Initial import

9 files changed, 1271 insertions, 0 deletions

diff --git a/arch/cris/arch-v32/lib/Makefile b/arch/cris/arch-v32/lib/Makefile
new file mode 100644
index 000000000..e91cf02f6
--- /dev/null
+++ b/arch/cris/arch-v32/lib/Makefile
@@ -0,0 +1,7 @@
+#
+# Makefile for Etrax-specific library files..
+#
+
+lib-y  = checksum.o checksumcopy.o string.o usercopy.o memset.o \
+	csumcpfruser.o delay.o strcmp.o
+
diff --git a/arch/cris/arch-v32/lib/checksum.S b/arch/cris/arch-v32/lib/checksum.S
new file mode 100644
index 000000000..4a72a94a4
--- /dev/null
+++ b/arch/cris/arch-v32/lib/checksum.S
@@ -0,0 +1,88 @@
+/*
+ * A fast checksum routine using movem
+ * Copyright (c) 1998-2007 Axis Communications AB
+ *
+ * csum_partial(const unsigned char * buff, int len, unsigned int sum)
+ */
+
+	.globl	csum_partial
+	.type   csum_partial,@function
+csum_partial:
+
+	;; r10 - src
+	;; r11 - length
+	;; r12 - checksum
+
+	;; Optimized for large packets
+	subq	10*4, $r11
+	blt	_word_loop
+	move.d	$r11, $acr
+
+	subq	9*4,$sp
+	clearf	c
+	movem	$r8,[$sp]
+
+	;; do a movem checksum
+
+_mloop:	movem	[$r10+],$r9	; read 10 longwords
+	;; Loop count without touching the c flag.
+	addoq	-10*4, $acr, $acr
+	;; perform dword checksumming on the 10 longwords
+
+	addc	$r0,$r12
+	addc	$r1,$r12
+	addc	$r2,$r12
+	addc	$r3,$r12
+	addc	$r4,$r12
+	addc	$r5,$r12
+	addc	$r6,$r12
+	addc	$r7,$r12
+	addc	$r8,$r12
+	addc	$r9,$r12
+
+	;; test $acr without trashing carry.
+	move.d	$acr, $acr
+	bpl	_mloop
+	;; r11 <= acr  is not really needed in the mloop, just using the dslot
+	;; to prepare for what is needed after mloop.
+	move.d	$acr, $r11
+
+	;; fold the last carry into r13
+	addc	0, $r12
+	movem	[$sp+],$r8	; restore regs
+
+_word_loop:
+	addq	10*4,$r11	; compensate for last loop underflowing length
+
+	moveq	-1,$r9		; put 0xffff in r9, faster than move.d 0xffff,r9
+	lsrq	16,$r9
+
+	move.d	$r12,$r13
+	lsrq	16,$r13		; r13 = checksum >> 16
+	and.d	$r9,$r12	; checksum = checksum & 0xffff
+
+_no_fold:
+	subq	2,$r11
+	blt	_no_words
+	add.d	$r13,$r12	; checksum += r13
+
+	;; checksum the rest of the words
+_wloop:	subq	2,$r11
+	bge	_wloop
+	addu.w	[$r10+],$r12
+
+_no_words:
+	addq	2,$r11
+	;; see if we have one odd byte more
+	bne	_do_byte
+	nop
+	ret
+	move.d	$r12,$r10
+
+_do_byte:
+	;; copy and checksum the last byte
+	addu.b	[$r10],$r12
+	ret
+	move.d	$r12,$r10
+
+	.size   csum_partial, .-csum_partial
diff --git a/arch/cris/arch-v32/lib/checksumcopy.S b/arch/cris/arch-v32/lib/checksumcopy.S
new file mode 100644
index 000000000..54e209f18
--- /dev/null
+++ b/arch/cris/arch-v32/lib/checksumcopy.S
@@ -0,0 +1,94 @@
+/*
+ * A fast checksum+copy routine using movem
+ * Copyright (c) 1998-2007 Axis Communications AB
+ *
+ * Authors:	Bjorn Wesen
+ *
+ * csum_partial_copy_nocheck(const char *src, char *dst,
+ *		             int len, unsigned int sum)
+ */
+
+	.globl	csum_partial_copy_nocheck
+	.type   csum_partial_copy_nocheck,@function
+csum_partial_copy_nocheck:
+
+	;; r10 - src
+	;; r11 - dst
+	;; r12 - length
+	;; r13 - checksum
+
+	;; Optimized for large packets
+	subq	10*4, $r12
+	blt	_word_loop
+	move.d	$r12, $acr
+
+	subq	9*4,$sp
+	clearf	c
+	movem	$r8,[$sp]
+
+	;; do a movem copy and checksum
+1:	;; A failing userspace access (the read) will have this as PC.
