X-Git-Url: https://git.gag.com/?p=fw%2Fstlink;a=blobdiff_plain;f=src%2Fstlink-common.c;h=03e1cce5228dd3f646c9c29a59f43128ca047555;hp=51f0f68b02472aad573ce4325696212b39330a23;hb=b95e4aa8a47c4caef0e525205945bdc61c5ee2c8;hpb=13c0da3bccf358150e11eab21d28d618db5ab25f

diff --git a/src/stlink-common.c b/src/stlink-common.c
index 51f0f68..03e1cce 100644
--- a/src/stlink-common.c
+++ b/src/stlink-common.c
@@ -1,4 +1,4 @@
-
+#define DEBUG_FLASH 0
 
 #include <stdarg.h>
 #include <stdio.h>
@@ -13,26 +13,18 @@
 
 
 #include "stlink-common.h"
+#include "uglylogging.h"
 
-void D(stlink_t *sl, char *txt) {
-    if (sl->verbose > 1)
-        fputs(txt, stderr);
-}
-
-void DD(stlink_t *sl, char *format, ...) {
-    if (sl->verbose > 0) {
-        va_list list;
-        va_start(list, format);
-        vfprintf(stderr, format, list);
-        va_end(list);
-    }
-}
-
+#define LOG_TAG __FILE__
+#define DLOG(format, args...)         ugly_log(UDEBUG, LOG_TAG, format, ## args)
+#define ILOG(format, args...)         ugly_log(UINFO, LOG_TAG, format, ## args)
+#define WLOG(format, args...)         ugly_log(UWARN, LOG_TAG, format, ## args)
+#define fatal(format, args...)        ugly_log(UFATAL, LOG_TAG, format, ## args)
 
 /* todo: stm32l15xxx flash memory, pm0062 manual */
 
 /* stm32f FPEC flash controller interface, pm0063 manual */
-
+// TODO - all of this needs to be abstracted out....
 #define FLASH_REGS_ADDR 0x40022000
 #define FLASH_REGS_SIZE 0x28
 
@@ -67,8 +59,10 @@ void DD(stlink_t *sl, char *format, ...) {
 #define STM32L_FLASH_PRGKEYR (STM32L_FLASH_REGS_ADDR + 0x10)
 #define STM32L_FLASH_OPTKEYR (STM32L_FLASH_REGS_ADDR + 0x14)
 #define STM32L_FLASH_SR (STM32L_FLASH_REGS_ADDR + 0x18)
-#define STM32L_FLASH_OBR (STM32L_FLASH_REGS_ADDR + 0x0c)
+#define STM32L_FLASH_OBR (STM32L_FLASH_REGS_ADDR + 0x1c)
 #define STM32L_FLASH_WRPR (STM32L_FLASH_REGS_ADDR + 0x20)
+#define FLASH_L1_FPRG 10
+#define FLASH_L1_PROG 3
 
 
 //STM32F4
@@ -129,27 +123,27 @@ uint32_t read_uint32(const unsigned char *c, const int pt) {
 }
 
 static uint32_t __attribute__((unused)) read_flash_rdp(stlink_t *sl) {
-    stlink_read_mem32(sl, FLASH_WRPR, sizeof (uint32_t));
-    return (*(uint32_t*) sl->q_buf) & 0xff;
+    return stlink_read_debug32(sl, FLASH_WRPR) & 0xff;
 }
 
 static inline uint32_t read_flash_wrpr(stlink_t *sl) {
-    stlink_read_mem32(sl, FLASH_WRPR, sizeof (uint32_t));
-    return *(uint32_t*) sl->q_buf;
+    return stlink_read_debug32(sl, FLASH_WRPR);
 }
 
 static inline uint32_t read_flash_obr(stlink_t *sl) {
-    stlink_read_mem32(sl, FLASH_OBR, sizeof (uint32_t));
-    return *(uint32_t*) sl->q_buf;
+    return stlink_read_debug32(sl, FLASH_OBR);
 }
 
 static inline uint32_t read_flash_cr(stlink_t *sl) {
+        uint32_t res;
 	if(sl->chip_id==STM32F4_CHIP_ID)
-		stlink_read_mem32(sl, FLASH_F4_CR, sizeof (uint32_t));
+		res = stlink_read_debug32(sl, FLASH_F4_CR);
 	else
-		stlink_read_mem32(sl, FLASH_CR, sizeof (uint32_t));
-	fprintf(stdout, "CR:%X\n", *(uint32_t*) sl->q_buf);
-    return *(uint32_t*) sl->q_buf;
+		res = stlink_read_debug32(sl, FLASH_CR);
+#if DEBUG_FLASH
+	fprintf(stdout, "CR:0x%x\n", res);
+#endif
+	return res;
 }
 
 static inline unsigned int is_flash_locked(stlink_t *sl) {
@@ -167,16 +161,12 @@ static void unlock_flash(stlink_t *sl) {
        the FPEC block until next reset.
      */
     if(sl->chip_id==STM32F4_CHIP_ID) {
-        write_uint32(sl->q_buf, FLASH_KEY1);
-    	stlink_write_mem32(sl, FLASH_F4_KEYR, sizeof (uint32_t));
-		write_uint32(sl->q_buf, FLASH_KEY2);
-		stlink_write_mem32(sl, FLASH_F4_KEYR, sizeof (uint32_t));
+   	stlink_write_debug32(sl, FLASH_F4_KEYR, FLASH_KEY1);
+		stlink_write_debug32(sl, FLASH_F4_KEYR, FLASH_KEY2);
     }
 	else {
-        write_uint32(sl->q_buf, FLASH_KEY1);
-    	stlink_write_mem32(sl, FLASH_KEYR, sizeof (uint32_t));
-		write_uint32(sl->q_buf, FLASH_KEY2);
-		stlink_write_mem32(sl, FLASH_KEYR, sizeof (uint32_t));
+    	stlink_write_debug32(sl, FLASH_KEYR, FLASH_KEY1);
+		stlink_write_debug32(sl, FLASH_KEYR, FLASH_KEY2);
 	}
 
 }
@@ -186,24 +176,24 @@ static int unlock_flash_if(stlink_t *sl) {
 
     if (is_flash_locked(sl)) {
         unlock_flash(sl);
-        if (is_flash_locked(sl))
+        if (is_flash_locked(sl)) {
+            WLOG("Failed to unlock flash!\n");
             return -1;
+        }
     }
-
+    DLOG("Successfully unlocked flash\n");
     return 0;
 }
 
 static void lock_flash(stlink_t *sl) {
     if(sl->chip_id==STM32F4_CHIP_ID) {
     	const uint32_t n = read_flash_cr(sl) | (1 << FLASH_F4_CR_LOCK);
-        write_uint32(sl->q_buf, n);
-    	stlink_write_mem32(sl, FLASH_F4_CR, sizeof (uint32_t));
+    	stlink_write_debug32(sl, FLASH_F4_CR, n);
     }
     else {
         /* write to 1 only. reset by hw at unlock sequence */
         const uint32_t n = read_flash_cr(sl) | (1 << FLASH_CR_LOCK);
-        write_uint32(sl->q_buf, n);
-        stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+        stlink_write_debug32(sl, FLASH_CR, n);
     }
 }
 
