flash/stm32h7x: add support of STM32H7Ax/H7Bx devices

this new device has the following features:
 - single core cortex-M7
 - 2MB flash - dual bank
    - page size 8k
    - write protection grouped by 4 sectors
    - write block size 128 bits (16 bytes)

the bit definition of FLASH_CR is different than STM32H74x,
that's why we introduced a helper to compute the FLASH_CR value

Change-Id: I4da10cde8dd215b1b0f2645f0efdba9d198038d1
Signed-off-by: Tarek BOCHKATI <tarek.bouchkati@gmail.com>
Reviewed-on: http://openocd.zylin.com/5441
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>

2 files changed, 47 insertions, 39 deletions

diff --git a/contrib/loaders/flash/stm32/stm32h7x.S b/contrib/loaders/flash/stm32/stm32h7x.S
index beb8fdbd..a4317229 100644
--- a/contrib/loaders/flash/stm32/stm32h7x.S
+++ b/contrib/loaders/flash/stm32/stm32h7x.S
@@ -25,29 +25,35 @@
  * Code limitations:
  * The workarea must have size multiple of 4 bytes, since R/W
  * operations are all at 32 bits.
- * The workarea must be big enough to contain 32 bytes of data,
- * thus the minimum size is (rp, wp, data) = 4 + 4 + 32 = 40 bytes.
+ * The workarea must be big enough to contain rp, wp and data, thus the minumum
+ * workarea size is: min_wa_size = sizeof(rp, wp, data) = 4 + 4 + sizeof(data).
+ *  - for 0x450 devices: sizeof(data) = 32 bytes, thus min_wa_size = 40 bytes.
+ *  - for 0x480 devices: sizeof(data) = 16 bytes, thus min_wa_size = 24 bytes.
  * To benefit from concurrent host write-to-buffer and target
  * write-to-flash, the workarea must be way bigger than the minimum.
- */
+ *
+ * To avoid confusions the write word size is got from .block_size member of
+ * struct stm32h7x_part_info defined in stm32h7x.c
+*/
 
 /*
  * Params :
  * r0 = workarea start, status (out)
  * r1 = workarea end
  * r2 = target address
- * r3 = count (256 bit words)
- * r4 = flash reg base
+ * r3 = count (of write words)
+ * r4 = size of write word
+ * r5 = flash reg base
  *
  * Clobbered:
- * r5 - rp
- * r6 - wp, status, tmp
- * r7 - loop index, tmp
+ * r6 - rp
+ * r7 - wp, status, tmp
+ * r8 - loop index, tmp
  */
 
 #define STM32_FLASH_CR_OFFSET	0x0C	/* offset of CR register in FLASH struct */
 #define STM32_FLASH_SR_OFFSET	0x10	/* offset of SR register in FLASH struct */
-#define STM32_CR_PROG			0x00000032	/* PSIZE64 | PG */
+#define STM32_CR_PROG			0x00000002	/* PG */
 #define STM32_SR_QW_MASK		0x00000004	/* QW */
 #define STM32_SR_ERROR_MASK		0x07ee0000	/* DBECCERR | SNECCERR | RDSERR | RDPERR | OPERR
 											   | INCERR | STRBERR | PGSERR | WRPERR */
@@ -55,54 +61,55 @@
 	.thumb_func
 	.global _start
 _start:
-	ldr		r5, [r0, #4]		/* read rp */
+	ldr		r6, [r0, #4]		/* read rp */
 
 wait_fifo:
-	ldr		r6, [r0, #0]		/* read wp */
-	cbz		r6, exit			/* abort if wp == 0, status = 0 */
-	subs	r6, r6, r5			/* number of bytes available for read in r6 */
+	ldr		r7, [r0, #0]		/* read wp */
+	cbz		r7, exit			/* abort if wp == 0, status = 0 */
+	subs	r7, r7, r6			/* number of bytes available for read in r7 */
 	ittt	mi					/* if wrapped around */
-	addmi	r6, r1				/* add size of buffer */
-	submi	r6, r0
-	submi	r6, #8
-	cmp		r6, #32				/* wait until 32 bytes are available */
+	addmi	r7, r1				/* add size of buffer */
+	submi	r7, r0
+	submi	r7, #8
+	cmp		r7, r4				/* wait until data buffer is full */
 	bcc		wait_fifo
 
-	mov		r6, #STM32_CR_PROG
-	str		r6, [r4, #STM32_FLASH_CR_OFFSET]
+	mov		r7, #STM32_CR_PROG
+	str		r7, [r5, #STM32_FLASH_CR_OFFSET]
 
