[MIPS] csum_partial and copy in parallel

Implement optimized asm version of csum_partial_copy_nocheck, csum_partial_copy_from_user and csum_and_copy_to_user which can do calculate and copy in parallel, based on memcpy.S. Signed-off-by: Atsushi Nemoto <anemo@mba.ocn.ne.jp> Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

[MIPS] csum_partial and copy in parallel
Implement optimized asm version of csum_partial_copy_nocheck, csum_partial_copy_from_user and csum_and_copy_to_user which can do calculate and copy in parallel, based on memcpy.S. Signed-off-by: Atsushi Nemoto <anemo@mba.ocn.ne.jp> Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Atsushi Nemoto · Ralf Baechle
1 parent 61e84f9987
Showing 5 changed files with 464 additions and 65 deletions Side-by-side Diff
arch/mips/kernel/mips_ksyms.c
arch/mips/lib/Makefile
arch/mips/lib/csum_partial.S
arch/mips/lib/csum_partial_copy.c
include/asm-mips/checksum.h
@@ -46,6 +46,8 @@
 EXPORT_SYMBOL(__strnlen_user_asm);
  
 EXPORT_SYMBOL(csum_partial);
+EXPORT_SYMBOL(csum_partial_copy_nocheck);
+EXPORT_SYMBOL(__csum_partial_copy_user);
  
 EXPORT_SYMBOL(invalid_pte_table);
@@ -2,7 +2,7 @@
 # Makefile for MIPS-specific library files..
 #
  
-lib-y	+= csum_partial.o csum_partial_copy.o memcpy.o promlib.o \
+lib-y	+= csum_partial.o memcpy.o promlib.o \
 	   strlen_user.o strncpy_user.o strnlen_user.o uncached.o
  
 obj-y	+= iomap.o
@@ -8,7 +8,9 @@
  * Copyright (C) 1998, 1999 Ralf Baechle
  * Copyright (C) 1999 Silicon Graphics, Inc.
  */
+#include <linux/errno.h>
 #include <asm/asm.h>
+#include <asm/asm-offsets.h>
 #include <asm/regdef.h>
  
 #ifdef CONFIG_64BIT
@@ -271,4 +273,444 @@
 	jr	ra
 	.set	noreorder
 	END(csum_partial)
+
+
+/*
+ * checksum and copy routines based on memcpy.S
+ *
+ *	csum_partial_copy_nocheck(src, dst, len, sum)
+ *	__csum_partial_copy_user(src, dst, len, sum, errp)
+ *
+ * See "Spec" in memcpy.S for details.  Unlike __copy_user, all
+ * function in this file use the standard calling convention.
+ */
+
+#define src a0
+#define dst a1
+#define len a2
+#define psum a3
+#define sum v0
+#define odd t8
+#define errptr t9
+
+/*
+ * The exception handler for loads requires that:
+ *  1- AT contain the address of the byte just past the end of the source
+ *     of the copy,
+ *  2- src_entry <= src < AT, and
+ *  3- (dst - src) == (dst_entry - src_entry),
+ * The _entry suffix denotes values when __copy_user was called.
+ *
+ * (1) is set up up by __csum_partial_copy_from_user and maintained by
+ *	not writing AT in __csum_partial_copy
+ * (2) is met by incrementing src by the number of bytes copied
+ * (3) is met by not doing loads between a pair of increments of dst and src
+ *
+ * The exception handlers for stores stores -EFAULT to errptr and return.
+ * These handlers do not need to overwrite any data.
+ */
+
+#define EXC(inst_reg,addr,handler)		\
+9:	inst_reg, addr;				\
+	.section __ex_table,"a";		\
+	PTR	9b, handler;			\
+	.previous
+
+#ifdef USE_DOUBLE
+
+#define LOAD   ld
+#define LOADL  ldl
+#define LOADR  ldr
+#define STOREL sdl
+#define STORER sdr
+#define STORE  sd
+#define ADD    daddu
+#define SUB    dsubu
+#define SRL    dsrl
+#define SLL    dsll
+#define SLLV   dsllv
+#define SRLV   dsrlv
+#define NBYTES 8
+#define LOG_NBYTES 3
+
+#else
+
+#define LOAD   lw
+#define LOADL  lwl
+#define LOADR  lwr
+#define STOREL swl
+#define STORER swr
+#define STORE  sw
+#define ADD    addu
+#define SUB    subu
+#define SRL    srl
+#define SLL    sll
+#define SLLV   sllv
+#define SRLV   srlv
+#define NBYTES 4
+#define LOG_NBYTES 2
+
+#endif /* USE_DOUBLE */
+
+#ifdef CONFIG_CPU_LITTLE_ENDIAN
+#define LDFIRST LOADR
+#define LDREST  LOADL
+#define STFIRST STORER
+#define STREST  STOREL
+#define SHIFT_DISCARD SLLV
+#define SHIFT_DISCARD_REVERT SRLV
+#else
+#define LDFIRST LOADL
+#define LDREST  LOADR
+#define STFIRST STOREL
+#define STREST  STORER
+#define SHIFT_DISCARD SRLV
+#define SHIFT_DISCARD_REVERT SLLV
+#endif
+
+#define FIRST(unit) ((unit)*NBYTES)
+#define REST(unit)  (FIRST(unit)+NBYTES-1)
+
+#define ADDRMASK (NBYTES-1)
+
+	.set	noat
+
+LEAF(__csum_partial_copy_user)
+	PTR_ADDU	AT, src, len	/* See (1) above. */
+#ifdef CONFIG_64BIT
+	move	errptr, a4
+#else
+	lw	errptr, 16(sp)
+#endif
+FEXPORT(csum_partial_copy_nocheck)
+	move	sum, zero
+	move	odd, zero
+	/*
+	 * Note: dst & src may be unaligned, len may be 0
+	 * Temps
+	 */
+	/*
+	 * The "issue break"s below are very approximate.