@@ -212,47 +202,48 @@ static void set_flash_cr_pg(stlink_t *sl) {
     if(sl->chip_id==STM32F4_CHIP_ID) {
 		uint32_t x = read_flash_cr(sl);
 		x |= (1 << FLASH_CR_PG);
-		write_uint32(sl->q_buf, x);
-    	stlink_write_mem32(sl, FLASH_F4_CR, sizeof (uint32_t));
+    	stlink_write_debug32(sl, FLASH_F4_CR, x);
     }
     else {
         const uint32_t n = 1 << FLASH_CR_PG;
-        write_uint32(sl->q_buf, n);
-        stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+        stlink_write_debug32(sl, FLASH_CR, n);
     }
 }
 
 static void __attribute__((unused)) clear_flash_cr_pg(stlink_t *sl) {
     const uint32_t n = read_flash_cr(sl) & ~(1 << FLASH_CR_PG);
-    write_uint32(sl->q_buf, n);
     if(sl->chip_id==STM32F4_CHIP_ID)
-    	stlink_write_mem32(sl, FLASH_F4_CR, sizeof (uint32_t));
+    	stlink_write_debug32(sl, FLASH_F4_CR, n);
     else
-        stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+        stlink_write_debug32(sl, FLASH_CR, n);
 }
 
 static void set_flash_cr_per(stlink_t *sl) {
     const uint32_t n = 1 << FLASH_CR_PER;
-    write_uint32(sl->q_buf, n);
-    stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+    stlink_write_debug32(sl, FLASH_CR, n);
 }
 
 static void __attribute__((unused)) clear_flash_cr_per(stlink_t *sl) {
     const uint32_t n = read_flash_cr(sl) & ~(1 << FLASH_CR_PER);
-    write_uint32(sl->q_buf, n);
-    stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+    stlink_write_debug32(sl, FLASH_CR, n);
 }
 
 static void set_flash_cr_mer(stlink_t *sl) {
-    const uint32_t n = 1 << FLASH_CR_MER;
-    write_uint32(sl->q_buf, n);
-    stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+    if(sl->chip_id == STM32F4_CHIP_ID)
+        stlink_write_debug32(sl, FLASH_F4_CR,
+                             stlink_read_debug32(sl, FLASH_F4_CR) | (1 << FLASH_CR_MER));
+    else 
+        stlink_write_debug32(sl, FLASH_CR,
+                             stlink_read_debug32(sl, FLASH_CR) | (1 << FLASH_CR_MER));
 }
 
 static void __attribute__((unused)) clear_flash_cr_mer(stlink_t *sl) {
-    const uint32_t n = read_flash_cr(sl) & ~(1 << FLASH_CR_MER);
-    write_uint32(sl->q_buf, n);
-    stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+    if(sl->chip_id == STM32F4_CHIP_ID)
+        stlink_write_debug32(sl, FLASH_F4_CR,
+                             stlink_read_debug32(sl, FLASH_F4_CR) & ~(1 << FLASH_CR_MER));
+    else 
+        stlink_write_debug32(sl, FLASH_CR,
+                             stlink_read_debug32(sl, FLASH_CR) & ~(1 << FLASH_CR_MER));
 }
 
 static void set_flash_cr_strt(stlink_t *sl) {
@@ -260,29 +251,27 @@ static void set_flash_cr_strt(stlink_t *sl) {
 	{
 		uint32_t x = read_flash_cr(sl);
 		x |= (1 << FLASH_F4_CR_STRT);
-		write_uint32(sl->q_buf, x);
-		stlink_write_mem32(sl, FLASH_F4_CR, sizeof (uint32_t));
+		stlink_write_debug32(sl, FLASH_F4_CR, x);
 	}
 	else {
-		/* assume come on the flash_cr_per path */
-	    const uint32_t n = (1 << FLASH_CR_PER) | (1 << FLASH_CR_STRT);
-	    write_uint32(sl->q_buf, n);
-	    stlink_write_mem32(sl, FLASH_CR, sizeof (uint32_t));
+	    stlink_write_debug32(
+                sl, FLASH_CR, 
+                stlink_read_debug32(sl,FLASH_CR) |(1 << FLASH_CR_STRT) );
 	}
 }
 
 static inline uint32_t read_flash_acr(stlink_t *sl) {
-    stlink_read_mem32(sl, FLASH_ACR, sizeof (uint32_t));
-    return *(uint32_t*) sl->q_buf;
+    return stlink_read_debug32(sl, FLASH_ACR);
 }
 
 static inline uint32_t read_flash_sr(stlink_t *sl) {
+	uint32_t res;
 	if(sl->chip_id==STM32F4_CHIP_ID)
-		stlink_read_mem32(sl, FLASH_F4_SR, sizeof (uint32_t));
+		res = stlink_read_debug32(sl, FLASH_F4_SR);
 	else
-		stlink_read_mem32(sl, FLASH_SR, sizeof (uint32_t));
-    //fprintf(stdout, "SR:%X\n", *(uint32_t*) sl->q_buf);
-	return *(uint32_t*) sl->q_buf;
+		res = stlink_read_debug32(sl, FLASH_SR);
+    //fprintf(stdout, "SR:0x%x\n", *(uint32_t*) sl->q_buf);
+    return res;
 }
 
 static inline unsigned int is_flash_busy(stlink_t *sl) {
@@ -298,14 +287,29 @@ static void wait_flash_busy(stlink_t *sl) {
         ;
 }
 
+static void wait_flash_busy_progress(stlink_t *sl) {
+    int i = 0;
+    fprintf(stdout, "Mass erasing");
+    fflush(stdout);
+    while (is_flash_busy(sl))
+    {
+        usleep(10000);
+        i++;
+        if (i % 100 == 0) {
+            fprintf(stdout, ".");
+            fflush(stdout);
+        }
+    }
+    fprintf(stdout, "\n");
+}
+
 static inline unsigned int is_flash_eop(stlink_t *sl) {
     return read_flash_sr(sl) & (1 << FLASH_SR_EOP);
 }
 
 static void __attribute__((unused)) clear_flash_sr_eop(stlink_t *sl) {
     const uint32_t n = read_flash_sr(sl) & ~(1 << FLASH_SR_EOP);
-    write_uint32(sl->q_buf, n);
-    stlink_write_mem32(sl, FLASH_SR, sizeof (uint32_t));
+    stlink_write_debug32(sl, FLASH_SR, n);
 }
 
 static void __attribute__((unused)) wait_flash_eop(stlink_t *sl) {
@@ -315,17 +319,17 @@ static void __attribute__((unused)) wait_flash_eop(stlink_t *sl) {
 }
 
 static inline void write_flash_ar(stlink_t *sl, uint32_t n) {
-    write_uint32(sl->q_buf, n);
-    stlink_write_mem32(sl, FLASH_AR, sizeof (uint32_t));
+    stlink_write_debug32(sl, FLASH_AR, n);
 }
 
 static inline void write_flash_cr_psiz(stlink_t *sl, uint32_t n) {
     uint32_t x = read_flash_cr(sl);
     x &= ~(0x03 << 8);
     x |= (n << 8);
-    fprintf(stdout, "PSIZ:%X %X\n", x, n);
-    write_uint32(sl->q_buf, x);
-    stlink_write_mem32(sl, FLASH_F4_CR, sizeof (uint32_t));
+#if DEBUG_FLASH
+    fprintf(stdout, "PSIZ:0x%x 0x%x\n", x, n);
+#endif
+    stlink_write_debug32(sl, FLASH_F4_CR, x);
 }
 
 
@@ -334,9 +338,10 @@ static inline void write_flash_cr_snb(stlink_t *sl, uint32_t n) {
     x &= ~FLASH_F4_CR_SNB_MASK;
     x |= (n << FLASH_F4_CR_SNB);
     x |= (1 << FLASH_F4_CR_SER);
-    fprintf(stdout, "SNB:%X %X\n", x, n);
-    write_uint32(sl->q_buf, x);
-    stlink_write_mem32(sl, FLASH_F4_CR, sizeof (uint32_t));
+#if DEBUG_FLASH
+    fprintf(stdout, "SNB:0x%x 0x%x\n", x, n);
+#endif
+    stlink_write_debug32(sl, FLASH_F4_CR, x);
 }
 
 #if 0 /* todo */
@@ -352,52 +357,45 @@ static void disable_flash_read_protection(stlink_t *sl) {
 // Delegates to the backends...
 
 void stlink_close(stlink_t *sl) {
-    D(sl, "\n*** stlink_close ***\n");
+    DLOG("*** stlink_close ***\n");
     sl->backend->close(sl);
     free(sl);
 }
 
 void stlink_exit_debug_mode(stlink_t *sl) {
-    D(sl, "\n*** stlink_exit_debug_mode ***\n");
+    DLOG("*** stlink_exit_debug_mode ***\n");
+    stlink_write_debug32(sl, DHCSR, DBGKEY);
     sl->backend->exit_debug_mode(sl);
 }
 
 void stlink_enter_swd_mode(stlink_t *sl) {
-    D(sl, "\n*** stlink_enter_swd_mode ***\n");
+    DLOG("*** stlink_enter_swd_mode ***\n");
     sl->backend->enter_swd_mode(sl);
 }
 
