Introduce unified API to zero memory

Introduce zeromem_dczva function on AArch64 that can handle unaligned
addresses and make use of DC ZVA instruction to zero a whole block at a
time. This zeroing takes place directly in the cache to speed it up
without doing external memory access.

Remove the zeromem16 function on AArch64 and replace it with an alias to
zeromem. This zeromem16 function is now deprecated.

Remove the 16-bytes alignment constraint on __BSS_START__ in
firmware-design.md as it is now not mandatory anymore (it used to comply
with zeromem16 requirements).

Change the 16-bytes alignment constraints in SP min's linker script to a
8-bytes alignment constraint as the AArch32 zeromem implementation is now
more efficient on 8-bytes aligned addresses.

Introduce zero_normalmem and zeromem helpers in platform agnostic header
that are implemented this way:
* AArch32:
	* zero_normalmem: zero using usual data access
	* zeromem: alias for zero_normalmem
* AArch64:
	* zero_normalmem: zero normal memory  using DC ZVA instruction
	                  (needs MMU enabled)
	* zeromem: zero using usual data access

Usage guidelines: in most cases, zero_normalmem should be preferred.

There are 2 scenarios where zeromem (or memset) must be used instead:
* Code that must run with MMU disabled (which means all memory is
  considered device memory for data accesses).
* Code that fills device memory with null bytes.

Optionally, the following rule can be applied if performance is
important:
* Code zeroing small areas (few bytes) that are not secrets should use
  memset to take advantage of compiler optimizations.

  Note: Code zeroing security-related critical information should use
  zero_normalmem/zeromem instead of memset to avoid removal by
  compilers' optimizations in some cases or misbehaving versions of GCC.

Fixes ARM-software/tf-issues#408

Change-Id: Iafd9663fc1070413c3e1904e54091cf60effaa82
Signed-off-by: Douglas Raillard <douglas.raillard@arm.com>

1 files changed, 349 insertions, 20 deletions

diff --git a/lib/aarch64/misc_helpers.S b/lib/aarch64/misc_helpers.S
index 574146f6..84265e0b 100644
--- a/lib/aarch64/misc_helpers.S
+++ b/lib/aarch64/misc_helpers.S
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2013-2014, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2013-2017, ARM Limited and Contributors. All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -37,6 +37,8 @@
 	.globl	eret
 	.globl	smc
 
