/*
* CPPI TX:
* ========
* TX is a lot more reasonable than RX; it doesn't need to run in
* irq-per-packet mode very often. RNDIS mode seems to behave too
* (except how it handles the exactly-N-packets case). Building a
* txdma queue with multiple requests (urb or usb_request) looks
* like it would work ... but fault handling would need much testing.
*
* The main issue with TX mode RNDIS relates to transfer lengths that
* are an exact multiple of the packet length. It appears that there's
* a hiccup in that case (maybe the DMA completes before the ZLP gets
* written?) boiling down to not being able to rely on CPPI writing any
* terminating zero length packet before the next transfer is written.
* So that's punted to PIO; better yet, gadget drivers can avoid it.
*
* Plus, there's allegedly an undocumented constraint that rndis transfer
* length be a multiple of 64 bytes ... but the chip doesn't act that
* way, and we really don't _want_ that behavior anyway.
*
* On TX, "transparent" mode works ... although experiments have shown
* problems trying to use the SOP/EOP bits in different USB packets.
*
* REVISIT try to handle terminating zero length packets using CPPI
* instead of doing it by PIO after an IRQ. (Meanwhile, make Ethernet
* links avoid that issue by forcing them to avoid zlps.)
*/
static void
cppi_next_tx_segment(struct musb *musb, struct cppi_channel *tx)
{
unsigned maxpacket = tx->maxpacket;
dma_addr_t addr = tx->buf_dma + tx->offset;
size_t length = tx->buf_len - tx->offset;
struct cppi_descriptor *bd;
unsigned n_bds;
unsigned i;
struct cppi_tx_stateram __iomem *tx_ram = tx->state_ram;
int rndis;
/* TX can use the CPPI "rndis" mode, where we can probably fit this
* transfer in one BD and one IRQ. The only time we would NOT want
* to use it is when hardware constraints prevent it, or if we'd
* trigger the "send a ZLP?" confusion.
*/
rndis = (maxpacket & 0x3f) == 0
&& length > maxpacket
&& length < 0xffff
&& (length % maxpacket) != 0;
if (rndis) {
maxpacket = length;
n_bds = 1;
} else {
if (length)
n_bds = DIV_ROUND_UP(length, maxpacket);
else
n_bds = 1;
n_bds = min(n_bds, (unsigned) NUM_TXCHAN_BD);
length = min(n_bds * maxpacket, length);
}
musb_dbg(musb, "TX DMA%d, pktSz %d %s bds %d dma 0x%llx len %u",
tx->index,
maxpacket,
rndis ? "rndis" : "transparent",
n_bds,
(unsigned long long)addr, length);
cppi_rndis_update(tx, 0, musb->ctrl_base, rndis);
/* assuming here that channel_program is called during
* transfer initiation ... current code maintains state
* for one outstanding request only (no queues, not even
* the implicit ones of an iso urb).
*/
bd = tx->freelist;
tx->head = bd;
tx->last_processed = NULL;
/* FIXME use BD pool like RX side does, and just queue
* the minimum number for this request.
*/
/* Prepare queue of BDs first, then hand it to hardware.
* All BDs except maybe the last should be of full packet
* size; for RNDIS there _is_ only that last packet.
