diff --git a/src/cryptoengine.c b/src/cryptoengine.c
index 6fad831..5d2415c 100644
--- a/src/cryptoengine.c
+++ b/src/cryptoengine.c
@@ -41,6 +41,20 @@ static ecc_key ecc;
 static struct vault_service svc_cache;
 static int svc_cache_id = -1;
 
+extern const uint8_t disk1_mbr[];
+
+#define CHACHA_BLOCK_SIZE 64
+#define CHACHA_BUFFER_SIZE (CHACHA_BLOCK_SIZE * 16)
+#define LBA_SIZE 512
+
+#define CHACHA_INLINE
+
+
+#ifndef CHACHA_INLINE
+static uint8_t chacha_tmp[CHACHA_BUFFER_SIZE];
+static uint32_t chacha_processed;
+#endif
+
 int flash_decrypt_read_sector(uint32_t sector, uint8_t *buf)
 {
     if (cryptoengine_check_vault() < 0)
@@ -51,8 +65,21 @@ int flash_decrypt_read_sector(uint32_t sector, uint8_t *buf)
         return -1;
     flash_read(DRIVE_FLASH_OFFSET + sector * SPI_FLASH_SECTOR_SIZE, buf,
             SPI_FLASH_SECTOR_SIZE);
-    wc_Chacha_SetIV(&cha, hdr_cache.host_seed, sector);
+    wc_Chacha_SetIV(&cha, hdr_cache.host_seed, sector * (SPI_FLASH_SECTOR_SIZE / CHACHA_BLOCK_SIZE));
+#ifndef CHACHA_INLINE
+    chacha_processed = 0;
+    while (chacha_processed < SPI_FLASH_SECTOR_SIZE) {
+        wc_Chacha_Process(&cha, chacha_tmp, buf + chacha_processed,
+                CHACHA_BUFFER_SIZE);
+        memcpy(buf + chacha_processed, chacha_tmp, CHACHA_BUFFER_SIZE);
+        chacha_processed += CHACHA_BUFFER_SIZE;
+    }
+#else
     wc_Chacha_Process(&cha, buf, buf, SPI_FLASH_SECTOR_SIZE);
+#endif
+    if (sector == 0) {
+        memcpy(buf, disk1_mbr, LBA_SIZE);
+    }
     return 0;
 }
 
@@ -64,8 +91,21 @@ int flash_encrypt_write_sector(uint32_t sector, uint8_t *buf)
         return -1;
     if ((sector * SPI_FLASH_SECTOR_SIZE) >= DRIVE_FLASH_SIZE)
         return -1;
-    wc_Chacha_SetIV(&cha, hdr_cache.host_seed, sector);
+    if (sector == 0) {
+        memcpy(buf, disk1_mbr, LBA_SIZE);
+    }
+    wc_Chacha_SetIV(&cha, hdr_cache.host_seed, sector * (SPI_FLASH_SECTOR_SIZE / CHACHA_BLOCK_SIZE));
+#ifndef CHACHA_INLINE
+    chacha_processed = 0;
+    while (chacha_processed < SPI_FLASH_SECTOR_SIZE) {
+        wc_Chacha_Process(&cha, chacha_tmp, buf + chacha_processed,
+                CHACHA_BUFFER_SIZE);
+        memcpy(buf + chacha_processed, chacha_tmp, CHACHA_BUFFER_SIZE);
+        chacha_processed += CHACHA_BUFFER_SIZE;
+    }
+#else
     wc_Chacha_Process(&cha, buf, buf, SPI_FLASH_SECTOR_SIZE);
+#endif
     flash_write(DRIVE_FLASH_OFFSET + sector * SPI_FLASH_SECTOR_SIZE, buf,
             SPI_FLASH_SECTOR_SIZE);
     return 0;
diff --git a/src/hid.c b/src/hid.c
index aa10fc1..53c0971 100644
--- a/src/hid.c
+++ b/src/hid.c
@@ -195,6 +195,8 @@ static void send_hid_report(uint8_t report_id, uint32_t unused)
     }
 }
 
+extern void ForceZero(const void* mem, word32 len);
+
 void hid_keys_string_send(const char *str)
 {
     int len = strlen(str);
diff --git a/src/msc_disk.c b/src/msc_disk.c
index 76a9879..444b199 100644
--- a/src/msc_disk.c
+++ b/src/msc_disk.c
@@ -45,12 +45,12 @@ struct disk_lba {
 
 static const uint8_t last_cluster_init_info[4] = { 0xF8, 0xFF, 0xFF, 0xFF };
 