+	.globl	zero_normalmem
+	.globl	zeromem
 	.globl	zeromem16
 	.globl	memcpy16
 
@@ -80,31 +82,358 @@ endfunc smc
  *
  * Initialise a memory region to 0.
  * The memory address must be 16-byte aligned.
+ * NOTE: This function is deprecated and zeromem should be used instead.
  * -----------------------------------------------------------------------
  */
-func zeromem16
+.equ	zeromem16, zeromem
+
+/* -----------------------------------------------------------------------
+ * void zero_normalmem(void *mem, unsigned int length);
+ *
+ * Initialise a region in normal memory to 0. This functions complies with the
+ * AAPCS and can be called from C code.
+ *
+ * NOTE: MMU must be enabled when using this function as it can only operate on
+ *       normal memory. It is intended to be mainly used from C code when MMU
+ *       is usually enabled.
+ * -----------------------------------------------------------------------
+ */
+.equ	zero_normalmem, zeromem_dczva
+
+/* -----------------------------------------------------------------------
+ * void zeromem(void *mem, unsigned int length);
+ *
+ * Initialise a region of device memory to 0. This functions complies with the
+ * AAPCS and can be called from C code.
+ *
+ * NOTE: When data caches and MMU are enabled, zero_normalmem can usually be
+ *       used instead for faster zeroing.
+ *
+ * -----------------------------------------------------------------------
+ */
+func zeromem
+	/* x2 is the address past the last zeroed address */
+	add	x2, x0, x1
+	/*
+	 * Uses the fallback path that does not use DC ZVA instruction and
+	 * therefore does not need enabled MMU
+	 */
+	b	.Lzeromem_dczva_fallback_entry
+endfunc zeromem
+
+/* -----------------------------------------------------------------------
+ * void zeromem_dczva(void *mem, unsigned int length);
+ *
+ * Fill a region of normal memory of size "length" in bytes with null bytes.
+ * MMU must be enabled and the memory be of
+ * normal type. This is because this function internally uses the DC ZVA
+ * instruction, which generates an Alignment fault if used on any type of
+ * Device memory (see section D3.4.9 of the ARMv8 ARM, issue k). When the MMU
+ * is disabled, all memory behaves like Device-nGnRnE memory (see section
+ * D4.2.8), hence the requirement on the MMU being enabled.
+ * NOTE: The code assumes that the block size as defined in DCZID_EL0
+ *       register is at least 16 bytes.
+ *
+ * -----------------------------------------------------------------------
+ */
+func zeromem_dczva
+
+	/*
+	 * The function consists of a series of loops that zero memory one byte
+	 * at a time, 16 bytes at a time or using the DC ZVA instruction to
+	 * zero aligned block of bytes, which is assumed to be more than 16.
+	 * In the case where the DC ZVA instruction cannot be used or if the
+	 * first 16 bytes loop would overflow, there is fallback path that does
+	 * not use DC ZVA.
+	 * Note: The fallback path is also used by the zeromem function that
+	 *       branches to it directly.
+	 *
+	 *              +---------+   zeromem_dczva
+	 *              |  entry  |
+	 *              +----+----+
+	 *                   |
+	 *                   v
+	 *              +---------+
+	 *              | checks  |>o-------+ (If any check fails, fallback)
+	 *              +----+----+         |
+	 *                   |              |---------------+
+	 *                   v              | Fallback path |
+	 *            +------+------+       |---------------+
+	 *            | 1 byte loop |       |
+	 *            +------+------+ .Lzeromem_dczva_initial_1byte_aligned_end
+	 *                   |              |
+	 *                   v              |
+	 *           +-------+-------+      |
+	 *           | 16 bytes loop |      |
+	 *           +-------+-------+      |
+	 *                   |              |
+	 *                   v              |
+	 *            +------+------+ .Lzeromem_dczva_blocksize_aligned
+	 *            | DC ZVA loop |       |
+	 *            +------+------+       |
+	 *       +--------+  |              |
+	 *       |        |  |              |
+	 *       |        v  v              |
+	 *       |   +-------+-------+ .Lzeromem_dczva_final_16bytes_aligned
+	 *       |   | 16 bytes loop |      |
+	 *       |   +-------+-------+      |
+	 *       |           |              |
+	 *       |           v              |
+	 *       |    +------+------+ .Lzeromem_dczva_final_1byte_aligned
+	 *       |    | 1 byte loop |       |
+	 *       |    +-------------+       |
+	 *       |           |              |
+	 *       |           v              |
+	 *       |       +---+--+           |
+	 *       |       | exit |           |
+	 *       |       +------+           |
+	 *       |			    |
+	 *       |           +--------------+    +------------------+ zeromem
+	 *       |           |  +----------------| zeromem function |
+	 *       |           |  |                +------------------+
+	 *       |           v  v
+	 *       |    +-------------+ .Lzeromem_dczva_fallback_entry
+	 *       |    | 1 byte loop |
+	 *       |    +------+------+
+	 *       |           |
+	 *       +-----------+
+	 */
+
+	/*
+	 * Readable names for registers
+	 *
+	 * Registers x0, x1 and x2 are also set by zeromem which
+	 * branches into the fallback path directly, so cursor, length and
+	 * stop_address should not be retargeted to other registers.
+	 */
+	cursor       .req x0 /* Start address and then current address */
+	length       .req x1 /* Length in bytes of the region to zero out */
