Commit | Line | Data |
---|---|---|
550a7375 FB |
1 | /* |
2 | * Copyright (C) 2005-2006 by Texas Instruments | |
3 | * | |
4 | * This file implements a DMA interface using TI's CPPI DMA. | |
5 | * For now it's DaVinci-only, but CPPI isn't specific to DaVinci or USB. | |
6 | * The TUSB6020, using VLYNQ, has CPPI that looks much like DaVinci. | |
7 | */ | |
8 | ||
91e9c4fe | 9 | #include <linux/platform_device.h> |
5a0e3ad6 | 10 | #include <linux/slab.h> |
550a7375 FB |
11 | #include <linux/usb.h> |
12 | ||
13 | #include "musb_core.h" | |
704a1485 | 14 | #include "musb_debug.h" |
550a7375 FB |
15 | #include "cppi_dma.h" |
16 | ||
17 | ||
18 | /* CPPI DMA status 7-mar-2006: | |
19 | * | |
20 | * - See musb_{host,gadget}.c for more info | |
21 | * | |
22 | * - Correct RX DMA generally forces the engine into irq-per-packet mode, | |
23 | * which can easily saturate the CPU under non-mass-storage loads. | |
24 | * | |
25 | * NOTES 24-aug-2006 (2.6.18-rc4): | |
26 | * | |
27 | * - peripheral RXDMA wedged in a test with packets of length 512/512/1. | |
28 | * evidently after the 1 byte packet was received and acked, the queue | |
29 | * of BDs got garbaged so it wouldn't empty the fifo. (rxcsr 0x2003, | |
30 | * and RX DMA0: 4 left, 80000000 8feff880, 8feff860 8feff860; 8f321401 | |
31 | * 004001ff 00000001 .. 8feff860) Host was just getting NAKed on tx | |
32 | * of its next (512 byte) packet. IRQ issues? | |
33 | * | |
34 | * REVISIT: the "transfer DMA" glue between CPPI and USB fifos will | |
35 | * evidently also directly update the RX and TX CSRs ... so audit all | |
36 | * host and peripheral side DMA code to avoid CSR access after DMA has | |
37 | * been started. | |
38 | */ | |
39 | ||
40 | /* REVISIT now we can avoid preallocating these descriptors; or | |
41 | * more simply, switch to a global freelist not per-channel ones. | |
42 | * Note: at full speed, 64 descriptors == 4K bulk data. | |
43 | */ | |
44 | #define NUM_TXCHAN_BD 64 | |
45 | #define NUM_RXCHAN_BD 64 | |
46 | ||
47 | static inline void cpu_drain_writebuffer(void) | |
48 | { | |
49 | wmb(); | |
50 | #ifdef CONFIG_CPU_ARM926T | |
51 | /* REVISIT this "should not be needed", | |
52 | * but lack of it sure seemed to hurt ... | |
53 | */ | |
54 | asm("mcr p15, 0, r0, c7, c10, 4 @ drain write buffer\n"); | |
55 | #endif | |
56 | } | |
57 | ||
58 | static inline struct cppi_descriptor *cppi_bd_alloc(struct cppi_channel *c) | |
59 | { | |
60 | struct cppi_descriptor *bd = c->freelist; | |
61 | ||
62 | if (bd) | |
63 | c->freelist = bd->next; | |
64 | return bd; | |
65 | } | |
66 | ||
67 | static inline void | |
68 | cppi_bd_free(struct cppi_channel *c, struct cppi_descriptor *bd) | |
69 | { | |
70 | if (!bd) | |
71 | return; | |
72 | bd->next = c->freelist; | |
73 | c->freelist = bd; | |
74 | } | |
75 | ||
76 | /* | |
77 | * Start DMA controller | |
78 | * | |
79 | * Initialize the DMA controller as necessary. | |
80 | */ | |
81 | ||
82 | /* zero out entire rx state RAM entry for the channel */ | |
83 | static void cppi_reset_rx(struct cppi_rx_stateram __iomem *rx) | |
84 | { | |
85 | musb_writel(&rx->rx_skipbytes, 0, 0); | |
86 | musb_writel(&rx->rx_head, 0, 0); | |
87 | musb_writel(&rx->rx_sop, 0, 0); | |