 // Force the core into the debug mode -> halted state.
 void stlink_force_debug(stlink_t *sl) {
-    D(sl, "\n*** stlink_force_debug_mode ***\n");
+    DLOG("*** stlink_force_debug_mode ***\n");
     sl->backend->force_debug(sl);
 }
 
 void stlink_exit_dfu_mode(stlink_t *sl) {
-    D(sl, "\n*** stlink_exit_dfu_mode ***\n");
+    DLOG("*** stlink_exit_dfu_mode ***\n");
     sl->backend->exit_dfu_mode(sl);
 }
 
 uint32_t stlink_core_id(stlink_t *sl) {
-    D(sl, "\n*** stlink_core_id ***\n");
+    DLOG("*** stlink_core_id ***\n");
     sl->backend->core_id(sl);
     if (sl->verbose > 2)
         stlink_print_data(sl);
+    DLOG("core_id = 0x%08x\n", sl->core_id);
     return sl->core_id;
 }
 
-void stlink_identify_device(stlink_t *sl) {
-	uint32_t core_id=stlink_core_id(sl);
-	stlink_read_mem32(sl, 0xE0042000, 4);
-    uint32_t chip_id = sl->q_buf[0] | (sl->q_buf[1] << 8) | (sl->q_buf[2] << 16) |
-            (sl->q_buf[3] << 24);
-    /* Fix chip_id for F4 */
-    if (((chip_id & 0xFFF) == 0x411) && (core_id == CORE_M4_R0)) {
-      printf("Fixing wrong chip_id for STM32F4 Rev A errata\n");
-      chip_id = 0x413;
-    }
-    sl->chip_id=chip_id;
-    sl->core_id=core_id;
+uint32_t stlink_chip_id(stlink_t *sl) {
+    uint32_t chip_id = stlink_read_debug32(sl, 0xE0042000);
+    return chip_id;
 }
 
 /**
@@ -406,8 +404,7 @@ void stlink_identify_device(stlink_t *sl) {
  * @param cpuid pointer to the result object
  */
 void stlink_cpu_id(stlink_t *sl, cortex_m3_cpuid_t *cpuid) {
-    stlink_read_mem32(sl, CM3_REG_CPUID, 4);
-    uint32_t raw = read_uint32(sl->q_buf, 0);
+    uint32_t raw = stlink_read_debug32(sl, CM3_REG_CPUID);
     cpuid->implementer_id = (raw >> 24) & 0x7f;
     cpuid->variant = (raw >> 20) & 0xf;
     cpuid->part = (raw >> 4) & 0xfff;
@@ -415,18 +412,79 @@ void stlink_cpu_id(stlink_t *sl, cortex_m3_cpuid_t *cpuid) {
     return;
 }
 
+/**
+ * reads and decodes the flash parameters, as dynamically as possible
+ * @param sl
+ * @return 0 for success, or -1 for unsupported core type.
+ */
+int stlink_load_device_params(stlink_t *sl) {
+    ILOG("Loading device parameters....\n");
+    const chip_params_t *params = NULL;
+    sl->core_id = stlink_core_id(sl);
+    uint32_t chip_id = stlink_chip_id(sl);
+    
+    sl->chip_id = chip_id & 0xfff;
+    /* Fix chip_id for F4 rev A errata , Read CPU ID, as CoreID is the same for F2/F4*/
+    if (sl->chip_id == 0x411) {
+        uint32_t cpuid = stlink_read_debug32(sl, 0xE000ED00);
+        if((cpuid  & 0xfff0) == 0xc240)
+            sl->chip_id = 0x413;
+    }
+
+    for(size_t i = 0; i < sizeof(devices) / sizeof(devices[0]); i++) {
+        if(devices[i].chip_id == sl->chip_id) {
+            params = &devices[i];
+            break;
+        }
+    }
+    if (params == NULL) {
+        WLOG("unknown chip id! %#x\n", chip_id);
+        return -1;
+    }
+    
+    // These are fixed...
+    sl->flash_base = STM32_FLASH_BASE;
+    sl->sram_base = STM32_SRAM_BASE;
+    
+    // read flash size from hardware, if possible...
+    if (sl->chip_id == STM32_CHIPID_F2) {
+        sl->flash_size = 0; // FIXME - need to work this out some other way, just set to max possible?
+    } else if (sl->chip_id == STM32_CHIPID_F4) {
+		sl->flash_size = 0x100000;			//todo: RM0090 error; size register same address as unique ID
+    } else {
+        uint32_t flash_size = stlink_read_debug32(sl, params->flash_size_reg) & 0xffff;
+        sl->flash_size = flash_size * 1024;
+    }
+    sl->flash_pgsz = params->flash_pagesize;
+    sl->sram_size = params->sram_size;
+    sl->sys_base = params->bootrom_base;
+    sl->sys_size = params->bootrom_size;
+    
+    ILOG("Device connected is: %s, id %#x\n", params->description, chip_id);
+    // TODO make note of variable page size here.....
+    ILOG("SRAM size: %#x bytes (%d KiB), Flash: %#x bytes (%d KiB) in pages of %zd bytes\n",
+        sl->sram_size, sl->sram_size / 1024, sl->flash_size, sl->flash_size / 1024, 
+        sl->flash_pgsz);
+    return 0;
+}
+
 void stlink_reset(stlink_t *sl) {
-    D(sl, "\n*** stlink_reset ***\n");
+    DLOG("*** stlink_reset ***\n");
     sl->backend->reset(sl);
 }
 
+void stlink_jtag_reset(stlink_t *sl, int value) {
+    DLOG("*** stlink_jtag_reset ***\n");
+    sl->backend->jtag_reset(sl, value);
+}
+
 void stlink_run(stlink_t *sl) {
-    D(sl, "\n*** stlink_run ***\n");
+    DLOG("*** stlink_run ***\n");
     sl->backend->run(sl);
 }
 
 void stlink_status(stlink_t *sl) {
-    D(sl, "\n*** stlink_status ***\n");
+    DLOG("*** stlink_status ***\n");
     sl->backend->status(sl);
     stlink_core_stat(sl);
 }
@@ -457,26 +515,36 @@ void _parse_version(stlink_t *sl, stlink_version_t *slv) {
 }
 
 void stlink_version(stlink_t *sl) {
-    D(sl, "*** looking up stlink version\n");
-    stlink_version_t slv;
+    DLOG("*** looking up stlink version\n");
     sl->backend->version(sl);
-    _parse_version(sl, &slv);
+    _parse_version(sl, &sl->version);
     
-    DD(sl, "st vid         = 0x%04x (expect 0x%04x)\n", slv.st_vid, USB_ST_VID);
-    DD(sl, "stlink pid     = 0x%04x\n", slv.stlink_pid);
-    DD(sl, "stlink version = 0x%x\n", slv.stlink_v);
-    DD(sl, "jtag version   = 0x%x\n", slv.jtag_v);
-    DD(sl, "swim version   = 0x%x\n", slv.swim_v);
-    if (slv.jtag_v == 0) {
-        DD(sl, "    notice: the firmware doesn't support a jtag/swd interface\n");
+    DLOG("st vid         = 0x%04x (expect 0x%04x)\n", sl->version.st_vid, USB_ST_VID);
+    DLOG("stlink pid     = 0x%04x\n", sl->version.stlink_pid);
+    DLOG("stlink version = 0x%x\n", sl->version.stlink_v);
+    DLOG("jtag version   = 0x%x\n", sl->version.jtag_v);
+    DLOG("swim version   = 0x%x\n", sl->version.swim_v);
+    if (sl->version.jtag_v == 0) {
+        DLOG("    notice: the firmware doesn't support a jtag/swd interface\n");
     }
-    if (slv.swim_v == 0) {
-        DD(sl, "    notice: the firmware doesn't support a swim interface\n");
+    if (sl->version.swim_v == 0) {
+        DLOG("    notice: the firmware doesn't support a swim interface\n");
     }
 }
 