+_mloop:	movem	[$r10+],$r9	; read 10 longwords
+	addoq	-10*4, $acr, $acr ; loop counter in latency cycle
+	movem	$r9,[$r11+]	; write 10 longwords
+
+	;; perform dword checksumming on the 10 longwords
+	addc	$r0,$r13
+	addc	$r1,$r13
+	addc	$r2,$r13
+	addc	$r3,$r13
+	addc	$r4,$r13
+	addc	$r5,$r13
+	addc	$r6,$r13
+	addc	$r7,$r13
+	addc	$r8,$r13
+	addc	$r9,$r13
+
+	;; test $acr, without trashing carry.
+	move.d	$acr, $acr
+	bpl	_mloop
+	;; r12 <= acr  is needed after mloop and in the exception handlers.
+	move.d	$acr, $r12
+
+	;; fold the last carry into r13
+	addc	0, $r13
+	movem	[$sp+],$r8	; restore regs
+
+_word_loop:
+	addq	10*4,$r12	; compensate for last loop underflowing length
+
+	;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below
+	;; r9 can be used as temporary.
+	move.d	$r13,$r9
+	lsrq	16,$r9		; r0 = checksum >> 16
+	and.d	0xffff,$r13	; checksum = checksum & 0xffff
+
+	subq	2, $r12
+	blt	_no_words
+	add.d	$r9,$r13	; checksum += r0
+
+	;; copy and checksum the rest of the words
+2:	;; A failing userspace access for the read below will have this as PC.
+_wloop:	move.w	[$r10+],$r9
+	addu.w	$r9,$r13
+	subq	2,$r12
+	bge	_wloop
+	move.w	$r9,[$r11+]
+
+_no_words:
+	addq	2,$r12
+	bne	_do_byte
+	nop
+	ret
+	move.d	$r13,$r10
+
+_do_byte:
+	;; copy and checksum the last byte
+3:	;; A failing userspace access for the read below will have this as PC.
+	move.b	[$r10],$r9
+	addu.b	$r9,$r13
+	move.b	$r9,[$r11]
+	ret
+	move.d	$r13,$r10
+
+	.size   csum_partial_copy_nocheck, . - csum_partial_copy_nocheck
diff --git a/arch/cris/arch-v32/lib/csumcpfruser.S b/arch/cris/arch-v32/lib/csumcpfruser.S
new file mode 100644
index 000000000..600ec16b9
--- /dev/null
+++ b/arch/cris/arch-v32/lib/csumcpfruser.S
@@ -0,0 +1,69 @@
+/*
+ * Add-on to transform csum_partial_copy_nocheck in checksumcopy.S into
+ * csum_partial_copy_from_user by adding exception records.
+ *
+ * Copyright (C) 2001, 2003 Axis Communications AB.
+ *
+ * Author: Hans-Peter Nilsson.
+ */
+
+#include <asm/errno.h>
+
+/* Same function body, but a different name.  If we just added exception
+   records to _csum_partial_copy_nocheck and made it generic, we wouldn't
+   know a user fault from a kernel fault and we would have overhead in
+   each kernel caller for the error-pointer argument.
+
+   unsigned int csum_partial_copy_from_user
+     (const char *src, char *dst, int len, unsigned int sum, int *errptr);
+
+   Note that the errptr argument is only set if we encounter an error.
+   It is conveniently located on the stack, so the normal function body
+   does not have to handle it.  */
+
+#define csum_partial_copy_nocheck csum_partial_copy_from_user
+
+/* There are local labels numbered 1, 2 and 3 present to mark the
+   different from-user accesses.  */
+#include "checksumcopy.S"
+
+	.section .fixup,"ax"
+
+;; Here from the movem loop; restore stack.
+4:
+	movem	[$sp+],$r8
+;; r12 is already decremented.  Add back chunk_size-2.
+	addq	40-2,$r12
+
+;; Here from the word loop; r12 is off by 2; add it back.
+5:
+	addq	2,$r12
+
+;; Here from a failing single byte.
+6:
+
+;; Signal in *errptr that we had a failing access.
+	move.d	[$sp],$acr
+	moveq	-EFAULT,$r9
+	subq	4,$sp
+	move.d	$r9,[$acr]
+
+;; Clear the rest of the destination area using memset.  Preserve the
+;; checksum for the readable bytes.
+	move.d	$r13,[$sp]
+	subq	4,$sp
+	move.d	$r11,$r10
+	move	$srp,[$sp]
+	jsr	memset
+	clear.d	$r11
+
+	move	[$sp+],$srp
+	ret
+	move.d	[$sp+],$r10
+
+	.previous
+	.section __ex_table,"a"
+	.dword 1b,4b
+	.dword 2b,5b
+	.dword 3b,6b
+	.previous
diff --git a/arch/cris/arch-v32/lib/delay.c b/arch/cris/arch-v32/lib/delay.c
new file mode 100644
index 000000000..39f1ac999
--- /dev/null
+++ b/arch/cris/arch-v32/lib/delay.c
@@ -0,0 +1,28 @@
+/*
+ * Precise Delay Loops for ETRAX FS
+ *
+ * Copyright (C) 2006 Axis Communications AB.
