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: Al Viro <viro@ZenIV.linux.org.uk>
Cc: <stable@vger.kernel.org> # v4.9+
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge 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