+	/* Reusing x1 as length is never used after block_mask is set */
+	block_mask   .req x1 /* Bitmask of the block size read in DCZID_EL0 */
+	stop_address .req x2 /* Address past the last zeroed byte */
+	block_size   .req x3 /* Size of a block in bytes as read in DCZID_EL0 */
+	tmp1         .req x4
+	tmp2         .req x5
+
 #if ASM_ASSERTION
-	tst	x0, #0xf
-	ASM_ASSERT(eq)
+	/*
+	 * Check for M bit (MMU enabled) of the current SCTLR_EL(1|3)
+	 * register value and panic if the MMU is disabled.
+	 */
+#if defined(IMAGE_BL1) || defined(IMAGE_BL31)
+	mrs	tmp1, sctlr_el3
+#else
+	mrs	tmp1, sctlr_el1
 #endif
-	add	x2, x0, x1
-/* zero 16 bytes at a time */
-z_loop16:
-	sub	x3, x2, x0
-	cmp	x3, #16
-	b.lt	z_loop1
-	stp	xzr, xzr, [x0], #16
-	b	z_loop16
-/* zero byte per byte */
-z_loop1:
-	cmp	x0, x2
-	b.eq	z_end
-	strb	wzr, [x0], #1
-	b	z_loop1
-z_end:
+
+	tst	tmp1, #SCTLR_M_BIT
+	ASM_ASSERT(ne)
+#endif /* ASM_ASSERTION */
+
+	/* stop_address is the address past the last to zero */
+	add	stop_address, cursor, length
+
+	/*
+	 * Get block_size = (log2(<block size>) >> 2) (see encoding of
+	 * dczid_el0 reg)
+	 */
+	mrs	block_size, dczid_el0
+
+	/*
+	 * Select the 4 lowest bits and convert the extracted log2(<block size
+	 * in words>) to <block size in bytes>
+	 */
+	ubfx	block_size, block_size, #0, #4
+	mov	tmp2, #(1 << 2)
+	lsl	block_size, tmp2, block_size
+
+#if ASM_ASSERTION
+	/*
+	 * Assumes block size is at least 16 bytes to avoid manual realignment
+	 * of the cursor at the end of the DCZVA loop.
+	 */
+	cmp	block_size, #16
+	ASM_ASSERT(hs)
+#endif
+	/*
+	 * Not worth doing all the setup for a region less than a block and
+	 * protects against zeroing a whole block when the area to zero is
+	 * smaller than that. Also, as it is assumed that the block size is at
+	 * least 16 bytes, this also protects the initial aligning loops from
+	 * trying to zero 16 bytes when length is less than 16.
+	 */
+	cmp	length, block_size
+	b.lo	.Lzeromem_dczva_fallback_entry
+
+	/*
+	 * Calculate the bitmask of the block alignment. It will never
+	 * underflow as the block size is between 4 bytes and 2kB.
+	 * block_mask = block_size - 1
+	 */
+	sub	block_mask, block_size, #1
+
+	/*
+	 * length alias should not be used after this point unless it is
+	 * defined as a register other than block_mask's.
+	 */
+	 .unreq length
+
+	/*
+	 * If the start address is already aligned to zero block size, go
+	 * straight to the cache zeroing loop. This is safe because at this
+	 * point, the length cannot be smaller than a block size.
+	 */
+	tst	cursor, block_mask
+	b.eq	.Lzeromem_dczva_blocksize_aligned
+
+	/*
+	 * Calculate the first block-size-aligned address. It is assumed that
+	 * the zero block size is at least 16 bytes. This address is the last
+	 * address of this initial loop.
+	 */
+	orr	tmp1, cursor, block_mask
+	add	tmp1, tmp1, #1
+
+	/*
+	 * If the addition overflows, skip the cache zeroing loops. This is
+	 * quite unlikely however.
+	 */
+	cbz	tmp1, .Lzeromem_dczva_fallback_entry
+
+	/*
+	 * If the first block-size-aligned address is past the last address,
+	 * fallback to the simpler code.
+	 */
+	cmp	tmp1, stop_address
+	b.hi	.Lzeromem_dczva_fallback_entry
+
+	/*
+	 * If the start address is already aligned to 16 bytes, skip this loop.
+	 * It is safe to do this because tmp1 (the stop address of the initial
+	 * 16 bytes loop) will never be greater than the final stop address.
+	 */
+	tst	cursor, #0xf
+	b.eq	.Lzeromem_dczva_initial_1byte_aligned_end
+
+	/* Calculate the next address aligned to 16 bytes */
+	orr	tmp2, cursor, #0xf
+	add	tmp2, tmp2, #1
+	/* If it overflows, fallback to the simple path (unlikely) */
+	cbz	tmp2, .Lzeromem_dczva_fallback_entry
+	/*
+	 * Next aligned address cannot be after the stop address because the
+	 * length cannot be smaller than 16 at this point.
+	 */
+
+	/* First loop: zero byte per byte */
+1:
+	strb	wzr, [cursor], #1
+	cmp	cursor, tmp2
+	b.ne	1b
+.Lzeromem_dczva_initial_1byte_aligned_end:
+
+	/*
+	 * Second loop: we need to zero 16 bytes at a time from cursor to tmp1
+	 * before being able to use the code that deals with block-size-aligned
+	 * addresses.
+	 */
+	cmp	cursor, tmp1
+	b.hs	2f
+1:
+	stp	xzr, xzr, [cursor], #16
+	cmp	cursor, tmp1
+	b.lo	1b
+2:
+
+	/*
+	 * Third loop: zero a block at a time using DC ZVA cache block zeroing
+	 * instruction.
+	 */
+.Lzeromem_dczva_blocksize_aligned:
+	/*
+	 * Calculate the last block-size-aligned address. If the result equals
+	 * to the start address, the loop will exit immediately.
+	 */
+	bic	tmp1, stop_address, block_mask
+
+	cmp	cursor, tmp1
+	b.hs	2f
+1:
+	/* Zero the block containing the cursor */
+	dc	zva, cursor
+	/* Increment the cursor by the size of a block */
+	add	cursor, cursor, block_size
+	cmp	cursor, tmp1
+	b.lo	1b
+2:
+
+	/*
+	 * Fourth loop: zero 16 bytes at a time and then byte per byte the
+	 * remaining area
+	 */
+.Lzeromem_dczva_final_16bytes_aligned:
+	/*
+	 * Calculate the last 16 bytes aligned address. It is assumed that the
+	 * block size will never be smaller than 16 bytes so that the current
+	 * cursor is aligned to at least 16 bytes boundary.
+	 */
+	bic	tmp1, stop_address, #15
+
+	cmp	cursor, tmp1
+	b.hs	2f
+1:
+	stp	xzr, xzr, [cursor], #16
+	cmp	cursor, tmp1
+	b.lo	1b
+2:
+
+	/* Fifth and final loop: zero byte per byte */
+.Lzeromem_dczva_final_1byte_aligned:
+	cmp	cursor, stop_address
+	b.eq	2f
+1:
+	strb	wzr, [cursor], #1
+	cmp	cursor, stop_address
+	b.ne	1b
+2:
 	ret
-endfunc zeromem16
 
