usb: massively improve writing logic

* Move logic around tracking transfers to asyncio.c.

* Add a "short buffer" holding 0-3 bytes between writes, so that the
  driver performs only 4-byte writes in the FIFO and a short write in
  the commit, if needed.
  - This is partially due to me thinking at some point that degrading
    writing size was impossible, but it might actually be possible by
    writing to FIFO/FIFO+2 or FIFO/FIFO+1/FIFO+2/FIFO+3.
  - In any case I think this new approach wins on performance.

* Get rid of unit_size since we now always use 4 bytes.

* Add a waiting function which is used in usb_close() (and once tested
  should be used in world switches too).

* Eliminate some of the special cases for the DCP, though not all (in
  particular I can't get the commit to rely on the BEMP interrupt yet,
  nor can I properly clear PID to NAK when unbinding).

CMakeLists.txt

@@ -123,12 +123,14 @@ set(SOURCES_COMMON
   src/tmu/sleep.c
   src/tmu/tmu.c
   # USB driver
+  src/usb/asyncio.c
   src/usb/classes/ff-bulk.c
   src/usb/configure.c
   src/usb/pipes.c
   src/usb/setup.c
   src/usb/string.c
   src/usb/usb.c
+  src/usb/write4.S
 )
 set(SOURCES_FX
   # Gray engine

TODO

@@ -16,6 +16,8 @@ Extensions on existing code:
 * core:      review forgotten globals and MPU addresses not in <gint/mpu/*.h>
 * core:      run destructors when a task-switch results in leaving the app
 * fs:        support read-only files backed with GetBlockAddress() on fx-CG
+* kernel:    SH4- or G-III-specific linker scripts?
+* keysc:     global shortcut SHIFT+0+EXIT for abort() as an infinite loop break
 
 Future directions:
 * Audio playback using TSWilliamson's libsnd method

include/gint/drivers/asyncio.h

@@ -11,7 +11,7 @@
 /* Data tracking the progress of a multi-part multi-round async I/O operation.
 
    * Multi-part refers to writes being constructed over several calls to
-     write(2) followed by a "commit" with sync(2) (for async file descriptors;
+     write(2) followed by a "commit" with fsync(2) (for async file descriptors;
      synchronous file descriptors are committed at every write).
    * Multi-round refers to the operation interacting multiple times with
      hardware in order to communicate the complete data.
@@ -29,16 +29,16 @@
            | DONE           write(2) | |
            | interrupt               | |
            |                         | | Data exhausted
-           |       sync(2): start    | v
+           |     fsync(2): start     | v
-     FLYING-COMMIT <------------ IN-PROGRESS
+      FLYING-SYNC <------------ IN-PROGRESS
                    transmission
 
    Initially the operation is in the IDLE state. When a write(2) is issued, it
    interacts with hardware then transitions to the IN-PROGRESS state, where it
-   remains for any subsequent write(2). A sync(2) will properly commit data to
+   remains for any subsequent write(2). A fsync(2) will properly commit data to
    the hardware, finish the operation and return to the IDLE state.
 
-   The FLYING-WRITE and FLYING-COMMIT states refer to waiting periods, after
+   The FLYING-WRITE and FLYING-SYNC states refer to waiting periods, after
    issuing hardware commands, during which hardware communicates. Usually an
    interrupt signals when hardware is ready to resume work.
 
@@ -47,6 +47,10 @@
    directly from IDLE or IN-PROGRESS, to PENDING, to IN-PROGRESS, without
    actually communicating with the outside world.
 
+   An asynchronous write(2) might return to the caller as soon as writing is
+   finished even if the operation is left in the FLYING-WRITE state, and it may
+   even return while the operation is in the WRITING state if the DMA is used.
+
    The invariants and meaning for each state are as follow:
 
    State          Characterization             Description
@@ -55,9 +59,9 @@
    PENDING        data_w && !flying_w \        Ready to write pending data
                     && round_size == 0
    WRITING        round_size > 0               CPU/DMA write to HW in progress
-   FLYING-WRITE   flying_w && !committed_w     HW transmission in progress
+   FLYING-WRITE   flying_w && type == WRITE    HW transmission in progress
-   IN-PROGRESS    data_w != NULL && !flying_w  Waiting for write(2) or sync(2)
+   IN-PROGRESS    !data_w && type == WRITE     Waiting for write(2) or fsync(2)
-   FLYING-COMMIT  flying_w && committed_w      HW commit in progress
+   FLYING-SYNC    flying_w && type == SYNC     HW commit in progress
    ============================================================================
 
    For a read:
@@ -90,20 +94,16 @@
 
    States can be checked and transitioned with the API functions below. */
 
-enum { ASYNCIO_NONE, ASYNCIO_READ, ASYNCIO_WRITE };
+enum { ASYNCIO_NONE, ASYNCIO_READ, ASYNCIO_WRITE, ASYNCIO_SYNC };
 
-typedef struct
+typedef volatile struct
 {
     /** User-facing information **/
 
-    /* Direction of I/O operation */
+    /* Type of I/O operation (read/write/fsync) */
     uint8_t type;
     /* Whether the DMA should be used for hardware access */
     bool dma;
-    /* Whether the data has been committed by sync(2) [write] */
-    bool committed_w;
-    /* Operation's unit size (meaning depends on hardware) */
-    uint8_t unit_size;
 
     union {
         /* Address of data to transfer, incremented gradually [write] */
@@ -124,17 +124,64 @@ typedef struct
        be asynchronous if it's using the DMA) */
     uint16_t round_size;
     /* Hardware resource being used for access (meaning depends on hardware).
-       Usually, this is assigned during hardware transactions, ie.:
+       Usually, this is assigned for the duration of hardware transaction.
-       - During a write, a controller is assigned when leaving the IDLE state
+       This value is user-managed and not modified by asyncio_op functions. */
-         and returned when re-entering the IDLE state.
-       - During a read, a controller is assigne when leaving the IDLE-EMPTY
-         state and returned when re-entering the IDLE-EMPTY state. */
     uint8_t controller;
     /* Whether a hardware operation is in progress ("flying" write states) */
+    // TODO: Do we actually set and maintain this member?!
     bool flying_w;
 