*/
for (i = 0; i < n_bds; ) {
if (++i < n_bds && bd->next)
bd->hw_next = bd->next->dma;
else
bd->hw_next = 0;
bd->hw_bufp = tx->buf_dma + tx->offset;
/* FIXME set EOP only on the last packet,
* SOP only on the first ... avoid IRQs
*/
if ((tx->offset + maxpacket) <= tx->buf_len) {
tx->offset += maxpacket;
bd->hw_off_len = maxpacket;
bd->hw_options = CPPI_SOP_SET | CPPI_EOP_SET
| CPPI_OWN_SET | maxpacket;
} else {
/* only this one may be a partial USB Packet */
u32 partial_len;
partial_len = tx->buf_len - tx->offset;
tx->offset = tx->buf_len;
bd->hw_off_len = partial_len;
bd->hw_options = CPPI_SOP_SET | CPPI_EOP_SET
| CPPI_OWN_SET | partial_len;
if (partial_len == 0)
bd->hw_options |= CPPI_ZERO_SET;
}
musb_dbg(musb, "TXBD %p: nxt %08x buf %08x len %04x opt %08x",
bd, bd->hw_next, bd->hw_bufp,
bd->hw_off_len, bd->hw_options);
/* update the last BD enqueued to the list */
tx->tail = bd;
bd = bd->next;
}
/* BDs live in DMA-coherent memory, but writes might be pending */
cpu_drain_writebuffer();
/* Write to the HeadPtr in state RAM to trigger */
musb_writel(&tx_ram->tx_head, 0, (u32)tx->freelist->dma);
cppi_dump_tx(5, tx, "/S");
}
irqreturn_t cppi_interrupt(int irq, void *dev_id)
{
struct musb *musb = dev_id;
struct cppi *cppi;
void __iomem *tibase;
struct musb_hw_ep *hw_ep = NULL;
u32 rx, tx;
int i, index;
unsigned long uninitialized_var(flags);
cppi = container_of(musb->dma_controller, struct cppi, controller);
if (cppi->irq)
spin_lock_irqsave(&musb->lock, flags);
tibase = musb->ctrl_base;
tx = musb_readl(tibase, DAVINCI_TXCPPI_MASKED_REG);
rx = musb_readl(tibase, DAVINCI_RXCPPI_MASKED_REG);
if (!tx && !rx) {
if (cppi->irq)
spin_unlock_irqrestore(&musb->lock, flags);
return IRQ_NONE;
}
musb_dbg(musb, "CPPI IRQ Tx%x Rx%x", tx, rx);
/* process TX channels */
for (index = 0; tx; tx = tx >> 1, index++) {
struct cppi_channel *tx_ch;
struct cppi_tx_stateram __iomem *tx_ram;
bool completed = false;
struct cppi_descriptor *bd;
if (!(tx & 1))
continue;
tx_ch = cppi->tx + index;
tx_ram = tx_ch->state_ram;
/* FIXME need a cppi_tx_scan() routine, which
* can also be called from abort code
*/
cppi_dump_tx(5, tx_ch, "/E");
bd = tx_ch->head;
/*
* If Head is null then this could mean that a abort interrupt
* that needs to be acknowledged.
*/
if (NULL == bd) {
musb_dbg(musb, "null BD");
musb_writel(&tx_ram->tx_complete, 0, 0);
continue;
}
/* run through all completed BDs */
for (i = 0; !completed && bd && i < NUM_TXCHAN_BD;
i++, bd = bd->next) {
u16 len;
/* catch latest BD writes from CPPI */
rmb();
if (bd->hw_options & CPPI_OWN_SET)
break;
musb_dbg(musb, "C/TXBD %p n %x b %x off %x opt %x",
bd, bd->hw_next, bd->hw_bufp,
bd->hw_off_len, bd->hw_options);
len = bd->hw_off_len & CPPI_BUFFER_LEN_MASK;
tx_ch->channel.actual_len += len;
tx_ch->last_processed = bd;
/* write completion register to acknowledge
* processing of completed BDs, and possibly
* release the IRQ; EOQ might not be set ...
*
* REVISIT use the same ack strategy as rx
*
* REVISIT have observed bit 18 set; huh??
*/
/* if ((bd->hw_options & CPPI_EOQ_MASK)) */
musb_writel(&tx_ram->tx_complete, 0, bd->dma);
/* stop scanning on end-of-segment */
if (bd->hw_next == 0)
completed = true;
}
/* on end of segment, maybe go to next one */
if (completed) {
/* cppi_dump_tx(4, tx_ch, "/complete"); */
/* transfer more, or report completion */
if (tx_ch->offset >= tx_ch->buf_len) {
tx_ch->head = NULL;
tx_ch->tail = NULL;
tx_ch->channel.status = MUSB_DMA_STATUS_FREE;
hw_ep = tx_ch->hw_ep;
musb_dma_completion(musb, index + 1, 1);
} else {
/* Bigger transfer than we could fit in
* that first batch of descriptors...
