parisc: Fix access fault handling in pa_memcpy()

pa_memcpy() is the major memcpy implementation in the parisc kernel which is
used to do any kind of userspace/kernel memory copies.

Al Viro noticed various bugs in the implementation of pa_mempcy(), most notably
that in case of faults it may report back to have copied more bytes than it
actually did.

Fixing those bugs is quite hard in the C-implementation, because the compiler
is messing around with the registers and we are not guaranteed that specific
variables are always in the same processor registers. This makes proper fault
handling complicated.

This patch implements pa_memcpy() in assembler. That way we have correct fault
handling and adding a 64-bit copy routine was quite easy.

Runtime tested with 32- and 64bit kernels.
Reported-by: NAl Viro <viro@ZenIV.linux.org.uk>
Cc: <stable@vger.kernel.org> # v4.9+
Signed-off-by: NJohn David Anglin <dave.anglin@bell.net>
Signed-off-by: NHelge Deller <deller@gmx.de>

arch/parisc/lib/lusercopy.S

@@ -5,6 +5,8 @@
  *    Copyright (C) 2000 Richard Hirst <rhirst with parisc-linux.org>
  *    Copyright (C) 2001 Matthieu Delahaye <delahaym at esiee.fr>
  *    Copyright (C) 2003 Randolph Chung <tausq with parisc-linux.org>
+ *    Copyright (C) 2017 Helge Deller <deller@gmx.de>
+ *    Copyright (C) 2017 John David Anglin <dave.anglin@bell.net>
  *
  *
  *    This program is free software; you can redistribute it and/or modify
@@ -132,4 +134,320 @@ ENDPROC_CFI(lstrnlen_user)
 
