arch/tile: finish enabling support for TILE-Gx 64-bit chip

This support was partially present in the existing code (look for
"__tilegx__" ifdefs) but with this change you can build a working
kernel using the TILE-Gx toolchain and ARCH=tilegx.

Most of these files are new, generally adding a foo_64.c file
where previously there was just a foo_32.c file.

The ARCH=tilegx directive redirects to arch/tile, not arch/tilegx,
using the existing SRCARCH mechanism in the top-level Makefile.

Changes to existing files:

- <asm/bitops.h> and <asm/bitops_32.h> changed to factor the
  include of <asm-generic/bitops/non-atomic.h> in the common header.

- <asm/compat.h> and arch/tile/kernel/compat.c changed to remove
  the "const" markers I had put on compat_sys_execve() when trying
  to match some recent similar changes to the non-compat execve.
  It turns out the compat version wasn't "upgraded" to use const.

- <asm/opcode-tile_64.h> and <asm/opcode_constants_64.h> were
  previously included accidentally, with the 32-bit contents.  Now
  they have the proper 64-bit contents.

Finally, I had to hack the existing hacky drivers/input/input-compat.h
to add yet another "#ifdef" for INPUT_COMPAT_TEST (same as x86_64).

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
Acked-by: Dmitry Torokhov <dmitry.torokhov@gmail.com> [drivers/input]

6 files changed, 3902 insertions, 2 deletions

diff --git a/arch/tile/kernel/futex_64.S b/arch/tile/kernel/futex_64.S
new file mode 100644
index 00000000000..f465d1eda20
--- /dev/null
+++ b/arch/tile/kernel/futex_64.S
@@ -0,0 +1,55 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Atomically access user memory, but use MMU to avoid propagating
+ * kernel exceptions.
+ */
+
+#include <linux/linkage.h>
+#include <asm/errno.h>
+#include <asm/futex.h>
+#include <asm/page.h>
+#include <asm/processor.h>
+
+/*
+ * Provide a set of atomic memory operations supporting <asm/futex.h>.
+ *
+ * r0: user address to manipulate
+ * r1: new value to write, or for cmpxchg, old value to compare against
+ * r2: (cmpxchg only) new value to write
+ *
+ * Return __get_user struct, r0 with value, r1 with error.
+ */
+#define FUTEX_OP(name, ...) \
+STD_ENTRY(futex_##name)			\
+	__VA_ARGS__;			\
+	{				\
+	 move   r1, zero;		\
+	 jrp    lr			\
+	};				\
+	STD_ENDPROC(futex_##name);	\
+	.pushsection __ex_table,"a";	\
+	.quad 1b, get_user_fault;	\
+	.popsection
+
+	.pushsection .fixup,"ax"
+get_user_fault:
+	{ movei r1, -EFAULT; jrp lr }
+	ENDPROC(get_user_fault)
+	.popsection
+
+FUTEX_OP(cmpxchg, mtspr CMPEXCH_VALUE, r1; 1: cmpexch4 r0, r0, r2)
+FUTEX_OP(set, 1: exch4 r0, r0, r1)
+FUTEX_OP(add, 1: fetchadd4 r0, r0, r1)
+FUTEX_OP(or, 1: fetchor4 r0, r0, r1)
+FUTEX_OP(andn, nor r1, r1, zero; 1: fetchand4 r0, r0, r1)
diff --git a/arch/tile/kernel/head_64.S b/arch/tile/kernel/head_64.S
new file mode 100644
index 00000000000..6bc3a932fe4
--- /dev/null
+++ b/arch/tile/kernel/head_64.S
@@ -0,0 +1,269 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE startup code.
+ */
+
+#include <linux/linkage.h>
+#include <linux/init.h>
+#include <asm/page.h>
+#include <asm/pgtable.h>
+#include <asm/thread_info.h>
+#include <asm/processor.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+#include <arch/spr_def.h>
+
+/*
+ * This module contains the entry code for kernel images. It performs the
+ * minimal setup needed to call the generic C routines.
+ */
+
+	__HEAD
+ENTRY(_start)
+	/* Notify the hypervisor of what version of the API we want */
+	{
+	  movei r1, TILE_CHIP
+	  movei r2, TILE_CHIP_REV
+	}
+	{
+	  moveli r0, _HV_VERSION
+	  jal hv_init
+	}
+	/* Get a reasonable default ASID in r0 */
+	{
+	  move r0, zero
+	  jal hv_inquire_asid
+	}
+
+	/*
+	 * Install the default page table.  The relocation required to
+	 * statically define the table is a bit too complex, so we have
+	 * to plug in the pointer from the L0 to the L1 table by hand.