+	 * Issue delays for dcache fills will perturb the schedule, as will
+	 * load queue full replay traps, etc.
+	 *
+	 * If len < NBYTES use byte operations.
+	 */
+	sltu	t2, len, NBYTES
+	and	t1, dst, ADDRMASK
+	bnez	t2, copy_bytes_checklen
+	 and	t0, src, ADDRMASK
+	andi	odd, dst, 0x1			/* odd buffer? */
+	bnez	t1, dst_unaligned
+	 nop
+	bnez	t0, src_unaligned_dst_aligned
+	/*
+	 * use delay slot for fall-through
+	 * src and dst are aligned; need to compute rem
+	 */
+both_aligned:
+	 SRL	t0, len, LOG_NBYTES+3    # +3 for 8 units/iter
+	beqz	t0, cleanup_both_aligned # len < 8*NBYTES
+	 nop
+	SUB	len, 8*NBYTES		# subtract here for bgez loop
+	.align	4
+1:
+EXC(	LOAD	t0, UNIT(0)(src),	l_exc)
+EXC(	LOAD	t1, UNIT(1)(src),	l_exc_copy)
+EXC(	LOAD	t2, UNIT(2)(src),	l_exc_copy)
+EXC(	LOAD	t3, UNIT(3)(src),	l_exc_copy)
+EXC(	LOAD	t4, UNIT(4)(src),	l_exc_copy)
+EXC(	LOAD	t5, UNIT(5)(src),	l_exc_copy)
+EXC(	LOAD	t6, UNIT(6)(src),	l_exc_copy)
+EXC(	LOAD	t7, UNIT(7)(src),	l_exc_copy)
+	SUB	len, len, 8*NBYTES
+	ADD	src, src, 8*NBYTES
+EXC(	STORE	t0, UNIT(0)(dst),	s_exc)
+	ADDC(sum, t0)
+EXC(	STORE	t1, UNIT(1)(dst),	s_exc)
+	ADDC(sum, t1)
+EXC(	STORE	t2, UNIT(2)(dst),	s_exc)
+	ADDC(sum, t2)
+EXC(	STORE	t3, UNIT(3)(dst),	s_exc)
+	ADDC(sum, t3)
+EXC(	STORE	t4, UNIT(4)(dst),	s_exc)
+	ADDC(sum, t4)
+EXC(	STORE	t5, UNIT(5)(dst),	s_exc)
+	ADDC(sum, t5)
+EXC(	STORE	t6, UNIT(6)(dst),	s_exc)
+	ADDC(sum, t6)
+EXC(	STORE	t7, UNIT(7)(dst),	s_exc)
+	ADDC(sum, t7)
+	bgez	len, 1b
+	 ADD	dst, dst, 8*NBYTES
+	ADD	len, 8*NBYTES		# revert len (see above)
+
+	/*
+	 * len == the number of bytes left to copy < 8*NBYTES
+	 */
+cleanup_both_aligned:
+#define rem t7
+	beqz	len, done
+	 sltu	t0, len, 4*NBYTES
+	bnez	t0, less_than_4units
+	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
+	/*
+	 * len >= 4*NBYTES
+	 */
+EXC(	LOAD	t0, UNIT(0)(src),	l_exc)
+EXC(	LOAD	t1, UNIT(1)(src),	l_exc_copy)
+EXC(	LOAD	t2, UNIT(2)(src),	l_exc_copy)
+EXC(	LOAD	t3, UNIT(3)(src),	l_exc_copy)
+	SUB	len, len, 4*NBYTES
+	ADD	src, src, 4*NBYTES
+EXC(	STORE	t0, UNIT(0)(dst),	s_exc)
+	ADDC(sum, t0)
+EXC(	STORE	t1, UNIT(1)(dst),	s_exc)
+	ADDC(sum, t1)
+EXC(	STORE	t2, UNIT(2)(dst),	s_exc)
+	ADDC(sum, t2)
+EXC(	STORE	t3, UNIT(3)(dst),	s_exc)
+	ADDC(sum, t3)
+	beqz	len, done
+	 ADD	dst, dst, 4*NBYTES
+less_than_4units:
+	/*
+	 * rem = len % NBYTES
+	 */
+	beq	rem, len, copy_bytes
+	 nop
+1:
+EXC(	LOAD	t0, 0(src),		l_exc)
+	ADD	src, src, NBYTES
+	SUB	len, len, NBYTES
+EXC(	STORE	t0, 0(dst),		s_exc)
+	ADDC(sum, t0)
+	bne	rem, len, 1b
+	 ADD	dst, dst, NBYTES
+
+	/*
+	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
+	 * A loop would do only a byte at a time with possible branch
+	 * mispredicts.  Can't do an explicit LOAD dst,mask,or,STORE
+	 * because can't assume read-access to dst.  Instead, use
+	 * STREST dst, which doesn't require read access to dst.
+	 *
+	 * This code should perform better than a simple loop on modern,
+	 * wide-issue mips processors because the code has fewer branches and
+	 * more instruction-level parallelism.