 	.procend
 
+
+
+/*
+ * unsigned long pa_memcpy(void *dstp, const void *srcp, unsigned long len)
+ *
+ * Inputs:
+ * - sr1 already contains space of source region
+ * - sr2 already contains space of destination region
+ *
+ * Returns:
+ * - number of bytes that could not be copied.
+ *   On success, this will be zero.
+ *
+ * This code is based on a C-implementation of a copy routine written by
+ * Randolph Chung, which in turn was derived from the glibc.
+ *
+ * Several strategies are tried to try to get the best performance for various
+ * conditions. In the optimal case, we copy by loops that copy 32- or 16-bytes
+ * at a time using general registers.  Unaligned copies are handled either by
+ * aligning the destination and then using shift-and-write method, or in a few
+ * cases by falling back to a byte-at-a-time copy.
+ *
+ * Testing with various alignments and buffer sizes shows that this code is
+ * often >10x faster than a simple byte-at-a-time copy, even for strangely
+ * aligned operands. It is interesting to note that the glibc version of memcpy
+ * (written in C) is actually quite fast already. This routine is able to beat
+ * it by 30-40% for aligned copies because of the loop unrolling, but in some
+ * cases the glibc version is still slightly faster. This lends more
+ * credibility that gcc can generate very good code as long as we are careful.
+ *
+ * Possible optimizations:
+ * - add cache prefetching
+ * - try not to use the post-increment address modifiers; they may create
+ *   additional interlocks. Assumption is that those were only efficient on old
+ *   machines (pre PA8000 processors)
+ */
+
+	dst = arg0
+	src = arg1
+	len = arg2
+	end = arg3
+	t1  = r19
+	t2  = r20
+	t3  = r21
+	t4  = r22
+	srcspc = sr1
+	dstspc = sr2
+
+	t0 = r1
+	a1 = t1
+	a2 = t2
+	a3 = t3
+	a0 = t4
+
+	save_src = ret0
+	save_dst = ret1
+	save_len = r31
+
+ENTRY_CFI(pa_memcpy)
+	.proc
+	.callinfo NO_CALLS
+	.entry
+
+	/* Last destination address */
+	add	dst,len,end
+
+	/* short copy with less than 16 bytes? */
+	cmpib,>>=,n 15,len,.Lbyte_loop
+
+	/* same alignment? */
+	xor	src,dst,t0
+	extru	t0,31,2,t1
+	cmpib,<>,n  0,t1,.Lunaligned_copy
+
+#ifdef CONFIG_64BIT
+	/* only do 64-bit copies if we can get aligned. */
+	extru	t0,31,3,t1
+	cmpib,<>,n  0,t1,.Lalign_loop32
+
+	/* loop until we are 64-bit aligned */
+.Lalign_loop64:
+	extru	dst,31,3,t1
+	cmpib,=,n	0,t1,.Lcopy_loop_16
+20:	ldb,ma	1(srcspc,src),t1
+21:	stb,ma	t1,1(dstspc,dst)
+	b	.Lalign_loop64
+	ldo	-1(len),len
+
+	ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done)
+
+	ldi	31,t0
+.Lcopy_loop_16:
+	cmpb,COND(>>=),n t0,len,.Lword_loop
+
+10:	ldd	0(srcspc,src),t1
+11:	ldd	8(srcspc,src),t2
+	ldo	16(src),src
+12:	std,ma	t1,8(dstspc,dst)
+13:	std,ma	t2,8(dstspc,dst)
+14:	ldd	0(srcspc,src),t1
+15:	ldd	8(srcspc,src),t2
+	ldo	16(src),src
+16:	std,ma	t1,8(dstspc,dst)
+17:	std,ma	t2,8(dstspc,dst)
+
+	ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(11b,.Lcopy16_fault)
+	ASM_EXCEPTIONTABLE_ENTRY(12b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(13b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(14b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(15b,.Lcopy16_fault)
+	ASM_EXCEPTIONTABLE_ENTRY(16b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(17b,.Lcopy_done)
+
+	b	.Lcopy_loop_16
+	ldo	-32(len),len
+
+.Lword_loop:
+	cmpib,COND(>>=),n 3,len,.Lbyte_loop
+20:	ldw,ma	4(srcspc,src),t1
+21:	stw,ma	t1,4(dstspc,dst)
+	b	.Lword_loop
+	ldo	-4(len),len
+
+	ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done)
+
+#endif /* CONFIG_64BIT */
+
+	/* loop until we are 32-bit aligned */
+.Lalign_loop32:
+	extru	dst,31,2,t1
+	cmpib,=,n	0,t1,.Lcopy_loop_4
+20:	ldb,ma	1(srcspc,src),t1
+21:	stb,ma	t1,1(dstspc,dst)
+	b	.Lalign_loop32
+	ldo	-1(len),len
+
+	ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done)
+
+
+.Lcopy_loop_4:
+	cmpib,COND(>>=),n 15,len,.Lbyte_loop
+
+10:	ldw	0(srcspc,src),t1
+11:	ldw	4(srcspc,src),t2
+12:	stw,ma	t1,4(dstspc,dst)
+13:	stw,ma	t2,4(dstspc,dst)
+14:	ldw	8(srcspc,src),t1
+15:	ldw	12(srcspc,src),t2
+	ldo	16(src),src
+16:	stw,ma	t1,4(dstspc,dst)