88 | musb_writel(&rx->rx_current, 0, 0); | |
89 | musb_writel(&rx->rx_buf_current, 0, 0); | |
90 | musb_writel(&rx->rx_len_len, 0, 0); | |
91 | musb_writel(&rx->rx_cnt_cnt, 0, 0); | |
92 | } | |
93 | ||
94 | /* zero out entire tx state RAM entry for the channel */ | |
95 | static void cppi_reset_tx(struct cppi_tx_stateram __iomem *tx, u32 ptr) | |
96 | { | |
97 | musb_writel(&tx->tx_head, 0, 0); | |
98 | musb_writel(&tx->tx_buf, 0, 0); | |
99 | musb_writel(&tx->tx_current, 0, 0); | |
100 | musb_writel(&tx->tx_buf_current, 0, 0); | |
101 | musb_writel(&tx->tx_info, 0, 0); | |
102 | musb_writel(&tx->tx_rem_len, 0, 0); | |
103 | /* musb_writel(&tx->tx_dummy, 0, 0); */ | |
104 | musb_writel(&tx->tx_complete, 0, ptr); | |
105 | } | |
106 | ||
107 | static void __init cppi_pool_init(struct cppi *cppi, struct cppi_channel *c) | |
108 | { | |
109 | int j; | |
110 | ||
111 | /* initialize channel fields */ | |
112 | c->head = NULL; | |
113 | c->tail = NULL; | |
114 | c->last_processed = NULL; | |
115 | c->channel.status = MUSB_DMA_STATUS_UNKNOWN; | |
116 | c->controller = cppi; | |
117 | c->is_rndis = 0; | |
118 | c->freelist = NULL; | |
119 | ||
120 | /* build the BD Free list for the channel */ | |
121 | for (j = 0; j < NUM_TXCHAN_BD + 1; j++) { | |
122 | struct cppi_descriptor *bd; | |
123 | dma_addr_t dma; | |
124 | ||
125 | bd = dma_pool_alloc(cppi->pool, GFP_KERNEL, &dma); | |
126 | bd->dma = dma; | |
127 | cppi_bd_free(c, bd); | |
128 | } | |
129 | } | |
130 | ||
131 | static int cppi_channel_abort(struct dma_channel *); | |
132 | ||
133 | static void cppi_pool_free(struct cppi_channel *c) | |
134 | { | |
135 | struct cppi *cppi = c->controller; | |
136 | struct cppi_descriptor *bd; | |
137 | ||
138 | (void) cppi_channel_abort(&c->channel); | |
139 | c->channel.status = MUSB_DMA_STATUS_UNKNOWN; | |
140 | c->controller = NULL; | |
141 | ||
142 | /* free all its bds */ | |
143 | bd = c->last_processed; | |
144 | do { | |
145 | if (bd) | |
146 | dma_pool_free(cppi->pool, bd, bd->dma); | |
147 | bd = cppi_bd_alloc(c); | |
148 | } while (bd); | |
149 | c->last_processed = NULL; | |
150 | } | |
151 | ||
152 | static int __init cppi_controller_start(struct dma_controller *c) | |
153 | { | |
154 | struct cppi *controller; | |
155 | void __iomem *tibase; | |
156 | int i; | |
157 | ||
158 | controller = container_of(c, struct cppi, controller); | |
159 | ||
160 | /* do whatever is necessary to start controller */ | |
161 | for (i = 0; i < ARRAY_SIZE(controller->tx); i++) { | |
162 | controller->tx[i].transmit = true; | |
163 | controller->tx[i].index = i; | |
164 | } | |
165 | for (i = 0; i < ARRAY_SIZE(controller->rx); i++) { | |
166 | controller->rx[i].transmit = false; | |
167 | controller->rx[i].index = i; | |
168 | } | |
169 | ||
170 | /* setup BD list on a per channel basis */ | |
171 | for (i = 0; i < ARRAY_SIZE(controller->tx); i++) | |
172 | cppi_pool_init(controller, controller->tx + i); | |
173 | for (i = 0; i < ARRAY_SIZE(controller->rx); i++) | |
174 | cppi_pool_init(controller, controller->rx + i); | |
175 | ||
176 | tibase = controller->tibase; | |
177 | INIT_LIST_HEAD(&controller->tx_complete); | |
178 | ||