+	 * We only do this on the first cpu to boot, though, since the
+	 * other CPUs should see a properly-constructed page table.
+	 */
+	{
+	  v4int_l r2, zero, r0    /* ASID for hv_install_context */
+	  moveli r4, hw1_last(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  shl16insli r4, r4, hw0(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  ld r1, r4               /* access_pte for hv_install_context */
+	}
+	{
+	  moveli r0, hw1_last(.Lsv_data_pmd - PAGE_OFFSET)
+	  moveli r6, hw1_last(temp_data_pmd - PAGE_OFFSET)
+	}
+	{
+	  /* After initializing swapper_pgprot, HV_PTE_GLOBAL is set. */
+	  bfextu r7, r1, HV_PTE_INDEX_GLOBAL, HV_PTE_INDEX_GLOBAL
+	  inv r4
+	}
+	bnez r7, .Lno_write
+	{
+	  shl16insli r0, r0, hw0(.Lsv_data_pmd - PAGE_OFFSET)
+	  shl16insli r6, r6, hw0(temp_data_pmd - PAGE_OFFSET)
+	}
+	{
+	  /* Cut off the low bits of the PT address. */
+	  shrui r6, r6, HV_LOG2_PAGE_TABLE_ALIGN
+	  /* Start with our access pte. */
+	  move r5, r1
+	}
+	{
+	  /* Stuff the address into the page table pointer slot of the PTE. */
+	  bfins r5, r6, HV_PTE_INDEX_PTFN, \
+			HV_PTE_INDEX_PTFN + HV_PTE_PTFN_BITS - 1
+	}
+	{
+	  /* Store the L0 data PTE. */
+	  st r0, r5
+	  addli r6, r6, (temp_code_pmd - temp_data_pmd) >> \
+			HV_LOG2_PAGE_TABLE_ALIGN
+	}
+	{
+	  addli r0, r0, .Lsv_code_pmd - .Lsv_data_pmd
+	  bfins r5, r6, HV_PTE_INDEX_PTFN, \
+			HV_PTE_INDEX_PTFN + HV_PTE_PTFN_BITS - 1
+	}
+	/* Store the L0 code PTE. */
+	st r0, r5
+
+.Lno_write:
+	moveli lr, hw2_last(1f)
+	{
+	  shl16insli lr, lr, hw1(1f)
+	  moveli r0, hw1_last(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  shl16insli lr, lr, hw0(1f)
+	  shl16insli r0, r0, hw0(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  move r3, zero
+	  j hv_install_context
+	}
+1:
+
+	/* Install the interrupt base. */
+	moveli r0, hw2_last(MEM_SV_START)
+	shl16insli r0, r0, hw1(MEM_SV_START)
+	shl16insli r0, r0, hw0(MEM_SV_START)
+	mtspr SPR_INTERRUPT_VECTOR_BASE_K, r0
+
+	/*
+	 * Get our processor number and save it away in SAVE_K_0.
+	 * Extract stuff from the topology structure: r4 = y, r6 = x,
+	 * r5 = width.  FIXME: consider whether we want to just make these
+	 * 64-bit values (and if so fix smp_topology write below, too).
+	 */
+	jal hv_inquire_topology
+	{
+	  v4int_l r5, zero, r1    /* r5 = width */
+	  shrui r4, r0, 32        /* r4 = y */
+	}
+	{
+	  v4int_l r6, zero, r0    /* r6 = x */
+	  mul_lu_lu r4, r4, r5
+	}
+	{
+	  add r4, r4, r6          /* r4 == cpu == y*width + x */
+	}
+
+#ifdef CONFIG_SMP
+	/*
+	 * Load up our per-cpu offset.  When the first (master) tile
+	 * boots, this value is still zero, so we will load boot_pc
+	 * with start_kernel, and boot_sp with init_stack + THREAD_SIZE.
+	 * The master tile initializes the per-cpu offset array, so that
+	 * when subsequent (secondary) tiles boot, they will instead load
+	 * from their per-cpu versions of boot_sp and boot_pc.