+17:	stw,ma	t2,4(dstspc,dst)
+
+	ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(11b,.Lcopy8_fault)
+	ASM_EXCEPTIONTABLE_ENTRY(12b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(13b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(14b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(15b,.Lcopy8_fault)
+	ASM_EXCEPTIONTABLE_ENTRY(16b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(17b,.Lcopy_done)
+
+	b	.Lcopy_loop_4
+	ldo	-16(len),len
+
+.Lbyte_loop:
+	cmpclr,COND(<>) len,%r0,%r0
+	b,n	.Lcopy_done
+20:	ldb	0(srcspc,src),t1
+	ldo	1(src),src
+21:	stb,ma	t1,1(dstspc,dst)
+	b	.Lbyte_loop
+	ldo	-1(len),len
+
+	ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done)
+
+.Lcopy_done:
+	bv	%r0(%r2)
+	sub	end,dst,ret0
+
+
+	/* src and dst are not aligned the same way. */
+	/* need to go the hard way */
+.Lunaligned_copy:
+	/* align until dst is 32bit-word-aligned */
+	extru	dst,31,2,t1
+	cmpib,COND(=),n	0,t1,.Lcopy_dstaligned
+20:	ldb	0(srcspc,src),t1
+	ldo	1(src),src
+21:	stb,ma	t1,1(dstspc,dst)
+	b	.Lunaligned_copy
+	ldo	-1(len),len
+
+	ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done)
+	ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done)
+
+.Lcopy_dstaligned:
+
+	/* store src, dst and len in safe place */
+	copy	src,save_src
+	copy	dst,save_dst
+	copy	len,save_len
+
+	/* len now needs give number of words to copy */
+	SHRREG	len,2,len
+
+	/*
+	 * Copy from a not-aligned src to an aligned dst using shifts.
+	 * Handles 4 words per loop.
+	 */
+
+	depw,z src,28,2,t0
+	subi 32,t0,t0
+	mtsar t0
+	extru len,31,2,t0
+	cmpib,= 2,t0,.Lcase2
+	/* Make src aligned by rounding it down.  */
+	depi 0,31,2,src
+
+	cmpiclr,<> 3,t0,%r0
+	b,n .Lcase3
+	cmpiclr,<> 1,t0,%r0
+	b,n .Lcase1
+.Lcase0:
+	cmpb,= %r0,len,.Lcda_finish
+	nop
+
+1:	ldw,ma 4(srcspc,src), a3
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+1:	ldw,ma 4(srcspc,src), a0
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	b,n .Ldo3
+.Lcase1:
+1:	ldw,ma 4(srcspc,src), a2
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+1:	ldw,ma 4(srcspc,src), a3
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	ldo -1(len),len
+	cmpb,=,n %r0,len,.Ldo0
+.Ldo4:
+1:	ldw,ma 4(srcspc,src), a0
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	shrpw a2, a3, %sar, t0
+1:	stw,ma t0, 4(dstspc,dst)
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done)
+.Ldo3:
+1:	ldw,ma 4(srcspc,src), a1
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	shrpw a3, a0, %sar, t0
+1:	stw,ma t0, 4(dstspc,dst)
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done)
+.Ldo2:
+1:	ldw,ma 4(srcspc,src), a2
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	shrpw a0, a1, %sar, t0
+1:	stw,ma t0, 4(dstspc,dst)
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done)
+.Ldo1:
+1:	ldw,ma 4(srcspc,src), a3
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	shrpw a1, a2, %sar, t0
+1:	stw,ma t0, 4(dstspc,dst)
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done)
+	ldo -4(len),len
+	cmpb,<> %r0,len,.Ldo4
+	nop
+.Ldo0:
+	shrpw a2, a3, %sar, t0
+1:	stw,ma t0, 4(dstspc,dst)
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done)
+
+.Lcda_rdfault:
+.Lcda_finish:
+	/* calculate new src, dst and len and jump to byte-copy loop */
+	sub	dst,save_dst,t0
+	add	save_src,t0,src
+	b	.Lbyte_loop
+	sub	save_len,t0,len
+
+.Lcase3:
+1:	ldw,ma 4(srcspc,src), a0
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+1:	ldw,ma 4(srcspc,src), a1
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	b .Ldo2
+	ldo 1(len),len
+.Lcase2:
+1:	ldw,ma 4(srcspc,src), a1
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+1:	ldw,ma 4(srcspc,src), a2
+	ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault)
+	b .Ldo1
+	ldo 2(len),len
+
+
+	/* fault exception fixup handlers: */
+#ifdef CONFIG_64BIT
+.Lcopy16_fault:
+10:	b	.Lcopy_done
+	std,ma	t1,8(dstspc,dst)
+	ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done)
+#endif
+
+.Lcopy8_fault:
+10:	b	.Lcopy_done
+	stw,ma	t1,4(dstspc,dst)
+	ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done)
+
+	.exit
+ENDPROC_CFI(pa_memcpy)
+	.procend
+
 	.end