+    /** Internal information **/
+
+    /* Number of bytes in short buffer (0..3) */
+    uint8_t shbuf_size;
+    /* Short buffer */
+    uint32_t shbuf;
+
 } asyncio_op_t;
 
-/* */
+//---
+// Initialization and query functions
+//---
+
+/* asyncio_op_clear(): Initialize/clear the storage for an I/O operation */
+void asyncio_op_clear(asyncio_op_t *op);
+
+/* asyncio_op_busy(): Check whether the transfer is busy for syscalls
+
+   This function checks whether the transfer is in a state where the CPU is
+   busy wrt. starting a new syscall, ie. read(2), write(2) or fsync(2). Returns
+   true if the CPU is busy and the call has to wait, false if the call can
+   proceed immediately. */
+bool asyncio_op_busy(asyncio_op_t const *op);
+
+//---
+// State transition functions
+//---
+
+/* asyncio_op_start_write(): Start a write call */
+void asyncio_op_start_write(asyncio_op_t *op, void const *data, size_t size,
+    bool use_dma, gint_call_t const *callback);
+
+/* asyncio_op_start_sync(): Transition a write I/O operation to a fsync call */
+void asyncio_op_start_sync(asyncio_op_t *op, gint_call_t const *callback);
+
+/* asyncio_op_finish_call(): Update state after a read/write/fsync call
+
+   This function should be called when the read(2)/write(2)/fsync(2) call last
+   started on the operation has concluded, including all of the hardware
+   effects. This isn't the moment when the syscall returns, rather it is the
+   moment when it completes its work. */
+void asyncio_op_finish_call(asyncio_op_t *op);
+
+//---
+// Write call functions
+//---
+
+/* asyncio_op_start_write_round(): Start a single-block write to hardware */
+void asyncio_op_start_write_round(asyncio_op_t *op, size_t size);
+
+/* asyncio_op_finish_write_round(): Finish a write round and advance data */
+void asyncio_op_finish_write_round(asyncio_op_t *op);
 
 #endif /* GINT_USB_ASYNCIO */

include/gint/usb.h

@@ -175,23 +175,22 @@ int usb_interface_pipe(usb_interface_t const *interface, int endpoint);
 
 /* usb_write_sync(): Synchronously write to a USB pipe
 
-   This functions writes (size) bytes of (data) into the specified pipe, by
+   This functions writes (size) bytes of (data) into the specified pipe. If the
-   units of (unit_size) bytes. The unit size must be 1, 2 or 4, and both (data)
+   data fits into the pipe, this function returns right away, and the data is
-   and (size) must be multiples of the unit size. In general, you should try to
+   *not* transmitted. Otherwise, data is written until the pipe is full, at
-   use the largest possible unit size, as it will be much faster. In a sequence
+   which point it is automatically transmitted. After the transfer, this
-   of writes that concludes with a commit, all the writes must use the same
+   function resumes writing, returning only once everything is written. Even
-   unit size.
+   then the last bytes will still not have been transmitted, to allow for other
+   writes to follow. After the last write in a sequence, use usb_commit_sync()
+   or usb_commit_async() to transmit the last bytes.
 
-   If the data fits into the pipe, this function returns right away, and the
+   If (use_dma=true), the write is performed with the DMA instead of the CPU.
-   data is *not* transmitted. Otherwise, data is written until the pipe is
+   This requires at least 4-byte alignment on:
-   full, at which point it is automatically transmitted. After the transfer,
+     1. The input data;
-   this function resumes writing, returning only once everything is written.
+     2. The size of this write;
-   Even then the last bytes will still not have been transmitted, to allow for
+     3. The amount of data previously written to the pipe not yet committed.
-   other writes to follow. After the last write in a sequence, use
+   This is because using the DMA does not allow any insertion of CPU logic to
-   usb_commit_sync() or usb_commit_async() to transmit the last bytes.
+   handle unaligned stuff.
-
-   If (use_dma=true), the write is performed with the DMA instead of the CPU,
-   which is generally faster.
 
    This function will use a FIFO controller to access the pipe. The FIFO
    controller will be reserved for further writes until the contents of the
@@ -204,17 +203,15 @@ int usb_interface_pipe(usb_interface_t const *interface, int endpoint);
    waits for the ressources to become available then proceeds normally.
 
    @pipe       Pipe to write into
-   @data       Source data (unit_size-aligned)
+   @data       Source data
-   @size       Size of source (multiple of unit_size)
+   @size       Size of source
-   @unit_size  FIFO access size (must be 1, 2, or 4)
    @dma        Whether to use the DMA to perform the write
    -> Returns an error code (0 on success). */
-int usb_write_sync(int pipe, void const *data, int size, int unit_size,
+int usb_write_sync(int pipe, void const *data, int size, bool use_dma);
-	bool use_dma);
 
 /* usb_write_sync_timeout(): Synchronously write, with a timeout */
 int usb_write_sync_timeout(int pipe, void const *data, int size,
-	int unit_size, bool use_dma, timeout_t const *timeout);
+	bool use_dma, timeout_t const *timeout);
 
 /* usb_write_async(): Asynchronously write to a USB pipe
 
@@ -236,14 +233,13 @@ int usb_write_sync_timeout(int pipe, void const *data, int size,
    is idle and USB_WRITE_BUSY otherwise.
 