+	 */
+	moveli r5, hw2_last(__per_cpu_offset)
+	shl16insli r5, r5, hw1(__per_cpu_offset)
+	shl16insli r5, r5, hw0(__per_cpu_offset)
+	shl3add r5, r4, r5
+	ld r5, r5
+	bnez r5, 1f
+
+	/*
+	 * Save the width and height to the smp_topology variable
+	 * for later use.
+	 */
+	moveli r0, hw2_last(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	shl16insli r0, r0, hw1(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	shl16insli r0, r0, hw0(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	st r0, r1
+1:
+#else
+	move r5, zero
+#endif
+
+	/* Load and go with the correct pc and sp. */
+	{
+	  moveli r1, hw2_last(boot_sp)
+	  moveli r0, hw2_last(boot_pc)
+	}
+	{
+	  shl16insli r1, r1, hw1(boot_sp)
+	  shl16insli r0, r0, hw1(boot_pc)
+	}
+	{
+	  shl16insli r1, r1, hw0(boot_sp)
+	  shl16insli r0, r0, hw0(boot_pc)
+	}
+	{
+	  add r1, r1, r5
+	  add r0, r0, r5
+	}
+	ld r0, r0
+	ld sp, r1
+	or r4, sp, r4
+	mtspr SPR_SYSTEM_SAVE_K_0, r4  /* save ksp0 + cpu */
+	addi sp, sp, -STACK_TOP_DELTA
+	{
+	  move lr, zero   /* stop backtraces in the called function */
+	  jr r0
+	}
+	ENDPROC(_start)
+
+__PAGE_ALIGNED_BSS
+	.align PAGE_SIZE
+ENTRY(empty_zero_page)
+	.fill PAGE_SIZE,1,0
+	END(empty_zero_page)
+
+	.macro PTE cpa, bits1
+	.quad HV_PTE_PAGE | HV_PTE_DIRTY | HV_PTE_PRESENT | HV_PTE_ACCESSED |\
+	      HV_PTE_GLOBAL | (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE) |\
+	      (\bits1) | (HV_CPA_TO_PFN(\cpa) << HV_PTE_INDEX_PFN)
+	.endm
+
+__PAGE_ALIGNED_DATA
+	.align PAGE_SIZE
+ENTRY(swapper_pg_dir)
+	.org swapper_pg_dir + HV_L0_INDEX(PAGE_OFFSET) * HV_PTE_SIZE
+.Lsv_data_pmd:
+	.quad 0  /* PTE temp_data_pmd - PAGE_OFFSET, 0 */
+	.org swapper_pg_dir + HV_L0_INDEX(MEM_SV_START) * HV_PTE_SIZE
+.Lsv_code_pmd:
+	.quad 0  /* PTE temp_code_pmd - PAGE_OFFSET, 0 */
+	.org swapper_pg_dir + HV_L0_SIZE
+	END(swapper_pg_dir)
+
+	.align HV_PAGE_TABLE_ALIGN
+ENTRY(temp_data_pmd)
+	/*
+	 * We fill the PAGE_OFFSET pmd with huge pages with
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions later.
+	 */
+	.set addr, 0
+	.rept HV_L1_ENTRIES
+	PTE addr, HV_PTE_READABLE | HV_PTE_WRITABLE
+	.set addr, addr + HV_PAGE_SIZE_LARGE
+	.endr
+	.org temp_data_pmd + HV_L1_SIZE
+	END(temp_data_pmd)
+
+	.align HV_PAGE_TABLE_ALIGN
+ENTRY(temp_code_pmd)
+	/*
+	 * We fill the MEM_SV_START pmd with huge pages with
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions later.
+	 */
+	.set addr, 0
+	.rept HV_L1_ENTRIES
+	PTE addr, HV_PTE_READABLE | HV_PTE_EXECUTABLE
+	.set addr, addr + HV_PAGE_SIZE_LARGE
+	.endr
+	.org temp_code_pmd + HV_L1_SIZE
+	END(temp_code_pmd)
+
+	/*
+	 * Isolate swapper_pgprot to its own cache line, since each cpu
+	 * starting up will read it using VA-is-PA and local homing.
+	 * This would otherwise likely conflict with other data on the cache
+	 * line, once we have set its permanent home in the page tables.