-static const uint8_t disk1_mbr[DISK_BLOCK_SIZE] = {
+const uint8_t disk1_mbr[DISK_BLOCK_SIZE] = {
     0xeb, 0x3c, 0x90, 0x6d, 0x6b, 0x66, 0x73, 0x2e, 0x66, 0x61, 0x74, 0x00,
     0x02, 0x04, 0x01, 0x00, 0x02, 0x00, 0x02, 0x00, 0x04, 0xf8, 0x01, 0x00,
     0x10, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x80, 0x00, 0x29, 0x9d, 0x0b, 0x6e, 0xd0, 0x4e, 0x4f, 0x20, 0x4e, 0x41,
-    0x4d, 0x45, 0x20, 0x20, 0x20, 0x20, 0x46, 0x41, 0x54, 0x31, 0x32, 0x20,
+    0x80, 0x00, 0x29, 0x9d, 0x0b, 0x6e, 0xd0, 'M',  'E',  'P',  ' ',  'V',
+    'a',  'u',  'l',  't',  0x20, 0x20, 0x46, 0x41, 0x54, 0x31, 0x32, 0x20,
     0x20, 0x20, 0x0e, 0x1f, 0xbe, 0x5b, 0x7c, 0xac, 0x22, 0xc0, 0x74, 0x0b,
     0x56, 0xb4, 0x0e, 0xbb, 0x07, 0x00, 0xcd, 0x10, 0x5e, 0xeb, 0xf0, 0x32,
     0xe4, 0xcd, 0x16, 0xcd, 0x19, 0xeb, 0xfe, 0x54, 0x68, 0x69, 0x73, 0x20,
@@ -107,14 +107,13 @@ static int sector_cache_load(uint32_t sector)
         sector_cached = -1;
         return -1;
     }
-    //ret = flash_decrypt_read_sector(sector, sector_cache[0].data);
-    ret = flash_read(sector * SPI_FLASH_SECTOR_SIZE + DRIVE_FLASH_OFFSET,
-            sector_cache[0].data, SPI_FLASH_SECTOR_SIZE);
+    ret = flash_decrypt_read_sector(sector, sector_cache[0].data);
     if (ret < 0) {
         sector_cached = -1;
         return ret;
     }
     sector_cached = sector;
+    return sector;
 }
 
 static inline void sector_cache_sched_commit(void)
@@ -130,10 +129,8 @@ static void sector_cache_commit(void)
         cache_busy = 0;
         return;
     }
-    //flash_encrypt_write_sector(sector, sector_cache[0].data);
     flash_sector_erase(sector * SPI_FLASH_SECTOR_SIZE + DRIVE_FLASH_OFFSET);
-    flash_write(sector * SPI_FLASH_SECTOR_SIZE + DRIVE_FLASH_OFFSET, 
-            sector_cache[0].data, SPI_FLASH_SECTOR_SIZE);
+    flash_encrypt_write_sector(sector, sector_cache[0].data);
     cache_busy = 0;
 }
 
@@ -163,7 +160,11 @@ static const struct drive Drives[N_DRIVES] = {
 
 uint8_t tud_msc_get_maxlun_cb(void)
 {
+#ifdef HAVE_VAULT
   return 2;
+#else
+  return 1;
+#endif
 }
 
 // Invoked when received SCSI_CMD_INQUIRY
@@ -240,6 +241,7 @@ int32_t tud_msc_read10_cb(uint8_t lun, uint32_t lba, uint32_t offset, void* buff
 
     const struct drive *drv;
     uint32_t sector, sec_off;
+    uint32_t remaining_bufsize = 0, total_bufsize = bufsize;
     if (lun >= N_DRIVES)
         return -1;
     drv = &Drives[lun];
@@ -257,21 +259,37 @@ int32_t tud_msc_read10_cb(uint8_t lun, uint32_t lba, uint32_t offset, void* buff
             memset(buffer, 0, bufsize);
             return bufsize;
         }
-        if (lba == 0) {
-            memcpy(buffer, disk1_mbr + offset, bufsize);
-            return bufsize;
-        }
         sector = lba / BLOCKS_PER_SECTOR;
         sec_off = lba - (sector * BLOCKS_PER_SECTOR);
          