+ *
+ */
+
+#include <hwregs/reg_map.h>
+#include <hwregs/reg_rdwr.h>
+#include <hwregs/timer_defs.h>
+#include <linux/types.h>
+#include <linux/delay.h>
+#include <linux/module.h>
+
+/*
+ * On ETRAX FS, we can check the free-running read-only 100MHz timer
+ * getting 32-bit 10ns precision, theoretically good for 42.94967295
+ * seconds.  Unsigned arithmetic and careful expression handles
+ * wrapping.
+ */
+
+void cris_delay10ns(u32 n10ns)
+{
+	u32 t0 = REG_RD(timer, regi_timer0, r_time);
+	while (REG_RD(timer, regi_timer0, r_time) - t0 < n10ns)
+		;
+}
+EXPORT_SYMBOL(cris_delay10ns);
diff --git a/arch/cris/arch-v32/lib/memset.c b/arch/cris/arch-v32/lib/memset.c
new file mode 100644
index 000000000..c94ea9b3e
--- /dev/null
+++ b/arch/cris/arch-v32/lib/memset.c
@@ -0,0 +1,259 @@
+/* A memset for CRIS.
+   Copyright (C) 1999-2005 Axis Communications.
+   All rights reserved.
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   1. Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+   2. Neither the name of Axis Communications nor the names of its
+      contributors may be used to endorse or promote products derived
+      from this software without specific prior written permission.
+
+   THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
+   COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
+   INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+   (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+   SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+   HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+   STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+   IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+   POSSIBILITY OF SUCH DAMAGE.  */
+
+/* FIXME: This file should really only be used for reference, as the
+   result is somewhat depending on gcc generating what we expect rather
+   than what we describe.  An assembly file should be used instead.  */
+
+/* Note the multiple occurrence of the expression "12*4", including the
+   asm.  It is hard to get it into the asm in a good way.  Thus better to
+   expose the problem everywhere: no macro.  */
+
+/* Assuming one cycle per dword written or read (ok, not really true; the
+   world is not ideal), and one cycle per instruction, then 43+3*(n/48-1)
+   <= 24+24*(n/48-1) so n >= 45.7; n >= 0.9; we win on the first full
+   48-byte block to set.  */
+
+#define MEMSET_BY_BLOCK_THRESHOLD (1 * 48)
+
+/* No name ambiguities in this file.  */
+__asm__ (".syntax no_register_prefix");
+
+void *memset(void *pdst, int c, unsigned int plen)
+{
+  /* Now we want the parameters in special registers.  Make sure the
+     compiler does something usable with this.  */
+
+  register char *return_dst __asm__ ("r10") = pdst;
+  register int n __asm__ ("r12") = plen;
+  register int lc __asm__ ("r11") = c;
+
+  /* Most apps use memset sanely.  Memsetting about 3..4 bytes or less get
+     penalized here compared to the generic implementation.  */
+
+  /* This is fragile performancewise at best.  Check with newer GCC
+     releases, if they compile cascaded "x |= x << 8" to sane code.  */
+  __asm__("movu.b %0,r13						\n\
+	   lslq 8,r13							\n\
+	   move.b %0,r13						\n\
+	   move.d r13,%0						\n\
+	   lslq 16,r13							\n\
+	   or.d r13,%0"
+          : "=r" (lc)		/* Inputs.  */
+	  : "0" (lc)		/* Outputs.  */
+	  : "r13");		/* Trash.  */
+
+  {
+    register char *dst __asm__ ("r13") = pdst;
+
+    if (((unsigned long) pdst & 3) != 0
+	/* Oops! n = 0 must be a valid call, regardless of alignment.  */
+	&& n >= 3)
+      {
+	if ((unsigned long) dst & 1)
+	  {
+	    *dst = (char) lc;
+	    n--;
+	    dst++;
+	  }
+
+	if ((unsigned long) dst & 2)
+	  {
+	    *(short *) dst = lc;
+	    n -= 2;
+	    dst += 2;
+	  }
+      }
+
+    /* Decide which setting method to use.  */
+    if (n >= MEMSET_BY_BLOCK_THRESHOLD)
+      {
+	/* It is not optimal to tell the compiler about clobbering any
+	   registers; that will move the saving/restoring of those registers
+	   to the function prologue/epilogue, and make non-block sizes
+	   suboptimal.  */
+	__asm__ volatile
+	  ("\
+	   ;; GCC does promise correct register allocations, but let's	\n\
+	   ;; make sure it keeps its promises.				\n\
+	   .ifnc %0-%1-%4,$r13-$r12-$r11				\n\
+	   .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\"	\n\
+	   .endif							\n\
+									\n\
+	   ;; Save the registers we'll clobber in the movem process	\n\
