diff --git a/sw/bootloader/src/exception.S b/sw/bootloader/src/exception.S
index daca4c0..0919434 100644
--- a/sw/bootloader/src/exception.S
+++ b/sw/bootloader/src/exception.S
@@ -175,13 +175,19 @@ exception_install:
     add $t5, $t5, $t4
 2:
     lw $t6, 0($t3)
-    sw $t6, 0($t4)
-    cache HIT_INVALIDATE_I, 0($t4)
     addiu $t3, 4
+    sw $t6, 0($t4)
     addiu $t4, 4
     bne $t4, $t5, 2b
     addiu $t1, VECTOR_SIZE
     bne $t1, $t2, 1b
+    li $t0, VECTOR_LOCATION
+    li $t1, (VECTOR_SIZE * VECTOR_NUM)
+    add $t1, $t0, $t1
+3:
+    cache HIT_INVALIDATE_I, 0($t0)
+    addiu $t0, CACHE_LINE_SIZE_I
+    bne $t0, $t1, 3b
     jr $ra
 
 
diff --git a/sw/bootloader/src/io.c b/sw/bootloader/src/io.c
index 39bc4ef..d3fa90a 100644
--- a/sw/bootloader/src/io.c
+++ b/sw/bootloader/src/io.c
@@ -1,4 +1,5 @@
 #include "io.h"
+#include "vr4300.h"
 
 
 static void cache_operation (uint8_t operation, uint8_t line_size, void *address, size_t length) {
@@ -13,16 +14,16 @@ static void cache_operation (uint8_t operation, uint8_t line_size, void *address
     }    
 }
 
-void cache_data_hit_invalidate (void *address, size_t length) {
-    cache_operation(0x11, 16, address, length);
+void cache_data_hit_writeback_invalidate (void *address, size_t length) {
+    cache_operation(HIT_WRITE_BACK_INVALIDATE_D, CACHE_LINE_SIZE_D, address, length);
 }
 
 void cache_data_hit_writeback (void *address, size_t length) {
-    cache_operation(0x19, 16, address, length);
+    cache_operation(HIT_WRITE_BACK_D, CACHE_LINE_SIZE_D, address, length);
 }
 
 void cache_inst_hit_invalidate (void *address, size_t length) {
-    cache_operation(0x10, 32, address, length);
+    cache_operation(HIT_INVALIDATE_I, CACHE_LINE_SIZE_I, address, length);
 }
 
 uint32_t io_read (io32_t *address) {
@@ -50,11 +51,11 @@ void pi_io_write (io32_t *address, uint32_t value) {
 }
 
 void pi_dma_read (io32_t *address, void *buffer, size_t length) {
+    cache_data_hit_writeback_invalidate(buffer, length);
     io_write(&PI->PADDR, (uint32_t) (PHYSICAL(address)));
     io_write(&PI->MADDR, (uint32_t) (PHYSICAL(buffer)));
     io_write(&PI->WDMA, length - 1);
     while (pi_busy());
-    cache_data_hit_invalidate(buffer, length);
 }
 
 void pi_dma_write (io32_t *address, void *buffer, size_t length) {
diff --git a/sw/bootloader/src/io.h b/sw/bootloader/src/io.h
index c3e11e5..03d3a18 100644
--- a/sw/bootloader/src/io.h
+++ b/sw/bootloader/src/io.h
@@ -286,7 +286,7 @@ void pi_dma_write (io32_t *address, void *buffer, size_t length);
 uint32_t si_busy (void);
 uint32_t si_io_read (io32_t *address);
 void si_io_write (io32_t *address, uint32_t value);
-void cache_data_hit_invalidate (void *address, size_t length);
+void cache_data_hit_writeback_invalidate (void *address, size_t length);
 void cache_data_hit_writeback (void *address, size_t length);
 void cache_inst_hit_invalidate (void *address, size_t length);
 
diff --git a/sw/bootloader/src/startup.S b/sw/bootloader/src/startup.S
index 04f640b..2b2fe03 100644
--- a/sw/bootloader/src/startup.S
+++ b/sw/bootloader/src/startup.S
@@ -40,7 +40,6 @@ bss_init:
 	la $a1, _ebss
 1:
 	sd $zero, 0($a0)
-    cache HIT_WRITE_BACK_D, 0($a0)
 	addiu $a0, 8
 	bltu $a0, $a1, 1b
 
diff --git a/sw/bootloader/src/vr4300.h b/sw/bootloader/src/vr4300.h
index e25ba45..58b3617 100644
--- a/sw/bootloader/src/vr4300.h
+++ b/sw/bootloader/src/vr4300.h
@@ -3,8 +3,11 @@
 
 
 #define HIT_INVALIDATE_I            ((4 << 2) | 0)
+#define HIT_WRITE_BACK_INVALIDATE_D ((5 << 2) | 1)
 #define HIT_WRITE_BACK_D            ((6 << 2) | 1)
 
+#define CACHE_LINE_SIZE_I           (32)
+#define CACHE_LINE_SIZE_D           (16)
 
 #define C0_BADVADDR                 $8
 #define C0_COUNT                    $9