+uint32_t stlink_read_debug32(stlink_t *sl, uint32_t addr) {
+    uint32_t data = sl->backend->read_debug32(sl, addr);
+    DLOG("*** stlink_read_debug32 %x is %#x\n", data, addr);
+    return data;
+}
+
+void stlink_write_debug32(stlink_t *sl, uint32_t addr, uint32_t data) {
+    DLOG("*** stlink_write_debug32 %x to %#x\n", data, addr);
+    sl->backend->write_debug32(sl, addr, data);
+}
+
 void stlink_write_mem32(stlink_t *sl, uint32_t addr, uint16_t len) {
-    D(sl, "\n*** stlink_write_mem32 ***\n");
+    DLOG("*** stlink_write_mem32 %u bytes to %#x\n", len, addr);
     if (len % 4 != 0) {
         fprintf(stderr, "Error: Data length doesn't have a 32 bit alignment: +%d byte.\n", len % 4);
         return;
@@ -485,7 +553,7 @@ void stlink_write_mem32(stlink_t *sl, uint32_t addr, uint16_t len) {
 }
 
 void stlink_read_mem32(stlink_t *sl, uint32_t addr, uint16_t len) {
-    D(sl, "\n*** stlink_read_mem32 ***\n");
+    DLOG("*** stlink_read_mem32 ***\n");
     if (len % 4 != 0) { // !!! never ever: fw gives just wrong values
         fprintf(stderr, "Error: Data length doesn't have a 32 bit alignment: +%d byte.\n",
                 len % 4);
@@ -495,23 +563,28 @@ void stlink_read_mem32(stlink_t *sl, uint32_t addr, uint16_t len) {
 }
 
 void stlink_write_mem8(stlink_t *sl, uint32_t addr, uint16_t len) {
-    D(sl, "\n*** stlink_write_mem8 ***\n");
+    DLOG("*** stlink_write_mem8 ***\n");
+    if (len > 0x40 ) { // !!! never ever: Writing more then 0x40 bytes gives unexpected behaviour
+        fprintf(stderr, "Error: Data length > 64: +%d byte.\n",
+                len);
+        return;
+    }
     sl->backend->write_mem8(sl, addr, len);
 }
 
 void stlink_read_all_regs(stlink_t *sl, reg *regp) {
-    D(sl, "\n*** stlink_read_all_regs ***\n");
+    DLOG("*** stlink_read_all_regs ***\n");
     sl->backend->read_all_regs(sl, regp);
 }
 
 void stlink_write_reg(stlink_t *sl, uint32_t reg, int idx) {
-    D(sl, "\n*** stlink_write_reg\n");
+    DLOG("*** stlink_write_reg\n");
     sl->backend->write_reg(sl, reg, idx);
 }
 
 void stlink_read_reg(stlink_t *sl, int r_idx, reg *regp) {
-    D(sl, "\n*** stlink_read_reg\n");
-    DD(sl, " (%d) ***\n", r_idx);
+    DLOG("*** stlink_read_reg\n");
+    DLOG(" (%d) ***\n", r_idx);
 
     if (r_idx > 20 || r_idx < 0) {
         fprintf(stderr, "Error: register index must be in [0..20]\n");
@@ -528,7 +601,7 @@ unsigned int is_core_halted(stlink_t *sl) {
 }
 
 void stlink_step(stlink_t *sl) {
-    D(sl, "\n*** stlink_step ***\n");
+    DLOG("*** stlink_step ***\n");
     sl->backend->step(sl);
 }
 
@@ -536,16 +609,16 @@ int stlink_current_mode(stlink_t *sl) {
     int mode = sl->backend->current_mode(sl);
     switch (mode) {
         case STLINK_DEV_DFU_MODE:
-            DD(sl, "stlink current mode: dfu\n");
+            DLOG("stlink current mode: dfu\n");
             return mode;
         case STLINK_DEV_DEBUG_MODE:
-            DD(sl, "stlink current mode: debug (jtag or swd)\n");
+            DLOG("stlink current mode: debug (jtag or swd)\n");
             return mode;
         case STLINK_DEV_MASS_MODE:
-            DD(sl, "stlink current mode: mass\n");
+            DLOG("stlink current mode: mass\n");
             return mode;
     }
-    DD(sl, "stlink mode: unknown!\n");
+    DLOG("stlink mode: unknown!\n");
     return STLINK_DEV_UNKNOWN_MODE;
 }
 
@@ -596,11 +669,11 @@ void stlink_core_stat(stlink_t *sl) {
     switch (sl->q_buf[0]) {
         case STLINK_CORE_RUNNING:
             sl->core_stat = STLINK_CORE_RUNNING;
-            DD(sl, "  core status: running\n");
+            DLOG("  core status: running\n");
             return;
         case STLINK_CORE_HALTED:
             sl->core_stat = STLINK_CORE_HALTED;
-            DD(sl, "  core status: halted\n");
+            DLOG("  core status: halted\n");
             return;
         default:
             sl->core_stat = STLINK_CORE_STAT_UNKNOWN;
@@ -609,7 +682,7 @@ void stlink_core_stat(stlink_t *sl) {
 }
 
 void stlink_print_data(stlink_t * sl) {
-    if (sl->q_len <= 0 || sl->verbose < 2)
+    if (sl->q_len <= 0 || sl->verbose < UDEBUG)
         return;
     if (sl->verbose > 2)
         fprintf(stdout, "data_len = %d 0x%x\n", sl->q_len, sl->q_len);
@@ -675,15 +748,21 @@ static void unmap_file(mapped_file_t * mf) {
     mf->len = 0;
 }
 
+/* Limit the block size to compare to 0x1800
+   Anything larger will stall the STLINK2
+   Maybe STLINK V1 needs smaller value!*/
 static int check_file(stlink_t* sl, mapped_file_t* mf, stm32_addr_t addr) {
     size_t off;
+    size_t n_cmp = sl->flash_pgsz;
+    if ( n_cmp > 0x1800)
+        n_cmp = 0x1800;
 
-    for (off = 0; off < mf->len; off += sl->flash_pgsz) {
+    for (off = 0; off < mf->len; off += n_cmp) {
         size_t aligned_size;
 
         /* adjust last page size */
-        size_t cmp_size = sl->flash_pgsz;
-        if ((off + sl->flash_pgsz) > mf->len)
+        size_t cmp_size = n_cmp;
+        if ((off + n_cmp) > mf->len)
             cmp_size = mf->len - off;
 
         aligned_size = cmp_size;
@@ -751,6 +830,11 @@ int stlink_fwrite_sram
 
     /* success */
     error = 0;
+    /* set stack*/
+    stlink_write_reg(sl, stlink_read_debug32(sl, addr    ),13);
+    /* Set PC to the reset routine*/
+    stlink_write_reg(sl, stlink_read_debug32(sl, addr + 4),15);
+    stlink_run(sl);
 
 on_error:
     unmap_file(&mf);
@@ -762,6 +846,8 @@ int stlink_fread(stlink_t* sl, const char* path, stm32_addr_t addr, size_t size)
 
     int error = -1;
     size_t off;
+    int num_empty = 0;
+    unsigned char erased_pattern =(sl->chip_id == STM32_CHIPID_L1_MEDIUM)?0:0xff;
 
     const int fd = open(path, O_RDWR | O_TRUNC | O_CREAT, 00700);
     if (fd == -1) {
@@ -769,10 +855,17 @@ int stlink_fread(stlink_t* sl, const char* path, stm32_addr_t addr, size_t size)
         return -1;
     }
 
+    if (size <1)
+	size = sl->flash_size;
+
+    if (size > sl->flash_size)
+	size = sl->flash_size;
+
     /* do the copy by 1k blocks */
     for (off = 0; off < size; off += 1024) {
         size_t read_size = 1024;
 	size_t rounded_size;
+	size_t index;
         if ((off + read_size) > size)
 	  read_size = size - off;
 
@@ -783,12 +876,21 @@ int stlink_fread(stlink_t* sl, const char* path, stm32_addr_t addr, size_t size)
 
         stlink_read_mem32(sl, addr + off, rounded_size);
 
+	for(index = 0; index < read_size; index ++) {
+	    if (sl->q_buf[index] == erased_pattern)
+		num_empty ++;
+	    else
+		num_empty = 0;
+	}
         if (write(fd, sl->q_buf, read_size) != (ssize_t) read_size) {
             fprintf(stderr, "write() != read_size\n");
             goto on_error;
         }
     }
 
+    /* Ignore NULL Bytes at end of file */
+    ftruncate(fd, size - num_empty);
+
     /* success */
     error = 0;
 
@@ -800,8 +902,16 @@ on_error:
 
 int write_buffer_to_sram(stlink_t *sl, flash_loader_t* fl, const uint8_t* buf, size_t size) {
     /* write the buffer right after the loader */
-    memcpy(sl->q_buf, buf, size);
-    stlink_write_mem8(sl, fl->buf_addr, size);
+    size_t chunk = size & ~0x3;
+    size_t rem   = size & 0x3;
+    if (chunk) {
+        memcpy(sl->q_buf, buf, chunk);
+        stlink_write_mem32(sl, fl->buf_addr, chunk);
+    }
+    if (rem) {
+        memcpy(sl->q_buf, buf+chunk, rem);
+        stlink_write_mem8(sl, (fl->buf_addr)+chunk, rem);
+    }
     return 0;
 }
 
