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  ...

bl1/bl1.ld.S

 /*
- * Copyright (c) 2013-2016, 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:
@@ -133,7 +133,8 @@ SECTIONS
 
     /*
      * The .bss section gets initialised to 0 at runtime.
-     * Its base address must be 16-byte aligned.
+     * Its base address should be 16-byte aligned for better performance of the
+     * zero-initialization code.
      */
     .bss : ALIGN(16) {
         __BSS_START__ = .;

bl1/bl1_fwu.c

 /*
- * Copyright (c) 2015-2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2015-2017, 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:
@@ -335,7 +335,7 @@ static int bl1_fwu_image_auth(unsigned int image_id,
 		 */
 		if (image_desc->state == IMAGE_STATE_COPIED) {
 			/* Clear the memory.*/
-			memset((void *)base_addr, 0, total_size);
+			zero_normalmem((void *)base_addr, total_size);
 			flush_dcache_range(base_addr, total_size);
 
 			/* Indicate that image can be copied again*/

bl2/aarch64/bl2_entrypoint.S

 /*
- * Copyright (c) 2013-2016, 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:
@@ -94,12 +94,12 @@ func bl2_entrypoint
 	 */
 	ldr	x0, =__BSS_START__
 	ldr	x1, =__BSS_SIZE__
-	bl	zeromem16
+	bl	zeromem
 
 #if USE_COHERENT_MEM
 	ldr	x0, =__COHERENT_RAM_START__
 	ldr	x1, =__COHERENT_RAM_UNALIGNED_SIZE__
-	bl	zeromem16
+	bl	zeromem
 #endif
 
 	/* --------------------------------------------

bl2/bl2.ld.S

 /*
- * 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:
@@ -113,7 +113,8 @@ SECTIONS
 
     /*
      * The .bss section gets initialised to 0 at runtime.
-     * Its base address must be 16-byte aligned.
+     * Its base address should be 16-byte aligned for better performance of the
+     * zero-initialization code.
      */
     .bss : ALIGN(16) {
         __BSS_START__ = .;

bl2u/aarch64/bl2u_entrypoint.S

 /*
- * Copyright (c) 2015-2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2015-2017, 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:
@@ -94,7 +94,7 @@ func bl2u_entrypoint
 	 */
 	ldr	x0, =__BSS_START__
 	ldr	x1, =__BSS_SIZE__
-	bl	zeromem16
+	bl	zeromem
 
 	/* --------------------------------------------
 	 * Allocate a stack whose memory will be marked

bl2u/bl2u.ld.S

 /*
- * Copyright (c) 2015, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2015-2017, 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:
@@ -100,7 +100,8 @@ SECTIONS
 
     /*
      * The .bss section gets initialised to 0 at runtime.
-     * Its base address must be 16-byte aligned.
+     * Its base address should be 16-byte aligned for better performance of the
+     * zero-initialization code.
      */
     .bss : ALIGN(16) {
         __BSS_START__ = .;

bl31/bl31.ld.S

 /*
- * Copyright (c) 2013-2016, 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:
@@ -158,7 +158,8 @@ SECTIONS
 
     /*
      * The .bss section gets initialised to 0 at runtime.
-     * Its base address must be 16-byte aligned.
+     * Its base address should be 16-byte aligned for better performance of the
+     * zero-initialization code.
      */
     .bss (NOLOAD) : ALIGN(16) {
         __BSS_START__ = .;

bl32/sp_min/sp_min.ld.S

 /*
- * Copyright (c) 2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2016-2017, 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:
@@ -134,9 +134,10 @@ SECTIONS
 
     /*
      * The .bss section gets initialised to 0 at runtime.
-     * Its base address must be 16-byte aligned.
+     * Its base address should be 8-byte aligned for better performance of the
+     * zero-initialization code.
      */
-    .bss (NOLOAD) : ALIGN(16) {
+    .bss (NOLOAD) : ALIGN(8) {
         __BSS_START__ = .;
         *(.bss*)
         *(COMMON)

bl32/tsp/aarch64/tsp_entrypoint.S

 /*
- * Copyright (c) 2013-2016, 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:
@@ -120,12 +120,12 @@ func tsp_entrypoint
 	 */
 	ldr	x0, =__BSS_START__
 	ldr	x1, =__BSS_SIZE__
-	bl	zeromem16
+	bl	zeromem
 
 #if USE_COHERENT_MEM
 	ldr	x0, =__COHERENT_RAM_START__
 	ldr	x1, =__COHERENT_RAM_UNALIGNED_SIZE__
-	bl	zeromem16
+	bl	zeromem
 #endif
 
 	/* --------------------------------------------

bl32/tsp/tsp.ld.S

 /*
- * 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:
@@ -104,7 +104,8 @@ SECTIONS
 