    @pipe       Pipe to write into
-   @data       Source data (unit_size-aligned)
+   @data       Source data
-   @size       Size of source (multiple of unit_size)
+   @size       Size of source
-   @unit_size  FIFO access size (must be 1, 2, or 4)
    @dma        Whether to use the DMA to perform the write
    @callback   Optional callback to invoke when the write completes
    -> Returns an error code (0 on success). */
-int usb_write_async(int pipe, void const *data, int size, int unit_size,
+int usb_write_async(int pipe, void const *data, int size, bool use_dma,
-	bool use_dma, gint_call_t callback);
+	gint_call_t callback);
 
 /* usb_commit_sync(): Synchronously commit a write
 

src/usb/asyncio.c

@@ -0,0 +1,70 @@
+#include <gint/drivers/asyncio.h>
+#include <string.h>
+
+void asyncio_op_clear(asyncio_op_t *op)
+{
+    memset((void *)op, 0, sizeof *op);
+}
+
+bool asyncio_op_busy(asyncio_op_t const *op)
+{
+    /* WAITING and READING states are busy */
+    if(op->type == ASYNCIO_READ)
+        return op->round_size || op->data_r != NULL;
+    /* WRITING, FLYING-WRITE, FLYING-COMMIT and PENDING states are busy */
+    if(op->type == ASYNCIO_WRITE)
+        return op->round_size || op->flying_w || op->data_w != NULL;
+
+    return false;
+}
+
+void asyncio_op_start_write(asyncio_op_t *op, void const *data, size_t size,
+    bool use_dma, gint_call_t const *callback)
+{
+    op->type = ASYNCIO_WRITE;
+    op->dma = use_dma;
+    op->data_w = data;
+    op->size = size;
+    op->callback = *callback;
+}
+
+void asyncio_op_start_write_round(asyncio_op_t *op, size_t size)
+{
+    op->round_size = size;
+}
+
+void asyncio_op_finish_write_round(asyncio_op_t *op)
+{
+    op->buffer_used += op->round_size;
+    op->data_w += op->round_size;
+    op->size -= op->round_size;
+    op->round_size = 0;
+}
+
+void asyncio_op_start_sync(asyncio_op_t *op, gint_call_t const *callback)
+{
+    if(op->type != ASYNCIO_WRITE)
+        return;
+
+    op->type = ASYNCIO_SYNC;
+    op->callback = *callback;
+}
+
+void asyncio_op_finish_call(asyncio_op_t *op)
+{
+    gint_call(op->callback);
+
+    /* Clean up the operation, unless it is a write, in which case keep
+       relevant states until the transaction finishes after a fsync(2). */
+    if(op->type == ASYNCIO_WRITE) {
+        op->dma = false;
+        op->data_w = NULL;
+        op->size = 0;
+        op->callback = GINT_CALL_NULL;
+        op->round_size = 0;
+        op->flying_w = false;
+    }
+    else {
+        asyncio_op_clear(op);
+    }
+}

src/usb/classes/ff-bulk-gray.c

@@ -22,10 +22,10 @@ static void capture_vram_gray(GUNUSED bool onscreen, char const *type)
 	subheader.pixel_format = htole32(USB_FXLINK_IMAGE_GRAY);
 
 	int pipe = usb_ff_bulk_output();
-	usb_write_sync(pipe, &header, sizeof header, 4, false);
+	usb_write_sync(pipe, &header, sizeof header, false);
-	usb_write_sync(pipe, &subheader, sizeof subheader, 4, false);
+	usb_write_sync(pipe, &subheader, sizeof subheader, false);
-	usb_write_sync(pipe, light, 1024, 4, false);
+	usb_write_sync(pipe, light, 1024, false);
-	usb_write_sync(pipe, dark, 1024, 4, false);
+	usb_write_sync(pipe, dark, 1024, false);
 	usb_commit_sync(pipe);
 }
 

src/usb/classes/ff-bulk.c

@@ -108,9 +108,9 @@ static void capture_vram(GUNUSED bool onscreen, char const *type)
 	subheader.pixel_format = htole32(format);
 
 	int pipe = usb_ff_bulk_output();
-	usb_write_sync(pipe, &header, sizeof header, 4, false);
+	usb_write_sync(pipe, &header, sizeof header, false);
-	usb_write_sync(pipe, &subheader, sizeof subheader, 4, false);
+	usb_write_sync(pipe, &subheader, sizeof subheader, false);
-	usb_write_sync(pipe, source, size, 4, false);
+	usb_write_sync(pipe, source, size, false);
 	usb_commit_sync(pipe);
 }
 
@@ -123,16 +123,12 @@ void usb_fxlink_text(char const *text, int size)
 {
 	if(size == 0) size = strlen(text);
 
-	int unit_size = 4;
-	if((uint32_t)text & 3 || size & 3) unit_size = 2;
-	if((uint32_t)text & 1 || size & 1) unit_size = 1;
-
 	usb_fxlink_header_t header;
 	usb_fxlink_fill_header(&header, "fxlink", "text", size);
 
 	int pipe = usb_ff_bulk_output();
-	usb_write_sync(pipe, &header, sizeof header, unit_size, false);
+	usb_write_sync(pipe, &header, sizeof header, false);
-	usb_write_sync(pipe, text, size, unit_size, false);
+	usb_write_sync(pipe, text, size, false);
 	usb_commit_sync(pipe);
 }
 

src/usb/pipes.c

@@ -6,7 +6,7 @@
 
 #include <string.h>
 
-#include "asyncio.h"
+#include <gint/drivers/asyncio.h>
 #include "usb_private.h"
 
