+	tilepro_bundle_bits result;
+	tilepro_bundle_bits op;
 
 	result = n & (~TILE_X1_MASK);
 
@@ -87,13 +78,13 @@ static inline tile_bundle_bits move_X1(tile_bundle_bits n, int dest, int src)
 	return result;
 }
 
-static inline tile_bundle_bits nop_X1(tile_bundle_bits n)
+static inline tilepro_bundle_bits nop_X1(tilepro_bundle_bits n)
 {
 	return move_X1(n, TREG_ZERO, TREG_ZERO);
 }
 
-static inline tile_bundle_bits addi_X1(
-	tile_bundle_bits n, int dest, int src, int imm)
+static inline tilepro_bundle_bits addi_X1(
+	tilepro_bundle_bits n, int dest, int src, int imm)
 {
 	n &= ~TILE_X1_MASK;
 
@@ -107,15 +98,26 @@ static inline tile_bundle_bits addi_X1(
 	return n;
 }
 
-static tile_bundle_bits rewrite_load_store_unaligned(
+static tilepro_bundle_bits rewrite_load_store_unaligned(
 	struct single_step_state *state,
-	tile_bundle_bits bundle,
+	tilepro_bundle_bits bundle,
 	struct pt_regs *regs,
 	enum mem_op mem_op,
 	int size, int sign_ext)
 {
 	unsigned char __user *addr;
 	int val_reg, addr_reg, err, val;
+	int align_ctl;
+
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
 
 	/* Get address and value registers */
 	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
@@ -160,7 +162,7 @@ static tile_bundle_bits rewrite_load_store_unaligned(
 	 * tilepro hardware would be doing, if it could provide us with the
 	 * actual bad address in an SPR, which it doesn't.
 	 */
-	if (unaligned_fixup == 0) {
+	if (align_ctl == 0) {
 		siginfo_t info = {
 			.si_signo = SIGBUS,
 			.si_code = BUS_ADRALN,
@@ -209,14 +211,14 @@ static tile_bundle_bits rewrite_load_store_unaligned(
 
 	if (err) {
 		siginfo_t info = {
-			.si_signo = SIGSEGV,
-			.si_code = SEGV_MAPERR,
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
 			.si_addr = addr
 		};
-		trace_unhandled_signal("segfault", regs,
-				       (unsigned long)addr, SIGSEGV);
+		trace_unhandled_signal("bad address for unaligned fixup", regs,
+				       (unsigned long)addr, SIGBUS);
 		force_sig_info(info.si_signo, &info, current);
-		return (tile_bundle_bits) 0;
+		return (tilepro_bundle_bits) 0;
 	}
 
 	if (unaligned_printk || unaligned_fixup_count == 0) {
@@ -285,7 +287,7 @@ void single_step_execve(void)
 	ti->step_state = NULL;
 }
 
-/**
+/*
  * single_step_once() - entry point when single stepping has been triggered.
  * @regs: The machine register state
  *
@@ -304,20 +306,31 @@ void single_step_execve(void)
  */
 void single_step_once(struct pt_regs *regs)
 {
-	extern tile_bundle_bits __single_step_ill_insn;
-	extern tile_bundle_bits __single_step_j_insn;
-	extern tile_bundle_bits __single_step_addli_insn;
-	extern tile_bundle_bits __single_step_auli_insn;
+	extern tilepro_bundle_bits __single_step_ill_insn;
+	extern tilepro_bundle_bits __single_step_j_insn;
+	extern tilepro_bundle_bits __single_step_addli_insn;
+	extern tilepro_bundle_bits __single_step_auli_insn;
 	struct thread_info *info = (void *)current_thread_info();
 	struct single_step_state *state = info->step_state;
 	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
-	tile_bundle_bits __user *buffer, *pc;
-	tile_bundle_bits bundle;
+	tilepro_bundle_bits __user *buffer, *pc;
+	tilepro_bundle_bits bundle;
 	int temp_reg;
 	int target_reg = TREG_LR;
 	int err;
 	enum mem_op mem_op = MEMOP_NONE;
 	int size = 0, sign_ext = 0;  /* happy compiler */
+	int align_ctl;
+
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
 
 	asm(
 "    .pushsection .rodata.single_step\n"
@@ -390,7 +403,7 @@ void single_step_once(struct pt_regs *regs)
 	if (regs->faultnum == INT_SWINT_1)
 		regs->pc -= 8;
 
-	pc = (tile_bundle_bits __user *)(regs->pc);
+	pc = (tilepro_bundle_bits __user *)(regs->pc);
 	if (get_user(bundle, pc) != 0) {
 		pr_err("Couldn't read instruction at %p trying to step\n", pc);
 		return;
@@ -627,9 +640,9 @@ void single_step_once(struct pt_regs *regs)
 
 	/*
 	 * Check if we need to rewrite an unaligned load/store.
-	 * Returning zero is a special value meaning we need to SIGSEGV.
+	 * Returning zero is a special value meaning we generated a signal.
 	 */
-	if (mem_op != MEMOP_NONE && unaligned_fixup >= 0) {
+	if (mem_op != MEMOP_NONE && align_ctl >= 0) {
 		bundle = rewrite_load_store_unaligned(state, bundle, regs,
 						      mem_op, size, sign_ext);
 		if (bundle == 0)
@@ -668,9 +681,9 @@ void single_step_once(struct pt_regs *regs)
 		}
 
 		/* End with a jump back to the next instruction */
-		delta = ((regs->pc + TILE_BUNDLE_SIZE_IN_BYTES) -
+		delta = ((regs->pc + TILEPRO_BUNDLE_SIZE_IN_BYTES) -
 			(unsigned long)buffer) >>
-			TILE_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
+			TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
 		bundle = __single_step_j_insn;
 		bundle |= create_JOffLong_X1(delta);
 		err |= __put_user(bundle, buffer++);
@@ -698,9 +711,6 @@ void single_step_once(struct pt_regs *regs)
 }
 
 #else
-#include <linux/smp.h>
-#include <linux/ptrace.h>
-#include <arch/spr_def.h>
 
 static DEFINE_PER_CPU(unsigned long, ss_saved_pc);
 
@@ -743,10 +753,10 @@ void gx_singlestep_handle(struct pt_regs *regs, int fault_num)
 	} else if ((*ss_pc != regs->pc) ||
 		   (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) {
 
-		ptrace_notify(SIGTRAP);
 		control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
 		control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
 		__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
+		send_sigtrap(current, regs);
 	}
 }