@@ -816,22 +926,24 @@ uint32_t calculate_F4_sectornum(uint32_t flashaddr){
 
 }
 
-uint32_t calculate_sectorsize(stlink_t *sl, uint32_t flashaddr){
+uint32_t stlink_calculate_pagesize(stlink_t *sl, uint32_t flashaddr){
 	if(sl->chip_id == STM32F4_CHIP_ID) {
 		uint32_t sector=calculate_F4_sectornum(flashaddr);
-		if (sector<4) return (0x4000);
-		else if(sector<5) return(0x10000);
-		else return(0x20000);
+		if (sector<4) sl->flash_pgsz=0x4000;
+		else if(sector<5) sl->flash_pgsz=0x10000;
+		else sl->flash_pgsz=0x20000;
 	}
-	else return (sl->flash_pgsz);
+	return (sl->flash_pgsz);
 }
 
-int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t page)
+/**
+ * Erase a page of flash, assumes sl is fully populated with things like chip/core ids
+ * @param sl stlink context
+ * @param flashaddr an address in the flash page to erase
+ * @return 0 on success -ve on failure
+ */
+int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t flashaddr)
 {
-  /* page an addr in the page to erase */
-
-  stlink_identify_device(sl);
-
   if (sl->chip_id == STM32F4_CHIP_ID)
   {
     /* wait for ongoing op to finish */
@@ -841,13 +953,11 @@ int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t page)
     unlock_flash_if(sl);
 
     /* select the page to erase */
-    //Page is passed to us as an addr, so calculate the actual page
-    uint32_t addr=page;
+    // calculate the actual page from the address
+    uint32_t sector=calculate_F4_sectornum(flashaddr);
 
-    page=calculate_F4_sectornum(addr);
-
-    fprintf(stderr, "Erasing Sector:%u SectorSize:%u\n", page, calculate_sectorsize(sl, addr));
-    write_flash_cr_snb(sl, page);
+    fprintf(stderr, "EraseFlash - Sector:0x%x Size:0x%x\n", sector, stlink_calculate_pagesize(sl, flashaddr));
+    write_flash_cr_snb(sl, sector);
 
     /* start erase operation */
     set_flash_cr_strt(sl);
@@ -858,55 +968,46 @@ int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t page)
     /* relock the flash */
     //todo: fails to program if this is in
     lock_flash(sl);
-	fprintf(stdout, "Erase Final CR:%X\n", read_flash_cr(sl));
-
+#if DEBUG_FLASH
+	fprintf(stdout, "Erase Final CR:0x%x\n", read_flash_cr(sl));
+#endif
   }
-
   else if (sl->core_id == STM32L_CORE_ID)
   {
 
     uint32_t val;
 
     /* disable pecr protection */
-    write_uint32(sl->q_buf, 0x89abcdef);
-    stlink_write_mem32(sl, STM32L_FLASH_PEKEYR, sizeof(uint32_t));
-    write_uint32(sl->q_buf, 0x02030405);
-    stlink_write_mem32(sl, STM32L_FLASH_PEKEYR, sizeof(uint32_t));
+    stlink_write_debug32(sl, STM32L_FLASH_PEKEYR, 0x89abcdef);
+    stlink_write_debug32(sl, STM32L_FLASH_PEKEYR, 0x02030405);
 
     /* check pecr.pelock is cleared */
-    stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    val = read_uint32(sl->q_buf, 0);
+    val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
     if (val & (1 << 0))
     {
-      fprintf(stderr, "pecr.pelock not clear (0x%x)\n", val);
+      WLOG("pecr.pelock not clear (%#x)\n", val);
       return -1;
     }
 
     /* unlock program memory */
-    write_uint32(sl->q_buf, 0x8c9daebf);
-    stlink_write_mem32(sl, STM32L_FLASH_PRGKEYR, sizeof(uint32_t));
-    write_uint32(sl->q_buf, 0x13141516);
-    stlink_write_mem32(sl, STM32L_FLASH_PRGKEYR, sizeof(uint32_t));
+    stlink_write_debug32(sl, STM32L_FLASH_PRGKEYR, 0x8c9daebf);
+    stlink_write_debug32(sl, STM32L_FLASH_PRGKEYR, 0x13141516);
 
     /* check pecr.prglock is cleared */
-    stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    val = read_uint32(sl->q_buf, 0);
+    val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
     if (val & (1 << 1))
     {
-      fprintf(stderr, "pecr.prglock not clear (0x%x)\n", val);
+      WLOG("pecr.prglock not clear (%#x)\n", val);
       return -1;
     }
 
     /* unused: unlock the option byte block */
 #if 0
-    write_uint32(sl->q_buf, 0xfbead9c8);
-    stlink_write_mem32(sl, STM32L_FLASH_OPTKEYR, sizeof(uint32_t));
-    write_uint32(sl->q_buf, 0x24252627);
-    stlink_write_mem32(sl, STM32L_FLASH_OPTKEYR, sizeof(uint32_t));
+    stlink_write_debug32(sl, STM32L_FLASH_OPTKEYR, 0xfbead9c8);
+    stlink_write_debug32(sl, STM32L_FLASH_OPTKEYR, 0x24252627);
 
     /* check pecr.optlock is cleared */
-    stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    val = read_uint32(sl->q_buf, 0);
+    val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
     if (val & (1 << 2))
     {
       fprintf(stderr, "pecr.prglock not clear\n");
@@ -916,25 +1017,37 @@ int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t page)
 
     /* set pecr.{erase,prog} */
     val |= (1 << 9) | (1 << 3);
-    write_uint32(sl->q_buf, val);
-    stlink_write_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
+    stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
+
+#if 0 /* fix_to_be_confirmed */
 
-    /* wait for sr.busy to be cleared */
-    while (1)
+    /* wait for sr.busy to be cleared
+       MP: Test shows that busy bit is not set here. Perhaps, PM0062 is
+       wrong and we do not need to wait here for clearing the busy bit.
+       TEXANE: ok, if experience says so and it works for you, we comment
+       it. If someone has a problem, please drop an email.
+     */
+    while ((stlink_read_debug32(sl, STM32L_FLASH_SR) & (1 << 0)) != 0)
     {
-      stlink_read_mem32(sl, STM32L_FLASH_SR, sizeof(uint32_t));
-      if ((read_uint32(sl->q_buf, 0) & (1 << 0)) == 0) break ;
     }
 
+#endif /* fix_to_be_confirmed */
+
     /* write 0 to the first word of the page to be erased */
-    memset(sl->q_buf, 0, sizeof(uint32_t));
-    stlink_write_mem32(sl, page, sizeof(uint32_t));
+    stlink_write_debug32(sl, flashaddr, 0);
+
+    /* MP: It is better to wait for clearing the busy bit after issuing
+    page erase command, even though PM0062 recommends to wait before it.
+    Test shows that a few iterations is performed in the following loop
+    before busy bit is cleared.*/
+    while ((stlink_read_debug32(sl, STM32L_FLASH_SR) & (1 << 0)) != 0)
+    {
+    }
 
     /* reset lock bits */
-    stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    val = read_uint32(sl->q_buf, 0) | (1 << 0) | (1 << 1) | (1 << 2);
-    write_uint32(sl->q_buf, val);
-    stlink_write_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
+    val = stlink_read_debug32(sl, STM32L_FLASH_PECR)
+        | (1 << 0) | (1 << 1) | (1 << 2);
+    stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
   }
   else if (sl->core_id == STM32VL_CORE_ID)
   {
@@ -948,7 +1061,7 @@ int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t page)
     set_flash_cr_per(sl);
 
     /* select the page to erase */
-    write_flash_ar(sl, page);
+    write_flash_ar(sl, flashaddr);
 
     /* start erase operation, reset by hw with bsy bit */
     set_flash_cr_strt(sl);
@@ -961,7 +1074,7 @@ int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t page)
   }
 
   else {
-    fprintf(stderr, "unknown device!\n");
+    WLOG("unknown coreid: %x\n", sl->core_id);
     return -1;
   }
 