+	   ;; on the stack.  Don't mention them to gcc, it will only be	\n\
+	   ;; upset.							\n\
+	   subq	   11*4,sp						\n\
+	   movem   r10,[sp]						\n\
+									\n\
+	   move.d  r11,r0						\n\
+	   move.d  r11,r1						\n\
+	   move.d  r11,r2						\n\
+	   move.d  r11,r3						\n\
+	   move.d  r11,r4						\n\
+	   move.d  r11,r5						\n\
+	   move.d  r11,r6						\n\
+	   move.d  r11,r7						\n\
+	   move.d  r11,r8						\n\
+	   move.d  r11,r9						\n\
+	   move.d  r11,r10						\n\
+									\n\
+	   ;; Now we've got this:					\n\
+	   ;; r13 - dst							\n\
+	   ;; r12 - n							\n\
+									\n\
+	   ;; Update n for the first loop				\n\
+	   subq	   12*4,r12						\n\
+0:									\n\
+"
+#ifdef __arch_common_v10_v32
+	   /* Cater to branch offset difference between v32 and v10.  We
+	      assume the branch below has an 8-bit offset.  */
+"	   setf\n"
+#endif
+"	   subq	  12*4,r12						\n\
+	   bge	   0b							\n\
+	   movem	r11,[r13+]					\n\
+									\n\
+	   ;; Compensate for last loop underflowing n.			\n\
+	   addq	  12*4,r12						\n\
+									\n\
+	   ;; Restore registers from stack.				\n\
+	   movem [sp+],r10"
+
+	   /* Outputs.	*/
+	   : "=r" (dst), "=r" (n)
+
+	   /* Inputs.  */
+	   : "0" (dst), "1" (n), "r" (lc));
+      }
+
+    /* An ad-hoc unroll, used for 4*12-1..16 bytes. */
+    while (n >= 16)
+      {
+	*(long *) dst = lc; dst += 4;
+	*(long *) dst = lc; dst += 4;
+	*(long *) dst = lc; dst += 4;
+	*(long *) dst = lc; dst += 4;
+	n -= 16;
+      }
+
+    switch (n)
+      {
+      case 0:
+        break;
+
+      case 1:
+        *dst = (char) lc;
+        break;
+
+      case 2:
+        *(short *) dst = (short) lc;
+        break;
+
+      case 3:
+        *(short *) dst = (short) lc; dst += 2;
+        *dst = (char) lc;
+        break;
+
+      case 4:
+        *(long *) dst = lc;
+        break;
+
+      case 5:
+        *(long *) dst = lc; dst += 4;
+        *dst = (char) lc;
+        break;
+
+      case 6:
+        *(long *) dst = lc; dst += 4;
+        *(short *) dst = (short) lc;
+        break;
+
+      case 7:
+        *(long *) dst = lc; dst += 4;
+        *(short *) dst = (short) lc; dst += 2;
+        *dst = (char) lc;
+        break;
+
+      case 8:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc;
+        break;
+
+      case 9:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *dst = (char) lc;
+        break;
+
+      case 10:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(short *) dst = (short) lc;
+        break;
+
+      case 11:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(short *) dst = (short) lc; dst += 2;
+        *dst = (char) lc;
+        break;
+
+      case 12:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc;
+        break;
+
+      case 13:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *dst = (char) lc;
+        break;
+
+      case 14:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(short *) dst = (short) lc;
+        break;
+
+      case 15:
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(long *) dst = lc; dst += 4;
+        *(short *) dst = (short) lc; dst += 2;
+        *dst = (char) lc;
+        break;
+      }
+  }
+
+  return return_dst;
+}
diff --git a/arch/cris/arch-v32/lib/strcmp.S b/arch/cris/arch-v32/lib/strcmp.S
new file mode 100644
index 000000000..8f7a1ee62
--- /dev/null
+++ b/arch/cris/arch-v32/lib/strcmp.S
@@ -0,0 +1,21 @@
+; strcmp.S -- CRISv32 version.
+; Copyright (C) 2008 AXIS Communications AB
+; Written by Edgar E. Iglesias
+;
+; This source code is licensed under the GNU General Public License,
+; Version 2.  See the file COPYING for more details.
+
+	.global	strcmp
+	.type	strcmp,@function
+strcmp:
+1:
+	move.b	[$r10+], $r12
+	seq	$r13
+	sub.b	[$r11+], $r12
+	or.b	$r12, $r13
+	beq	1b
+	nop
+
+	ret
+	movs.b	$r12, $r10
+	.size	strcmp, . - strcmp
diff --git a/arch/cris/arch-v32/lib/string.c b/arch/cris/arch-v32/lib/string.c
new file mode 100644
index 000000000..c7bd6ebdc
--- /dev/null
+++ b/arch/cris/arch-v32/lib/string.c
@@ -0,0 +1,236 @@
+/* A memcpy for CRIS.
+   Copyright (C) 1994-2005 Axis Communications.
+   All rights reserved.
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   1. Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+   2. Neither the name of Axis Communications nor the names of its
+      contributors may be used to endorse or promote products derived
+      from this software without specific prior written permission.