+	 */
+#define bits t2
+	beqz	len, done
+	 ADD	t1, dst, len	# t1 is just past last byte of dst
+	li	bits, 8*NBYTES
+	SLL	rem, len, 3	# rem = number of bits to keep
+EXC(	LOAD	t0, 0(src),		l_exc)
+	SUB	bits, bits, rem	# bits = number of bits to discard
+	SHIFT_DISCARD t0, t0, bits
+EXC(	STREST	t0, -1(t1),		s_exc)
+	SHIFT_DISCARD_REVERT t0, t0, bits
+	.set reorder
+	ADDC(sum, t0)
+	b	done
+	.set noreorder
+dst_unaligned:
+	/*
+	 * dst is unaligned
+	 * t0 = src & ADDRMASK
+	 * t1 = dst & ADDRMASK; T1 > 0
+	 * len >= NBYTES
+	 *
+	 * Copy enough bytes to align dst
+	 * Set match = (src and dst have same alignment)
+	 */
+#define match rem
+EXC(	LDFIRST	t3, FIRST(0)(src),	l_exc)
+	ADD	t2, zero, NBYTES
+EXC(	LDREST	t3, REST(0)(src),	l_exc_copy)
+	SUB	t2, t2, t1	# t2 = number of bytes copied
+	xor	match, t0, t1
+EXC(	STFIRST t3, FIRST(0)(dst),	s_exc)
+	SLL	t4, t1, 3		# t4 = number of bits to discard
+	SHIFT_DISCARD t3, t3, t4
+	/* no SHIFT_DISCARD_REVERT to handle odd buffer properly */
+	ADDC(sum, t3)
+	beq	len, t2, done
+	 SUB	len, len, t2
+	ADD	dst, dst, t2
+	beqz	match, both_aligned
+	 ADD	src, src, t2
+
+src_unaligned_dst_aligned:
+	SRL	t0, len, LOG_NBYTES+2    # +2 for 4 units/iter
+	beqz	t0, cleanup_src_unaligned
+	 and	rem, len, (4*NBYTES-1)   # rem = len % 4*NBYTES
+1:
+/*
+ * Avoid consecutive LD*'s to the same register since some mips
+ * implementations can't issue them in the same cycle.
+ * It's OK to load FIRST(N+1) before REST(N) because the two addresses
+ * are to the same unit (unless src is aligned, but it's not).
+ */
+EXC(	LDFIRST	t0, FIRST(0)(src),	l_exc)
+EXC(	LDFIRST	t1, FIRST(1)(src),	l_exc_copy)
+	SUB     len, len, 4*NBYTES
+EXC(	LDREST	t0, REST(0)(src),	l_exc_copy)
+EXC(	LDREST	t1, REST(1)(src),	l_exc_copy)
+EXC(	LDFIRST	t2, FIRST(2)(src),	l_exc_copy)
+EXC(	LDFIRST	t3, FIRST(3)(src),	l_exc_copy)
+EXC(	LDREST	t2, REST(2)(src),	l_exc_copy)
+EXC(	LDREST	t3, REST(3)(src),	l_exc_copy)
+	ADD	src, src, 4*NBYTES
+#ifdef CONFIG_CPU_SB1
+	nop				# improves slotting
+#endif
+EXC(	STORE	t0, UNIT(0)(dst),	s_exc)
+	ADDC(sum, t0)
+EXC(	STORE	t1, UNIT(1)(dst),	s_exc)
+	ADDC(sum, t1)
+EXC(	STORE	t2, UNIT(2)(dst),	s_exc)
+	ADDC(sum, t2)
+EXC(	STORE	t3, UNIT(3)(dst),	s_exc)
+	ADDC(sum, t3)
+	bne	len, rem, 1b
+	 ADD	dst, dst, 4*NBYTES
+
+cleanup_src_unaligned:
+	beqz	len, done
+	 and	rem, len, NBYTES-1  # rem = len % NBYTES
+	beq	rem, len, copy_bytes
+	 nop
+1:
+EXC(	LDFIRST t0, FIRST(0)(src),	l_exc)
+EXC(	LDREST	t0, REST(0)(src),	l_exc_copy)
+	ADD	src, src, NBYTES
+	SUB	len, len, NBYTES
+EXC(	STORE	t0, 0(dst),		s_exc)
+	ADDC(sum, t0)
+	bne	len, rem, 1b
+	 ADD	dst, dst, NBYTES
+
+copy_bytes_checklen:
+	beqz	len, done
+	 nop
+copy_bytes:
+	/* 0 < len < NBYTES  */
+#ifdef CONFIG_CPU_LITTLE_ENDIAN
+#define SHIFT_START 0
+#define SHIFT_INC 8
+#else
+#define SHIFT_START 8*(NBYTES-1)
+#define SHIFT_INC -8
+#endif
+	move	t2, zero	# partial word
+	li	t3, SHIFT_START	# shift
+/* use l_exc_copy here to return correct sum on fault */
+#define COPY_BYTE(N)			\
+EXC(	lbu	t0, N(src), l_exc_copy);	\
+	SUB	len, len, 1;		\
+EXC(	sb	t0, N(dst), s_exc);	\
+	SLLV	t0, t0, t3;		\
+	addu	t3, SHIFT_INC;		\
+	beqz	len, copy_bytes_done;	\
+	 or	t2, t0
+
+	COPY_BYTE(0)
+	COPY_BYTE(1)
+#ifdef USE_DOUBLE
+	COPY_BYTE(2)
+	COPY_BYTE(3)
+	COPY_BYTE(4)
+	COPY_BYTE(5)
+#endif
+EXC(	lbu	t0, NBYTES-2(src), l_exc_copy)
+	SUB	len, len, 1
+EXC(	sb	t0, NBYTES-2(dst), s_exc)
+	SLLV	t0, t0, t3
+	or	t2, t0
+copy_bytes_done:
+	ADDC(sum, t2)
+done:
+	/* fold checksum */
+#ifdef USE_DOUBLE
+	dsll32	v1, sum, 0
+	daddu	sum, v1
+	sltu	v1, sum, v1
+	dsra32	sum, sum, 0
+	addu	sum, v1
+#endif
+	sll	v1, sum, 16
+	addu	sum, v1
+	sltu	v1, sum, v1
+	srl	sum, sum, 16
+	addu	sum, v1
+
+	/* odd buffer alignment? */
+	beqz	odd, 1f
+	 nop
+	sll	v1, sum, 8
+	srl	sum, sum, 8
+	or	sum, v1
+	andi	sum, 0xffff
+1:
+	.set reorder
+	ADDC(sum, psum)
+	jr	ra
+	.set noreorder
+
+l_exc_copy:
+	/*
+	 * Copy bytes from src until faulting load address (or until a
+	 * lb faults)
+	 *
+	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
+	 * may be more than a byte beyond the last address.
+	 * Hence, the lb below may get an exception.