arch/parisc/lib/memcpy.c

@@ -2,7 +2,7 @@
  *    Optimized memory copy routines.
  *
  *    Copyright (C) 2004 Randolph Chung <tausq@debian.org>
- *    Copyright (C) 2013 Helge Deller <deller@gmx.de>
+ *    Copyright (C) 2013-2017 Helge Deller <deller@gmx.de>
  *
  *    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
@@ -21,474 +21,21 @@
  *    Portions derived from the GNU C Library
  *    Copyright (C) 1991, 1997, 2003 Free Software Foundation, Inc.
  *
- * Several strategies are tried to try to get the best performance for various
- * conditions. In the optimal case, we copy 64-bytes in an unrolled loop using 
- * fp regs. This is followed by loops that copy 32- or 16-bytes at a time using
- * general registers.  Unaligned copies are handled either by aligning the 
- * destination and then using shift-and-write method, or in a few cases by 
- * falling back to a byte-at-a-time copy.
- *
- * I chose to implement this in C because it is easier to maintain and debug,
- * and in my experiments it appears that the C code generated by gcc (3.3/3.4
- * at the time of writing) is fairly optimal. Unfortunately some of the 
- * semantics of the copy routine (exception handling) is difficult to express
- * in C, so we have to play some tricks to get it to work.
- *
- * All the loads and stores are done via explicit asm() code in order to use
- * the right space registers. 
- * 
- * Testing with various alignments and buffer sizes shows that this code is 
- * often >10x faster than a simple byte-at-a-time copy, even for strangely
- * aligned operands. It is interesting to note that the glibc version
- * of memcpy (written in C) is actually quite fast already. This routine is 
- * able to beat it by 30-40% for aligned copies because of the loop unrolling, 
- * but in some cases the glibc version is still slightly faster. This lends 
- * more credibility that gcc can generate very good code as long as we are 
- * careful.
- *
- * TODO:
- * - cache prefetching needs more experimentation to get optimal settings
- * - try not to use the post-increment address modifiers; they create additional
- *   interlocks
- * - replace byte-copy loops with stybs sequences
  */
 
-#ifdef __KERNEL__
 #include <linux/module.h>
 #include <linux/compiler.h>
 #include <linux/uaccess.h>
-#define s_space "%%sr1"
-#define d_space "%%sr2"
-#else
-#include "memcpy.h"
-#define s_space "%%sr0"
-#define d_space "%%sr0"
-#define pa_memcpy new2_copy
-#endif
 
 DECLARE_PER_CPU(struct exception_data, exception_data);
 
-#define preserve_branch(label)	do {					\
-	volatile int dummy = 0;						\
-	/* The following branch is never taken, it's just here to  */	\
-	/* prevent gcc from optimizing away our exception code. */ 	\
-	if (unlikely(dummy != dummy))					\
-		goto label;						\
-} while (0)
-
 #define get_user_space() (segment_eq(get_fs(), KERNEL_DS) ? 0 : mfsp(3))
 #define get_kernel_space() (0)
 