 #define USB SH7305_USB
@@ -69,8 +69,9 @@ void usb_pipe_clear(int pipe)
 	USB.PIPECTR[pipe-1].ACLRM = 1;
 	USB.PIPECTR[pipe-1].ACLRM = 0;
 
-	/* Clear the sequence bit (important after a world switch since we restore
+	/* Clear the sequence bit (important after a world switch since we
-	   hardware registers but the host connection is starting from scratch!) */
+	   restore hardware registers but the connection to the hosts restarts
+	   from scratch!) */
 	USB.PIPECTR[pipe-1].SQCLR = 1;
 	usb_while(USB.PIPECTR[pipe-1].SQMON != 0);
 }
@@ -106,41 +107,30 @@ static fifo_t fifo_find_available_controller(int pipe)
 }
 
 /* fifo_bind(): Bind a FIFO to a pipe in reading or writing mode */
-static void fifo_bind(fifo_t ct, int pipe, int mode, int size)
+static void fifo_bind(fifo_t ct, int pipe, int mode)
 {
-	size = (size - (size == 4) - 1) & 3;
+	int reading = (mode == FIFO_READ);
+	int writing = (mode == FIFO_WRITE);
 
-	if(pipe == 0) {
+	/* RCNT=0 REW=0 MBW=2 BIGEND=1 ISEL=mode CURPIPE=0 */
-		if(USB.CFIFOSEL.ISEL == 1 && USB.DCPCTR.PID == 1)
+	if(ct == CF) {
-			return;
+		USB.CFIFOSEL.word = 0x0900 | (mode << 5);
-
-		if(mode == FIFO_WRITE)
-			USB.DCPCTR.PID = PID_BUF;
-		/* RCNT=0 REW=0 MBW=size BIGEND=1 ISEL=mode CURPIPE=0 */
-		USB.CFIFOSEL.word = 0x0100 | (mode << 5) | (size << 10);
 		usb_while(!USB.CFIFOCTR.FRDY || USB.CFIFOSEL.ISEL != mode);
-		return;
 	}
-
+	/* RCNT=0 REW=0 DCLRM=reading DREQE=0 MBW=2 BIGEND=1 CURPIPE=pipe */
-	__typeof__(USB.D0FIFOSEL) sel;
-	sel.RCNT = 0;
-	sel.REW = 0;
-	sel.DCLRM = (mode == FIFO_READ);
-	sel.DREQE = 0;
-	sel.MBW = size;
-	sel.BIGEND = 1;
-	sel.CURPIPE = pipe;
-
 	if(ct == D0F) {
-		USB.D0FIFOSEL.word = sel.word;
+		USB.D0FIFOSEL.word = 0x0900 | (reading << 13) | pipe;
 		usb_while(!USB.D0FIFOCTR.FRDY || USB.PIPECFG.DIR != mode);
 	}
 	if(ct == D1F) {
-		USB.D1FIFOSEL.word = sel.word;
+		USB.D1FIFOSEL.word = 0x0900 | (reading << 13) | pipe;
 		usb_while(!USB.D1FIFOCTR.FRDY || USB.PIPECFG.DIR != mode);
 	}
 
 	/* Enable USB comunication! */
+	if(pipe == 0 && writing)
+		USB.DCPCTR.PID = PID_BUF;
+	if(pipe != 0)
 		USB.PIPECTR[pipe-1].PID = PID_BUF;
 }
 
@@ -148,6 +138,12 @@ static void fifo_bind(fifo_t ct, int pipe, int mode, int size)
 static void fifo_unbind(fifo_t ct)
 {
 	int pipe = -1;
+	/* TODO: USB (DCP normalization): NAK when unbinding?
+	if(ct == CF) {
+		USB.DCPCTR.CCPL = 0;
+		USB.DCPCTR.PID = PID_NAK;
+		usb_while(!USB.DCPCTR.PBUSY);
+	} */
 	if(ct == D0F) pipe = USB.D0FIFOSEL.CURPIPE;
 	if(ct == D1F) pipe = USB.D1FIFOSEL.CURPIPE;
 	if(pipe <= 0)
@@ -171,52 +167,69 @@ static void fifo_unbind(fifo_t ct)
 // Writing operations
 //---
 
-/* Current operation waiting to be performed on each pipe. There are two
+/* The writing logic is asynchronous, which makes it sometimes hard to track.
-   possible states for a pipe's transfer data:
+   The series of call for a write I/O is zero or more usb_write_async()
-   -> Either there is a transfer going on, in which case (data != NULL),
+   followed by a usb_commit_async():
-      (size != 0), and (buffer_used) has no meaning.
-   -> Either there is no transfer going on, and (data = NULL), (size = 0).
 
-   A controller is assigned to t->controller when a write first occurs until
+     write_io ::= usb_write_async* usb_commit_async
-   the pipe is fully committed. (ct = NOF) indicates an unused pipe, while
-   (ct != NOF) indicates that stuff has been written and is waiting a commit.
 
-   Additionally, between a call to write_round() and the corresponding
+   A usb_write_async() will write to the hardware buffer as many times as it
-   finish_write(), the (round_size) attribute is set to a non-zero value
+   takes to exhaust the input, including 1 time if the hardware buffer can hold
-   indicating how many bytes are waiting for write completion. */
+   the entire input and 0 times if there is no input. Each _round_ consists of
+   a call to write_round() to copy with the CPU or start the copy with the DMA,
+   and a call to finish_round() when the copy is finished.
 
