diff --git a/include/ultra64/bcp.h b/include/ultra64/bcp.h
index 78c4584fe7..f66fee2f02 100644
--- a/include/ultra64/bcp.h
+++ b/include/ultra64/bcp.h
@@ -190,6 +190,16 @@
  */
 #define PI_NAND_CTRL_REG (PI_BASE_REG + 0x48)
 
+/**
+ * PI internal buffer DMA read length. Writes initiate a DMA from RDRAM to the PI buffer.
+ */
+#define PI_EX_RD_LEN_REG (PI_BASE_REG + 0x58)
+
+/**
+ * PI internal buffer DMA write length. Writes initiate a DMA from the PI buffer to RDRAM.
+ */
+#define PI_EX_WR_LEN_REG (PI_BASE_REG + 0x5C)
+
 /**
  * [31:16] Box ID
  *   [31:30] Hardware Revision? (osInitialize checks this and sets __osBbIsBb to 2 if != 0)
diff --git a/src/libultra/io/devmgr.c b/src/libultra/io/devmgr.c
index d0e7d5289a..05dcb99519 100644
--- a/src/libultra/io/devmgr.c
+++ b/src/libultra/io/devmgr.c
@@ -2,6 +2,8 @@
 #include "ultra64/internal.h"
 #include "ultra64/leodrive.h"
 
+#define TEMP_BUFFER ((void*)0x80600000)
+
 // os.h
 #define LEO_BLOCK_MODE 1
 #define LEO_TRACK_MODE 2
@@ -14,6 +16,9 @@ void __osDevMgrMain(void* arg) {
     s32 ret;
     OSDevMgr* dm = (OSDevMgr*)arg;
     s32 messageSend;
+#ifdef BBPLAYER
+    s32 check = false;
+#endif
 
     while (true) {
         osRecvMesg(dm->cmdQueue, (OSMesg*)&ioMesg, OS_MESG_BLOCK);
@@ -78,6 +83,13 @@ void __osDevMgrMain(void* arg) {
             switch (ioMesg->hdr.type) {
                 case OS_MESG_TYPE_DMAREAD:
                     osRecvMesg(dm->acsQueue, &dummy, OS_MESG_BLOCK);
+#ifdef BBPLAYER
+                    if (__osBbIsBb == 1 && ((u32)ioMesg->dramAddr & 0x7F) >= 0x60) {
+                        check = true;
+                        ret = dm->dma(OS_READ, ioMesg->devAddr, TEMP_BUFFER, ioMesg->size);
+                        break;
+                    }
+#endif
                     ret = dm->dma(OS_READ, ioMesg->devAddr, ioMesg->dramAddr, ioMesg->size);
                     break;
                 case OS_MESG_TYPE_DMAWRITE:
@@ -86,6 +98,13 @@ void __osDevMgrMain(void* arg) {
                     break;
                 case OS_MESG_TYPE_EDMAREAD:
                     osRecvMesg(dm->acsQueue, &dummy, OS_MESG_BLOCK);
+#ifdef BBPLAYER
+                    if (__osBbIsBb == 1 && ((u32)ioMesg->dramAddr & 0x7F) >= 0x60) {
+                        check = true;
+                        ret = dm->edma(ioMesg->piHandle, OS_READ, ioMesg->devAddr, TEMP_BUFFER, ioMesg->size);
+                        break;
+                    }
+#endif
                     ret = dm->edma(ioMesg->piHandle, OS_READ, ioMesg->devAddr, ioMesg->dramAddr, ioMesg->size);
                     break;
                 case OS_MESG_TYPE_EDMAWRITE:
@@ -103,6 +122,14 @@ void __osDevMgrMain(void* arg) {
 
             if (ret == 0) {
                 osRecvMesg(dm->evtQueue, &em, OS_MESG_BLOCK);
+#ifdef BBPLAYER
+                if (__osBbIsBb == 1 && check) {
+                    osInvalDCache(TEMP_BUFFER, (ioMesg->size + DCACHE_LINEMASK) & ~DCACHE_LINEMASK);
+                    bcopy(TEMP_BUFFER, ioMesg->dramAddr, ioMesg->size);
+                    check = false;
+                    osWritebackDCache(ioMesg->dramAddr, ioMesg->size);
+                }
+#endif
                 osSendMesg(ioMesg->hdr.retQueue, (OSMesg)ioMesg, OS_MESG_NOBLOCK);
                 osSendMesg(dm->acsQueue, NULL, OS_MESG_NOBLOCK);
             }
diff --git a/src/libultra/io/epirawdma.c b/src/libultra/io/epirawdma.c
index ff6bb9a653..dbc072fbb2 100644
--- a/src/libultra/io/epirawdma.c
+++ b/src/libultra/io/epirawdma.c
@@ -1,8 +1,18 @@
 #include "global.h"
+#include "ultra64/bcp.h"
 
 s32 __osEPiRawStartDma(OSPiHandle* handle, s32 direction, u32 cartAddr, void* dramAddr, size_t size) {
-    s32 status;
+#ifdef BBPLAYER
+    u64 dummybuf[2];
+#endif
+    u32 status;
     OSPiHandle* curHandle;
+#ifdef BBPLAYER
+    u32 buffer;
+    u32 pgsize;
+    u16* adr;
+    u32 i;
+#endif
 
     status = IO_READ(PI_STATUS_REG);
     while (status & (PI_STATUS_DMA_BUSY | PI_STATUS_IO_BUSY)) {
@@ -53,9 +63,102 @@ s32 __osEPiRawStartDma(OSPiHandle* handle, s32 direction, u32 cartAddr, void* dr
         curHandle->pulse = handle->pulse;
     }
 