*/
cppi_next_tx_segment(musb, tx_ch);
}
} else
tx_ch->head = bd;
}
/* Start processing the RX block */
for (index = 0; rx; rx = rx >> 1, index++) {
if (rx & 1) {
struct cppi_channel *rx_ch;
rx_ch = cppi->rx + index;
/* let incomplete dma segments finish */
if (!cppi_rx_scan(cppi, index))
continue;
/* start another dma segment if needed */
if (rx_ch->channel.actual_len != rx_ch->buf_len
&& rx_ch->channel.actual_len
== rx_ch->offset) {
cppi_next_rx_segment(musb, rx_ch, 1);
continue;
}
/* all segments completed! */
rx_ch->channel.status = MUSB_DMA_STATUS_FREE;
hw_ep = rx_ch->hw_ep;
core_rxirq_disable(tibase, index + 1);
musb_dma_completion(musb, index + 1, 0);
}
}
/* write to CPPI EOI register to re-enable interrupts */
musb_writel(tibase, DAVINCI_CPPI_EOI_REG, 0);
if (cppi->irq)
spin_unlock_irqrestore(&musb->lock, flags);
return IRQ_HANDLED;
}
/*
* Context: controller irqlocked, endpoint selected
*/
static int cppi_channel_abort(struct dma_channel *channel)
{
struct cppi_channel *cppi_ch;
struct cppi *controller;
void __iomem *mbase;
void __iomem *tibase;
void __iomem *regs;
u32 value;
struct cppi_descriptor *queue;
cppi_ch = container_of(channel, struct cppi_channel, channel);
controller = cppi_ch->controller;
switch (channel->status) {
case MUSB_DMA_STATUS_BUS_ABORT:
case MUSB_DMA_STATUS_CORE_ABORT:
/* from RX or TX fault irq handler */
case MUSB_DMA_STATUS_BUSY:
/* the hardware needs shutting down */
regs = cppi_ch->hw_ep->regs;
break;
case MUSB_DMA_STATUS_UNKNOWN:
case MUSB_DMA_STATUS_FREE:
return 0;
default:
return -EINVAL;
}
if (!cppi_ch->transmit && cppi_ch->head)
cppi_dump_rxq(3, "/abort", cppi_ch);
mbase = controller->mregs;
tibase = controller->tibase;
queue = cppi_ch->head;
cppi_ch->head = NULL;
cppi_ch->tail = NULL;
/* REVISIT should rely on caller having done this,
* and caller should rely on us not changing it.
* peripheral code is safe ... check host too.
*/
musb_ep_select(mbase, cppi_ch->index + 1);
if (cppi_ch->transmit) {
struct cppi_tx_stateram __iomem *tx_ram;
/* REVISIT put timeouts on these controller handshakes */
cppi_dump_tx(6, cppi_ch, " (teardown)");
/* teardown DMA engine then usb core */
do {
value = musb_readl(tibase, DAVINCI_TXCPPI_TEAR_REG);
} while (!(value & CPPI_TEAR_READY));
musb_writel(tibase, DAVINCI_TXCPPI_TEAR_REG, cppi_ch->index);
tx_ram = cppi_ch->state_ram;
do {
value = musb_readl(&tx_ram->tx_complete, 0);
} while (0xFFFFFFFC != value);
/* FIXME clean up the transfer state ... here?
* the completion routine should get called with
* an appropriate status code.