179 | /* initialise tx/rx channel head pointers to zero */ | |
180 | for (i = 0; i < ARRAY_SIZE(controller->tx); i++) { | |
181 | struct cppi_channel *tx_ch = controller->tx + i; | |
182 | struct cppi_tx_stateram __iomem *tx; | |
183 | ||
184 | INIT_LIST_HEAD(&tx_ch->tx_complete); | |
185 | ||
186 | tx = tibase + DAVINCI_TXCPPI_STATERAM_OFFSET(i); | |
187 | tx_ch->state_ram = tx; | |
188 | cppi_reset_tx(tx, 0); | |
189 | } | |
190 | for (i = 0; i < ARRAY_SIZE(controller->rx); i++) { | |
191 | struct cppi_channel *rx_ch = controller->rx + i; | |
192 | struct cppi_rx_stateram __iomem *rx; | |
193 | ||
194 | INIT_LIST_HEAD(&rx_ch->tx_complete); | |
195 | ||
196 | rx = tibase + DAVINCI_RXCPPI_STATERAM_OFFSET(i); | |
197 | rx_ch->state_ram = rx; | |
198 | cppi_reset_rx(rx); | |
199 | } | |
200 | ||
201 | /* enable individual cppi channels */ | |
202 | musb_writel(tibase, DAVINCI_TXCPPI_INTENAB_REG, | |
203 | DAVINCI_DMA_ALL_CHANNELS_ENABLE); | |
204 | musb_writel(tibase, DAVINCI_RXCPPI_INTENAB_REG, | |
205 | DAVINCI_DMA_ALL_CHANNELS_ENABLE); | |
206 | ||
207 | /* enable tx/rx CPPI control */ | |
208 | musb_writel(tibase, DAVINCI_TXCPPI_CTRL_REG, DAVINCI_DMA_CTRL_ENABLE); | |
209 | musb_writel(tibase, DAVINCI_RXCPPI_CTRL_REG, DAVINCI_DMA_CTRL_ENABLE); | |
210 | ||
211 | /* disable RNDIS mode, also host rx RNDIS autorequest */ | |
212 | musb_writel(tibase, DAVINCI_RNDIS_REG, 0); | |
213 | musb_writel(tibase, DAVINCI_AUTOREQ_REG, 0); | |
214 | ||
215 | return 0; | |
216 | } | |
217 | ||
218 | /* | |
219 | * Stop DMA controller | |
220 | * | |
221 | * De-Init the DMA controller as necessary. | |
222 | */ | |
223 | ||
224 | static int cppi_controller_stop(struct dma_controller *c) | |
225 | { | |
226 | struct cppi *controller; | |
227 | void __iomem *tibase; | |
228 | int i; | |
229 | ||
230 | controller = container_of(c, struct cppi, controller); | |
231 | ||
232 | tibase = controller->tibase; | |
233 | /* DISABLE INDIVIDUAL CHANNEL Interrupts */ | |
234 | musb_writel(tibase, DAVINCI_TXCPPI_INTCLR_REG, | |
235 | DAVINCI_DMA_ALL_CHANNELS_ENABLE); | |
236 | musb_writel(tibase, DAVINCI_RXCPPI_INTCLR_REG, | |
237 | DAVINCI_DMA_ALL_CHANNELS_ENABLE); | |
238 | ||
239 | DBG(1, "Tearing down RX and TX Channels\n"); | |
240 | for (i = 0; i < ARRAY_SIZE(controller->tx); i++) { | |
241 | /* FIXME restructure of txdma to use bds like rxdma */ | |
242 | controller->tx[i].last_processed = NULL; | |
243 | cppi_pool_free(controller->tx + i); | |
244 | } | |
245 | for (i = 0; i < ARRAY_SIZE(controller->rx); i++) | |
246 | cppi_pool_free(controller->rx + i); | |
247 | ||
248 | /* in Tx Case proper teardown is supported. We resort to disabling | |
249 | * Tx/Rx CPPI after cleanup of Tx channels. Before TX teardown is | |
250 | * complete TX CPPI cannot be disabled. | |
251 | */ | |
252 | /*disable tx/rx cppi */ | |
253 | musb_writel(tibase, DAVINCI_TXCPPI_CTRL_REG, DAVINCI_DMA_CTRL_DISABLE); | |
254 | musb_writel(tibase, DAVINCI_RXCPPI_CTRL_REG, DAVINCI_DMA_CTRL_DISABLE); | |
255 | ||
256 | return 0; | |
257 | } | |
258 | ||
259 | /* While dma channel is allocated, we only want the core irqs active | |
260 | * for fault reports, otherwise we'd get irqs that we don't care about. | |