-/* Multi-round operations to be continued whenever buffers are ready */
+   If the round fills the buffer, finish_round() is triggered by the BEMP
-GBSS static asyncio_op_t volatile pipe_transfers[10];
+   interrupt after the hardware finishes transferring. Otherwise finish_round()
+   is triggered directly when writing finishes.
+
+     complete_round ::= write_round
+                        <Finish writing with CPU or DMA>
+                        <USB module auto-commits pipe>
+                        <BEMP interrupt after transmission>
+                        finish_round
+
+     partial_round ::= write_round
+                       <Finish writing with CPU or DMA>
+                       finish_round
+
+   Note that the "<Finish writing>" event is asynchronous if the DMA is used. A
+   full write will take zero or more complete rounds followed by zero or one
+   partial round before finish_call() is called:
+
+     usb_write_async ::= complete_round* partial_round? finish_call
+
+   And a commit will trigger a transmission of whatever is left in the buffer
+   (including nothing) and wait for the BEMP interrupt.
+
+     usb_commit_async ::= <Manually commit pipe>
+                          <BEMP interrut after transmission>
+                          finish_call
+
+   Most functions can execute either in the main thread or within an interrupt
+   handler. */
+GBSS static asyncio_op_t pipe_transfers[10];
 
 void usb_pipe_init_transfers(void)
 {
-	memset((void *)pipe_transfers, 0, sizeof pipe_transfers);
+	for(int i = 0; i < 10; i++)
+		asyncio_op_clear(&pipe_transfers[i]);
 }
 
-static void write_8(uint8_t const *data, int size, uint8_t volatile *FIFO)
+void usb_wait_all_transfers(bool await_long_writes)
 {
-	for(int i = 0; i < size; i++) *FIFO = data[i];
+	while(1) {
+		bool all_done = true;
+		for(int i = 0; i < 10; i++) {
+			asyncio_op_t const *t = &pipe_transfers[i];
+			all_done &= !asyncio_op_busy(t);
+			if(await_long_writes)
+				all_done &= (t->type != ASYNCIO_WRITE);
 		}
-static void write_16(uint16_t const *data, int size, uint16_t volatile *FIFO)
+		if(all_done)
-{
+			return;
-	for(int i = 0; i < size; i++) *FIFO = data[i];
+		sleep();
 	}
-static void write_32(uint32_t const *data, int size, uint32_t volatile *FIFO)
-{
-	for(int i = 0; i < size; i++) *FIFO = data[i];
-}
-
-/* Check whether a pipe is busy with a multi-round write or a transfer */
-GINLINE static bool pipe_busy(int pipe)
-{
-	/* Multi-round write still not finished */
-	if(pipe_transfers[pipe].data_w) return true;
-	/* Transfer in progress */
-	if(pipe && !USB.PIPECTR[pipe-1].BSTS) return true;
-	/* Callback for a just-finished transfer not yet called */
-	if(pipe_transfers[pipe].round_size) return true;
-	/* All good */
-	return false;
 }
 
 /* Size of a pipe's buffer area, in bytes */
@@ -229,91 +242,70 @@ static int pipe_bufsize(int pipe)
 	return (USB.PIPEBUF.BUFSIZE + 1) * 64;
 }
 
-/* finish_transfer(): Finish a multi-round write transfer
+/* This function is called when a read/write/fsync call and its associated
-
+   hardware interactions all complete. */
-   This function is called when the final round of a transfer has completed,
+static void finish_call(asyncio_op_t *t, int pipe)
-   either by the handler of the BEMP interrupt or by the usb_commit_async()
-   function if the pipe is being committed when empty. */
-static void finish_transfer(asyncio_op_t volatile *t, int pipe)
 {
-	/* Free the FIFO controller */
+	/* Unbind the USB controller used for the call, except for writes since
+	   the USB module requires us to keep it until the final commit */
+	if(t->type != ASYNCIO_WRITE) {
 		fifo_unbind(t->controller);
 		t->controller = NOF;
-
-	/* Mark the transfer as unused */
-	t->committed_w = false;
-	t->buffer_used = 0;
-
-	/* Disable the interrupt */
-	if(pipe != 0)
-		USB.BEMPENB &= ~(1 << pipe);
-
-	gint_call(t->callback);
-	USB_TRACE("finish_transfer()");
 	}
 
-/* finish_round(): Update transfer logic after a write round completes
+	/* Disable interrupts */
+	if((t->type == ASYNCIO_WRITE || t->type == ASYNCIO_SYNC) && pipe != 0)
+		USB.BEMPENB &= ~(1 << pipe);
 
-   This function is called when a write round completes, either by the handler
+	asyncio_op_finish_call(t);
-   of the BEMP interrupt if the round filled the FIFO, or by the handler of the
+	USB_TRACE("finish_call()");
-   DMA transfer or the write_round() function itself if it didn't.
+}
 
-   It the current write operation has finished with this round, this function
+/* This function is called when a round of writing has completed, including all
-   invokes the write_async callback. */
+   hardware interactions. If the FIFO got filled by the writing, this is after
-static void finish_round(asyncio_op_t volatile *t, int pipe)
+   the transmission and BEMP interrupt; otherwise this is when the CPU/DMA
+   finished writing. */
+static void finish_round(asyncio_op_t *t, int pipe)
 {
-	/* Update the pointer as a result of the newly-finished write */
+//	USB_LOG("[PIPE%d] finish_round() for %d bytes\n", pipe, t->round_size);
-	t->buffer_used += t->round_size;
+	asyncio_op_finish_write_round(t);
-	t->data_w += t->round_size;
-	t->size -= t->round_size;
-	t->round_size = 0;
 