+	 *
+	 * Assumes src < THREAD_BUADDR($28)
+	 */
+	LOAD	t0, TI_TASK($28)
+	 li	t2, SHIFT_START
+	LOAD	t0, THREAD_BUADDR(t0)
+1:
+EXC(	lbu	t1, 0(src),	l_exc)
+	ADD	src, src, 1
+	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
+	SLLV	t1, t1, t2
+	addu	t2, SHIFT_INC
+	ADDC(sum, t1)
+	bne	src, t0, 1b
+	 ADD	dst, dst, 1
+l_exc:
+	LOAD	t0, TI_TASK($28)
+	 nop
+	LOAD	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
+	 nop
+	SUB	len, AT, t0		# len number of uncopied bytes
+	/*
+	 * Here's where we rely on src and dst being incremented in tandem,
+	 *   See (3) above.
+	 * dst += (fault addr - src) to put dst at first byte to clear
+	 */
+	ADD	dst, t0			# compute start address in a1
+	SUB	dst, src
+	/*
+	 * Clear len bytes starting at dst.  Can't call __bzero because it
+	 * might modify len.  An inefficient loop for these rare times...
+	 */
+	beqz	len, done
+	 SUB	src, len, 1
+1:	sb	zero, 0(dst)
+	ADD	dst, dst, 1
+	bnez	src, 1b
+	 SUB	src, src, 1
+	li	v1, -EFAULT
+	b	done
+	 sw	v1, (errptr)
+
+s_exc:
+	li	v0, -1 /* invalid checksum */
+	li	v1, -EFAULT
+	jr	ra
+	 sw	v1, (errptr)
+	END(__csum_partial_copy_user)
-/*
- * This file is subject to the terms and conditions of the GNU General Public
- * License.  See the file "COPYING" in the main directory of this archive
- * for more details.
- *
- * Copyright (C) 1994, 1995 Waldorf Electronics GmbH
- * Copyright (C) 1998, 1999 Ralf Baechle
- */
-#include <linux/kernel.h>
-#include <linux/module.h>
-#include <linux/types.h>
-#include <asm/byteorder.h>
-#include <asm/string.h>
-#include <asm/uaccess.h>
-#include <net/checksum.h>
-
-/*
- * copy while checksumming, otherwise like csum_partial
- */
-__wsum csum_partial_copy_nocheck(const void *src,
-	void *dst, int len, __wsum sum)
-{
-	/*
-	 * It's 2:30 am and I don't feel like doing it real ...
-	 * This is lots slower than the real thing (tm)
-	 */
-	sum = csum_partial(src, len, sum);
-	memcpy(dst, src, len);
-
-	return sum;
-}
-
-EXPORT_SYMBOL(csum_partial_copy_nocheck);
-
-/*
- * Copy from userspace and compute checksum.  If we catch an exception
- * then zero the rest of the buffer.
- */
-__wsum csum_partial_copy_from_user (const void __user *src,
-	void *dst, int len, __wsum sum, int *err_ptr)
-{
-	int missing;
-
-	might_sleep();
-	missing = copy_from_user(dst, src, len);
-	if (missing) {
-		memset(dst + len - missing, 0, missing);
-		*err_ptr = -EFAULT;
-	}
-
-	return csum_partial(dst, len, sum);
-}
@@ -29,31 +29,38 @@
  */
 __wsum csum_partial(const void *buff, int len, __wsum sum);
  
+__wsum __csum_partial_copy_user(const void *src, void *dst,
+				int len, __wsum sum, int *err_ptr);
+
 /*
  * this is a new version of the above that records errors it finds in *errp,
  * but continues and zeros the rest of the buffer.
  */
-__wsum csum_partial_copy_from_user(const void __user *src,
-					 void *dst, int len,
-					 __wsum sum, int *errp);
+static inline
+__wsum csum_partial_copy_from_user(const void __user *src, void *dst, int len,
+				   __wsum sum, int *err_ptr)
+{
+	might_sleep();
+	return __csum_partial_copy_user((__force void *)src, dst,
+					len, sum, err_ptr);
+}
  
 /*
  * Copy and checksum to user
  */
 #define HAVE_CSUM_COPY_USER
-static inline __wsum csum_and_copy_to_user (const void *src, void __user *dst,
-						  int len, __wsum sum,
-						  int *err_ptr)
+static inline
+__wsum csum_and_copy_to_user(const void *src, void __user *dst, int len,
+			     __wsum sum, int *err_ptr)
 {
 	might_sleep();
-	sum = csum_partial(src, len, sum);
-
-	if (copy_to_user(dst, src, len)) {
+	if (access_ok(VERIFY_WRITE, dst, len))
+		return __csum_partial_copy_user(src, (__force void *)dst,
+						len, sum, err_ptr);
+	if (len)
 		*err_ptr = -EFAULT;
-		return (__force __wsum)-1;
-	}
  
-	return sum;
+	return (__force __wsum)-1; /* invalid checksum */
 }
  
 /*
...	...	@@ -46,6 +46,8 @@
46	46	EXPORT_SYMBOL(__strnlen_user_asm);
47	47
48	48	EXPORT_SYMBOL(csum_partial);
	49	+EXPORT_SYMBOL(csum_partial_copy_nocheck);
	50	+EXPORT_SYMBOL(__csum_partial_copy_user);
49	51
50	52	EXPORT_SYMBOL(invalid_pte_table);
...	...	@@ -2,7 +2,7 @@
2	2	# Makefile for MIPS-specific library files..
3	3	#
4	4
5		-lib-y += csum_partial.o csum_partial_copy.o memcpy.o promlib.o \
	5	+lib-y += csum_partial.o memcpy.o promlib.o \
6	6	strlen_user.o strncpy_user.o strnlen_user.o uncached.o
7	7
8	8	obj-y += iomap.o
...	...	@@ -8,7 +8,9 @@
8	8	* Copyright (C) 1998, 1999 Ralf Baechle
9	9	* Copyright (C) 1999 Silicon Graphics, Inc.