@@ -971,26 +1084,42 @@ int stlink_erase_flash_page(stlink_t *sl, stm32_addr_t page)
 }
 
 int stlink_erase_flash_mass(stlink_t *sl) {
-    /* wait for ongoing op to finish */
-    wait_flash_busy(sl);
-
-    /* unlock if locked */
-    unlock_flash_if(sl);
-
-    /* set the mass erase bit */
-    set_flash_cr_mer(sl);
-
-    /* start erase operation, reset by hw with bsy bit */
-    set_flash_cr_strt(sl);
-
-    /* wait for completion */
-    wait_flash_busy(sl);
-
-    /* relock the flash */
-    lock_flash(sl);
-
-    /* todo: verify the erased memory */
-
+     if (sl->chip_id == STM32_CHIPID_L1_MEDIUM) {
+	 /* erase each page */
+	 int i = 0, num_pages = sl->flash_size/sl->flash_pgsz;
+	 for (i = 0; i < num_pages; i++) {
+	     /* addr must be an addr inside the page */
+	     stm32_addr_t addr = sl->flash_base + i * sl->flash_pgsz;
+	     if (stlink_erase_flash_page(sl, addr) == -1) {
+		 WLOG("Failed to erase_flash_page(%#zx) == -1\n", addr);
+		 return -1;
+	     }
+	     fprintf(stdout,"\rFlash page at %5d/%5d erased", i, num_pages);
+	     fflush(stdout);
+	 }
+	 fprintf(stdout, "\n");
+     }
+     else {
+	 /* wait for ongoing op to finish */
+	 wait_flash_busy(sl);
+	 
+	 /* unlock if locked */
+	 unlock_flash_if(sl);
+	 
+	 /* set the mass erase bit */
+	 set_flash_cr_mer(sl);
+	 
+	 /* start erase operation, reset by hw with bsy bit */
+	 set_flash_cr_strt(sl);
+	 
+	 /* wait for completion */
+	 wait_flash_busy_progress(sl);
+	 
+	 /* relock the flash */
+	 lock_flash(sl);
+	 
+	 /* todo: verify the erased memory */
+     }
     return 0;
 }
 
@@ -999,13 +1128,13 @@ int init_flash_loader(stlink_t *sl, flash_loader_t* fl) {
 
     /* allocate the loader in sram */
     if (write_loader_to_sram(sl, &fl->loader_addr, &size) == -1) {
-        fprintf(stderr, "write_loader_to_sram() == -1\n");
+        WLOG("Failed to write flash loader to sram!\n");
         return -1;
     }
 
     /* allocate a one page buffer in sram right after loader */
     fl->buf_addr = fl->loader_addr + size;
-
+    ILOG("Successfully loaded flash loader in sram\n");
     return 0;
 }
 
@@ -1068,7 +1197,7 @@ int write_loader_to_sram(stlink_t *sl, stm32_addr_t* addr, size_t* size) {
     }
     else
     {
-      fprintf(stderr, "unknown coreid: %x\n", sl->core_id);
+      WLOG("unknown coreid, not sure what flash loader to use, aborting!: %x\n", sl->core_id);
       return -1;
     }
 
@@ -1098,42 +1227,137 @@ int stlink_fcheck_flash(stlink_t *sl, const char* path, stm32_addr_t addr) {
     return res;
 }
 
-int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned len) {
+/**
+ * Verify addr..addr+len is binary identical to base...base+len
+ * @param sl stlink context
+ * @param address stm device address
+ * @param data host side buffer to check against
+ * @param length how much
+ * @return 0 for success, -ve for failure
+ */
+int stlink_verify_write_flash(stlink_t *sl, stm32_addr_t address, uint8_t *data, unsigned length) {
     size_t off;
-    flash_loader_t fl;
+    if (sl->chip_id == STM32_CHIPID_F4) {
+        DLOG("(FIXME)Skipping verification for F4, not enough ram (yet)\n");
+        return 0;
+    }
+    ILOG("Starting verification of write complete\n");
+    for (off = 0; off < length; off += sl->flash_pgsz) {
+        size_t aligned_size;
 
-    fprintf(stdout, "WriteFlash - addr:%x len:%x\n", addr, len);
+        /* adjust last page size */
+        size_t cmp_size = sl->flash_pgsz;
+        if ((off + sl->flash_pgsz) > length)
+            cmp_size = length - off;
 
-    stlink_identify_device(sl);
+        aligned_size = cmp_size;
+        if (aligned_size & (4 - 1))
+            aligned_size = (cmp_size + 4) & ~(4 - 1);
 
-    fprintf(stdout, "CoreID:%X ChipID:%X\n", sl->core_id, sl->chip_id);
+        stlink_read_mem32(sl, address + off, aligned_size);
 
+        if (memcmp(sl->q_buf, data + off, cmp_size)) {
+            WLOG("Verification of flash failed at offset: %zd\n", off);
+            return -1;
+        }
+    }
+    ILOG("Flash written and verified! jolly good!\n");
+    return 0;
+
+}
+
+int stm32l1_write_half_pages(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned num_half_pages)
+{        
+    unsigned int count;
+    uint32_t val;
+    flash_loader_t fl;
+
+    ILOG("Starting Half page flash write for STM32L core id\n");
+    /* flash loader initialization */
+    if (init_flash_loader(sl, &fl) == -1) {
+        WLOG("init_flash_loader() == -1\n");
+        return -1;
+    }
+    /* Unlock already done */
+    val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
+    val |= (1 << FLASH_L1_FPRG);
+    stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
+    
+    val |= (1 << FLASH_L1_PROG);
+    stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
+    while ((stlink_read_debug32(sl, STM32L_FLASH_SR) & (1 << 0)) != 0) {}
+
+#define L1_WRITE_BLOCK_SIZE 0x80
+    for (count = 0; count  < num_half_pages; count ++) {
+        if (run_flash_loader(sl, &fl, addr + count * L1_WRITE_BLOCK_SIZE, base + count * L1_WRITE_BLOCK_SIZE, L1_WRITE_BLOCK_SIZE) == -1) {
+            WLOG("l1_run_flash_loader(%#zx) failed! == -1\n", addr + count * L1_WRITE_BLOCK_SIZE);
+            val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
+            val &= ~((1 << FLASH_L1_FPRG) |(1 << FLASH_L1_PROG));
+            stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
+            return -1;
+        }
+        /* wait for sr.busy to be cleared */
+        if (sl->verbose >= 1) {
+            /* show progress. writing procedure is slow
+               and previous errors are misleading */
+            fprintf(stdout, "\r%3u/%u halfpages written", count, num_half_pages);
+            fflush(stdout);
+        }
+        while ((stlink_read_debug32(sl, STM32L_FLASH_SR) & (1 << 0)) != 0) {
+        }
+    }
+    val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
+    val &= ~(1 << FLASH_L1_PROG);
+    stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
+    val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
+    val &= ~(1 << FLASH_L1_FPRG);
+    stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
+
+    return 0;
+}
+
+int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned len) {
+    size_t off;
+    flash_loader_t fl;
+    ILOG("Attempting to write %d (%#x) bytes to stm32 address: %u (%#x)\n",
+        len, len, addr, addr);
     /* check addr range is inside the flash */
+    stlink_calculate_pagesize(sl, addr);
     if (addr < sl->flash_base) {
-        fprintf(stderr, "addr too low\n");
+        WLOG("addr too low %#x < %#x\n", addr, sl->flash_base);
         return -1;
     } else if ((addr + len) < addr) {
-        fprintf(stderr, "addr overruns\n");
+        WLOG("addr overruns\n");
         return -1;
     } else if ((addr + len) > (sl->flash_base + sl->flash_size)) {
-        fprintf(stderr, "addr too high\n");
+        WLOG("addr too high\n");
         return -1;
     } else if ((addr & 1) || (len & 1)) {
-        fprintf(stderr, "unaligned addr or size\n");
+        WLOG("unaligned addr or size\n");
         return -1;
     } else if (addr & (sl->flash_pgsz - 1)) {
-        fprintf(stderr, "addr not a multiple of pagesize, not supported\n");
+        WLOG("addr not a multiple of pagesize, not supported\n");
         return -1;
     }
 