 	/* Account for auto-transfers */
 	if(t->buffer_used == pipe_bufsize(pipe))
 		t->buffer_used = 0;
 
-	/* At the end, free the FIFO and invoke the callback. Hold the
-	   controller until the pipe is committed */
-	if(t->size == 0) {
-		t->data_w = NULL;
-		gint_call(t->callback);
-	}
-
 	USB_TRACE("finish_round()");
-}
 
+	if(t->size == 0)
+		finish_call(t, pipe);
+}
 /* write_round(): Write up to a FIFO's worth of data to a pipe
 
    If this is a partial round (FIFO not going to be full), finish_round() is
    invoked after the write. Otherwise the FIFO is transmitted automatically and
    the BEMP handler will call finish_round() after the transfer. */
-static void write_round(asyncio_op_t volatile *t, int pipe)
+static void write_round(asyncio_op_t *t, int pipe)
 {
 	fifo_t ct = t->controller;
-	void volatile *FIFO = NULL;
 
+	void volatile *FIFO = NULL;
 	if(ct == CF)  FIFO = &USB.CFIFO;
 	if(ct == D0F) FIFO = &USB.D0FIFO;
 	if(ct == D1F) FIFO = &USB.D1FIFO;
-	fifo_bind(ct, pipe, FIFO_WRITE, t->unit_size);
 
 	/* Amount of data that can be transferred in a single run */
-	int available = pipe_bufsize(pipe) - (pipe == 0 ? 0 : t->buffer_used);
+	int available = pipe_bufsize(pipe) - t->buffer_used;
 	int size = min(t->size, available);
-	t->round_size = size;
 
 	/* If this is a partial write (size < available), call finish_round()
 	   after the copy to notify the user that the pipe is ready. Otherwise,
 	   a USB transfer will occur and the BEMP handler will do it. */
 	bool partial = (size < available);
 
+	asyncio_op_start_write_round(t, size);
+
 	if(t->dma)
 	{
-		/* TODO: USB: Can we use 32-byte DMA transfers? */
-		int block_size = DMA_1B;
-		if(t->unit_size == 2) block_size = DMA_2B, size >>= 1;
-		if(t->unit_size == 4) block_size = DMA_4B, size >>= 2;
-
 		gint_call_t callback = partial ?
 			GINT_CALL(finish_round, (void *)t, pipe) :
 			GINT_CALL_NULL;
@@ -321,60 +313,56 @@ static void write_round(asyncio_op_t volatile *t, int pipe)
 		/* Use DMA channel 3 for D0F and 4 for D1F */
 		int channel = (ct == D0F) ? 3 : 4;
 
-		bool ok = dma_transfer_async(channel, block_size, size,
+		/* TODO: USB: Can we use 32-byte DMA transfers? */
+		bool ok = dma_transfer_async(channel, DMA_4B, size >> 2,
 			t->data_w, DMA_INC, (void *)FIFO, DMA_FIXED, callback);
 		if(!ok) USB_LOG("DMA async failed on channel %d!\n", channel);
 	}
 	else
 	{
-		if(t->unit_size == 1) write_8(t->data_w, size, FIFO);
+		usb_pipe_write4(t->data_w, size, &t->shbuf, &t->shbuf_size,
-		if(t->unit_size == 2) write_16(t->data_w, size >> 1, FIFO);
+			FIFO);
-		if(t->unit_size == 4) write_32(t->data_w, size >> 2, FIFO);
 		if(partial) finish_round(t, pipe);
 	}
 
 	USB_TRACE("write_round()");
 }
 
-int usb_write_async(int pipe, void const *data, int size, int unit_size,
+int usb_write_async(int pipe, void const *data, int size, bool use_dma,
-	bool use_dma, gint_call_t callback)
+	gint_call_t callback)
 {
-	if(pipe_busy(pipe)) return USB_WRITE_BUSY;
+	asyncio_op_t *t = &pipe_transfers[pipe];
+	if(asyncio_op_busy(t))
+		return USB_WRITE_BUSY;
 
-	asyncio_op_t volatile *t = &pipe_transfers[pipe];
+	/* If this if the first write of a series, find a controller. */
-	if(!data || !size) return 0;
-
-	/* Re-use the controller from a previous write if there is one,
-	   otherwise try to get a new free one */
 	/* TODO: usb_write_async(): TOC/TOU race on controller being free */
-	fifo_t ct = t->controller;
+	if(t->controller == NOF) {
-	if(ct == NOF) ct = fifo_find_available_controller(pipe);
+		fifo_t ct = fifo_find_available_controller(pipe);
-	if(ct == NOF) return USB_WRITE_NOFIFO;
+		if(ct == NOF)
-
+			return USB_WRITE_NOFIFO;
-	t->data_w = data;
+		fifo_bind(ct, pipe, FIFO_WRITE);
-	t->size = size;
-	t->unit_size = (pipe == 0) ? 1 : unit_size;
-	t->dma = use_dma;
-	t->committed_w = false;
 		t->controller = ct;
-	t->callback = callback;
+	}
+
+	asyncio_op_start_write(t, data, size, use_dma, &callback);
 