+	 */
+	__INITDATA
+	.align CHIP_L2_LINE_SIZE()
+ENTRY(swapper_pgprot)
+	.quad HV_PTE_PRESENT | (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE)
+	.align CHIP_L2_LINE_SIZE()
+	END(swapper_pgprot)
diff --git a/arch/tile/kernel/intvec_64.S b/arch/tile/kernel/intvec_64.S
new file mode 100644
index 00000000000..79c93e10ba2
--- /dev/null
+++ b/arch/tile/kernel/intvec_64.S
@@ -0,0 +1,1231 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Linux interrupt vectors.
+ */
+
+#include <linux/linkage.h>
+#include <linux/errno.h>
+#include <linux/unistd.h>
+#include <asm/ptrace.h>
+#include <asm/thread_info.h>
+#include <asm/irqflags.h>
+#include <asm/asm-offsets.h>
+#include <asm/types.h>
+#include <hv/hypervisor.h>
+#include <arch/abi.h>
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+#ifdef CONFIG_PREEMPT
+# error "No support for kernel preemption currently"
+#endif
+
+#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
+
+#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
+
+
+	.macro  push_reg reg, ptr=sp, delta=-8
+	{
+	 st     \ptr, \reg
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg reg, ptr=sp, delta=8
+	{
+	 ld     \reg, \ptr
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=8
+	{
+	 move   \zreg, zero
+	 ld     \reg, \ptr
+	 addi   \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  push_extra_callee_saves reg
+	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
+	push_reg r51, \reg
+	push_reg r50, \reg
+	push_reg r49, \reg
+	push_reg r48, \reg
+	push_reg r47, \reg
+	push_reg r46, \reg
+	push_reg r45, \reg
+	push_reg r44, \reg
+	push_reg r43, \reg
+	push_reg r42, \reg
+	push_reg r41, \reg
+	push_reg r40, \reg
+	push_reg r39, \reg
+	push_reg r38, \reg
+	push_reg r37, \reg
+	push_reg r36, \reg
+	push_reg r35, \reg
+	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
+	.endm
+
+	.macro  panic str
+	.pushsection .rodata, "a"
+1:
+	.asciz  "\str"
+	.popsection
+	{
+	 moveli r0, hw2_last(1b)
+	}
+	{
+	 shl16insli r0, r0, hw1(1b)
+	}
+	{
+	 shl16insli r0, r0, hw0(1b)
+	 jal    panic
+	}
+	.endm
+
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+intvec_feedback:
+	.popsection
+#endif
+
+	/*
+	 * Default interrupt handler.
+	 *
+	 * vecnum is where we'll put this code.
+	 * c_routine is the C routine we'll call.
+	 *
+	 * The C routine is passed two arguments:
+	 * - A pointer to the pt_regs state.
+	 * - The interrupt vector number.
+	 *
+	 * The "processing" argument specifies the code for processing
+	 * the interrupt. Defaults to "handle_interrupt".
+	 */
+	.macro  int_hand vecnum, vecname, c_routine, processing=handle_interrupt
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	/* Temporarily save a register so we have somewhere to work. */
+
+	mtspr   SPR_SYSTEM_SAVE_K_1, r0
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+
+	andi    r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+
+	.ifc    \vecnum, INT_DOUBLE_FAULT
+	/*
+	 * For double-faults from user-space, fall through to the normal
+	 * register save and stack setup path.  Otherwise, it's the
+	 * hypervisor giving us one last chance to dump diagnostics, and we
+	 * branch to the kernel_double_fault routine to do so.
+	 */
+	beqz    r0, 1f
+	j       _kernel_double_fault
+1:
+	.else
+	/*
+	 * If we're coming from user-space, then set sp to the top of
+	 * the kernel stack.  Otherwise, assume sp is already valid.
+	 */
+	{
+	 bnez   r0, 0f
+	 move   r0, sp
+	}
+	.endif
+
+	.ifc    \c_routine, do_page_fault
+	/*
+	 * The page_fault handler may be downcalled directly by the
+	 * hypervisor even when Linux is running and has ICS set.
+	 *
+	 * In this case the contents of EX_CONTEXT_K_1 reflect the
+	 * previous fault and can't be relied on to choose whether or
+	 * not to reinitialize the stack pointer.  So we add a test
+	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
+	 * and if so we don't reinitialize sp, since we must be coming
+	 * from Linux.  (In fact the precise case is !(val & ~1),
+	 * but any Linux PC has to have the high bit set.)