+	/* Fallback for unaligned start addresses */
+.Lzeromem_dczva_fallback_entry:
+	/*
+	 * If the start address is already aligned to 16 bytes, skip this loop.
+	 */
+	tst	cursor, #0xf
+	b.eq	.Lzeromem_dczva_final_16bytes_aligned
+
+	/* Calculate the next address aligned to 16 bytes */
+	orr	tmp1, cursor, #15
+	add	tmp1, tmp1, #1
+	/* If it overflows, fallback to byte per byte zeroing */
+	cbz	tmp1, .Lzeromem_dczva_final_1byte_aligned
+	/* If the next aligned address is after the stop address, fall back */
+	cmp	tmp1, stop_address
+	b.hs	.Lzeromem_dczva_final_1byte_aligned
+
+	/* Fallback entry loop: zero byte per byte */
+1:
+	strb	wzr, [cursor], #1
+	cmp	cursor, tmp1
+	b.ne	1b
+
+	b	.Lzeromem_dczva_final_16bytes_aligned
+
+	.unreq	cursor
+	/*
+	 * length is already unreq'ed to reuse the register for another
+	 * variable.
+	 */
+	.unreq	stop_address
+	.unreq	block_size
+	.unreq	block_mask
+	.unreq	tmp1
+	.unreq	tmp2
+endfunc zeromem_dczva
 
 /* --------------------------------------------------------------------------
  * void memcpy16(void *dest, const void *src, unsigned int length)