 	/* Set up the Buffer Empty interrupt to refill the buffer when it gets
 	   empty, and be notified when the transfer completes. */
-	if(pipe) USB.BEMPENB |= (1 << pipe);
+	USB.BEMPENB |= (1 << pipe);
 
 	write_round(t, pipe);
 	return 0;
 }
 
-int usb_write_sync_timeout(int pipe, void const *data, int size, int unit_size,
+int usb_write_sync_timeout(int pipe, void const *data, int size, bool use_dma,
-	bool use_dma, timeout_t const *timeout)
+	timeout_t const *timeout)
 {
 	volatile int flag = 0;
 
 	while(1)
 	{
-		int rc = usb_write_async(pipe, data, size, unit_size, use_dma,
+		int rc = usb_write_async(pipe, data, size, use_dma,
 			GINT_CALL_SET(&flag));
 		if(rc == 0)
 			break;
@@ -396,39 +384,45 @@ int usb_write_sync_timeout(int pipe, void const *data, int size, int unit_size,
 	return 0;
 }
 
-int usb_write_sync(int pipe, void const *data, int size, int unit, bool dma)
+int usb_write_sync(int pipe, void const *data, int size, bool dma)
 {
-	return usb_write_sync_timeout(pipe, data, size, unit, dma, NULL);
+	return usb_write_sync_timeout(pipe, data, size, dma, NULL);
 }
 
 int usb_commit_async(int pipe, gint_call_t callback)
 {
-	asyncio_op_t volatile *t = &pipe_transfers[pipe];
+	asyncio_op_t *t = &pipe_transfers[pipe];
-	if(pipe_busy(pipe)) return USB_COMMIT_BUSY;
+	if(asyncio_op_busy(t))
+		return USB_COMMIT_BUSY;
+	if(t->type != ASYNCIO_WRITE || t->controller == NOF)
+		return USB_COMMIT_INACTIVE;
 
-	if(t->controller == NOF) return USB_COMMIT_INACTIVE;
+	/* Flush any remaining bytes in the short buffer. This cannot fill the
+	   buffer and create an auto-transmission situation; instead the module
+	   remains idle after this write. This is because we only use 32-bit
+	   writes, therefore at worst the buffer is 4 bytes away from being
+	   full, and will not be filled by an extra 0-3 bytes. */
+	void volatile *FIFO = NULL;
+	if(t->controller == CF)  FIFO = &USB.CFIFO;
+	if(t->controller == D0F) FIFO = &USB.D0FIFO;
+	if(t->controller == D1F) FIFO = &USB.D1FIFO;
+	usb_pipe_flush4(t->shbuf, t->shbuf_size, FIFO);
 
-	t->committed_w = true;
+	/* Switch from WRITE to SYNC type; this influences the BEMP handler and
-	t->callback = callback;
+	   the final finish_call() */
+	asyncio_op_start_sync(t, &callback);
 
-	/* TODO: Handle complex commits on the DCP */
+	/* TODO: Figure out why previous attempts to use BEMP to finish commit
-	if(pipe == 0)
+	   TODO| calls on the DCP failed with a freeze */
-	{
+	if(pipe == 0) {
-		finish_transfer(t, pipe);
 		USB.CFIFOCTR.BVAL = 1;
-		return 0;
+		finish_call(t, pipe);
-	}
-
-	/* Committing an empty pipe ends the transfer on the spot */
-	if(t->buffer_used == 0)
-	{
-		finish_transfer(t, pipe);
 		return 0;
 	}
 
 	/* Set BVAL=1 and inform the BEMP handler of the commitment with the
-	   committed_w flag; the handler will invoke finish_transfer() */
+	   SYNC type; the handler will invoke finish_call() */
-	fifo_bind(t->controller, pipe, FIFO_WRITE, t->unit_size);
+	USB.BEMPENB |= (1 << pipe);
 	if(t->controller == D0F) USB.D0FIFOCTR.BVAL = 1;
 	if(t->controller == D1F) USB.D1FIFOCTR.BVAL = 1;
 	USB_LOG("[PIPE%d] Committed transfer\n", pipe);
@@ -471,11 +465,11 @@ void usb_commit_sync(int pipe)
 /* usb_pipe_write_bemp(): Callback for the BEMP interrupt on a pipe */
 void usb_pipe_write_bemp(int pipe)
 {
-	asyncio_op_t volatile *t = &pipe_transfers[pipe];
+	asyncio_op_t *t = &pipe_transfers[pipe];
 
-	if(t->committed_w)
+	if(t->type == ASYNCIO_SYNC)
 	{
-		finish_transfer(t, pipe);
+		finish_call(t, pipe);
 	}
 	else
 	{

src/usb/setup.c

@@ -8,7 +8,7 @@
 
 #define USB SH7305_USB
 
-#define dcp_write(data, size) usb_write_sync(0, data, size, 1, false)
+#define dcp_write(data, size) usb_write_sync(0, data, size, false)
 
 //---
 // SETUP requests

src/usb/usb.c

@@ -223,6 +223,9 @@ void usb_open_wait(void)
 
 void usb_close(void)
 {
+	usb_wait_all_transfers(false);
+	usb_pipe_init_transfers();
+
 	intc_priority(INTC_USB, 0);
 	hpoweroff();
 	USB_LOG("---- usb_close ----\n");
@@ -277,7 +280,7 @@ static void usb_interrupt_handler(void)
 		uint16_t status = USB.BEMPSTS;
 		USB.BEMPSTS = 0;
 
-		for(int i = 1; i <= 9; i++)
+		for(int i = 0; i <= 9; i++)
 		{
 			if(status & (1 << i)) usb_pipe_write_bemp(i);
 		}

src/usb/usb_private.h

@@ -132,6 +132,46 @@ void usb_pipe_write_bemp(int pipe);
 /* usb_pipe_init_transfers(): Initialize transfer information */
 void usb_pipe_init_transfers(void);
 