+	 *
+	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
+	 * any path that turns into a downcall to one of our TLB handlers.
+	 *
+	 * FIXME: if we end up never using this path, perhaps we should
+	 * prevent the hypervisor from generating downcalls in this case.
+	 * The advantage of getting a downcall is we can panic in Linux.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_2
+	{
+	 bltz   r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
+	 move   r0, sp
+	}
+	.endif
+
+
+	/*
+	 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
+	 * the current stack top in the higher bits.  So we recover
+	 * our stack top by just masking off the low bits, then
+	 * point sp at the top aligned address on the actual stack page.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_0
+	mm      r0, zero, LOG2_THREAD_SIZE, 63
+
+0:
+	/*
+	 * Align the stack mod 64 so we can properly predict what
+	 * cache lines we need to write-hint to reduce memory fetch
+	 * latency as we enter the kernel.  The layout of memory is
+	 * as follows, with cache line 0 at the lowest VA, and cache
+	 * line 8 just below the r0 value this "andi" computes.
+	 * Note that we never write to cache line 8, and we skip
+	 * cache lines 1-3 for syscalls.
+	 *
+	 *    cache line 8: ptregs padding (two words)
+	 *    cache line 7: sp, lr, pc, ex1, faultnum, orig_r0, flags, cmpexch
+	 *    cache line 6: r46...r53 (tp)
+	 *    cache line 5: r38...r45
+	 *    cache line 4: r30...r37
+	 *    cache line 3: r22...r29
+	 *    cache line 2: r14...r21
+	 *    cache line 1: r6...r13
+	 *    cache line 0: 2 x frame, r0..r5
+	 */
+	andi    r0, r0, -64
+
+	/*
+	 * Push the first four registers on the stack, so that we can set
+	 * them to vector-unique values before we jump to the common code.
+	 *
+	 * Registers are pushed on the stack as a struct pt_regs,
+	 * with the sp initially just above the struct, and when we're
+	 * done, sp points to the base of the struct, minus
+	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
+	 *
+	 * This routine saves just the first four registers, plus the
+	 * stack context so we can do proper backtracing right away,
+	 * and defers to handle_interrupt to save the rest.
+	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
+	 */
+	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
+	wh64    r0   /* cache line 7 */
+	{
+	 st     r0, lr
+	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	}
+	{
+	 st     r0, sp
+	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
+	}
+	wh64    sp   /* cache line 6 */
+	{
+	 st     sp, r52
+	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
+	}
+	wh64    sp   /* cache line 0 */
+	{
+	 st     sp, r1
+	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
+	}
+	{
+	 st     sp, r2
+	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
+	}
+	{
+	 st     sp, r3
+	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_0
+	.ifc \processing,handle_syscall
+	/*
+	 * Bump the saved PC by one bundle so that when we return, we won't
+	 * execute the same swint instruction again.  We need to do this while
+	 * we're in the critical section.
+	 */
+	addi    r0, r0, 8
+	.endif
+	{
+	 st     sp, r0
+	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+	{
+	 st     sp, r0
+	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	/*
+	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
+	 * so that it gets passed through unchanged to the handler routine.
+	 * Note that the .if conditional confusingly spans bundles.
+	 */
+	 .ifc \processing,handle_syscall
+	 movei  r0, \vecnum
+	}
+	{
+	 st     sp, r0
+	 .else
+	 movei  r1, \vecnum
+	}
+	{
+	 st     sp, r1
+	 .endif
+	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
+	}
+	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
+	{
+	 st     sp, r0
+	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 8
+	}
+	{
+	 st     sp, zero        /* write zero into "Next SP" frame pointer */
+	 addi   sp, sp, -8      /* leave SP pointing at bottom of frame */
+	}
+	.ifc \processing,handle_syscall
+	j       handle_syscall
+	.else
+	/* Capture per-interrupt SPR context to registers. */
+	.ifc \c_routine, do_page_fault
+	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
+	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
+	.else
+	.ifc \vecnum, INT_ILL_TRANS
+	mfspr   r2, ILL_TRANS_REASON
+	.else
+	.ifc \vecnum, INT_DOUBLE_FAULT
+	mfspr   r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
+	.else
+	.ifc \c_routine, do_trap
+	mfspr   r2, GPV_REASON
+	.else
+	.ifc \c_routine, op_handle_perf_interrupt
+	mfspr   r2, PERF_COUNT_STS
+#if CHIP_HAS_AUX_PERF_COUNTERS()
+	.else
+	.ifc \c_routine, op_handle_aux_perf_interrupt
+	mfspr   r2, AUX_PERF_COUNT_STS
+	.endif
+#endif
+	.endif
+	.endif
+	.endif
+	.endif
+	.endif
+	/* Put function pointer in r0 */
+	moveli  r0, hw2_last(\c_routine)
+	shl16insli r0, r0, hw1(\c_routine)
+	{
+	 shl16insli r0, r0, hw0(\c_routine)
+	 j       \processing
+	}
+	.endif
+	ENDPROC(intvec_\vecname)
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+	.org    (\vecnum << 5)
+	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt1, 1 << 8)
+	jrp     lr
+	.popsection
+#endif
+
+	.endm
+
+
+	/*
+	 * Save the rest of the registers that we didn't save in the actual
+	 * vector itself.  We can't use r0-r10 inclusive here.