     /*
      * The .bss section gets initialised to 0 at runtime.
-     * Its base address must be 16-byte aligned.
+     * Its base address should be 16-byte aligned for better performance of the
+     * zero-initialization code.
      */
     .bss : ALIGN(16) {
         __BSS_START__ = .;

common/bl_common.c

 /*
- * Copyright (c) 2013-2016, 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:
@@ -347,7 +347,7 @@ static int load_auth_image_internal(unsigned int image_id,
 				 image_data->image_size);
 	if (rc != 0) {
 		/* Authentication error, zero memory and flush it right away. */
-		memset((void *)image_data->image_base, 0x00,
+		zero_normalmem((void *)image_data->image_base,
 		       image_data->image_size);
 		flush_dcache_range(image_data->image_base,
 				   image_data->image_size);
@@ -543,7 +543,7 @@ static int load_auth_image_internal(meminfo_t *mem_layout,
 				 image_data->image_size);
 	if (rc != 0) {
 		/* Authentication error, zero memory and flush it right away. */
-		memset((void *)image_data->image_base, 0x00,
+		zero_normalmem((void *)image_data->image_base,
 		       image_data->image_size);
 		flush_dcache_range(image_data->image_base,
 				   image_data->image_size);

docs/firmware-design.md

@@ -1342,7 +1342,7 @@ All BL images share the following requirements:
 
 The following linker symbols are defined for this purpose:
 
-*   `__BSS_START__`          Must be aligned on a 16-byte boundary.
+*   `__BSS_START__`
 *   `__BSS_SIZE__`
 *   `__COHERENT_RAM_START__` Must be aligned on a page-size boundary.
 *   `__COHERENT_RAM_END__`   Must be aligned on a page-size boundary.

include/common/aarch64/el3_common_macros.S

 /*
- * Copyright (c) 2015-2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2015-2017, 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:
@@ -252,12 +252,12 @@
 
 		ldr	x0, =__BSS_START__
 		ldr	x1, =__BSS_SIZE__
-		bl	zeromem16
+		bl	zeromem
 
 #if USE_COHERENT_MEM
 		ldr	x0, =__COHERENT_RAM_START__
 		ldr	x1, =__COHERENT_RAM_UNALIGNED_SIZE__
-		bl	zeromem16
+		bl	zeromem
 #endif
 
 #ifdef IMAGE_BL1

include/lib/utils.h

@@ -80,4 +80,35 @@
 # define ULL(_x)	(_x##ull)
 #endif
 
+/*
+ * C code should be put in this part of the header to avoid breaking ASM files
+ * or linker scripts including it.
+ */
+#if !(defined(__LINKER__) || defined(__ASSEMBLY__))
+
+#include <types.h>
+
+/*
+ * Fill a region of normal memory of size "length" in bytes with zero bytes.
+ *
+ * WARNING: This function can only operate on normal memory. This means that
+ *          the MMU must be enabled when using this function. Otherwise, use
+ *          zeromem.
+ */
+void zero_normalmem(void *mem, u_register_t length);
+
+/*
+ * Fill a region of memory of size "length" in bytes with null bytes.
+ *
+ * Unlike zero_normalmem, this function has no restriction on the type of
+ * memory targeted and can be used for any device memory as well as normal
+ * memory. This function must be used instead of zero_normalmem when MMU is
+ * disabled.
+ *
+ * NOTE: When data cache and MMU are enabled, prefer zero_normalmem for faster
+ *       zeroing.
+ */
+void zeromem(void *mem, u_register_t length);
+#endif /* !(defined(__LINKER__) || defined(__ASSEMBLY__)) */
+
 #endif /* __UTILS_H__ */

lib/aarch32/misc_helpers.S

 /*
- * Copyright (c) 2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2016-2017, 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:
@@ -34,6 +34,7 @@
 
 	.globl	smc
 	.globl	zeromem
+	.globl	zero_normalmem
 	.globl	memcpy4
 	.globl	disable_mmu_icache_secure
 	.globl	disable_mmu_secure
@@ -50,30 +51,108 @@ func smc
 endfunc smc
 