+
+   THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
+   COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
+   INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+   (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+   SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+   HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+   STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+   IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+   POSSIBILITY OF SUCH DAMAGE.  */
+
+/* FIXME: This file should really only be used for reference, as the
+   result is somewhat depending on gcc generating what we expect rather
+   than what we describe.  An assembly file should be used instead.  */
+
+#include <stddef.h>
+
+/* Break even between movem and move16 is really at 38.7 * 2, but
+   modulo 44, so up to the next multiple of 44, we use ordinary code.  */
+#define MEMCPY_BY_BLOCK_THRESHOLD (44 * 2)
+
+/* No name ambiguities in this file.  */
+__asm__ (".syntax no_register_prefix");
+
+void *
+memcpy(void *pdst, const void *psrc, size_t pn)
+{
+  /* Now we want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+     As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     If gcc was allright, it really would need no temporaries, and no
+     stack space to save stuff on.  */
+
+  register void *return_dst __asm__ ("r10") = pdst;
+  register unsigned char *dst __asm__ ("r13") = pdst;
+  register unsigned const char *src __asm__ ("r11") = psrc;
+  register int n __asm__ ("r12") = pn;
+
+  /* When src is aligned but not dst, this makes a few extra needless
+     cycles.  I believe it would take as many to check that the
+     re-alignment was unnecessary.  */
+  if (((unsigned long) dst & 3) != 0
+      /* Don't align if we wouldn't copy more than a few bytes; so we
+	 don't have to check further for overflows.  */
+      && n >= 3)
+  {
+    if ((unsigned long) dst & 1)
+      {
+	n--;
+	*dst = *src;
+	src++;
+	dst++;
+      }
+
+    if ((unsigned long) dst & 2)
+      {
+	n -= 2;
+	*(short *) dst = *(short *) src;
+	src += 2;
+	dst += 2;
+      }
+  }
+
+  /* Decide which copying method to use.  */
+  if (n >= MEMCPY_BY_BLOCK_THRESHOLD)
+    {
+      /* It is not optimal to tell the compiler about clobbering any
+	 registers; that will move the saving/restoring of those registers
+	 to the function prologue/epilogue, and make non-movem sizes
+	 suboptimal.  */
+      __asm__ volatile
+	("\
+	 ;; GCC does promise correct register allocations, but let's	\n\
+	 ;; make sure it keeps its promises.				\n\
+	 .ifnc %0-%1-%2,$r13-$r11-$r12					\n\
+	 .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\"	\n\
+	 .endif								\n\
+									\n\
+	 ;; Save the registers we'll use in the movem process		\n\
+	 ;; on the stack.						\n\
+	 subq	11*4,sp							\n\
+	 movem	r10,[sp]						\n\
+									\n\
+	 ;; Now we've got this:						\n\
+	 ;; r11 - src							\n\
+	 ;; r13 - dst							\n\
+	 ;; r12 - n							\n\
+									\n\
+	 ;; Update n for the first loop.				\n\
+	 subq	 44,r12							\n\
+0:									\n\
+"
+#ifdef __arch_common_v10_v32
+	 /* Cater to branch offset difference between v32 and v10.  We
+	    assume the branch below has an 8-bit offset.  */
+"	 setf\n"
+#endif
+"	 movem	[r11+],r10						\n\
+	 subq	44,r12							\n\
+	 bge	 0b							\n\
+	 movem	r10,[r13+]						\n\
+									\n\
+	 ;; Compensate for last loop underflowing n.			\n\
+	 addq	44,r12							\n\
+									\n\
+	 ;; Restore registers from stack.				\n\
+	 movem [sp+],r10"
+
+	 /* Outputs.  */
+	 : "=r" (dst), "=r" (src), "=r" (n)
+
+	 /* Inputs.  */
+	 : "0" (dst), "1" (src), "2" (n));
+    }
+
+  while (n >= 16)
+    {
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+
+      n -= 16;
+    }
+
+  switch (n)
+    {
+    case 0:
+      break;
+
+    case 1:
+      *dst = *src;
+      break;
+
+    case 2:
+      *(short *) dst = *(short *) src;
+      break;
+
+    case 3:
+      *(short *) dst = *(short *) src; dst += 2; src += 2;
+      *dst = *src;
+      break;
+
+    case 4:
+      *(long *) dst = *(long *) src;
+      break;
+
+    case 5:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *dst = *src;
+      break;
+
+    case 6:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(short *) dst = *(short *) src;
+      break;
+
+    case 7:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(short *) dst = *(short *) src; dst += 2; src += 2;
+      *dst = *src;
+      break;
+
+    case 8:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src;
+      break;
+
+    case 9:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *dst = *src;
+      break;
+
+    case 10:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(short *) dst = *(short *) src;
+      break;
+
+    case 11:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(short *) dst = *(short *) src; dst += 2; src += 2;
+      *dst = *src;
+      break;
+
+    case 12:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src;
+      break;
+
+    case 13:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *dst = *src;
+      break;
+
+    case 14:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(short *) dst = *(short *) src;
+      break;
+
+    case 15:
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(long *) dst = *(long *) src; dst += 4; src += 4;
+      *(short *) dst = *(short *) src; dst += 2; src += 2;
+      *dst = *src;
+      break;
+    }
+
+  return return_dst;
+}
diff --git a/arch/cris/arch-v32/lib/usercopy.c b/arch/cris/arch-v32/lib/usercopy.c
new file mode 100644
index 000000000..f0f335d8a
--- /dev/null
+++ b/arch/cris/arch-v32/lib/usercopy.c
@@ -0,0 +1,469 @@
+/*
+ * User address space access functions.