+	 */
+	.macro  finish_interrupt_save, function
+
+	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
+	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
+	{
+	 .ifc \function,handle_syscall
+	 st     r52, r0
+	 .else
+	 st     r52, zero
+	 .endif
+	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
+	}
+	st      r52, tp
+	{
+	 mfspr  tp, CMPEXCH_VALUE
+	 PTREGS_PTR(r52, PTREGS_OFFSET_CMPEXCH)
+	}
+
+	/*
+	 * For ordinary syscalls, we save neither caller- nor callee-
+	 * save registers, since the syscall invoker doesn't expect the
+	 * caller-saves to be saved, and the called kernel functions will
+	 * take care of saving the callee-saves for us.
+	 *
+	 * For interrupts we save just the caller-save registers.  Saving
+	 * them is required (since the "caller" can't save them).  Again,
+	 * the called kernel functions will restore the callee-save
+	 * registers for us appropriately.
+	 *
+	 * On return, we normally restore nothing special for syscalls,
+	 * and just the caller-save registers for interrupts.
+	 *
+	 * However, there are some important caveats to all this:
+	 *
+	 * - We always save a few callee-save registers to give us
+	 *   some scratchpad registers to carry across function calls.
+	 *
+	 * - fork/vfork/etc require us to save all the callee-save
+	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
+	 *
+	 * - We always save r0..r5 and r10 for syscalls, since we need
+	 *   to reload them a bit later for the actual kernel call, and
+	 *   since we might need them for -ERESTARTNOINTR, etc.
+	 *
+	 * - Before invoking a signal handler, we save the unsaved
+	 *   callee-save registers so they are visible to the
+	 *   signal handler or any ptracer.
+	 *
+	 * - If the unsaved callee-save registers are modified, we set
+	 *   a bit in pt_regs so we know to reload them from pt_regs
+	 *   and not just rely on the kernel function unwinding.
+	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
+	 */
+	{
+	 st     r52, tp
+	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
+	}
+	wh64    r52    /* cache line 4 */
+	push_reg r33, r52
+	push_reg r32, r52
+	push_reg r31, r52
+	.ifc \function,handle_syscall
+	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
+	push_reg TREG_SYSCALL_NR_NAME, r52, \
+	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
+	.else
+
+	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
+	wh64    r52   /* cache line 3 */
+	push_reg r29, r52
+	push_reg r28, r52
+	push_reg r27, r52
+	push_reg r26, r52
+	push_reg r25, r52
+	push_reg r24, r52
+	push_reg r23, r52
+	push_reg r22, r52
+	wh64    r52   /* cache line 2 */
+	push_reg r21, r52
+	push_reg r20, r52
+	push_reg r19, r52
+	push_reg r18, r52
+	push_reg r17, r52
+	push_reg r16, r52
+	push_reg r15, r52
+	push_reg r14, r52
+	wh64    r52   /* cache line 1 */
+	push_reg r13, r52
+	push_reg r12, r52
+	push_reg r11, r52
+	push_reg r10, r52
+	push_reg r9, r52
+	push_reg r8, r52
+	push_reg r7, r52
+	push_reg r6, r52
+
+	.endif
+
+	push_reg r5, r52
+	st      r52, r4
+
+	/* Load tp with our per-cpu offset. */
+#ifdef CONFIG_SMP
+	{
+	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
+	 moveli r21, hw2_last(__per_cpu_offset)
+	}
+	{
+	 shl16insli r21, r21, hw1(__per_cpu_offset)
+	 bfextu r20, r20, 0, LOG2_THREAD_SIZE-1
+	}
+	shl16insli r21, r21, hw0(__per_cpu_offset)
+	shl3add r20, r20, r21
+	ld      tp, r20
+#else
+	move    tp, zero
+#endif
+
+	/*
+	 * If we will be returning to the kernel, we will need to
+	 * reset the interrupt masks to the state they had before.
