[MIPS] Fixup copy_from_user_inatomic

From the 01408c4939479ec46c15aa7ef6e2406be50eeeca log message:

The problem is that when we write to a file, the copy from userspace to
pagecache is first done with preemption disabled, so if the source
address is not immediately available the copy fails *and* *zeros* *the*
*destination*.

This is a problem because a concurrent read (which admittedly is an odd
thing to do) might see zeros rather that was there before the write, or
what was there after, or some mixture of the two (any of these being a
reasonable thing to see).

If the copy did fail, it will immediately be retried with preemption
re-enabled so any transient problem with accessing the source won't
cause an error.

The first copying does not need to zero any uncopied bytes, and doing
so causes the problem.  It uses copy_from_user_atomic rather than
copy_from_user so the simple expedient is to change copy_from_user_atomic
to *not* zero out bytes on failure.

< --- end cite --- >

This patch finally implements at least a not so pretty solution by
duplicating the relevant part of __copy_user.

Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

2 files changed, 437 insertions, 1 deletions

diff --git a/arch/mips/lib/Makefile b/arch/mips/lib/Makefile
index 9e5d985936b3..2453ea244cb8 100644
--- a/arch/mips/lib/Makefile
+++ b/arch/mips/lib/Makefile
@@ -2,7 +2,7 @@
 # Makefile for MIPS-specific library files..
 #
 
-lib-y	+= csum_partial.o memcpy.o memset.o promlib.o \
+lib-y	+= csum_partial.o memcpy.o memcpy-inatomic.o memset.o promlib.o \
 	   strlen_user.o strncpy_user.o strnlen_user.o uncached.o
 