+ * The non-inlined parts of asm-cris/uaccess.h are here.
+ *
+ * Copyright (C) 2000, 2003 Axis Communications AB.
+ *
+ * Written by Hans-Peter Nilsson.
+ * Pieces used from memcpy, originally by Kenny Ranerup long time ago.
+ */
+
+#include <asm/uaccess.h>
+
+/* Asm:s have been tweaked (within the domain of correctness) to give
+   satisfactory results for "gcc version 3.2.1 Axis release R53/1.53-v32".
+
+   Check regularly...
+
+   Note that for CRISv32, the PC saved at a bus-fault is the address
+   *at* the faulting instruction, with a special case for instructions
+   in delay slots: then it's the address of the branch.  Note also that
+   in contrast to v10, a postincrement in the instruction is *not*
+   performed at a bus-fault; the register is seen having the original
+   value in fault handlers.  */
+
+
+/* Copy to userspace.  This is based on the memcpy used for
+   kernel-to-kernel copying; see "string.c".  */
+
+unsigned long __copy_user(void __user *pdst, const void *psrc, unsigned long pn)
+{
+  /* We want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+     As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     FIXME: Comment for old gcc version.  Check.
+     If gcc was alright, it really would need no temporaries, and no
+     stack space to save stuff on. */
+
+  register char *dst __asm__ ("r13") = pdst;
+  register const char *src __asm__ ("r11") = psrc;
+  register int n __asm__ ("r12") = pn;
+  register int retn __asm__ ("r10") = 0;
+
+
+  /* When src is aligned but not dst, this makes a few extra needless
+     cycles.  I believe it would take as many to check that the
+     re-alignment was unnecessary.  */
+  if (((unsigned long) dst & 3) != 0
+      /* Don't align if we wouldn't copy more than a few bytes; so we
+	 don't have to check further for overflows.  */
+      && n >= 3)
+  {
+    if ((unsigned long) dst & 1)
+    {
+      __asm_copy_to_user_1 (dst, src, retn);
+      n--;
+    }
+
+    if ((unsigned long) dst & 2)
+    {
+      __asm_copy_to_user_2 (dst, src, retn);
+      n -= 2;
+    }
+  }
+
+  /* Movem is dirt cheap.  The overheap is low enough to always use the
+     minimum possible block size as the threshold.  */
+  if (n >= 44)
+  {
+    /* For large copies we use 'movem'.  */
+
+    /* It is not optimal to tell the compiler about clobbering any
+       registers; that will move the saving/restoring of those registers
+       to the function prologue/epilogue, and make non-movem sizes
+       suboptimal.  */
+    __asm__ volatile ("\
+        ;; Check that the register asm declaration got right.		\n\
+        ;; The GCC manual explicitly says TRT will happen.		\n\
+	.ifnc %0%1%2%3,$r13$r11$r12$r10					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll use in the movem process		\n\
+	;; on the stack.						\n\
+	subq	11*4,$sp						\n\
+	movem	$r10,[$sp]						\n\
+									\n\
+	;; Now we've got this:						\n\
+	;; r11 - src							\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+	;; Update n for the first loop					\n\
+	subq	44,$r12							\n\
+0:									\n\
+	movem	[$r11+],$r10						\n\
+	subq   44,$r12							\n\
+1:	bge	0b							\n\
+	movem	$r10,[$r13+]						\n\
+3:									\n\
+	addq   44,$r12  ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+	movem [$sp+],$r10						\n\
+2:									\n\
+	.section .fixup,\"ax\"						\n\
+4:									\n\
+; When failing on any of the 1..44 bytes in a chunk, we adjust back the	\n\
+; source pointer and just drop through	to the by-16 and by-4 loops to	\n\
+; get the correct number of failing bytes.  This necessarily means a	\n\
+; few extra exceptions, but invalid user pointers shouldn't happen in	\n\
+; time-critical code anyway.						\n\
+	jump 3b								\n\
+	subq 44,$r11							\n\
+									\n\
+	.previous							\n\
+	.section __ex_table,\"a\"					\n\
+	.dword 1b,4b							\n\
+	.previous"
+
+     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
+     /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
+
+  }
+
+  while (n >= 16)
+  {
+    __asm_copy_to_user_16 (dst, src, retn);
+    n -= 16;
+  }
+
+  /* Having a separate by-four loops cuts down on cache footprint.