10	10	*/
	11	+#include <linux/errno.h>
11	12	#include <asm/asm.h>
	13	+#include <asm/asm-offsets.h>
12	14	#include <asm/regdef.h>
13	15
14	16	#ifdef CONFIG_64BIT
...	...	@@ -271,4 +273,444 @@
271	273	jr ra
272	274	.set noreorder
273	275	END(csum_partial)
	276	+
	277	+
	278	+/*
	279	+ * checksum and copy routines based on memcpy.S
	280	+ *
	281	+ * csum_partial_copy_nocheck(src, dst, len, sum)
	282	+ * __csum_partial_copy_user(src, dst, len, sum, errp)
	283	+ *
	284	+ * See "Spec" in memcpy.S for details. Unlike __copy_user, all
	285	+ * function in this file use the standard calling convention.
	286	+ */
	287	+
	288	+#define src a0
	289	+#define dst a1
	290	+#define len a2
	291	+#define psum a3
	292	+#define sum v0
	293	+#define odd t8
	294	+#define errptr t9
	295	+
	296	+/*
	297	+ * The exception handler for loads requires that:
	298	+ * 1- AT contain the address of the byte just past the end of the source
	299	+ * of the copy,
	300	+ * 2- src_entry <= src < AT, and
	301	+ * 3- (dst - src) == (dst_entry - src_entry),
	302	+ * The _entry suffix denotes values when __copy_user was called.
	303	+ *
	304	+ * (1) is set up up by __csum_partial_copy_from_user and maintained by
	305	+ * not writing AT in __csum_partial_copy
	306	+ * (2) is met by incrementing src by the number of bytes copied
	307	+ * (3) is met by not doing loads between a pair of increments of dst and src
	308	+ *
	309	+ * The exception handlers for stores stores -EFAULT to errptr and return.
	310	+ * These handlers do not need to overwrite any data.
	311	+ */
	312	+
	313	+#define EXC(inst_reg,addr,handler) \
	314	+9: inst_reg, addr; \
	315	+ .section __ex_table,"a"; \
	316	+ PTR 9b, handler; \
	317	+ .previous
	318	+
	319	+#ifdef USE_DOUBLE
	320	+
	321	+#define LOAD ld
	322	+#define LOADL ldl
	323	+#define LOADR ldr
	324	+#define STOREL sdl
	325	+#define STORER sdr
	326	+#define STORE sd
	327	+#define ADD daddu
	328	+#define SUB dsubu
	329	+#define SRL dsrl
	330	+#define SLL dsll
	331	+#define SLLV dsllv
	332	+#define SRLV dsrlv
	333	+#define NBYTES 8
	334	+#define LOG_NBYTES 3
	335	+
	336	+#else
	337	+
	338	+#define LOAD lw
	339	+#define LOADL lwl
	340	+#define LOADR lwr
	341	+#define STOREL swl
	342	+#define STORER swr
	343	+#define STORE sw
	344	+#define ADD addu
	345	+#define SUB subu
	346	+#define SRL srl
	347	+#define SLL sll
	348	+#define SLLV sllv
	349	+#define SRLV srlv
	350	+#define NBYTES 4
	351	+#define LOG_NBYTES 2
	352	+
	353	+#endif /* USE_DOUBLE */
	354	+
	355	+#ifdef CONFIG_CPU_LITTLE_ENDIAN
	356	+#define LDFIRST LOADR
	357	+#define LDREST LOADL
	358	+#define STFIRST STORER
	359	+#define STREST STOREL
	360	+#define SHIFT_DISCARD SLLV
	361	+#define SHIFT_DISCARD_REVERT SRLV
	362	+#else
	363	+#define LDFIRST LOADL
	364	+#define LDREST LOADR
	365	+#define STFIRST STOREL
	366	+#define STREST STORER
	367	+#define SHIFT_DISCARD SRLV
	368	+#define SHIFT_DISCARD_REVERT SLLV
	369	+#endif
	370	+
	371	+#define FIRST(unit) ((unit)*NBYTES)
	372	+#define REST(unit) (FIRST(unit)+NBYTES-1)
	373	+
	374	+#define ADDRMASK (NBYTES-1)
	375	+
	376	+ .set noat
	377	+
	378	+LEAF(__csum_partial_copy_user)
	379	+ PTR_ADDU AT, src, len /* See (1) above. */
	380	+#ifdef CONFIG_64BIT
	381	+ move errptr, a4
	382	+#else
	383	+ lw errptr, 16(sp)
	384	+#endif
	385	+FEXPORT(csum_partial_copy_nocheck)
	386	+ move sum, zero
	387	+ move odd, zero
	388	+ /*
	389	+ * Note: dst & src may be unaligned, len may be 0
	390	+ * Temps
	391	+ */
	392	+ /*
	393	+ * The "issue break"s below are very approximate.
	394	+ * Issue delays for dcache fills will perturb the schedule, as will
	395	+ * load queue full replay traps, etc.
	396	+ *
	397	+ * If len < NBYTES use byte operations.