 obj-y			+= iomap.o
diff --git a/arch/mips/lib/memcpy-inatomic.S b/arch/mips/lib/memcpy-inatomic.S
new file mode 100644
index 000000000000..3a534b2baa0f
--- /dev/null
+++ b/arch/mips/lib/memcpy-inatomic.S
@@ -0,0 +1,436 @@
+/*
+ * 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.
+ *
+ * Unified implementation of memcpy, memmove and the __copy_user backend.
+ *
+ * Copyright (C) 1998, 99, 2000, 01, 2002 Ralf Baechle (ralf@gnu.org)
+ * Copyright (C) 1999, 2000, 01, 2002 Silicon Graphics, Inc.
+ * Copyright (C) 2002 Broadcom, Inc.
+ *   memcpy/copy_user author: Mark Vandevoorde
+ *
+ * Mnemonic names for arguments to memcpy/__copy_user
+ */
+
+/*
+ * Hack to resolve longstanding prefetch issue
+ *
+ * Prefetching may be fatal on some systems if we're prefetching beyond the
+ * end of memory on some systems.  It's also a seriously bad idea on non
+ * dma-coherent systems.
+ */
+#if !defined(CONFIG_DMA_COHERENT) || !defined(CONFIG_DMA_IP27)
+#undef CONFIG_CPU_HAS_PREFETCH
+#endif
+#ifdef CONFIG_MIPS_MALTA
+#undef CONFIG_CPU_HAS_PREFETCH
+#endif
+
+#include <asm/asm.h>
+#include <asm/asm-offsets.h>
+#include <asm/regdef.h>
+
+#define dst a0
+#define src a1
+#define len a2
+
+/*
+ * Spec
+ *
+ * memcpy copies len bytes from src to dst and sets v0 to dst.
+ * It assumes that
+ *   - src and dst don't overlap
+ *   - src is readable
+ *   - dst is writable
+ * memcpy uses the standard calling convention
+ *
+ * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
+ * the number of uncopied bytes due to an exception caused by a read or write.
+ * __copy_user assumes that src and dst don't overlap, and that the call is
+ * implementing one of the following:
+ *   copy_to_user
+ *     - src is readable  (no exceptions when reading src)
+ *   copy_from_user
+ *     - dst is writable  (no exceptions when writing dst)
+ * __copy_user uses a non-standard calling convention; see
+ * include/asm-mips/uaccess.h
+ *
+ * When an exception happens on a load, the handler must
+ # ensure that all of the destination buffer is overwritten to prevent
+ * leaking information to user mode programs.
+ */
+
+/*
+ * Implementation
+ */
+
+/*
+ * 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 uaccess.h and maintained by not writing AT in copy_user
+ * (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 adjust len (if necessary) and return.
+ * These handlers do not need to overwrite any data.
+ *
+ * For __rmemcpy and memmove an exception is always a kernel bug, therefore
+ * they're not protected.
+ */
+
+#define EXC(inst_reg,addr,handler)		\
+9:	inst_reg, addr;				\
+	.section __ex_table,"a";		\
+	PTR	9b, handler;			\
+	.previous
+
+/*
+ * Only on the 64-bit kernel we can made use of 64-bit registers.
+ */
+#ifdef CONFIG_64BIT
+#define USE_DOUBLE
+#endif
+
+#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 SRA    dsra
+#define SLL    dsll
+#define SLLV   dsllv
+#define SRLV   dsrlv
+#define NBYTES 8
+#define LOG_NBYTES 3
+
+/*
+ * As we are sharing code base with the mips32 tree (which use the o32 ABI
+ * register definitions). We need to redefine the register definitions from
+ * the n64 ABI register naming to the o32 ABI register naming.
+ */
+#undef t0
+#undef t1
+#undef t2
+#undef t3
+#define t0	$8
+#define t1	$9
+#define t2	$10
+#define t3	$11
+#define t4	$12
+#define t5	$13
+#define t6	$14
+#define t7	$15
+
+#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 SRA    sra
+#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
+#else
+#define LDFIRST LOADL
+#define LDREST  LOADR
+#define STFIRST STOREL
+#define STREST  STORER
+#define SHIFT_DISCARD SRLV
+#endif
+
+#define FIRST(unit) ((unit)*NBYTES)
+#define REST(unit)  (FIRST(unit)+NBYTES-1)
+#define UNIT(unit)  FIRST(unit)
+
+#define ADDRMASK (NBYTES-1)
+
+	.text
+	.set	noreorder
+	.set	noat
+
+/*
+ * A combined memcpy/__copy_user
+ * __copy_user sets len to 0 for success; else to an upper bound of
+ * the number of uncopied bytes.
+ * memcpy sets v0 to dst.
+ */
+	.align	5
+LEAF(__copy_user_inatomic)
+	/*
+	 * Note: dst & src may be unaligned, len may be 0
+	 * Temps
+	 */
+#define rem t8
+
+	/*
+	 * 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.
+	 */
+	PREF(	0, 0(src) )
+	PREF(	1, 0(dst) )
+	sltu	t2, len, NBYTES
+	and	t1, dst, ADDRMASK
+	PREF(	0, 1*32(src) )
+	PREF(	1, 1*32(dst) )
+	bnez	t2, copy_bytes_checklen
+	 and	t0, src, ADDRMASK
+	PREF(	0, 2*32(src) )
+	PREF(	1, 2*32(dst) )
+	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
+	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
+	PREF(	0, 3*32(src) )
+	PREF(	1, 3*32(dst) )
+	.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)
+	SUB	len, len, 8*NBYTES
+EXC(	LOAD	t4, UNIT(4)(src),	l_exc_copy)
+EXC(	LOAD	t7, UNIT(5)(src),	l_exc_copy)
+	STORE	t0, UNIT(0)(dst)
+	STORE	t1, UNIT(1)(dst)
+EXC(	LOAD	t0, UNIT(6)(src),	l_exc_copy)
+EXC(	LOAD	t1, UNIT(7)(src),	l_exc_copy)
+	ADD	src, src, 8*NBYTES
+	ADD	dst, dst, 8*NBYTES
+	STORE	t2, UNIT(-6)(dst)
+	STORE	t3, UNIT(-5)(dst)
+	STORE	t4, UNIT(-4)(dst)
+	STORE	t7, UNIT(-3)(dst)
+	STORE	t0, UNIT(-2)(dst)
+	STORE	t1, UNIT(-1)(dst)
+	PREF(	0, 8*32(src) )
+	PREF(	1, 8*32(dst) )
+	bne	len, rem, 1b
+	 nop
+
+	/*
+	 * len == rem == the number of bytes left to copy < 8*NBYTES
+	 */
+cleanup_both_aligned:
+	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
+	STORE	t0, UNIT(0)(dst)
+	STORE	t1, UNIT(1)(dst)
+	STORE	t2, UNIT(2)(dst)
+	STORE	t3, UNIT(3)(dst)
+	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
+	STORE	t0, 0(dst)
+	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
+	STREST	t0, -1(t1)
+	jr	ra
+	 move	len, zero
+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
+	STFIRST t3, FIRST(0)(dst)
+	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
+	PREF(	0, 3*32(src) )
+	beqz	t0, cleanup_src_unaligned
+	 and	rem, len, (4*NBYTES-1)   # rem = len % 4*NBYTES
+	PREF(	1, 3*32(dst) )
+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)
+	PREF(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
+	ADD	src, src, 4*NBYTES
+#ifdef CONFIG_CPU_SB1
+	nop				# improves slotting
+#endif
+	STORE	t0, UNIT(0)(dst)
+	STORE	t1, UNIT(1)(dst)
+	STORE	t2, UNIT(2)(dst)
+	STORE	t3, UNIT(3)(dst)
+	PREF(	1, 9*32(dst) )     	# 1 is PREF_STORE (not streamed)
+	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
+	STORE	t0, 0(dst)
+	bne	len, rem, 1b
+	 ADD	dst, dst, NBYTES
+
+copy_bytes_checklen:
+	beqz	len, done
+	 nop
+copy_bytes:
+	/* 0 < len < NBYTES  */
+#define COPY_BYTE(N)			\
+EXC(	lb	t0, N(src), l_exc);	\
+	SUB	len, len, 1;		\
+	beqz	len, done;		\
+	 sb	t0, N(dst)
+
+	COPY_BYTE(0)
+	COPY_BYTE(1)
+#ifdef USE_DOUBLE
+	COPY_BYTE(2)
+	COPY_BYTE(3)
+	COPY_BYTE(4)
+	COPY_BYTE(5)
+#endif
+EXC(	lb	t0, NBYTES-2(src), l_exc)
+	SUB	len, len, 1
+	jr	ra
+	 sb	t0, NBYTES-2(dst)
+done:
+	jr	ra
+	 nop
+	END(__copy_user_inatomic)
+
+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)
+	 nop
+	LOAD	t0, THREAD_BUADDR(t0)
+1:
+EXC(	lb	t1, 0(src),	l_exc)
+	ADD	src, src, 1
+	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
+	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
+	jr	ra
+	 nop