-	mov		r7, #8				/* program by 8 words = 32 bytes */
+	mov		r8, #4
+	udiv	r8, r4, r8			/* number of words is size of write word devided by 4*/
 write_flash:
 	dsb
-	ldr		r6, [r5], #0x04		/* read one word from src, increment ptr */
-	str		r6, [r2], #0x04		/* write one word to dst, increment ptr */
+	ldr		r7, [r6], #0x04		/* read one word from src, increment ptr */
+	str		r7, [r2], #0x04		/* write one word to dst, increment ptr */
 	dsb
-	cmp		r5, r1				/* if rp >= end of buffer ... */
+	cmp		r6, r1				/* if rp >= end of buffer ... */
 	it		cs
-	addcs	r5, r0, #8			/* ... then wrap at buffer start */
-	subs	r7, r7, #1			/* decrement loop index */
+	addcs	r6, r0, #8			/* ... then wrap at buffer start */
+	subs	r8, r8, #1			/* decrement loop index */
 	bne		write_flash			/* loop if not done */
 
 busy:
-	ldr		r6, [r4, #STM32_FLASH_SR_OFFSET]
-	tst		r6, #STM32_SR_QW_MASK
+	ldr		r7, [r5, #STM32_FLASH_SR_OFFSET]
+	tst		r7, #STM32_SR_QW_MASK
 	bne		busy				/* operation in progress, wait ... */
 
-	ldr		r7, =STM32_SR_ERROR_MASK
-	tst		r6, r7
+	ldr		r8, =STM32_SR_ERROR_MASK
+	tst		r7, r8
 	bne		error				/* fail... */
 
-	str		r5, [r0, #4]		/* store rp */
+	str		r6, [r0, #4]		/* store rp */
 	subs	r3, r3, #1			/* decrement count */
 	bne		wait_fifo			/* loop if not done */
 	b		exit
 
 error:
-	movs	r7, #0
-	str		r7, [r0, #4]		/* set rp = 0 on error */
+	movs	r8, #0
+	str		r8, [r0, #4]		/* set rp = 0 on error */
 
 exit:
-	mov		r0, r6				/* return status in r0 */
+	mov		r0, r7				/* return status in r0 */
 	bkpt	#0x00
 
 	.pool
diff --git a/contrib/loaders/flash/stm32/stm32h7x.inc b/contrib/loaders/flash/stm32/stm32h7x.inc
index ec14de0e..015644ff 100644
--- a/contrib/loaders/flash/stm32/stm32h7x.inc
+++ b/contrib/loaders/flash/stm32/stm32h7x.inc
@@ -1,7 +1,8 @@
 /* Autogenerated with ../../../../src/helper/bin2char.sh */
-0x45,0x68,0x06,0x68,0x36,0xb3,0x76,0x1b,0x42,0xbf,0x76,0x18,0x36,0x1a,0x08,0x3e,
-0x20,0x2e,0xf6,0xd3,0x4f,0xf0,0x32,0x06,0xe6,0x60,0x4f,0xf0,0x08,0x07,0xbf,0xf3,
-0x4f,0x8f,0x55,0xf8,0x04,0x6b,0x42,0xf8,0x04,0x6b,0xbf,0xf3,0x4f,0x8f,0x8d,0x42,
-0x28,0xbf,0x00,0xf1,0x08,0x05,0x01,0x3f,0xf1,0xd1,0x26,0x69,0x16,0xf0,0x04,0x0f,
-0xfb,0xd1,0x05,0x4f,0x3e,0x42,0x03,0xd1,0x45,0x60,0x01,0x3b,0xd9,0xd1,0x01,0xe0,
-0x00,0x27,0x47,0x60,0x30,0x46,0x00,0xbe,0x00,0x00,0xee,0x07,
+0x46,0x68,0x07,0x68,0x6f,0xb3,0xbf,0x1b,0x42,0xbf,0x7f,0x18,0x3f,0x1a,0x08,0x3f,
+0xa7,0x42,0xf6,0xd3,0x4f,0xf0,0x02,0x07,0xef,0x60,0x4f,0xf0,0x04,0x08,0xb4,0xfb,
+0xf8,0xf8,0xbf,0xf3,0x4f,0x8f,0x56,0xf8,0x04,0x7b,0x42,0xf8,0x04,0x7b,0xbf,0xf3,
+0x4f,0x8f,0x8e,0x42,0x28,0xbf,0x00,0xf1,0x08,0x06,0xb8,0xf1,0x01,0x08,0xf0,0xd1,
+0x2f,0x69,0x17,0xf0,0x04,0x0f,0xfb,0xd1,0xdf,0xf8,0x1c,0x80,0x17,0xea,0x08,0x0f,
+0x03,0xd1,0x46,0x60,0x01,0x3b,0xd4,0xd1,0x03,0xe0,0x5f,0xf0,0x00,0x08,0xc0,0xf8,
+0x04,0x80,0x38,0x46,0x00,0xbe,0x00,0x00,0x00,0x00,0xee,0x07,