	398	+ */
	399	+ sltu t2, len, NBYTES
	400	+ and t1, dst, ADDRMASK
	401	+ bnez t2, copy_bytes_checklen
	402	+ and t0, src, ADDRMASK
	403	+ andi odd, dst, 0x1 /* odd buffer? */
	404	+ bnez t1, dst_unaligned
	405	+ nop
	406	+ bnez t0, src_unaligned_dst_aligned
	407	+ /*
	408	+ * use delay slot for fall-through
	409	+ * src and dst are aligned; need to compute rem
	410	+ */
	411	+both_aligned:
	412	+ SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
	413	+ beqz t0, cleanup_both_aligned # len < 8*NBYTES
	414	+ nop
	415	+ SUB len, 8*NBYTES # subtract here for bgez loop
	416	+ .align 4
	417	+1:
	418	+EXC( LOAD t0, UNIT(0)(src), l_exc)
	419	+EXC( LOAD t1, UNIT(1)(src), l_exc_copy)
	420	+EXC( LOAD t2, UNIT(2)(src), l_exc_copy)
	421	+EXC( LOAD t3, UNIT(3)(src), l_exc_copy)
	422	+EXC( LOAD t4, UNIT(4)(src), l_exc_copy)
	423	+EXC( LOAD t5, UNIT(5)(src), l_exc_copy)
	424	+EXC( LOAD t6, UNIT(6)(src), l_exc_copy)
	425	+EXC( LOAD t7, UNIT(7)(src), l_exc_copy)
	426	+ SUB len, len, 8*NBYTES
	427	+ ADD src, src, 8*NBYTES
	428	+EXC( STORE t0, UNIT(0)(dst), s_exc)
	429	+ ADDC(sum, t0)
	430	+EXC( STORE t1, UNIT(1)(dst), s_exc)
	431	+ ADDC(sum, t1)
	432	+EXC( STORE t2, UNIT(2)(dst), s_exc)
	433	+ ADDC(sum, t2)
	434	+EXC( STORE t3, UNIT(3)(dst), s_exc)
	435	+ ADDC(sum, t3)
	436	+EXC( STORE t4, UNIT(4)(dst), s_exc)
	437	+ ADDC(sum, t4)
	438	+EXC( STORE t5, UNIT(5)(dst), s_exc)
	439	+ ADDC(sum, t5)
	440	+EXC( STORE t6, UNIT(6)(dst), s_exc)
	441	+ ADDC(sum, t6)
	442	+EXC( STORE t7, UNIT(7)(dst), s_exc)
	443	+ ADDC(sum, t7)
	444	+ bgez len, 1b
	445	+ ADD dst, dst, 8*NBYTES
	446	+ ADD len, 8*NBYTES # revert len (see above)
	447	+
	448	+ /*
	449	+ * len == the number of bytes left to copy < 8*NBYTES
	450	+ */
	451	+cleanup_both_aligned:
	452	+#define rem t7
	453	+ beqz len, done
	454	+ sltu t0, len, 4*NBYTES
	455	+ bnez t0, less_than_4units
	456	+ and rem, len, (NBYTES-1) # rem = len % NBYTES
	457	+ /*
	458	+ * len >= 4*NBYTES
	459	+ */
	460	+EXC( LOAD t0, UNIT(0)(src), l_exc)
	461	+EXC( LOAD t1, UNIT(1)(src), l_exc_copy)
	462	+EXC( LOAD t2, UNIT(2)(src), l_exc_copy)
	463	+EXC( LOAD t3, UNIT(3)(src), l_exc_copy)
	464	+ SUB len, len, 4*NBYTES
	465	+ ADD src, src, 4*NBYTES
	466	+EXC( STORE t0, UNIT(0)(dst), s_exc)
	467	+ ADDC(sum, t0)
	468	+EXC( STORE t1, UNIT(1)(dst), s_exc)
	469	+ ADDC(sum, t1)
	470	+EXC( STORE t2, UNIT(2)(dst), s_exc)
	471	+ ADDC(sum, t2)
	472	+EXC( STORE t3, UNIT(3)(dst), s_exc)
	473	+ ADDC(sum, t3)
	474	+ beqz len, done
	475	+ ADD dst, dst, 4*NBYTES
	476	+less_than_4units:
	477	+ /*
	478	+ * rem = len % NBYTES
	479	+ */
	480	+ beq rem, len, copy_bytes
	481	+ nop
	482	+1:
	483	+EXC( LOAD t0, 0(src), l_exc)
	484	+ ADD src, src, NBYTES
	485	+ SUB len, len, NBYTES
	486	+EXC( STORE t0, 0(dst), s_exc)
	487	+ ADDC(sum, t0)
	488	+ bne rem, len, 1b
	489	+ ADD dst, dst, NBYTES
	490	+
	491	+ /*
	492	+ * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
	493	+ * A loop would do only a byte at a time with possible branch
	494	+ * mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
	495	+ * because can't assume read-access to dst. Instead, use
	496	+ * STREST dst, which doesn't require read access to dst.
	497	+ *
	498	+ * This code should perform better than a simple loop on modern,
	499	+ * wide-issue mips processors because the code has fewer branches and
	500	+ * more instruction-level parallelism.
	501	+ */
	502	+#define bits t2
	503	+ beqz len, done
	504	+ ADD t1, dst, len # t1 is just past last byte of dst
	505	+ li bits, 8*NBYTES
	506	+ SLL rem, len, 3 # rem = number of bits to keep
	507	+EXC( LOAD t0, 0(src), l_exc)
	508	+ SUB bits, bits, rem # bits = number of bits to discard
	509	+ SHIFT_DISCARD t0, t0, bits
	510	+EXC( STREST t0, -1(t1), s_exc)
	511	+ SHIFT_DISCARD_REVERT t0, t0, bits
	512	+ .set reorder
	513	+ ADDC(sum, t0)
	514	+ b done
	515	+ .set noreorder
	516	+dst_unaligned:
	517	+ /*
	518	+ * dst is unaligned
	519	+ * t0 = src & ADDRMASK
	520	+ * t1 = dst & ADDRMASK; T1 > 0
	521	+ * len >= NBYTES
	522	+ *
	523	+ * Copy enough bytes to align dst
	524	+ * Set match = (src and dst have same alignment)
	525	+ */
	526	+#define match rem
	527	+EXC( LDFIRST t3, FIRST(0)(src), l_exc)
	528	+ ADD t2, zero, NBYTES
	529	+EXC( LDREST t3, REST(0)(src), l_exc_copy)
	530	+ SUB t2, t2, t1 # t2 = number of bytes copied
	531	+ xor match, t0, t1
	532	+EXC( STFIRST t3, FIRST(0)(dst), s_exc)
	533	+ SLL t4, t1, 3 # t4 = number of bits to discard
	534	+ SHIFT_DISCARD t3, t3, t4
	535	+ /* no SHIFT_DISCARD_REVERT to handle odd buffer properly */
	536	+ ADDC(sum, t3)
	537	+ beq len, t2, done
	538	+ SUB len, len, t2
	539	+ ADD dst, dst, t2
	540	+ beqz match, both_aligned
	541	+ ADD src, src, t2
	542	+
	543	+src_unaligned_dst_aligned:
	544	+ SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
	545	+ beqz t0, cleanup_src_unaligned
	546	+ and rem, len, (4NBYTES-1) # rem = len % 4NBYTES
	547	+1:
	548	+/*
	549	+ * Avoid consecutive LD*'s to the same register since some mips
	550	+ * implementations can't issue them in the same cycle.