+	 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
+	 */
+	mfspr   r32, SPR_EX_CONTEXT_K_1
+	{
+	 andi   r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
+	}
+	beqzt   r32, 1f       /* zero if from user space */
+	IRQS_DISABLED(r32)    /* zero if irqs enabled */
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
+#endif
+1:
+	.ifnc \function,handle_syscall
+	/* Record the fact that we saved the caller-save registers above. */
+	ori     r32, r32, PT_FLAGS_CALLER_SAVES
+	.endif
+	st      r21, r32
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	/*
+	 * Notify the feedback routines that we were in the
+	 * appropriate fixed interrupt vector area.  Note that we
+	 * still have ICS set at this point, so we can't invoke any
+	 * atomic operations or we will panic.  The feedback
+	 * routines internally preserve r0..r10 and r30 up.
+	 */
+	.ifnc \function,handle_syscall
+	shli    r20, r1, 5
+	.else
+	moveli  r20, INT_SWINT_1 << 5
+	.endif
+	moveli  r21, hw2_last(intvec_feedback)
+	shl16insli r21, r21, hw1(intvec_feedback)
+	shl16insli r21, r21, hw0(intvec_feedback)
+	add     r20, r20, r21
+	jalr    r20
+
+	/* And now notify the feedback routines that we are here. */
+	FEEDBACK_ENTER(\function)
+#endif
+
+	/*
+	 * we've captured enough state to the stack (including in
+	 * particular our EX_CONTEXT state) that we can now release
+	 * the interrupt critical section and replace it with our
+	 * standard "interrupts disabled" mask value.  This allows
+	 * synchronous interrupts (and profile interrupts) to punch
+	 * through from this point onwards.
+	 */
+	.ifc \function,handle_nmi
+	IRQ_DISABLE_ALL(r20)
+	.else
+	IRQ_DISABLE(r20, r21)
+	.endif
+	mtspr   INTERRUPT_CRITICAL_SECTION, zero
+
+	/*
+	 * Prepare the first 256 stack bytes to be rapidly accessible
+	 * without having to fetch the background data.
+	 */
+	addi    r52, sp, -64
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	wh64    r52
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+	.ifnc \function,handle_nmi
+	/*
+	 * We finally have enough state set up to notify the irq
+	 * tracing code that irqs were disabled on entry to the handler.
+	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
+	 * For syscalls, we already have the register state saved away
+	 * on the stack, so we don't bother to do any register saves here,
+	 * and later we pop the registers back off the kernel stack.
+	 * For interrupt handlers, save r0-r3 in callee-saved registers.
+	 */
+	.ifnc \function,handle_syscall
+	{ move r30, r0; move r31, r1 }
+	{ move r32, r2; move r33, r3 }
+	.endif
+	TRACE_IRQS_OFF
+	.ifnc \function,handle_syscall
+	{ move r0, r30; move r1, r31 }
+	{ move r2, r32; move r3, r33 }
+	.endif
+	.endif
+#endif
+
+	.endm
+
+	/*
+	 * Redispatch a downcall.
+	 */
+	.macro  dc_dispatch vecnum, vecname
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	j       hv_downcall_dispatch
+	ENDPROC(intvec_\vecname)
+	.endm
+
+	/*
+	 * Common code for most interrupts.  The C function we're eventually
+	 * going to is in r0, and the faultnum is in r1; the original
+	 * values for those registers are on the stack.
+	 */
+	.pushsection .text.handle_interrupt,"ax"
+handle_interrupt:
+	finish_interrupt_save handle_interrupt
+
+	/* Jump to the C routine; it should enable irqs as soon as possible. */
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_interrupt)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_interrupt)
+
+/*
+ * This routine takes a boolean in r30 indicating if this is an NMI.
+ * If so, we also expect a boolean in r31 indicating whether to
+ * re-enable the oprofile interrupts.