*/
value = musb_readw(regs, MUSB_TXCSR);
value &= ~MUSB_TXCSR_DMAENAB;
value |= MUSB_TXCSR_FLUSHFIFO;
musb_writew(regs, MUSB_TXCSR, value);
musb_writew(regs, MUSB_TXCSR, value);
/*
* 1. Write to completion Ptr value 0x1(bit 0 set)
* (write back mode)
* 2. Wait for abort interrupt and then put the channel in
* compare mode by writing 1 to the tx_complete register.
*/
cppi_reset_tx(tx_ram, 1);
cppi_ch->head = NULL;
musb_writel(&tx_ram->tx_complete, 0, 1);
cppi_dump_tx(5, cppi_ch, " (done teardown)");
/* REVISIT tx side _should_ clean up the same way
* as the RX side ... this does no cleanup at all!
*/
} else /* RX */ {
u16 csr;
/* NOTE: docs don't guarantee any of this works ... we
* expect that if the usb core stops telling the cppi core
* to pull more data from it, then it'll be safe to flush
* current RX DMA state iff any pending fifo transfer is done.
*/
core_rxirq_disable(tibase, cppi_ch->index + 1);
/* for host, ensure ReqPkt is never set again */
if (is_host_active(cppi_ch->controller->controller.musb)) {
value = musb_readl(tibase, DAVINCI_AUTOREQ_REG);
value &= ~((0x3) << (cppi_ch->index * 2));
musb_writel(tibase, DAVINCI_AUTOREQ_REG, value);
}
csr = musb_readw(regs, MUSB_RXCSR);
/* for host, clear (just) ReqPkt at end of current packet(s) */
if (is_host_active(cppi_ch->controller->controller.musb)) {
csr |= MUSB_RXCSR_H_WZC_BITS;
csr &= ~MUSB_RXCSR_H_REQPKT;
} else
csr |= MUSB_RXCSR_P_WZC_BITS;
/* clear dma enable */
csr &= ~(MUSB_RXCSR_DMAENAB);
musb_writew(regs, MUSB_RXCSR, csr);
csr = musb_readw(regs, MUSB_RXCSR);
/* Quiesce: wait for current dma to finish (if not cleanup).
* We can't use bit zero of stateram->rx_sop, since that
* refers to an entire "DMA packet" not just emptying the
* current fifo. Most segments need multiple usb packets.
*/
if (channel->status == MUSB_DMA_STATUS_BUSY)
udelay(50);
/* scan the current list, reporting any data that was
* transferred and acking any IRQ
*/
cppi_rx_scan(controller, cppi_ch->index);
/* clobber the existing state once it's idle
*
* NOTE: arguably, we should also wait for all the other
* RX channels to quiesce (how??) and then temporarily
* disable RXCPPI_CTRL_REG ... but it seems that we can
* rely on the controller restarting from state ram, with
* only RXCPPI_BUFCNT state being bogus. BUFCNT will
* correct itself after the next DMA transfer though.
*
* REVISIT does using rndis mode change that?