+    // Make sure we've loaded the context with the chip details
+    stlink_core_id(sl);
     /* erase each page */
-    for (off = 0; off < len; off += calculate_sectorsize(sl, addr + off) ) {
+    int page_count = 0;
+    for (off = 0; off < len; off += stlink_calculate_pagesize(sl, addr + off)) {
         /* addr must be an addr inside the page */
         if (stlink_erase_flash_page(sl, addr + off) == -1) {
-           fprintf(stderr, "erase_flash_page(0x%zx) == -1\n", addr + off);
-	    return -1;
+            WLOG("Failed to erase_flash_page(%#zx) == -1\n", addr + off);
+            return -1;
         }
+        fprintf(stdout,"\rFlash page at addr: 0x%08lx erased",
+		(unsigned long)addr + off);
+        fflush(stdout);
+        page_count++;
     }
+    fprintf(stdout,"\n");
+    ILOG("Finished erasing %d pages of %d (%#x) bytes\n", 
+        page_count, sl->flash_pgsz, sl->flash_pgsz);
 
     if (sl->chip_id == STM32F4_CHIP_ID) {
     	/* todo: check write operation */
@@ -1141,6 +1365,7 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
     	/* First unlock the cr */
     	unlock_flash_if(sl);
 
+	/* TODO: Check that Voltage range is 2.7 - 3.6 V */
     	/* set parallelisim to 32 bit*/
     	write_flash_cr_psiz(sl, 2);
 
@@ -1150,6 +1375,7 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
 #define PROGRESS_CHUNK_SIZE 0x1000
     	/* write a word in program memory */
     	for (off = 0; off < len; off += sizeof(uint32_t)) {
+    		uint32_t data;
     		if (sl->verbose >= 1) {
     			if ((off & (PROGRESS_CHUNK_SIZE - 1)) == 0) {
     				/* show progress. writing procedure is slow
@@ -1160,8 +1386,8 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
     			}
     		}
 
-    		memcpy(sl->q_buf, (const void*)(base + off), sizeof(uint32_t));
-    		stlink_write_mem32(sl, addr + off, sizeof(uint32_t));
+    		write_uint32((unsigned char*) &data, *(uint32_t*) (base + off));
+    		stlink_write_debug32(sl, addr + off, data);
 
     		/* wait for sr.busy to be cleared */
     	    wait_flash_busy(sl);
@@ -1169,7 +1395,10 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
     	}
     	/* Relock flash */
     	lock_flash(sl);
-		fprintf(stdout, "Final CR:%X\n", read_flash_cr(sl));
+
+#if 0 /* todo: debug mode */
+	fprintf(stdout, "Final CR:0x%x\n", read_flash_cr(sl));
+#endif
 
 
 
@@ -1177,7 +1406,6 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
 
     else if (sl->core_id == STM32L_CORE_ID)    {
     	/* use fast word write. todo: half page. */
-
     	uint32_t val;
 
 #if 0 /* todo: check write operation */
@@ -1189,59 +1417,61 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
 #endif /* todo: check write operation */
 
     	/* disable pecr protection */
-    	write_uint32(sl->q_buf, 0x89abcdef);
-    	stlink_write_mem32(sl, STM32L_FLASH_PEKEYR, sizeof(uint32_t));
-    	write_uint32(sl->q_buf, 0x02030405);
-    	stlink_write_mem32(sl, STM32L_FLASH_PEKEYR, sizeof(uint32_t));
+    	stlink_write_debug32(sl, STM32L_FLASH_PEKEYR, 0x89abcdef);
+    	stlink_write_debug32(sl, STM32L_FLASH_PEKEYR, 0x02030405);
 
     	/* check pecr.pelock is cleared */
-    	stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    	val = read_uint32(sl->q_buf, 0);
+    	val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
     	if (val & (1 << 0)) {
     		fprintf(stderr, "pecr.pelock not clear\n");
     		return -1;
     	}
 
     	/* unlock program memory */
-    	write_uint32(sl->q_buf, 0x8c9daebf);
-    	stlink_write_mem32(sl, STM32L_FLASH_PRGKEYR, sizeof(uint32_t));
-    	write_uint32(sl->q_buf, 0x13141516);
-    	stlink_write_mem32(sl, STM32L_FLASH_PRGKEYR, sizeof(uint32_t));
+    	stlink_write_debug32(sl, STM32L_FLASH_PRGKEYR, 0x8c9daebf);
+    	stlink_write_debug32(sl, STM32L_FLASH_PRGKEYR, 0x13141516);
 
     	/* check pecr.prglock is cleared */
-    	stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    	val = read_uint32(sl->q_buf, 0);
+    	val = stlink_read_debug32(sl, STM32L_FLASH_PECR);
     	if (val & (1 << 1)) {
     		fprintf(stderr, "pecr.prglock not clear\n");
     		return -1;
     	}
+	off = 0;
+        if (len > L1_WRITE_BLOCK_SIZE) {
+            if (stm32l1_write_half_pages(sl, addr, base, len/L1_WRITE_BLOCK_SIZE) == -1){
+		/* This may happen on a blank device! */
+                WLOG("\nwrite_half_pages failed == -1\n");
+	    }
+	    else{
+		off = (len /L1_WRITE_BLOCK_SIZE)*L1_WRITE_BLOCK_SIZE;
+	    }
+	}
 
-    	/* write a word in program memory */
-    	for (off = 0; off < len; off += sizeof(uint32_t)) {
-    		if (sl->verbose >= 1) {
-    			if ((off & (sl->flash_pgsz - 1)) == 0) {
-    				/* show progress. writing procedure is slow
-				   and previous errors are misleading */
-    				const uint32_t pgnum = off / sl->flash_pgsz;
-    				const uint32_t pgcount = len / sl->flash_pgsz;
-    				fprintf(stdout, "%u pages written out of %u\n", pgnum, pgcount);
-    			}
-    		}
+    	/* write remainingword in program memory */
+    	for ( ; off < len; off += sizeof(uint32_t)) {
+    		uint32_t data;
+		if (off > 254)
+		    fprintf(stdout, "\r");
 
-    		memcpy(sl->q_buf, (const void*)(base + off), sizeof(uint32_t));
-    		stlink_write_mem32(sl, addr + off, sizeof(uint32_t));
+		if ((off % sl->flash_pgsz) > (sl->flash_pgsz -5)) {
+		    fprintf(stdout, "\r%3u/%u pages written", 
+			    off/sl->flash_pgsz, len/sl->flash_pgsz);
+		    fflush(stdout);
+		}
+
+    		write_uint32((unsigned char*) &data, *(uint32_t*) (base + off));
+    		stlink_write_debug32(sl, addr + off, data);
 
     		/* wait for sr.busy to be cleared */
-    		while (1) {
-    			stlink_read_mem32(sl, STM32L_FLASH_SR, sizeof(uint32_t));
-    			if ((read_uint32(sl->q_buf, 0) & (1 << 0)) == 0) break ;
+    		while ((stlink_read_debug32(sl, STM32L_FLASH_SR) & (1 << 0)) != 0) {
     		}
 
 #if 0 /* todo: check redo write operation */
 
     		/* check written bytes. todo: should be on a per page basis. */
-    		stlink_read_mem32(sl, addr + off, sizeof(uint32_t));
-    		if (memcmp(sl->q_buf, base + off, sizeof(uint32_t))) {
+    		data = stlink_read_debug32(sl, addr + off);
+    		if (data == *(uint32_t*)(base + off)) {
     			/* re erase the page and redo the write operation */
     			uint32_t page;
     			uint32_t val;
@@ -1258,14 +1488,13 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
     			if (off % sl->flash_pgsz) off &= ~(sl->flash_pgsz - 1);
     			page = addr + off;
 
-    			fprintf(stderr, "invalid write @%x(%x): %x != %x. retrying.\n",
+    			fprintf(stderr, "invalid write @0x%x(0x%x): 0x%x != 0x%x. retrying.\n",
     					page, addr + off, read_uint32(base + off, 0), read_uint32(sl->q_buf, 0));
 
     			/* reset lock bits */
-    			stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    			val = read_uint32(sl->q_buf, 0) | (1 << 0) | (1 << 1) | (1 << 2);
-    			write_uint32(sl->q_buf, val);
-    			stlink_write_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
+    			val = stlink_read_debug32(sl, STM32L_FLASH_PECR)
+                             | (1 << 0) | (1 << 1) | (1 << 2);
+    			stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
 
     			stlink_erase_flash_page(sl, page);
 