 /* -----------------------------------------------------------------------
- * void zeromem(void *mem, unsigned int length);
+ * void zeromem(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.
  *
- * Initialise a memory region to 0.
- * The memory address and length must be 4-byte aligned.
  * -----------------------------------------------------------------------
  */
 func zeromem
-#if ASM_ASSERTION
-	tst	r0, #0x3
-	ASM_ASSERT(eq)
-	tst	r1, #0x3
-	ASM_ASSERT(eq)
-#endif
-	add	r2, r0, r1
-	mov	r1, #0
-z_loop:
-	cmp	r2, r0
-	beq	z_end
-	str	r1, [r0], #4
-	b	z_loop
-z_end:
+	/*
+	 * Readable names for registers
+	 *
+	 * Registers r0, r1 and r2 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 r0 /* Start address and then current address */
+	length       .req r1 /* Length in bytes of the region to zero out */
+	/*
+	 * Reusing the r1 register as length is only used at the beginning of
+	 * the function.
+	 */
+	stop_address .req r1  /* Address past the last zeroed byte */
+	zeroreg1     .req r2  /* Source register filled with 0 */
+	zeroreg2     .req r3  /* Source register filled with 0 */
+	tmp	     .req r12 /* Temporary scratch register */
+
+	mov	zeroreg1, #0
+
+	/* stop_address is the address past the last to zero */
+	add	stop_address, cursor, length
+
+	/*
+	 * Length cannot be used anymore as it shares the same register with
+	 * stop_address.
+	 */
+	.unreq	length
+
+	/*
+	 * If the start address is already aligned to 8 bytes, skip this loop.
+	 */
+	tst	cursor, #(8-1)
+	beq	.Lzeromem_8bytes_aligned
+
+	/* Calculate the next address aligned to 8 bytes */
+	orr	tmp, cursor, #(8-1)
+	adds	tmp, tmp, #1
+	/* If it overflows, fallback to byte per byte zeroing */
+	beq	.Lzeromem_1byte_aligned
+	/* If the next aligned address is after the stop address, fall back */
+	cmp	tmp, stop_address
+	bhs	.Lzeromem_1byte_aligned
+
+	/* zero byte per byte */
+1:
+	strb	zeroreg1, [cursor], #1
+	cmp	cursor, tmp
+	bne	1b
+
+	/* zero 8 bytes at a time */
+.Lzeromem_8bytes_aligned:
+
+	/* Calculate the last 8 bytes aligned address. */
+	bic	tmp, stop_address, #(8-1)
+
+	cmp	cursor, tmp
+	bhs	2f
+
+	mov	zeroreg2, #0
+1:
+	stmia	cursor!, {zeroreg1, zeroreg2}
+	cmp	cursor, tmp
+	blo	1b
+2:
+
+	/* zero byte per byte */
+.Lzeromem_1byte_aligned:
+	cmp	cursor, stop_address
+	beq	2f
+1:
+	strb	zeroreg1, [cursor], #1
+	cmp	cursor, stop_address
+	bne	1b
+2:
 	bx	lr
+
+	.unreq	cursor
+	/*
+	 * length is already unreq'ed to reuse the register for another
+	 * variable.
+	 */
+	.unreq	stop_address
+	.unreq	zeroreg1
+	.unreq	zeroreg2
+	.unreq	tmp
 endfunc zeromem
 
+/*
+ * AArch32 does not have special ways of zeroing normal memory as AArch64 does
+ * using the DC ZVA instruction, so we just alias zero_normalmem to zeromem.
+ */
+.equ	zero_normalmem, zeromem
+
 /* --------------------------------------------------------------------------
  * void memcpy4(void *dest, const void *src, unsigned int length)
  *

lib/aarch64/misc_helpers.S

 /*
- * 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)

plat/arm/css/common/css_bl2_setup.c

 /*
- * Copyright (c) 2015, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2015-2017, 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:
@@ -98,7 +98,7 @@ void bl2_platform_setup(void)
 	 *  - restoring the SCP boot configuration.
 	 */
 	VERBOSE("BL2: Restoring SCP reset data in Trusted SRAM\n");
-	memset((void *) ARM_TRUSTED_SRAM_BASE, 0, 128);
+	zero_normalmem((void *)ARM_TRUSTED_SRAM_BASE, 128);
 	mmio_write_32(SCP_BOOT_CFG_ADDR, scp_boot_config);
 }
 #endif /* EL3_PAYLOAD_BASE */

plat/mediatek/mt6795/bl31.ld.S

 /*
- * Copyright (c) 2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2016-2017, 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:
@@ -113,7 +113,8 @@ SECTIONS
 
     /*
      * The .bss section gets initialised to 0 at runtime.
-     * Its base address must be 16-byte aligned.
+     * Its base address should be 16-byte aligned for better performance of the
+     * zero-initialization code.
      */
     .bss (NOLOAD) : ALIGN(16) {
         __BSS_START__ = .;

plat/nvidia/tegra/common/drivers/memctrl/memctrl.c

 /*
- * Copyright (c) 2015, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2015-2017, 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:
@@ -37,8 +37,6 @@
 #include <tegra_def.h>
 #include <xlat_tables.h>
 
-extern void zeromem16(void *mem, unsigned int length);
-
 #define TEGRA_GPU_RESET_REG_OFFSET	0x28c
 #define  GPU_RESET_BIT			(1 << 24)
 
@@ -114,13 +112,13 @@ static void tegra_clear_videomem(uintptr_t non_overlap_area_start,
 	 * Perform cache maintenance to ensure that the non-overlapping area is
 	 * zeroed out. The first invalidation of this range ensures that
 	 * possible evictions of dirty cache lines do not interfere with the
-	 * 'zeromem16' operation. Other CPUs could speculatively prefetch the
+	 * 'zeromem' operation. Other CPUs could speculatively prefetch the
 	 * main memory contents of this area between the first invalidation and
-	 * the 'zeromem16' operation. The second invalidation ensures that any
+	 * the 'zeromem' operation. The second invalidation ensures that any
 	 * such cache lines are removed as well.
 	 */
 	inv_dcache_range(non_overlap_area_start, non_overlap_area_size);
-	zeromem16((void *)non_overlap_area_start, non_overlap_area_size);
+	zeromem((void *)non_overlap_area_start, non_overlap_area_size);
 	inv_dcache_range(non_overlap_area_start, non_overlap_area_size);
 }