+/* usb_wait_all_transfers(): Wait for all transfers to finish
+
+   This function waits for all current operations on the pipes to finish their
+   current read/write/fsync call. Once the waiting period is finished, the
+   calls are guaranteed to be finished, but write transactions might not (as
+   they require multiple calls finishing with a fsync(2)).
+
+   If `await_long_writes` is set, this function also waits for all writes to be
+   committed, which only makes sense if said writes are executed in a thread
+   that is able to run while this is waiting. */
+void usb_wait_all_transfers(bool await_long_writes);
+
+/* usb_pipe_write4(): Copy arbitrary ranges of memory to a 4-byte USB FIFO
+
+   This function copies arbitrarily-aligned data of any size into a 4-byte
+   USB FIFO register. It rearranges data so as to perform only 4-byte aligned
+   writes. If the data size isn't a multiple of 4 bytes, it stores the
+   remainder into a short buffer (holding between 0 and 3 bytes), to be
+   combined with fresh data on the next call. The remainder of the buffer can
+   be discharged eventually with usb_pipe_flush4().
+
+   @data         Data to write into the FIFO
+   @size         Number of bytes to write
+   @buffer       Address of short buffer
+   @buffer_size  Address of short buffer's size tracker
+   @FIFO         FIFO to output to */
+void usb_pipe_write4(void const *data, int size, uint32_t volatile *buffer,
+   uint8_t volatile *buffer_size, uint32_t volatile *FIFO);
+
+/* usb_pipe_flush4(): Flush usb_pipe_write4()'s short buffer
+
+   This function is used after a sequence of usb_pipe_write4() to flush the
+   last few bytes remaining in the short buffer.
+
+   @buffer       Contents of short buffer
+   @buffer_size  Short buffer's size tracker
+   @FIFO         FIFO to output to */
+void usb_pipe_flush4(uint32_t buffer, int buffer_size,
+   uint32_t volatile *FIFO);
+
 //---
 // Timeout waits
 //---
@@ -171,7 +211,7 @@ enum {
 void usb_req_setup(void);
 
 //---
-// Enumerations and stuff
+// Enumerated constants
 //---
 
 enum {

src/usb/write4.S

@@ -0,0 +1,136 @@
+.global	_usb_pipe_write4
+.global	_usb_pipe_flush4
+
+#define _fifo r3
+#define _data r4
+#define _size r5
+#define _buf r6
+#define _bufsize r7
+
+_usb_pipe_write4:
+	/* Skip to writing the data if the buffer's empty. This test is free
+	   because having it simplifies a later while loop into a do/while. */
+	mov.b	@_bufsize, r1
+	mov	#4, r0
+	mov.l	@r15, _fifo
+	tst	r1, r1
+	bt.s	.write_data
+	sub	r1, r0		/* Bytes required to fill the buffer */
+
+	/* If we can't even fill the buffer, skip to the end push. */
+	mov.l	@_buf, r2
+	cmp/gt	_size, r0
+	bt	.push_buffer
+
+	/* Precompute the amount of data left after filling the buffer */
+	sub	r0, _size
+
+	/* Fill the buffer by reading unaligned bytes */
+1:	mov.b	@_data+, r1
+	shll8	r2
+	dt	r0
+	extu.b	r1, r1
+	bf.s	1b
+	or	r1, r2
+
+	/* Commit the filled buffer */
+	mov.l	r2, @_fifo
+
+.write_data:
+	/* Check if we have enough data to run this loop */
+	/* TODO: For small sizes use another loop, so we can unroll? */
+	mov	#4, r0
+	cmp/gt	_size, r0
+	bt	4f
+
+	/* Determine whether we need to use unaligned reads */
+	mov	#3, r0
+	tst	r0, _data
+	mov	_size, r1
+	bt.s	3f
+	shlr2	r1
+
+	/* Unaligned write loop */
+2:	movua.l	@_data+, r0
+	dt	r1
+	bf.s	2b
+	mov.l	r0, @_fifo
+
+	bra	4f
+	nop
+
+	/* Aligned write loop */
+3:	mov.l	@_data+, r0
+	dt	r1
+	bf.s	3b
+	mov.l	r0, @_fifo
+
+4:	mov	#3, r0
+	and	r0, _size
+
+	mov	#0, r2
+	mov	#0, r1
+
+.push_buffer:
+	/* Here r1 = buffer size, r2 = buffer contents, _size = data left */
+
+	/* Check if there is any data left to push into the short buffer */
+	tst	_size, _size
+	mov	r1, r0
+	add	_size, r0
+	bt.s	.end
+	mov.b	r0, @_bufsize
+
+	/* Push loop */
+5:	mov.b	@_data+, r1
+	shll8	r2
+	dt	_size
+	extu.b	r1, r1
+	bf.s	5b
+	or	r1, r2
+
+.end:
+	rts
+	mov.l	r2, @_buf
+
+#undef _fifo
+#undef _data
+#undef _size
+#undef _buf
+#undef _bufsize
+
+/* --- */
+
+#define _buf r4
+#define _bufsize r5
+#define _fifo r6
+
+_usb_pipe_flush4:
+	/* Jump table. We skip 4*_bufsize bytes, which lands us right on labels
+	   0:, 1:, 2: or 3: depending on the value of _bufsize. */
+	shll2	_bufsize
+	braf	_bufsize
+	mov	_buf, r0
+
+0:	/* No extra data to write out */
+	rts
+	nop
+
+1:	/* Single byte */
+	rts
+	mov.b	r0, @_fifo
+
+2:	/* Two bytes */
+	rts
+	mov.w	r0, @_fifo
+
+3:	/* Three bytes */
+	mov	r0, r1
+	shlr8	r1
+	mov.w	r1, @_fifo
+	rts
+	mov.b	r0, @(2, _fifo)
+
+#undef _buf
+#undef _bufsize
+#undef _fifo