	551	+ * It's OK to load FIRST(N+1) before REST(N) because the two addresses
	552	+ * are to the same unit (unless src is aligned, but it's not).
	553	+ */
	554	+EXC( LDFIRST t0, FIRST(0)(src), l_exc)
	555	+EXC( LDFIRST t1, FIRST(1)(src), l_exc_copy)
	556	+ SUB len, len, 4*NBYTES
	557	+EXC( LDREST t0, REST(0)(src), l_exc_copy)
	558	+EXC( LDREST t1, REST(1)(src), l_exc_copy)
	559	+EXC( LDFIRST t2, FIRST(2)(src), l_exc_copy)
	560	+EXC( LDFIRST t3, FIRST(3)(src), l_exc_copy)
	561	+EXC( LDREST t2, REST(2)(src), l_exc_copy)
	562	+EXC( LDREST t3, REST(3)(src), l_exc_copy)
	563	+ ADD src, src, 4*NBYTES
	564	+#ifdef CONFIG_CPU_SB1
	565	+ nop # improves slotting
	566	+#endif
	567	+EXC( STORE t0, UNIT(0)(dst), s_exc)
	568	+ ADDC(sum, t0)
	569	+EXC( STORE t1, UNIT(1)(dst), s_exc)
	570	+ ADDC(sum, t1)
	571	+EXC( STORE t2, UNIT(2)(dst), s_exc)
	572	+ ADDC(sum, t2)
	573	+EXC( STORE t3, UNIT(3)(dst), s_exc)
	574	+ ADDC(sum, t3)
	575	+ bne len, rem, 1b
	576	+ ADD dst, dst, 4*NBYTES
	577	+
	578	+cleanup_src_unaligned:
	579	+ beqz len, done
	580	+ and rem, len, NBYTES-1 # rem = len % NBYTES
	581	+ beq rem, len, copy_bytes
	582	+ nop
	583	+1:
	584	+EXC( LDFIRST t0, FIRST(0)(src), l_exc)
	585	+EXC( LDREST t0, REST(0)(src), l_exc_copy)
	586	+ ADD src, src, NBYTES
	587	+ SUB len, len, NBYTES
	588	+EXC( STORE t0, 0(dst), s_exc)
	589	+ ADDC(sum, t0)
	590	+ bne len, rem, 1b
	591	+ ADD dst, dst, NBYTES
	592	+
	593	+copy_bytes_checklen:
	594	+ beqz len, done
	595	+ nop
	596	+copy_bytes:
	597	+ /* 0 < len < NBYTES */
	598	+#ifdef CONFIG_CPU_LITTLE_ENDIAN
	599	+#define SHIFT_START 0
	600	+#define SHIFT_INC 8
	601	+#else
	602	+#define SHIFT_START 8*(NBYTES-1)
	603	+#define SHIFT_INC -8
	604	+#endif
	605	+ move t2, zero # partial word
	606	+ li t3, SHIFT_START # shift
	607	+/* use l_exc_copy here to return correct sum on fault */
	608	+#define COPY_BYTE(N) \
	609	+EXC( lbu t0, N(src), l_exc_copy); \
	610	+ SUB len, len, 1; \
	611	+EXC( sb t0, N(dst), s_exc); \
	612	+ SLLV t0, t0, t3; \
	613	+ addu t3, SHIFT_INC; \
	614	+ beqz len, copy_bytes_done; \
	615	+ or t2, t0
	616	+
	617	+ COPY_BYTE(0)
	618	+ COPY_BYTE(1)
	619	+#ifdef USE_DOUBLE
	620	+ COPY_BYTE(2)
	621	+ COPY_BYTE(3)
	622	+ COPY_BYTE(4)
	623	+ COPY_BYTE(5)
	624	+#endif
	625	+EXC( lbu t0, NBYTES-2(src), l_exc_copy)
	626	+ SUB len, len, 1
	627	+EXC( sb t0, NBYTES-2(dst), s_exc)
	628	+ SLLV t0, t0, t3
	629	+ or t2, t0
	630	+copy_bytes_done:
	631	+ ADDC(sum, t2)
	632	+done:
	633	+ /* fold checksum */
	634	+#ifdef USE_DOUBLE
	635	+ dsll32 v1, sum, 0
	636	+ daddu sum, v1
	637	+ sltu v1, sum, v1
	638	+ dsra32 sum, sum, 0
	639	+ addu sum, v1
	640	+#endif
	641	+ sll v1, sum, 16
	642	+ addu sum, v1
	643	+ sltu v1, sum, v1
	644	+ srl sum, sum, 16
	645	+ addu sum, v1
	646	+
	647	+ /* odd buffer alignment? */
	648	+ beqz odd, 1f
	649	+ nop
	650	+ sll v1, sum, 8
	651	+ srl sum, sum, 8
	652	+ or sum, v1
	653	+ andi sum, 0xffff
	654	+1:
	655	+ .set reorder
	656	+ ADDC(sum, psum)
	657	+ jr ra
	658	+ .set noreorder
	659	+
	660	+l_exc_copy:
	661	+ /*
	662	+ * Copy bytes from src until faulting load address (or until a
	663	+ * lb faults)
	664	+ *
	665	+ * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
	666	+ * may be more than a byte beyond the last address.
