Index: firmware/target/arm/ata-sd-pp.c
===================================================================
--- firmware/target/arm/ata-sd-pp.c	(Revision 17762)
+++ firmware/target/arm/ata-sd-pp.c	(Arbeitskopie)
@@ -478,6 +478,30 @@
     } while (--cnt > 0); /* tail loop is faster */
 }
 
+#ifdef SANSA_E200
+static inline void copy_write_sectors_usb(const unsigned char** buf)
+{
+    /* No need to check for alignment as the USB buffer will always be aligned. */
+    asm volatile (
+        "ldmia  %[buf]!, { r2, r4, r6, r8 } \r\n"
+        "mov    r3, r2, lsr #16             \r\n" 
+        "mov    r5, r4, lsr #16             \r\n" 
+        "mov    r7, r6, lsr #16             \r\n" 
+        "mov    r9, r8, lsr #16             \r\n" 
+        "stmia  %[data], { r2-r9 }          \r\n"
+        "ldmia  %[buf]!, { r2, r4, r6, r8 } \r\n"
+        "mov    r3, r2, lsr #16             \r\n" 
+        "mov    r5, r4, lsr #16             \r\n" 
+        "mov    r7, r6, lsr #16             \r\n" 
+        "mov    r9, r8, lsr #16             \r\n" 
+        "stmia  %[data], { r2-r9 }          \r\n"
+        : [buf]"+&r"(*buf)
+        : [data]"r"(&DATA_REG)
+        : "r2", "r3", "r4", "r5", "r6", "r7", "r8", "r9"
+    );
+}
+#endif
+
 static int sd_select_bank(unsigned char bank)
 {
     unsigned char card_data[512];
@@ -975,26 +999,56 @@
 
     buf_end = outbuf + count * currcard->block_size - 2*FIFO_LEN;
 
-    for (buf = outbuf; buf <= buf_end;)
+#ifdef SANSA_E200
+    if (usb_exclusive_ata())
     {
-        if (buf == buf_end)
+        /* Duplicating the loop to avoid calling usb_exclusive_ata() inside. */
+        for (buf = outbuf; buf <= buf_end;)
         {
-            /* Set SD_STATE_REG to PRG for the last buffer fill */
-            SD_STATE_REG = PRG;
-        }
+            if (buf == buf_end)
+            {
+                /* Set SD_STATE_REG to PRG for the last buffer fill */
+                SD_STATE_REG = PRG;
+            }
 
-        udelay(2); /* needed here (loop is too fast :-) */
+            /* Wait for the FIFO to empty */
+            /* Not using sd_poll_status here because yielding */
+            /* may disturb the timing. */
+            while ((STATUS_REG & FIFO_EMPTY) == 0);
+            
+            udelay(3); /* Needed on e200. */
+                       /* Too short or too long delay will raise */
+                       /* the 'two bytes inserted' bug. */
+                       /* Too short delay may cause lockup as well. */
 
-        /* Wait for the FIFO to empty */
-        if (sd_poll_status(FIFO_EMPTY, 0x80000))
-        {
-            copy_write_sectors(&buf); /* Copy one chunk of 16 words */
+            copy_write_sectors_usb(&buf);
             /* TODO: Switch bank if necessary */
-            continue;
         }
+	}
+    else
+#endif
+	{
+        for (buf = outbuf; buf <= buf_end;)
+        {
+            if (buf == buf_end)
+            {
+                /* Set SD_STATE_REG to PRG for the last buffer fill */
+                SD_STATE_REG = PRG;
+            }
 
-        ret = -EC_FIFO_WR_EMPTY;
-        goto ata_write_error;
+            udelay(2); /* needed here (loop is too fast :-) */
+
+            /* Wait for the FIFO to empty */
+            if (sd_poll_status(FIFO_EMPTY, 0x80000))
+            {
+                copy_write_sectors(&buf); /* Copy one chunk of 16 words */
+                /* TODO: Switch bank if necessary */
+                continue;
+            }
+
+            ret = -EC_FIFO_WR_EMPTY;
+            goto ata_write_error;
+        }
     }
 
     last_disk_activity = current_tick;