+ */
+STD_ENTRY(interrupt_return)
+	/* If we're resuming to kernel space, don't check thread flags. */
+	{
+	 bnez   r30, .Lrestore_all  /* NMIs don't special-case user-space */
+	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
+	}
+	ld      r29, r29
+	andi    r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	{
+	 beqzt  r29, .Lresume_userspace
+	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
+	}
+
+	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
+	moveli  r27, hw2_last(_cpu_idle_nap)
+	{
+	 ld     r28, r29
+	 shl16insli r27, r27, hw1(_cpu_idle_nap)
+	}
+	{
+	 shl16insli r27, r27, hw0(_cpu_idle_nap)
+	}
+	{
+	 cmpeq  r27, r27, r28
+	}
+	{
+	 blbc   r27, .Lrestore_all
+	 addi   r28, r28, 8
+	}
+	st      r29, r28
+	j       .Lrestore_all
+
+.Lresume_userspace:
+	FEEDBACK_REENTER(interrupt_return)
+
+	/*
+	 * Disable interrupts so as to make sure we don't
+	 * miss an interrupt that sets any of the thread flags (like
+	 * need_resched or sigpending) between sampling and the iret.
+	 * Routines like schedule() or do_signal() may re-enable
+	 * interrupts before returning.
+	 */
+	IRQ_DISABLE(r20, r21)
+	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
+
+	/* Get base of stack in r32; note r30/31 are used as arguments here. */
+	GET_THREAD_INFO(r32)
+
+
+	/* Check to see if there is any work to do before returning to user. */
+	{
+	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
+	 moveli r1, hw1_last(_TIF_ALLWORK_MASK)
+	}
+	{
+	 ld     r29, r29
+	 shl16insli r1, r1, hw0(_TIF_ALLWORK_MASK)
+	}
+	and     r1, r29, r1
+	beqzt   r1, .Lrestore_all
+
+	/*
+	 * Make sure we have all the registers saved for signal
+	 * handling or single-step.  Call out to C code to figure out
+	 * exactly what we need to do for each flag bit, then if
+	 * necessary, reload the flags and recheck.
+	 */
+	push_extra_callee_saves r0
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_work_pending
+	}
+	bnez    r0, .Lresume_userspace
+
+	/*
+	 * In the NMI case we
+	 * omit the call to single_process_check_nohz, which normally checks
+	 * to see if we should start or stop the scheduler tick, because
+	 * we can't call arbitrary Linux code from an NMI context.
+	 * We always call the homecache TLB deferral code to re-trigger
+	 * the deferral mechanism.
+	 *
+	 * The other chunk of responsibility this code has is to reset the
+	 * interrupt masks appropriately to reset irqs and NMIs.  We have
+	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
+	 * lockdep-type stuff, but we can't set ICS until afterwards, since
+	 * ICS can only be used in very tight chunks of code to avoid
+	 * tripping over various assertions that it is off.
+	 */
+.Lrestore_all:
+	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
+	{
+	 ld      r0, r0
+	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
+	}
+	{
+	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
+	 ld     r32, r32
+	}
+	bnez    r0, 1f
+	j       2f
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use blbct below
+#endif
+1:	blbct   r32, 2f
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	beqzt   r30, .Lrestore_regs
+	j       3f
+2:	TRACE_IRQS_ON
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	IRQ_ENABLE(r20, r21)
+	beqzt   r30, .Lrestore_regs
+3:
+
+
+	/*
+	 * We now commit to returning from this interrupt, since we will be
+	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
+	 * frame.  No calls should be made to any other code after this point.
+	 * This code should only be entered with ICS set.
+	 * r32 must still be set to ptregs.flags.
+	 * We launch loads to each cache line separately first, so we can
+	 * get some parallelism out of the memory subsystem.
+	 * We start zeroing caller-saved registers throughout, since
+	 * that will save some cycles if this turns out to be a syscall.
+	 */
+.Lrestore_regs:
+	FEEDBACK_REENTER(interrupt_return)   /* called from elsewhere */
+
+	/*
+	 * Rotate so we have one high bit and one low bit to test.
+	 * - low bit says whether to restore all the callee-saved registers,
+	 *   or just r30-r33, and r52 up.
+	 * - high bit (i.e. sign bit) says whether to restore all the
+	 *   caller-saved registers, or just r0.
+	 */
+#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
+# error Rotate trick does not work :-)
+#endif