-        if ( sector != sector_cached) {
-            if (!cache_busy)
-                sector_cache_load(sector);
-            return 0;
-        }
-        memcpy(buffer, sector_cache[sec_off].data + offset, bufsize);
-        asm volatile("DMB");
-        return bufsize;
+        do {
+            if ( sector != sector_cached) {
+                if (!cache_busy) {
+                    if (sector_cache_load(sector) != sector)
+                        return -1;
+                } else {
+                    if (total_bufsize == bufsize)
+                        return -1;
+                    return total_bufsize - bufsize;
+                }
+            }
+            if (sec_off * DISK_BLOCK_SIZE + offset + bufsize > SPI_FLASH_SECTOR_SIZE) {
+                remaining_bufsize = bufsize;
+                bufsize = SPI_FLASH_SECTOR_SIZE - (sec_off * DISK_BLOCK_SIZE + offset);
+                remaining_bufsize -= bufsize;
+            }
+            memcpy(buffer, sector_cache[sec_off].data + offset, bufsize);
+            asm volatile("DMB");
+            if (remaining_bufsize > 0) {
+                sector++;
+                sec_off = 0;
+                bufsize = remaining_bufsize;
+                remaining_bufsize = 0;
+            } else {
+                break;
+            }
+        } while (bufsize > 0);
+        return total_bufsize;
     }
 }
 
@@ -290,27 +308,49 @@ bool tud_msc_is_writable_cb (uint8_t lun)
 int32_t tud_msc_write10_cb(uint8_t lun, uint32_t lba, uint32_t offset, uint8_t* buffer, uint32_t bufsize)
 {
     uint32_t sector, sec_off;
+    uint32_t remaining_bufsize = 0;
+    uint32_t total_bufsize = bufsize;
     if (lun == 0)
         return bufsize;
-    /* 
-    if (lba == 0)
-        return bufsize;
-    */
     if ( lba >= Drives[lun].n_lba)
         return -1;
     if (fsm_get() < VAULT_MAIN_MENU)
         return -1;
     sector = lba / BLOCKS_PER_SECTOR;
     sec_off = lba - (sector * BLOCKS_PER_SECTOR);
-    if ( sector != sector_cached) {
-        while(cache_busy)
-            sleep_ms(1);
-        sector_cache_load(sector);
-    }
-    memcpy(sector_cache[sec_off].data + offset, buffer, bufsize);
-    sector_cache_sched_commit();
+    do {
+        if ( sector != sector_cached) {
+            while(cache_busy)
+                sleep_ms(1);
+            if (sector_cache_load(sector) < 0)
+                return total_bufsize;
+        }
+        if (bufsize + sec_off * DISK_BLOCK_SIZE + offset > SPI_FLASH_SECTOR_SIZE) {
+            remaining_bufsize = bufsize;
+            bufsize = SPI_FLASH_SECTOR_SIZE - (sec_off * DISK_BLOCK_SIZE + offset);
+            remaining_bufsize -= bufsize;
+        } else {
+            remaining_bufsize = 0;
+        }
+        if (bufsize == 0)
+            return total_bufsize;
+        memcpy(sector_cache[sec_off].data + offset, buffer, bufsize);
+        sector_cache_sched_commit();
+        if (remaining_bufsize > 0) {
+            return total_bufsize;
+            /*
+            sector++;
+            sec_off = 0;
+            bufsize = remaining_bufsize;
+            remaining_bufsize = 0;
+            */
+        } else {
+            bufsize = 0;
+            break;
+        }
+    } while (bufsize > 0);
     asm volatile("DMB");
-    return bufsize;
+    return total_bufsize;
 }
 
 // Callback invoked when received an SCSI command not in built-in list below
@@ -369,7 +409,8 @@ void disk1_format(void)
     memset(block, 0, DISK_BLOCK_SIZE);
     for (i = 1; i < 5; i++) {
         if (sector_cached != i) {
-            sector_cache_load(i);
+            if (sector_cache_load(i) < 0)
+                break;
         }
         for (j = 0; j < BLOCKS_PER_SECTOR; j++)
             memcpy(sector_cache[j].data, block, DISK_BLOCK_SIZE);