*/
cppi_reset_rx(cppi_ch->state_ram);
/* next DMA request _should_ load cppi head ptr */
/* ... we don't "free" that list, only mutate it in place. */
cppi_dump_rx(5, cppi_ch, " (done abort)");
/* clean up previously pending bds */
cppi_bd_free(cppi_ch, cppi_ch->last_processed);
cppi_ch->last_processed = NULL;
while (queue) {
struct cppi_descriptor *tmp = queue->next;
cppi_bd_free(cppi_ch, queue);
queue = tmp;
}
}
channel->status = MUSB_DMA_STATUS_FREE;
cppi_ch->buf_dma = 0;
cppi_ch->offset = 0;
cppi_ch->buf_len = 0;
cppi_ch->maxpacket = 0;
return 0;
}
void cppi_completion(struct musb *musb, u32 rx, u32 tx)
{
void __iomem *tibase;
int i, index;
struct cppi *cppi;
struct musb_hw_ep *hw_ep = NULL;
cppi = container_of(musb->dma_controller, struct cppi, controller);
tibase = musb->ctrl_base;
/* process TX channels */
for (index = 0; tx; tx = tx >> 1, index++) {
struct cppi_channel *tx_ch;
struct cppi_tx_stateram __iomem *tx_ram;
bool completed = false;
struct cppi_descriptor *bd;
if (!(tx & 1))
continue;
tx_ch = cppi->tx + index;
tx_ram = tx_ch->state_ram;
/* FIXME need a cppi_tx_scan() routine, which
* can also be called from abort code
*/
cppi_dump_tx(5, tx_ch, "/E");
bd = tx_ch->head;
if (NULL == bd) {
DBG(1, "null BD\n");
continue;
}
/* run through all completed BDs */
for (i = 0; !completed && bd && i < NUM_TXCHAN_BD;
i++, bd = bd->next) {
u16 len;
/* catch latest BD writes from CPPI */
rmb();
if (bd->hw_options & CPPI_OWN_SET)
break;
DBG(5, "C/TXBD %p n %x b %x off %x opt %x\n",
bd, bd->hw_next, bd->hw_bufp,
bd->hw_off_len, bd->hw_options);
len = bd->hw_off_len & CPPI_BUFFER_LEN_MASK;
tx_ch->channel.actual_len += len;
tx_ch->last_processed = bd;
/* write completion register to acknowledge
* processing of completed BDs, and possibly
* release the IRQ; EOQ might not be set ...
*
* REVISIT use the same ack strategy as rx
*
* REVISIT have observed bit 18 set; huh??
*/
/* if ((bd->hw_options & CPPI_EOQ_MASK)) */
musb_writel(&tx_ram->tx_complete, 0, bd->dma);
/* stop scanning on end-of-segment */
if (bd->hw_next == 0)
completed = true;
}
/* on end of segment, maybe go to next one */
if (completed) {
/* cppi_dump_tx(4, tx_ch, "/complete"); */
/* transfer more, or report completion */
if (tx_ch->offset >= tx_ch->buf_len) {
tx_ch->head = NULL;
tx_ch->tail = NULL;
tx_ch->channel.status = MUSB_DMA_STATUS_FREE;
hw_ep = tx_ch->hw_ep;
/* Peripheral role never repurposes the
* endpoint, so immediate completion is
* safe. Host role waits for the fifo
* to empty (TXPKTRDY irq) before going
* to the next queued bulk transfer.
*/
if (is_host_active(cppi->musb)) {
#if 0
/* WORKAROUND because we may
* not always get TXKPTRDY ...
*/
int csr;
csr = musb_readw(hw_ep->regs,
MUSB_TXCSR);
if (csr & MUSB_TXCSR_TXPKTRDY)
#endif
completed = false;
}
if (completed)
musb_dma_completion(musb, index + 1, 1);
} else {
/* Bigger transfer than we could fit in
* that first batch of descriptors...
*/
cppi_next_tx_segment(musb, tx_ch);
}
} else
tx_ch->head = bd;
}
/* Start processing the RX block */
for (index = 0; rx; rx = rx >> 1, index++) {
if (rx & 1) {
struct cppi_channel *rx_ch;
rx_ch = cppi->rx + index;
/* let incomplete dma segments finish */
if (!cppi_rx_scan(cppi, index))
continue;
/* start another dma segment if needed */
if (rx_ch->channel.actual_len != rx_ch->buf_len
&& rx_ch->channel.actual_len
== rx_ch->offset) {
cppi_next_rx_segment(musb, rx_ch, 1);
continue;
}
/* all segments completed! */
rx_ch->channel.status = MUSB_DMA_STATUS_FREE;
hw_ep = rx_ch->hw_ep;
core_rxirq_disable(tibase, index + 1);
musb_dma_completion(musb, index + 1, 0);
}
}
/* write to CPPI EOI register to re-enable interrupts */
musb_writel(tibase, DAVINCI_CPPI_EOI_REG, 0);
}