+     FIXME:  Test with and without; increasing switch to be 0..15.  */
+  while (n >= 4)
+  {
+    __asm_copy_to_user_4 (dst, src, retn);
+    n -= 4;
+  }
+
+  switch (n)
+  {
+    case 0:
+      break;
+    case 1:
+      __asm_copy_to_user_1 (dst, src, retn);
+      break;
+    case 2:
+      __asm_copy_to_user_2 (dst, src, retn);
+      break;
+    case 3:
+      __asm_copy_to_user_3 (dst, src, retn);
+      break;
+  }
+
+  return retn;
+}
+EXPORT_SYMBOL(__copy_user);
+
+/* Copy from user to kernel, zeroing the bytes that were inaccessible in
+   userland.  The return-value is the number of bytes that were
+   inaccessible.  */
+unsigned long __copy_user_zeroing(void *pdst, const void __user *psrc,
+				  unsigned long pn)
+{
+  /* We want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+     As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     FIXME: Comment for old gcc version.  Check.
+     If gcc was alright, it really would need no temporaries, and no
+     stack space to save stuff on.  */
+
+  register char *dst __asm__ ("r13") = pdst;
+  register const char *src __asm__ ("r11") = psrc;
+  register int n __asm__ ("r12") = pn;
+  register int retn __asm__ ("r10") = 0;
+
+  /* The best reason to align src is that we then know that a read-fault
+     was for aligned bytes; there's no 1..3 remaining good bytes to
+     pickle.  */
+  if (((unsigned long) src & 3) != 0)
+  {
+    if (((unsigned long) src & 1) && n != 0)
+    {
+      __asm_copy_from_user_1 (dst, src, retn);
+      n--;
+    }
+
+    if (((unsigned long) src & 2) && n >= 2)
+    {
+      __asm_copy_from_user_2 (dst, src, retn);
+      n -= 2;
+    }
+
+    /* We only need one check after the unalignment-adjustments, because
+       if both adjustments were done, either both or neither reference
+       had an exception.  */
+    if (retn != 0)
+      goto copy_exception_bytes;
+  }
+
+  /* Movem is dirt cheap.  The overheap is low enough to always use the
+     minimum possible block size as the threshold.  */
+  if (n >= 44)
+  {
+    /* It is not optimal to tell the compiler about clobbering any
+       registers; that will move the saving/restoring of those registers
+       to the function prologue/epilogue, and make non-movem sizes
+       suboptimal.  */
+    __asm__ volatile ("\
+	.ifnc %0%1%2%3,$r13$r11$r12$r10					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll use in the movem process		\n\
+	;; on the stack.						\n\
+	subq	11*4,$sp						\n\
+	movem	$r10,[$sp]						\n\
+									\n\
+	;; Now we've got this:						\n\
+	;; r11 - src							\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+	;; Update n for the first loop					\n\
+	subq	44,$r12							\n\
+0:									\n\
+	movem	[$r11+],$r10						\n\
+									\n\
+	subq   44,$r12							\n\
+	bge	0b							\n\
+	movem	$r10,[$r13+]						\n\
+									\n\
+4:									\n\
+	addq   44,$r12  ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+	movem [$sp+],$r10						\n\
+	.section .fixup,\"ax\"						\n\
+									\n\
+;; Do not jump back into the loop if we fail.  For some uses, we get a	\n\
+;; page fault somewhere on the line.  Without checking for page limits,	\n\
+;; we don't know where, but we need to copy accurately and keep an	\n\
+;; accurate count; not just clear the whole line.  To do that, we fall	\n\
+;; down in the code below, proceeding with smaller amounts.  It should	\n\
+;; be kept in mind that we have to cater to code like what at one time	\n\
+;; was in fs/super.c:							\n\
+;;  i = size - copy_from_user((void *)page, data, size);		\n\
+;; which would cause repeated faults while clearing the remainder of	\n\
+;; the SIZE bytes at PAGE after the first fault.			\n\
+;; A caveat here is that we must not fall through from a failing page	\n\
+;; to a valid page.							\n\
+									\n\
+3:									\n\
+	jump	4b ;; Fall through, pretending the fault didn't happen.	\n\
+	nop								\n\
+									\n\
+	.previous							\n\
+	.section __ex_table,\"a\"					\n\
+	.dword 0b,3b							\n\
+	.previous"
+
+     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
+     /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
+  }
+
+  /* Either we directly start copying here, using dword copying in a loop,
+     or we copy as much as possible with 'movem' and then the last block
+     (<44 bytes) is copied here.  This will work since 'movem' will have
+     updated src, dst and n.  (Except with failing src.)