	667	+ * Hence, the lb below may get an exception.
	668	+ *
	669	+ * Assumes src < THREAD_BUADDR($28)
	670	+ */
	671	+ LOAD t0, TI_TASK($28)
	672	+ li t2, SHIFT_START
	673	+ LOAD t0, THREAD_BUADDR(t0)
	674	+1:
	675	+EXC( lbu t1, 0(src), l_exc)
	676	+ ADD src, src, 1
	677	+ sb t1, 0(dst) # can't fault -- we're copy_from_user
	678	+ SLLV t1, t1, t2
	679	+ addu t2, SHIFT_INC
	680	+ ADDC(sum, t1)
	681	+ bne src, t0, 1b
	682	+ ADD dst, dst, 1
	683	+l_exc:
	684	+ LOAD t0, TI_TASK($28)
	685	+ nop
	686	+ LOAD t0, THREAD_BUADDR(t0) # t0 is just past last good address
	687	+ nop
	688	+ SUB len, AT, t0 # len number of uncopied bytes
	689	+ /*
	690	+ * Here's where we rely on src and dst being incremented in tandem,
	691	+ * See (3) above.
	692	+ * dst += (fault addr - src) to put dst at first byte to clear
	693	+ */
	694	+ ADD dst, t0 # compute start address in a1
	695	+ SUB dst, src
	696	+ /*
	697	+ * Clear len bytes starting at dst. Can't call __bzero because it
	698	+ * might modify len. An inefficient loop for these rare times...
	699	+ */
	700	+ beqz len, done
	701	+ SUB src, len, 1
	702	+1: sb zero, 0(dst)
	703	+ ADD dst, dst, 1
	704	+ bnez src, 1b
	705	+ SUB src, src, 1
	706	+ li v1, -EFAULT
	707	+ b done
	708	+ sw v1, (errptr)
	709	+
	710	+s_exc:
	711	+ li v0, -1 /* invalid checksum */
	712	+ li v1, -EFAULT
	713	+ jr ra
	714	+ sw v1, (errptr)
	715	+ END(__csum_partial_copy_user)
1		-/*
2		- * This file is subject to the terms and conditions of the GNU General Public
3		- * License. See the file "COPYING" in the main directory of this archive
4		- * for more details.
5		- *
6		- * Copyright (C) 1994, 1995 Waldorf Electronics GmbH
7		- * Copyright (C) 1998, 1999 Ralf Baechle
8		- */
9		-#include <linux/kernel.h>
10		-#include <linux/module.h>
11		-#include <linux/types.h>
12		-#include <asm/byteorder.h>
13		-#include <asm/string.h>
14		-#include <asm/uaccess.h>
15		-#include <net/checksum.h>
16		-
17		-/*
18		- * copy while checksumming, otherwise like csum_partial
19		- */
20		-__wsum csum_partial_copy_nocheck(const void *src,
21		- void *dst, int len, __wsum sum)
22		-{
23		- /*
24		- * It's 2:30 am and I don't feel like doing it real ...
25		- * This is lots slower than the real thing (tm)
26		- */
27		- sum = csum_partial(src, len, sum);
28		- memcpy(dst, src, len);
29		-
30		- return sum;
31		-}
32		-
33		-EXPORT_SYMBOL(csum_partial_copy_nocheck);
34		-
35		-/*
36		- * Copy from userspace and compute checksum. If we catch an exception
37		- * then zero the rest of the buffer.
38		- */
39		-__wsum csum_partial_copy_from_user (const void __user *src,
40		- void dst, int len, __wsum sum, int err_ptr)
41		-{
42		- int missing;
43		-
44		- might_sleep();
45		- missing = copy_from_user(dst, src, len);
46		- if (missing) {
47		- memset(dst + len - missing, 0, missing);
48		- *err_ptr = -EFAULT;
49		- }
50		-
51		- return csum_partial(dst, len, sum);
52		-}
...	...	@@ -29,31 +29,38 @@
29	29	*/
30	30	__wsum csum_partial(const void *buff, int len, __wsum sum);
31	31
	32	+__wsum __csum_partial_copy_user(const void src, void dst,
	33	+ int len, __wsum sum, int *err_ptr);
	34	+
32	35	/*
33	36	* this is a new version of the above that records errors it finds in *errp,
34	37	* but continues and zeros the rest of the buffer.
35	38	*/
36		-__wsum csum_partial_copy_from_user(const void __user *src,
37		- void *dst, int len,
38		- __wsum sum, int *errp);
	39	+static inline
	40	+__wsum csum_partial_copy_from_user(const void __user src, void dst, int len,
	41	+ __wsum sum, int *err_ptr)
	42	+{
	43	+ might_sleep();
	44	+ return __csum_partial_copy_user((__force void *)src, dst,
	45	+ len, sum, err_ptr);
	46	+}
39	47
40	48	/*
41	49	* Copy and checksum to user
42	50	*/
43	51	#define HAVE_CSUM_COPY_USER
44		-static inline __wsum csum_and_copy_to_user (const void src, void __user dst,
45		- int len, __wsum sum,
46		- int *err_ptr)
	52	+static inline
	53	+__wsum csum_and_copy_to_user(const void src, void __user dst, int len,
	54	+ __wsum sum, int *err_ptr)
47	55	{
48	56	might_sleep();
49		- sum = csum_partial(src, len, sum);
50		-
51		- if (copy_to_user(dst, src, len)) {
	57	+ if (access_ok(VERIFY_WRITE, dst, len))
	58	+ return __csum_partial_copy_user(src, (__force void *)dst,
	59	+ len, sum, err_ptr);
	60	+ if (len)
52	61	*err_ptr = -EFAULT;
53		- return (__force __wsum)-1;
54		- }
55	62
56		- return sum;
	63	+ return (__force __wsum)-1; /* invalid checksum */
57	64	}
58	65
59	66	/*