diff --git a/TODO b/TODO
index a382c3e..ac8b04e 100644
--- a/TODO
+++ b/TODO
@@ -7,7 +7,7 @@ For the 2.0.0 release:
 Crucial, missing things.
 ! core:      the four basic memory functions (with automated tests)
 ! core:      use gint_switch() to handle TLB misses
-! bfile:     solve stability issues
+! core:      recheck SH3 compatibility
 
 Issues.
 * #3         make drawing functions parameterized
@@ -16,6 +16,7 @@ Issues.
 * #10        support fx-CG 20
 
 Complementary elements on existing code.
+* gray:      add gprint()
 * gray:      double-buffer gray settings and unify d* with g*
 * display:   deprecate image_t and rename it bopti_image_t
 * topti:     support unicode fonts
diff --git a/src/core/setup.c b/src/core/setup.c
index b46b5f0..052fefe 100644
--- a/src/core/setup.c
+++ b/src/core/setup.c
@@ -162,18 +162,18 @@ static void gint_switch_out(void)
 	gint_ctx_save(&gint_ctx);
 
 	/* Restore the system context */
-	for(gint_driver_t *drv = &edrv; (--drv) >= &bdrv;)
+	gint_ctx_restore(&sys_ctx);
+	for(gint_driver_t *drv = &bdrv; drv < &edrv; drv++)
 	{
 		if(!drv->ctx_save || !drv->ctx_restore) continue;
 		drv->ctx_restore(drv->sys_ctx);
 	}
-	gint_ctx_restore(&sys_ctx);
 }
 
 static void gint_switch_in(void)
 {
 	/* Save system state again */
-	for(gint_driver_t *drv = &bdrv; drv < &edrv; drv++)
+	for(gint_driver_t *drv = &edrv; (--drv) >= &bdrv;)
 	{
 		if(!drv->ctx_save || !drv->ctx_restore) continue;
 		drv->ctx_save(drv->sys_ctx);
@@ -181,12 +181,12 @@ static void gint_switch_in(void)
 	gint_ctx_save(&sys_ctx);
 
 	/* Restore all drivers to their gint state */
+	gint_ctx_restore(&gint_ctx);
 	for(gint_driver_t *drv = &bdrv; drv < &edrv; drv++)
 	{
 		if(!drv->ctx_save || !drv->ctx_restore) continue;
 		drv->ctx_restore(drv->gint_ctx);
 	}
-	gint_ctx_restore(&gint_ctx);
 }
 
 /* gint_switch() - temporarily switch out of gint */
@@ -200,6 +200,24 @@ void gint_switch(void (*function)(void))
 
 	gint_setvbr(system_vbr, gint_switch_out);
 	if(function) function();
+
+	/* Remap all of the ROM */
+	extern uint32_t brom, srom;
+	volatile void *b = &brom;
+	int32_t s = (int32_t)&srom;
+	GUNUSED uint8_t x;
+	while(s > 0)
+	{
+		x = *(volatile uint8_t *)b;
+		b += 1024;
+		s -= 1024;
+	}
+
+	/* Wait for the OS to finish working */
+	for(gint_driver_t *drv = &edrv; (--drv) >= &bdrv;)
+	{
+		if(drv->wait) drv->wait();
+	}
 	gint_setvbr((uint32_t)&gint_vbr, gint_switch_in);
 }
 
diff --git a/src/core/start.c b/src/core/start.c
index 09b330a..2936d45 100644
--- a/src/core/start.c
+++ b/src/core/start.c
@@ -156,6 +156,9 @@ int start(int isappli, int optnum)
 	/* Initialize all drivers by saving the system settings */
 	for(drv = &bdrv; drv < &edrv; drv++)
 	{
+		/* Wait for hardware availability */
+		if(drv->wait) drv->wait();
+
 		/* Hook for old SH3 fx9860g machines */
 		if(isSH3() && drv->driver_sh3) drv->driver_sh3();
 
diff --git a/src/core/syscalls.S b/src/core/syscalls.S
index 5db34e6..fefc997 100644
--- a/src/core/syscalls.S
+++ b/src/core/syscalls.S
@@ -45,6 +45,7 @@
 	mov.l	1f, r0			;\
 	jmp	@r2			;\
 	nop				;\
+.align 4				;\
 1:	.long	id
 
 #ifdef FX9860G
@@ -123,10 +124,12 @@ _realloc:
 /* BFile driver */
 
 _BFile_Remove:
+	mov #0, r5
 	syscall(0x1db4)
 _BFile_Create:
 	syscall(0x1dae)
 _BFile_Open:
+	mov #0, r6
 	syscall(0x1da3)
 _BFile_Close:
 	syscall(0x1da4)
diff --git a/src/dma/dma.c b/src/dma/dma.c
index 263e53e..e4a480f 100644
--- a/src/dma/dma.c
+++ b/src/dma/dma.c
@@ -138,9 +138,10 @@ void dma_transfer_wait(int channel)
 	if(ch->SAR >= 0xe5007000 && ch->SAR < 0xe5204000) onchip = 1;
 	if(ch->DAR >= 0xe5007000 && ch->DAR < 0xe5204000) onchip = 1;
 
-	while(ch->CHCR.DE)
+	while(ch->CHCR.DE && !ch->CHCR.TE)
 	{
-		if(!onchip) sleep();
+		/* Sleep only if the interrupt is enabled to wake us up */
+		if(!onchip && ch->CHCR.IE) sleep();
 	}
 }
 
@@ -228,8 +229,8 @@ static void wait(void)
 typedef struct
 {
 	channel_t ch[6];
-	uint16_t OR;
 	int clock;
+	uint16_t OR;
 
 } GPACKED(4) ctx_t;
 
@@ -257,8 +258,6 @@ static void ctx_save(void *buf)
 
 static void ctx_restore(void *buf)
 {
-	dma_transfer_wait(-1);
-
 	ctx_t *ctx = buf;
 
 	/* Restore the supply status of the DMA0 clock first. If the DMA was
@@ -267,6 +266,9 @@ static void ctx_restore(void *buf)
 	   power it up again for the writes to registers to have any effect */
 	POWER.MSTPCR0.DMAC0 = ctx->clock;
 
+	/* Disable the DMA while editing */
+	DMA.OR.DME = 0;
+
 	for(int i = 0; i < 6; i++)
 	{
 		channel_t *ch = dma_channel(i);
@@ -275,7 +277,6 @@ static void ctx_restore(void *buf)
 		ch->TCR        = ctx->ch[i].TCR;
 		ch->CHCR.lword = ctx->ch[i].CHCR.lword;
 	}
-
 	DMA.OR.word = ctx->OR;
 }