/* Release a CPPI Channel. */
static void cppi_channel_release(struct dma_channel *channel)
{
struct cppi_channel *c;
void __iomem *tibase;
/* REVISIT: for paranoia, check state and abort if needed... */
c = container_of(channel, struct cppi_channel, channel);
tibase = c->controller->tibase;
if (!c->hw_ep)
DBG(1, "releasing idle DMA channel %p\n", c);
else if (!c->transmit)
core_rxirq_enable(tibase, c->index + 1);
/* for now, leave its cppi IRQ enabled (we won't trigger it) */
c->hw_ep = NULL;
channel->status = MUSB_DMA_STATUS_UNKNOWN;
}
/**
* cppi_next_rx_segment - dma read for the next chunk of a buffer
* @musb: the controller
* @rx: dma channel
* @onepacket: true unless caller treats short reads as errors, and
* performs fault recovery above usbcore.
* Context: controller irqlocked
*
* See above notes about why we can't use multi-BD RX queues except in
* rare cases (mass storage class), and can never use the hardware "rndis"
* mode (since it's not a "true" RNDIS mode) with complete safety..
*
* It's ESSENTIAL that callers specify "onepacket" mode unless they kick in
* code to recover from corrupted datastreams after each short transfer.
*/
static void
cppi_next_rx_segment(struct musb *musb, struct cppi_channel *rx, int onepacket)
{
unsigned maxpacket = rx->maxpacket;
dma_addr_t addr = rx->buf_dma + rx->offset;
size_t length = rx->buf_len - rx->offset;
struct cppi_descriptor *bd, *tail;
unsigned n_bds;
unsigned i;
void __iomem *tibase = musb->ctrl_base;
int is_rndis = 0;
struct cppi_rx_stateram __iomem *rx_ram = rx->state_ram;
struct cppi_descriptor *d;
if (onepacket) {
/* almost every USB driver, host or peripheral side */
n_bds = 1;
/* maybe apply the heuristic above */
if (cppi_rx_rndis
&& is_peripheral_active(musb)
&& length > maxpacket
&& (length & ~0xffff) == 0
&& (length & 0x0fff) != 0
&& (length & (maxpacket - 1)) == 0) {
maxpacket = length;
is_rndis = 1;
}
} else {
/* virtually nothing except mass storage class */
if (length > 0xffff) {
n_bds = 0xffff / maxpacket;
length = n_bds * maxpacket;
} else {
n_bds = DIV_ROUND_UP(length, maxpacket);
}
if (n_bds == 1)
onepacket = 1;
else
n_bds = min(n_bds, (unsigned) NUM_RXCHAN_BD);
}
/* In host mode, autorequest logic can generate some IN tokens; it's
* tricky since we can't leave REQPKT set in RXCSR after the transfer
* finishes. So: multipacket transfers involve two or more segments.
* And always at least two IRQs ... RNDIS mode is not an option.
*/
if (is_host_active(musb))
n_bds = cppi_autoreq_update(rx, tibase, onepacket, n_bds);
cppi_rndis_update(rx, 1, musb->ctrl_base, is_rndis);
length = min(n_bds * maxpacket, length);
musb_dbg(musb, "RX DMA%d seg, maxp %d %s bds %d (cnt %d) "
"dma 0x%llx len %u %u/%u",
rx->index, maxpacket,
onepacket
? (is_rndis ? "rndis" : "onepacket")
: "multipacket",
n_bds,
musb_readl(tibase,
DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4))
& 0xffff,
(unsigned long long)addr, length,
rx->channel.actual_len, rx->buf_len);
/* only queue one segment at a time, since the hardware prevents
* correct queue shutdown after unexpected short packets
*/
bd = cppi_bd_alloc(rx);
rx->head = bd;
/* Build BDs for all packets in this segment */
for (i = 0, tail = NULL; bd && i < n_bds; i++, tail = bd) {
u32 bd_len;
if (i) {
bd = cppi_bd_alloc(rx);
if (!bd)
break;
tail->next = bd;
tail->hw_next = bd->dma;
}
bd->hw_next = 0;
/* all but the last packet will be maxpacket size */
if (maxpacket < length)
bd_len = maxpacket;
else
bd_len = length;
bd->hw_bufp = addr;
addr += bd_len;
rx->offset += bd_len;
bd->hw_off_len = (0 /*offset*/ << 16) + bd_len;
bd->buflen = bd_len;
bd->hw_options = CPPI_OWN_SET | (i == 0 ? length : 0);
length -= bd_len;
}
/* we always expect at least one reusable BD! */
if (!tail) {
WARNING("rx dma%d -- no BDs? need %d\n", rx->index, n_bds);
return;
} else if (i < n_bds)
WARNING("rx dma%d -- only %d of %d BDs\n", rx->index, i, n_bds);
tail->next = NULL;
tail->hw_next = 0;
bd = rx->head;
rx->tail = tail;
/* short reads and other faults should terminate this entire
* dma segment. we want one "dma packet" per dma segment, not
* one per USB packet, terminating the whole queue at once...
* NOTE that current hardware seems to ignore SOP and EOP.
*/
bd->hw_options |= CPPI_SOP_SET;
tail->hw_options |= CPPI_EOP_SET;
for (d = rx->head; d; d = d->next)
cppi_dump_rxbd("S", d);
/* in case the preceding transfer left some state... */
tail = rx->last_processed;
if (tail) {
tail->next = bd;
tail->hw_next = bd->dma;
}
core_rxirq_enable(tibase, rx->index + 1);
/* BDs live in DMA-coherent memory, but writes might be pending */
cpu_drain_writebuffer();
/* REVISIT specs say to write this AFTER the BUFCNT register
* below ... but that loses badly.
*/
musb_writel(&rx_ram->rx_head, 0, bd->dma);
/* bufferCount must be at least 3, and zeroes on completion
* unless it underflows below zero, or stops at two, or keeps
* growing ... grr.
*/
i = musb_readl(tibase,
DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4))
& 0xffff;
if (!i)
musb_writel(tibase,
DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4),
n_bds + 2);
else if (n_bds > (i - 3))
musb_writel(tibase,
DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4),
n_bds - (i - 3));
i = musb_readl(tibase,
DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4))
& 0xffff;
if (i < (2 + n_bds)) {
musb_dbg(musb, "bufcnt%d underrun - %d (for %d)",
rx->index, i, n_bds);
musb_writel(tibase,
DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4),
n_bds + 2);
}
cppi_dump_rx(4, rx, "/S");
}