-#define MERGE(w0, sh_1, w1, sh_2)  ({					\
-	unsigned int _r;						\
-	asm volatile (							\
-	"mtsar %3\n"							\
-	"shrpw %1, %2, %%sar, %0\n"					\
-	: "=r"(_r)							\
-	: "r"(w0), "r"(w1), "r"(sh_2)					\
-	);								\
-	_r;								\
-})
-#define THRESHOLD	16
-
-#ifdef DEBUG_MEMCPY
-#define DPRINTF(fmt, args...) do { printk(KERN_DEBUG "%s:%d:%s ", __FILE__, __LINE__, __func__ ); printk(KERN_DEBUG fmt, ##args ); } while (0)
-#else
-#define DPRINTF(fmt, args...)
-#endif
-
-#define def_load_ai_insn(_insn,_sz,_tt,_s,_a,_t,_e)	\
-	__asm__ __volatile__ (				\
-	"1:\t" #_insn ",ma " #_sz "(" _s ",%1), %0\n\t"	\
-	ASM_EXCEPTIONTABLE_ENTRY(1b,_e)			\
-	: _tt(_t), "+r"(_a)				\
-	: 						\
-	: "r8")
-
-#define def_store_ai_insn(_insn,_sz,_tt,_s,_a,_t,_e) 	\
-	__asm__ __volatile__ (				\
-	"1:\t" #_insn ",ma %1, " #_sz "(" _s ",%0)\n\t"	\
-	ASM_EXCEPTIONTABLE_ENTRY(1b,_e)			\
-	: "+r"(_a) 					\
-	: _tt(_t)					\
-	: "r8")
-
-#define ldbma(_s, _a, _t, _e) def_load_ai_insn(ldbs,1,"=r",_s,_a,_t,_e)
-#define stbma(_s, _t, _a, _e) def_store_ai_insn(stbs,1,"r",_s,_a,_t,_e)
-#define ldwma(_s, _a, _t, _e) def_load_ai_insn(ldw,4,"=r",_s,_a,_t,_e)
-#define stwma(_s, _t, _a, _e) def_store_ai_insn(stw,4,"r",_s,_a,_t,_e)
-#define flddma(_s, _a, _t, _e) def_load_ai_insn(fldd,8,"=f",_s,_a,_t,_e)
-#define fstdma(_s, _t, _a, _e) def_store_ai_insn(fstd,8,"f",_s,_a,_t,_e)
-
-#define def_load_insn(_insn,_tt,_s,_o,_a,_t,_e) 	\
-	__asm__ __volatile__ (				\
-	"1:\t" #_insn " " #_o "(" _s ",%1), %0\n\t"	\
-	ASM_EXCEPTIONTABLE_ENTRY(1b,_e)			\
-	: _tt(_t) 					\
-	: "r"(_a)					\
-	: "r8")
-
-#define def_store_insn(_insn,_tt,_s,_t,_o,_a,_e) 	\
-	__asm__ __volatile__ (				\
-	"1:\t" #_insn " %0, " #_o "(" _s ",%1)\n\t" 	\
-	ASM_EXCEPTIONTABLE_ENTRY(1b,_e)			\
-	: 						\
-	: _tt(_t), "r"(_a)				\
-	: "r8")
-
-#define ldw(_s,_o,_a,_t,_e)	def_load_insn(ldw,"=r",_s,_o,_a,_t,_e)
-#define stw(_s,_t,_o,_a,_e) 	def_store_insn(stw,"r",_s,_t,_o,_a,_e)
-
-#ifdef  CONFIG_PREFETCH
-static inline void prefetch_src(const void *addr)
-{
-	__asm__("ldw 0(" s_space ",%0), %%r0" : : "r" (addr));
-}
-
-static inline void prefetch_dst(const void *addr)
-{
-	__asm__("ldd 0(" d_space ",%0), %%r0" : : "r" (addr));
-}
-#else
-#define prefetch_src(addr) do { } while(0)
-#define prefetch_dst(addr) do { } while(0)
-#endif
-
-#define PA_MEMCPY_OK		0
-#define PA_MEMCPY_LOAD_ERROR	1
-#define PA_MEMCPY_STORE_ERROR	2
-
-/* Copy from a not-aligned src to an aligned dst, using shifts. Handles 4 words
- * per loop.  This code is derived from glibc. 
- */
-static noinline unsigned long copy_dstaligned(unsigned long dst,
-					unsigned long src, unsigned long len)
-{
-	/* gcc complains that a2 and a3 may be uninitialized, but actually
-	 * they cannot be.  Initialize a2/a3 to shut gcc up.
-	 */
-	register unsigned int a0, a1, a2 = 0, a3 = 0;
-	int sh_1, sh_2;
-
-	/* prefetch_src((const void *)src); */
-
-	/* Calculate how to shift a word read at the memory operation
-	   aligned srcp to make it aligned for copy.  */
-	sh_1 = 8 * (src % sizeof(unsigned int));
-	sh_2 = 8 * sizeof(unsigned int) - sh_1;