@@ -1277,106 +1506,96 @@ int stlink_write_flash(stlink_t *sl, stm32_addr_t addr, uint8_t* base, unsigned
 
 #endif /* todo: check redo write operation */
     	}
+        fprintf(stdout, "\n");
     	/* reset lock bits */
-    	stlink_read_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-    	val = read_uint32(sl->q_buf, 0) | (1 << 0) | (1 << 1) | (1 << 2);
-    	write_uint32(sl->q_buf, val);
-    	stlink_write_mem32(sl, STM32L_FLASH_PECR, sizeof(uint32_t));
-	}
-
-
-
-
-    else if (sl->core_id == STM32VL_CORE_ID) {
-    	/* flash loader initialization */
-    	if (init_flash_loader(sl, &fl) == -1) {
-    		fprintf(stderr, "init_flash_loader() == -1\n");
-    		return -1;
-    	}
-    	/* write each page. above WRITE_BLOCK_SIZE fails? */
-#define WRITE_BLOCK_SIZE 0x40
-      for (off = 0; off < len; off += WRITE_BLOCK_SIZE)      {
-    	  /* adjust last write size */
-    	  size_t size = WRITE_BLOCK_SIZE;
-    	  if ((off + WRITE_BLOCK_SIZE) > len) size = len - off;
-
-    	  /* unlock and set programming mode */
-    	  unlock_flash_if(sl);
-    	  set_flash_cr_pg(sl);
-
-    	  if (run_flash_loader(sl, &fl, addr + off, base + off, size) == -1) {
-    		  fprintf(stderr, "run_flash_loader(0x%zx) == -1\n", addr + off);
-    		  return -1;
-    	  }
-    	  lock_flash(sl);
-      }
-    }
-
-
-
-
-    else
-    {
-      fprintf(stderr, "unknown device!\n");
-      return -1;
-    }
-
-
-
-
-
-#if(0)
-    //todo: F4 Can't stlink_read_mem32 an entire sector, not enough ram!
-    for (off = 0; off < len; off += sl->flash_pgsz) {
-        size_t aligned_size;
-
-        /* adjust last page size */
-        size_t cmp_size = sl->flash_pgsz;
-        if ((off + sl->flash_pgsz) > len)
-            cmp_size = len - off;
-
-        aligned_size = cmp_size;
-        if (aligned_size & (4 - 1))
-            aligned_size = (cmp_size + 4) & ~(4 - 1);
-
-		fprintf(stdout, "AlignedSize:%x\n", aligned_size);
-        stlink_read_mem32(sl, addr + off, aligned_size);
-
-        if (memcmp(sl->q_buf, base + off, cmp_size))
+    	val = stlink_read_debug32(sl, STM32L_FLASH_PECR)
+             | (1 << 0) | (1 << 1) | (1 << 2);
+    	stlink_write_debug32(sl, STM32L_FLASH_PECR, val);
+    } else if (sl->core_id == STM32VL_CORE_ID) {
+        ILOG("Starting Flash write for VL core id\n");
+        /* flash loader initialization */
+        if (init_flash_loader(sl, &fl) == -1) {
+            WLOG("init_flash_loader() == -1\n");
             return -1;
-    }
-#endif
+        }
 
-    return 0;
+        int write_block_count = 0;
+        for (off = 0; off < len; off += sl->flash_pgsz) {
+            /* adjust last write size */
+            size_t size = sl->flash_pgsz;
+            if ((off + sl->flash_pgsz) > len) size = len - off;
+
+            /* unlock and set programming mode */
+            unlock_flash_if(sl);
+            set_flash_cr_pg(sl);
+            //DLOG("Finished setting flash cr pg, running loader!\n");
+            if (run_flash_loader(sl, &fl, addr + off, base + off, size) == -1) {
+                WLOG("run_flash_loader(%#zx) failed! == -1\n", addr + off);
+                return -1;
+            }
+            lock_flash(sl);
+            if (sl->verbose >= 1) {
+                /* show progress. writing procedure is slow
+                   and previous errors are misleading */
+	      fprintf(stdout, "\r%3u/%lu pages written", write_block_count++, (unsigned long)len/sl->flash_pgsz);
+                fflush(stdout);
+            }
+        }
+        fprintf(stdout, "\n");
+    } else {
+        WLOG("unknown coreid, not sure how to write: %x\n", sl->core_id);
+        return -1;
+    }
+    
+    return stlink_verify_write_flash(sl, addr, base, len);
 }
 
-
-
-
+/**
+ * Write the given binary file into flash at address "addr"
+ * @param sl
+ * @param path readable file path, should be binary image
+ * @param addr where to start writing
+ * @return 0 on success, -ve on failure.
+ */
 int stlink_fwrite_flash(stlink_t *sl, const char* path, stm32_addr_t addr) {
     /* write the file in flash at addr */
-
     int err;
+    unsigned int num_empty = 0, index;
+    unsigned char erased_pattern =(sl->chip_id == STM32_CHIPID_L1_MEDIUM)?0:0xff;
     mapped_file_t mf = MAPPED_FILE_INITIALIZER;
-
     if (map_file(&mf, path) == -1) {
-        fprintf(stderr, "map_file() == -1\n");
+        WLOG("map_file() == -1\n");
         return -1;
     }
-
+    for(index = 0; index < mf.len; index ++) {
+	if (mf.base[index] == erased_pattern)
+	    num_empty ++;
+	else
+	    num_empty = 0;
+    }
+    if(num_empty != 0) {
+	ILOG("Ignoring %d bytes of Zeros at end of file\n",num_empty);
+	mf.len -= num_empty;
+    }
     err = stlink_write_flash(sl, addr, mf.base, mf.len);
-
+    /* set stack*/
+    stlink_write_reg(sl, stlink_read_debug32(sl, addr    ),13);
+    /* Set PC to the reset routine*/
+    stlink_write_reg(sl, stlink_read_debug32(sl, addr + 4),15);
+    stlink_run(sl);
     unmap_file(&mf);
-
     return err;
 }
 
 int run_flash_loader(stlink_t *sl, flash_loader_t* fl, stm32_addr_t target, const uint8_t* buf, size_t size) {
 
     reg rr;
-
+    int i = 0;
+    DLOG("Running flash loader, write address:%#x, size: %zd\n", target, size);
+    // FIXME This can never return -1
     if (write_buffer_to_sram(sl, fl, buf, size) == -1) {
-        fprintf(stderr, "write_buffer_to_sram() == -1\n");
+        // IMPOSSIBLE!
+        WLOG("write_buffer_to_sram() == -1\n");
         return -1;
     }
 
@@ -1389,7 +1608,6 @@ int run_flash_loader(stlink_t *sl, flash_loader_t* fl, stm32_addr_t target, cons
       stlink_write_reg(sl, target, 0); /* target */
       stlink_write_reg(sl, fl->buf_addr, 1); /* source */
       stlink_write_reg(sl, count, 2); /* count (32 bits words) */
-      stlink_write_reg(sl, 0, 3); /* output count */
       stlink_write_reg(sl, fl->loader_addr, 15); /* pc register */
 
     } else if (sl->core_id == STM32VL_CORE_ID) {
@@ -1405,16 +1623,24 @@ int run_flash_loader(stlink_t *sl, flash_loader_t* fl, stm32_addr_t target, cons
       stlink_write_reg(sl, fl->loader_addr, 15); /* pc register */
 
     } else {
-      fprintf(stderr, "unknown coreid: %x\n", sl->core_id);
+      fprintf(stderr, "unknown coreid: 0x%x\n", sl->core_id);
       return -1;
     }
 
     /* run loader */
-    stlink_step(sl);
+    stlink_run(sl);
 
     /* wait until done (reaches breakpoint) */
-    while (is_core_halted(sl) == 0) ;
+    while ((is_core_halted(sl) == 0) && (i <1000))
+    {
+        i++;
+    }
 
+    if ( i > 999) {
+        fprintf(stderr, "run error\n");
+        return -1;
+    }
+        
     /* check written byte count */
     if (sl->core_id == STM32L_CORE_ID) {
 
@@ -1437,7 +1663,7 @@ int run_flash_loader(stlink_t *sl, flash_loader_t* fl, stm32_addr_t target, cons
 
     } else {
 
-      fprintf(stderr, "unknown coreid: %x\n", sl->core_id);
+      fprintf(stderr, "unknown coreid: 0x%x\n", sl->core_id);
       return -1;
 
     }