+
+     Since we want to keep src accurate, we can't use
+     __asm_copy_from_user_N with N != (1, 2, 4); it updates dst and
+     retn, but not src (by design; it's value is ignored elsewhere).  */
+
+  while (n >= 4)
+  {
+    __asm_copy_from_user_4 (dst, src, retn);
+    n -= 4;
+
+    if (retn)
+      goto copy_exception_bytes;
+  }
+
+  /* If we get here, there were no memory read faults.  */
+  switch (n)
+  {
+    /* These copies are at least "naturally aligned" (so we don't have
+       to check each byte), due to the src alignment code before the
+       movem loop.  The *_3 case *will* get the correct count for retn.  */
+    case 0:
+      /* This case deliberately left in (if you have doubts check the
+	 generated assembly code).  */
+      break;
+    case 1:
+      __asm_copy_from_user_1 (dst, src, retn);
+      break;
+    case 2:
+      __asm_copy_from_user_2 (dst, src, retn);
+      break;
+    case 3:
+      __asm_copy_from_user_3 (dst, src, retn);
+      break;
+  }
+
+  /* If we get here, retn correctly reflects the number of failing
+     bytes.  */
+  return retn;
+
+copy_exception_bytes:
+  /* We already have "retn" bytes cleared, and need to clear the
+     remaining "n" bytes.  A non-optimized simple byte-for-byte in-line
+     memset is preferred here, since this isn't speed-critical code and
+     we'd rather have this a leaf-function than calling memset.  */
+  {
+    char *endp;
+    for (endp = dst + n; dst < endp; dst++)
+      *dst = 0;
+  }
+
+  return retn + n;
+}
+EXPORT_SYMBOL(__copy_user_zeroing);
+
+/* Zero userspace.  */
+unsigned long __do_clear_user(void __user *pto, unsigned long pn)
+{
+  /* We want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+      As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     FIXME: Comment for old gcc version.  Check.
+     If gcc was alright, it really would need no temporaries, and no
+     stack space to save stuff on. */
+
+  register char *dst __asm__ ("r13") = pto;
+  register int n __asm__ ("r12") = pn;
+  register int retn __asm__ ("r10") = 0;
+
+
+  if (((unsigned long) dst & 3) != 0
+     /* Don't align if we wouldn't copy more than a few bytes.  */
+      && n >= 3)
+  {
+    if ((unsigned long) dst & 1)
+    {
+      __asm_clear_1 (dst, retn);
+      n--;
+    }
+
+    if ((unsigned long) dst & 2)
+    {
+      __asm_clear_2 (dst, retn);
+      n -= 2;
+    }
+  }
+
+  /* Decide which copying method to use.
+     FIXME: This number is from the "ordinary" kernel memset.  */
+  if (n >= 48)
+  {
+    /* For large clears we use 'movem' */
+
+    /* It is not optimal to tell the compiler about clobbering any
+       call-saved registers; that will move the saving/restoring of
+       those registers to the function prologue/epilogue, and make
+       non-movem sizes suboptimal.
+
+       This method is not foolproof; it assumes that the "asm reg"
+       declarations at the beginning of the function really are used
+       here (beware: they may be moved to temporary registers).
+       This way, we do not have to save/move the registers around into
+       temporaries; we can safely use them straight away.
+
+      If you want to check that the allocation was right; then
+      check the equalities in the first comment.  It should say
+      something like "r13=r13, r11=r11, r12=r12". */
+    __asm__ volatile ("\
+	.ifnc %0%1%2,$r13$r12$r10					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll clobber in the movem process	\n\
+	;; on the stack.  Don't mention them to gcc, it will only be	\n\
+	;; upset.							\n\
+	subq	11*4,$sp						\n\
+	movem	$r10,[$sp]						\n\
+									\n\
+	clear.d $r0							\n\
+	clear.d $r1							\n\
+	clear.d $r2							\n\
+	clear.d $r3							\n\
+	clear.d $r4							\n\
+	clear.d $r5							\n\
+	clear.d $r6							\n\
+	clear.d $r7							\n\
+	clear.d $r8							\n\
+	clear.d $r9							\n\
+	clear.d $r10							\n\
+	clear.d $r11							\n\
+									\n\
+	;; Now we've got this:						\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+	;; Update n for the first loop					\n\
+	subq	12*4,$r12						\n\
+0:									\n\
+	subq   12*4,$r12						\n\
+1:									\n\
+	bge	0b							\n\
+	movem	$r11,[$r13+]						\n\
+									\n\
+	addq   12*4,$r12 ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+	movem [$sp+],$r10						\n\
+2:									\n\
+	.section .fixup,\"ax\"						\n\
+3:									\n\
+	movem [$sp],$r10						\n\
+	addq 12*4,$r10							\n\
+	addq 12*4,$r13							\n\
+	movem $r10,[$sp]						\n\
+	jump 0b								\n\
+	clear.d $r10							\n\
+									\n\
+	.previous							\n\
+	.section __ex_table,\"a\"					\n\
+	.dword 1b,3b							\n\
+	.previous"
+
+     /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
+     /* Inputs */ : "0" (dst), "1" (n), "2" (retn)
+     /* Clobber */ : "r11");
+  }
+
+  while (n >= 16)
+  {
+    __asm_clear_16 (dst, retn);
+    n -= 16;
+  }
+
+  /* Having a separate by-four loops cuts down on cache footprint.
+     FIXME:  Test with and without; increasing switch to be 0..15.  */
+  while (n >= 4)
+  {
+    __asm_clear_4 (dst, retn);
+    n -= 4;
+  }
+
+  switch (n)
+  {
+    case 0:
+      break;
+    case 1:
+      __asm_clear_1 (dst, retn);
+      break;
+    case 2:
+      __asm_clear_2 (dst, retn);
+      break;
+    case 3:
+      __asm_clear_3 (dst, retn);
+      break;
+  }
+
+  return retn;
+}
+EXPORT_SYMBOL(__do_clear_user);