-
-	/* Make src aligned by rounding it down.  */
-	src &= -sizeof(unsigned int);
-
-	switch (len % 4)
-	{
-		case 2:
-			/* a1 = ((unsigned int *) src)[0];
-			   a2 = ((unsigned int *) src)[1]; */
-			ldw(s_space, 0, src, a1, cda_ldw_exc);
-			ldw(s_space, 4, src, a2, cda_ldw_exc);
-			src -= 1 * sizeof(unsigned int);
-			dst -= 3 * sizeof(unsigned int);
-			len += 2;
-			goto do1;
-		case 3:
-			/* a0 = ((unsigned int *) src)[0];
-			   a1 = ((unsigned int *) src)[1]; */
-			ldw(s_space, 0, src, a0, cda_ldw_exc);
-			ldw(s_space, 4, src, a1, cda_ldw_exc);
-			src -= 0 * sizeof(unsigned int);
-			dst -= 2 * sizeof(unsigned int);
-			len += 1;
-			goto do2;
-		case 0:
-			if (len == 0)
-				return PA_MEMCPY_OK;
-			/* a3 = ((unsigned int *) src)[0];
-			   a0 = ((unsigned int *) src)[1]; */
-			ldw(s_space, 0, src, a3, cda_ldw_exc);
-			ldw(s_space, 4, src, a0, cda_ldw_exc);
-			src -=-1 * sizeof(unsigned int);
-			dst -= 1 * sizeof(unsigned int);
-			len += 0;
-			goto do3;
-		case 1:
-			/* a2 = ((unsigned int *) src)[0];
-			   a3 = ((unsigned int *) src)[1]; */
-			ldw(s_space, 0, src, a2, cda_ldw_exc);
-			ldw(s_space, 4, src, a3, cda_ldw_exc);
-			src -=-2 * sizeof(unsigned int);
-			dst -= 0 * sizeof(unsigned int);
-			len -= 1;
-			if (len == 0)
-				goto do0;
-			goto do4;			/* No-op.  */
-	}
-
-	do
-	{
-		/* prefetch_src((const void *)(src + 4 * sizeof(unsigned int))); */
-do4:
-		/* a0 = ((unsigned int *) src)[0]; */
-		ldw(s_space, 0, src, a0, cda_ldw_exc);
-		/* ((unsigned int *) dst)[0] = MERGE (a2, sh_1, a3, sh_2); */
-		stw(d_space, MERGE (a2, sh_1, a3, sh_2), 0, dst, cda_stw_exc);
-do3:
-		/* a1 = ((unsigned int *) src)[1]; */
-		ldw(s_space, 4, src, a1, cda_ldw_exc);
-		/* ((unsigned int *) dst)[1] = MERGE (a3, sh_1, a0, sh_2); */
-		stw(d_space, MERGE (a3, sh_1, a0, sh_2), 4, dst, cda_stw_exc);
-do2:
-		/* a2 = ((unsigned int *) src)[2]; */
-		ldw(s_space, 8, src, a2, cda_ldw_exc);
-		/* ((unsigned int *) dst)[2] = MERGE (a0, sh_1, a1, sh_2); */
-		stw(d_space, MERGE (a0, sh_1, a1, sh_2), 8, dst, cda_stw_exc);
-do1:
-		/* a3 = ((unsigned int *) src)[3]; */
-		ldw(s_space, 12, src, a3, cda_ldw_exc);
-		/* ((unsigned int *) dst)[3] = MERGE (a1, sh_1, a2, sh_2); */
-		stw(d_space, MERGE (a1, sh_1, a2, sh_2), 12, dst, cda_stw_exc);
-
-		src += 4 * sizeof(unsigned int);
-		dst += 4 * sizeof(unsigned int);
-		len -= 4;
-	}
-	while (len != 0);
-
-do0:
-	/* ((unsigned int *) dst)[0] = MERGE (a2, sh_1, a3, sh_2); */
-	stw(d_space, MERGE (a2, sh_1, a3, sh_2), 0, dst, cda_stw_exc);
-
-	preserve_branch(handle_load_error);
-	preserve_branch(handle_store_error);
-
-	return PA_MEMCPY_OK;
-
-handle_load_error:
-	__asm__ __volatile__ ("cda_ldw_exc:\n");
-	return PA_MEMCPY_LOAD_ERROR;
-
-handle_store_error:
-	__asm__ __volatile__ ("cda_stw_exc:\n");
-	return PA_MEMCPY_STORE_ERROR;
-}
-
-
-/* Returns PA_MEMCPY_OK, PA_MEMCPY_LOAD_ERROR or PA_MEMCPY_STORE_ERROR.
- * In case of an access fault the faulty address can be read from the per_cpu
- * exception data struct. */
-static noinline unsigned long pa_memcpy_internal(void *dstp, const void *srcp,