+#ifdef BBPLAYER
+    if (direction == OS_READ) {
+        // Device page size in bytes
+        pgsize = 1;
+        for (i = 1; i <= (u32)handle->pageSize + 2; i++) {
+            pgsize *= 2;
+        }
+
+        // If the initial cart address mod pgsize is at the last u16 in the page,
+        // need to do some manual DMAs?
+        if ((cartAddr & (pgsize - 1)) == pgsize - sizeof(u16)) {
+            // Read 32 bits starting 2 bytes before the target DMA address,
+            // so that the lower 16 bits of the result are the first 2 bytes
+            // of the requested data.
+            __osEPiRawReadIo(handle, cartAddr - sizeof(u16), &buffer);
+
+            // Poke the lower 16 bits into the destination address
+            adr = (u16*)PHYS_TO_K1(dramAddr);
+            *(adr++) = (u16)buffer;
+
+            // Update DMA parameters
+            cartAddr += sizeof(u16);
+            dramAddr = adr;
+            size -= sizeof(u16);
+
+            // If the remaining size is >= 4
+            if (size >= sizeof(u32)) {
+                // Read another 32 bits at the cart addr
+                __osEPiRawReadIo(handle, cartAddr, &buffer);
+
+                // Store all 32 bits to RAM
+                adr = (u16*)dramAddr;
+                *(adr++) = buffer >> 16;
+                *(adr++) = (u16)buffer;
+
+                // Update DMA parameters again
+                cartAddr += sizeof(u32);
+                dramAddr = adr;
+                size -= sizeof(u32);
+
+                // If we're not at the end of the DMA
+                if (size != 0) {
+                    // Read 32 bits again
+                    __osEPiRawReadIo(handle, cartAddr, &buffer);
+
+                    // Store just the upper 16 bits
+                    adr = (u16*)PHYS_TO_K1(dramAddr);
+                    *(adr++) = buffer >> 16;
+
+                    // Update DMA parameters once more
+                    cartAddr += sizeof(u16);
+                    dramAddr = adr;
+                    size -= sizeof(u16);
+                }
+            }
+        }
+
+        // If the end cart address mod pgsize is just 2 bytes into a page or the remaining data size is just 1x u16
+        if (((((cartAddr + size) & (pgsize - 1)) == sizeof(u16)) | (size == sizeof(u16))) != 0) {
+            if ((cartAddr + size) & 2) {
+                // Read 32 bits at end - 2, store the upper 16 bits
+                __osEPiRawReadIo(handle, cartAddr + size - sizeof(u16), &buffer);
+                adr = (u16*)PHYS_TO_K1(dramAddr) + (size - sizeof(u16)) / sizeof(u16);
+                *adr = buffer >> 16;
+            } else {
+                // Read 32 bits at end - 4, store the lower 16 bits
+                __osEPiRawReadIo(handle, cartAddr + size - sizeof(u32), &buffer);
+                adr = (u16*)PHYS_TO_K1(dramAddr) + (size - sizeof(u16)) / sizeof(u16);
+                *adr = (u16)buffer;
+            }
+            size -= sizeof(u16);
+        }
+
+        if (size == 0) {
+            // If size ended up 0 following the adjustments, run an 8-byte dummy DMA anyway
+            size = 8;
+            dramAddr = (void*)dummybuf;
+            cartAddr = 0;
+        }
+    }
+#endif
+
     IO_WRITE(PI_DRAM_ADDR_REG, osVirtualToPhysical(dramAddr));
     IO_WRITE(PI_CART_ADDR_REG, K1_TO_PHYS(handle->baseAddress | cartAddr));
 
+#ifdef BBPLAYER
+    if (direction != OS_READ && direction != OS_WRITE) {
+        return -1;
+    }
+
+    if ((handle->baseAddress | cartAddr) <= 0x400) {
+        IO_WRITE((direction == OS_READ) ? PI_EX_WR_LEN_REG : PI_EX_RD_LEN_REG, size - 1);
+    } else {
+        IO_WRITE((direction == OS_READ) ? PI_WR_LEN_REG : PI_RD_LEN_REG, size - 1);
+    }
+#else
     switch (direction) {
         case OS_READ:
             IO_WRITE(PI_WR_LEN_REG, size - 1);
@@ -66,5 +169,6 @@ s32 __osEPiRawStartDma(OSPiHandle* handle, s32 direction, u32 cartAddr, void* dr
         default:
             return -1;
     }
+#endif
     return 0;
 }
diff --git a/src/libultra/io/pimgr.c b/src/libultra/io/pimgr.c
index e8aa98d01b..7a0b68f8dc 100644
--- a/src/libultra/io/pimgr.c
+++ b/src/libultra/io/pimgr.c
@@ -40,8 +40,8 @@ void osCreatePiManager(OSPri pri, OSMesgQueue* cmdQueue, OSMesg* cmdBuf, s32 cmd
         prevInt = __osDisableInt();
 
         __osPiDevMgr.active = true;
-        __osPiDevMgr.cmdQueue = cmdQueue;
         __osPiDevMgr.thread = &piThread;
+        __osPiDevMgr.cmdQueue = cmdQueue;
         __osPiDevMgr.evtQueue = &piEventQueue;
         __osPiDevMgr.acsQueue = &__osPiAccessQueue;
         __osPiDevMgr.dma = __osPiRawStartDma;