Index: firmware/target/arm/ata-sd-pp.c
===================================================================
--- firmware/target/arm/ata-sd-pp.c	(Revision 19889)
+++ firmware/target/arm/ata-sd-pp.c	(Arbeitskopie)
@@ -54,6 +54,7 @@
 #define DATA_REG        (*(volatile unsigned int *)(0x70008280))
 
 /* STATUS_REG bits */
+#define DATA_TRAN_DONE  (1 << 11)
 #define DATA_DONE       (1 << 12)
 #define CMD_DONE        (1 << 13)
 #define ERROR_BITS      (0x3f)
@@ -89,6 +90,8 @@
 #define EC_COMMAND              19
 #define NUM_EC                  20
 
+
+static int wait_loop;
 /* for compatibility */
 static long last_disk_activity = -1;
 
@@ -423,15 +426,46 @@
 /* Writes have to be kept slow for now */
 static inline void copy_write_sectors(const unsigned char** buf)
 {
-    int cnt = FIFO_LEN;
+    int cnt = FIFO_LEN - 1;
     unsigned t;
+    long time;
 
-    do
+    time = USEC_TIMER + 3;
+    if  (((intptr_t)*buf & 3) == 0)
     {
+        asm volatile (
+            "ldmia  %[buf]!, { r3, r5, r7, r9 } \r\n"
+            "mov    r4, r3, lsr #16             \r\n" 
+            "mov    r6, r5, lsr #16             \r\n" 
+            "mov    r8, r7, lsr #16             \r\n" 
+            "mov    r10, r9, lsr #16            \r\n" 
+            "stmia  %[data], { r3-r10 }         \r\n"
+            "ldmia  %[buf]!, { r3, r5, r7, r9 } \r\n"
+            "mov    r4, r3, lsr #16             \r\n" 
+            "mov    r6, r5, lsr #16             \r\n" 
+            "mov    r8, r7, lsr #16             \r\n" 
+            "mov    %[t], r9, lsr #16           \r\n" 
+            "stmia  %[data], { r3-r9 }          \r\n"
+            : [buf]"+&r"(*buf), [t]"=&r"(t) 
+            : [data]"r"(&DATA_REG)
+            : "r3", "r4", "r5", "r6", "r7", "r8", "r9", "r10"
+        );
+    }
+    else
+    {
+        do
+        {
+            t  = *(*buf)++;
+            t |= *(*buf)++ << 8;
+            DATA_REG = t;
+        } while (--cnt > 0); /* tail loop is faster */
         t  = *(*buf)++;
         t |= *(*buf)++ << 8;
-        DATA_REG = t;
-    } while (--cnt > 0); /* tail loop is faster */
+    }
+    /* Don't write the last word before at least 3 usec have elapsed since FIFO_EMPTY */
+    /* This prevents the 'two bytes inserted' bug. */
+    while (!TIME_AFTER(USEC_TIMER, time));
+    DATA_REG = t;
 }
 
 static int sd_select_bank(unsigned char bank)
@@ -920,12 +954,12 @@
     if(currcard->ocr & (1<<30) )
     {
         /* SDHC */
-        ret = sd_command(SD_WRITE_MULTIPLE_BLOCK, start, NULL, 0x1c2d);
+        ret = sd_command(SD_WRITE_MULTIPLE_BLOCK, start, NULL, 0xd);
     }
     else
 #endif
     {
-        ret = sd_command(SD_WRITE_MULTIPLE_BLOCK, start*BLOCK_SIZE, NULL, 0x1c2d);
+        ret = sd_command(SD_WRITE_MULTIPLE_BLOCK, start*BLOCK_SIZE, NULL, 0xd);
     }
     if (ret < 0)
         goto sd_write_error;
@@ -940,18 +974,24 @@
             SD_STATE_REG = SD_PRG;
         }
 
-        udelay(2); /* needed here (loop is too fast :-) */
+        { int cnt = wait_loop;
 
+          asm volatile (
+          "1:                        \n"
+          "subs   %[cnt], %[cnt], #1 \n"
+          "bgt    1b                 \n"
+          : [cnt]"+&r"(cnt) );
+        }
+
+        copy_write_sectors(&buf); /* Copy one chunk of 16 words */
+        /* TODO: Switch bank if necessary */
+
         /* Wait for the FIFO to empty */
-        if (sd_poll_status(FIFO_EMPTY, 0x80000))
+        if(sd_poll_status(FIFO_EMPTY, 0x80000) == 0)
         {
-            copy_write_sectors(&buf); /* Copy one chunk of 16 words */
-            /* TODO: Switch bank if necessary */
-            continue;
+            ret = -EC_FIFO_WR_EMPTY;
+            goto sd_write_error;
         }
-
-        ret = -EC_FIFO_WR_EMPTY;
-        goto sd_write_error;
     }
 
     last_disk_activity = current_tick;
Index: firmware/usbstack/usb_storage.c
===================================================================
--- firmware/usbstack/usb_storage.c	(Revision 19889)
+++ firmware/usbstack/usb_storage.c	(Arbeitskopie)
@@ -39,9 +39,12 @@
  * by not overlapping storage_write_sector() with USB transfers. This does reduce
  * write performance, so we only do it for the affected DAPs
  */
-#if (CONFIG_STORAGE & STORAGE_SD)
+
+#if (CONFIG_STORAGE & STORAGE_SD) && !defined(CPU_PP)
 #define SERIALIZE_WRITES
 #endif
+
+
 /* Enable the following define to export only the SD card slot. This
  * is useful for USBCV MSC tests, as those are destructive.
  * This won't work right if the device doesn't have a card slot.