-					unsigned long len)
-{
-	register unsigned long src, dst, t1, t2, t3;
-	register unsigned char *pcs, *pcd;
-	register unsigned int *pws, *pwd;
-	register double *pds, *pdd;
-	unsigned long ret;
-
-	src = (unsigned long)srcp;
-	dst = (unsigned long)dstp;
-	pcs = (unsigned char *)srcp;
-	pcd = (unsigned char *)dstp;
-
-	/* prefetch_src((const void *)srcp); */
-
-	if (len < THRESHOLD)
-		goto byte_copy;
-
-	/* Check alignment */
-	t1 = (src ^ dst);
-	if (unlikely(t1 & (sizeof(double)-1)))
-		goto unaligned_copy;
-
-	/* src and dst have same alignment. */
-
-	/* Copy bytes till we are double-aligned. */
-	t2 = src & (sizeof(double) - 1);
-	if (unlikely(t2 != 0)) {
-		t2 = sizeof(double) - t2;
-		while (t2 && len) {
-			/* *pcd++ = *pcs++; */
-			ldbma(s_space, pcs, t3, pmc_load_exc);
-			len--;
-			stbma(d_space, t3, pcd, pmc_store_exc);
-			t2--;
-		}
-	}
-
-	pds = (double *)pcs;
-	pdd = (double *)pcd;
-
-#if 0
-	/* Copy 8 doubles at a time */
-	while (len >= 8*sizeof(double)) {
-		register double r1, r2, r3, r4, r5, r6, r7, r8;
-		/* prefetch_src((char *)pds + L1_CACHE_BYTES); */
-		flddma(s_space, pds, r1, pmc_load_exc);
-		flddma(s_space, pds, r2, pmc_load_exc);
-		flddma(s_space, pds, r3, pmc_load_exc);
-		flddma(s_space, pds, r4, pmc_load_exc);
-		fstdma(d_space, r1, pdd, pmc_store_exc);
-		fstdma(d_space, r2, pdd, pmc_store_exc);
-		fstdma(d_space, r3, pdd, pmc_store_exc);
-		fstdma(d_space, r4, pdd, pmc_store_exc);
-
-#if 0
-		if (L1_CACHE_BYTES <= 32)
-			prefetch_src((char *)pds + L1_CACHE_BYTES);
-#endif
-		flddma(s_space, pds, r5, pmc_load_exc);
-		flddma(s_space, pds, r6, pmc_load_exc);
-		flddma(s_space, pds, r7, pmc_load_exc);
-		flddma(s_space, pds, r8, pmc_load_exc);
-		fstdma(d_space, r5, pdd, pmc_store_exc);
-		fstdma(d_space, r6, pdd, pmc_store_exc);
-		fstdma(d_space, r7, pdd, pmc_store_exc);
-		fstdma(d_space, r8, pdd, pmc_store_exc);
-		len -= 8*sizeof(double);
-	}
-#endif
-
-	pws = (unsigned int *)pds;
-	pwd = (unsigned int *)pdd;
-
-word_copy:
-	while (len >= 8*sizeof(unsigned int)) {
-		register unsigned int r1,r2,r3,r4,r5,r6,r7,r8;
-		/* prefetch_src((char *)pws + L1_CACHE_BYTES); */
-		ldwma(s_space, pws, r1, pmc_load_exc);
-		ldwma(s_space, pws, r2, pmc_load_exc);
-		ldwma(s_space, pws, r3, pmc_load_exc);
-		ldwma(s_space, pws, r4, pmc_load_exc);
-		stwma(d_space, r1, pwd, pmc_store_exc);
-		stwma(d_space, r2, pwd, pmc_store_exc);
-		stwma(d_space, r3, pwd, pmc_store_exc);
-		stwma(d_space, r4, pwd, pmc_store_exc);
-
-		ldwma(s_space, pws, r5, pmc_load_exc);
-		ldwma(s_space, pws, r6, pmc_load_exc);
-		ldwma(s_space, pws, r7, pmc_load_exc);
-		ldwma(s_space, pws, r8, pmc_load_exc);
-		stwma(d_space, r5, pwd, pmc_store_exc);
-		stwma(d_space, r6, pwd, pmc_store_exc);
-		stwma(d_space, r7, pwd, pmc_store_exc);
-		stwma(d_space, r8, pwd, pmc_store_exc);
-		len -= 8*sizeof(unsigned int);
-	}
-
-	while (len >= 4*sizeof(unsigned int)) {
-		register unsigned int r1,r2,r3,r4;
-		ldwma(s_space, pws, r1, pmc_load_exc);
-		ldwma(s_space, pws, r2, pmc_load_exc);
-		ldwma(s_space, pws, r3, pmc_load_exc);
-		ldwma(s_space, pws, r4, pmc_load_exc);
-		stwma(d_space, r1, pwd, pmc_store_exc);
-		stwma(d_space, r2, pwd, pmc_store_exc);
-		stwma(d_space, r3, pwd, pmc_store_exc);
-		stwma(d_space, r4, pwd, pmc_store_exc);
-		len -= 4*sizeof(unsigned int);
-	}
-
-	pcs = (unsigned char *)pws;
-	pcd = (unsigned char *)pwd;
-
-byte_copy:
-	while (len) {
-		/* *pcd++ = *pcs++; */
-		ldbma(s_space, pcs, t3, pmc_load_exc);
-		stbma(d_space, t3, pcd, pmc_store_exc);
-		len--;
-	}
-
-	return PA_MEMCPY_OK;
-
-unaligned_copy:
-	/* possibly we are aligned on a word, but not on a double... */
-	if (likely((t1 & (sizeof(unsigned int)-1)) == 0)) {
-		t2 = src & (sizeof(unsigned int) - 1);
-
-		if (unlikely(t2 != 0)) {
-			t2 = sizeof(unsigned int) - t2;
-			while (t2) {
-				/* *pcd++ = *pcs++; */
-				ldbma(s_space, pcs, t3, pmc_load_exc);
-				stbma(d_space, t3, pcd, pmc_store_exc);
-				len--;
-				t2--;
-			}
-		}
-
-		pws = (unsigned int *)pcs;
-		pwd = (unsigned int *)pcd;
-		goto word_copy;
-	}
-
-	/* Align the destination.  */
-	if (unlikely((dst & (sizeof(unsigned int) - 1)) != 0)) {
-		t2 = sizeof(unsigned int) - (dst & (sizeof(unsigned int) - 1));
-		while (t2) {
-			/* *pcd++ = *pcs++; */
-			ldbma(s_space, pcs, t3, pmc_load_exc);
-			stbma(d_space, t3, pcd, pmc_store_exc);
-			len--;
-			t2--;
-		}
-		dst = (unsigned long)pcd;
-		src = (unsigned long)pcs;
-	}
-
-	ret = copy_dstaligned(dst, src, len / sizeof(unsigned int));
-	if (ret)
-		return ret;
-
-	pcs += (len & -sizeof(unsigned int));
-	pcd += (len & -sizeof(unsigned int));
-	len %= sizeof(unsigned int);
-
-	preserve_branch(handle_load_error);
-	preserve_branch(handle_store_error);
-
-	goto byte_copy;
-
-handle_load_error:
-	__asm__ __volatile__ ("pmc_load_exc:\n");
-	return PA_MEMCPY_LOAD_ERROR;
-
-handle_store_error:
-	__asm__ __volatile__ ("pmc_store_exc:\n");
-	return PA_MEMCPY_STORE_ERROR;
-}
-
-
 /* Returns 0 for success, otherwise, returns number of bytes not transferred. */
-static unsigned long pa_memcpy(void *dstp, const void *srcp, unsigned long len)
-{
-	unsigned long ret, fault_addr, reference;
-	struct exception_data *d;
-
-	ret = pa_memcpy_internal(dstp, srcp, len);
-	if (likely(ret == PA_MEMCPY_OK))
-		return 0;
-
-	/* if a load or store fault occured we can get the faulty addr */
-	d = this_cpu_ptr(&exception_data);
-	fault_addr = d->fault_addr;
-
-	/* error in load or store? */
-	if (ret == PA_MEMCPY_LOAD_ERROR)
-		reference = (unsigned long) srcp;
-	else
-		reference = (unsigned long) dstp;
+extern unsigned long pa_memcpy(void *dst, const void *src,
+				unsigned long len);
 
-	DPRINTF("pa_memcpy: fault type = %lu, len=%lu fault_addr=%lu ref=%lu\n",
-		ret, len, fault_addr, reference);
-
-	if (fault_addr >= reference)
-		return len - (fault_addr - reference);
-	else
-		return len;
-}
-
-#ifdef __KERNEL__
 unsigned long __copy_to_user(void __user *dst, const void *src,
 			     unsigned long len)
 {
@@ -537,5 +84,3 @@ long probe_kernel_read(void *dst, const void *src, size_t size)
 
 	return __probe_kernel_read(dst, src, size);
 }
-
-#endif