261 | * Except for TX irqs, where dma done != fifo empty and reusable ... | |
262 | * | |
263 | * NOTE: docs don't say either way, but irq masking **enables** irqs. | |
264 | * | |
265 | * REVISIT same issue applies to pure PIO usage too, and non-cppi dma... | |
266 | */ | |
267 | static inline void core_rxirq_disable(void __iomem *tibase, unsigned epnum) | |
268 | { | |
269 | musb_writel(tibase, DAVINCI_USB_INT_MASK_CLR_REG, 1 << (epnum + 8)); | |
270 | } | |
271 | ||
272 | static inline void core_rxirq_enable(void __iomem *tibase, unsigned epnum) | |
273 | { | |
274 | musb_writel(tibase, DAVINCI_USB_INT_MASK_SET_REG, 1 << (epnum + 8)); | |
275 | } | |
276 | ||
277 | ||
278 | /* | |
279 | * Allocate a CPPI Channel for DMA. With CPPI, channels are bound to | |
280 | * each transfer direction of a non-control endpoint, so allocating | |
281 | * (and deallocating) is mostly a way to notice bad housekeeping on | |
282 | * the software side. We assume the irqs are always active. | |
283 | */ | |
284 | static struct dma_channel * | |
285 | cppi_channel_allocate(struct dma_controller *c, | |
286 | struct musb_hw_ep *ep, u8 transmit) | |
287 | { | |
288 | struct cppi *controller; | |
289 | u8 index; | |
290 | struct cppi_channel *cppi_ch; | |
291 | void __iomem *tibase; | |
292 | ||
293 | controller = container_of(c, struct cppi, controller); | |
294 | tibase = controller->tibase; | |
295 | ||
296 | /* ep0 doesn't use DMA; remember cppi indices are 0..N-1 */ | |
297 | index = ep->epnum - 1; | |
298 | ||
299 | /* return the corresponding CPPI Channel Handle, and | |
300 | * probably disable the non-CPPI irq until we need it. | |
301 | */ | |
302 | if (transmit) { | |
303 | if (index >= ARRAY_SIZE(controller->tx)) { | |
304 | DBG(1, "no %cX%d CPPI channel\n", 'T', index); | |
305 | return NULL; | |
306 | } | |
307 | cppi_ch = controller->tx + index; | |
308 | } else { | |
309 | if (index >= ARRAY_SIZE(controller->rx)) { | |
310 | DBG(1, "no %cX%d CPPI channel\n", 'R', index); | |
311 | return NULL; | |
312 | } | |
313 | cppi_ch = controller->rx + index; | |
314 | core_rxirq_disable(tibase, ep->epnum); | |
315 | } | |
316 | ||
317 | /* REVISIT make this an error later once the same driver code works | |
318 | * with the other DMA engine too | |
319 | */ | |
320 | if (cppi_ch->hw_ep) | |
321 | DBG(1, "re-allocating DMA%d %cX channel %p\n", | |
322 | index, transmit ? 'T' : 'R', cppi_ch); | |
323 | cppi_ch->hw_ep = ep; | |
324 | cppi_ch->channel.status = MUSB_DMA_STATUS_FREE; | |
325 | ||
326 | DBG(4, "Allocate CPPI%d %cX\n", index, transmit ? 'T' : 'R'); | |
327 | return &cppi_ch->channel; | |
328 | } | |
329 | ||
330 | /* Release a CPPI Channel. */ | |
331 | static void cppi_channel_release(struct dma_channel *channel) | |
332 | { | |
333 | struct cppi_channel *c; | |
334 | void __iomem *tibase; | |
335 | ||
336 | /* REVISIT: for paranoia, check state and abort if needed... */ | |
337 | ||
338 | c = container_of(channel, struct cppi_channel, channel); | |
339 | tibase = c->controller->tibase; | |
340 | if (!c->hw_ep) | |
341 | DBG(1, "releasing idle DMA channel %p\n", c); | |
342 | else if (!c->transmit) | |
343 | core_rxirq_enable(tibase, c->index + 1); | |
344 | ||
345 | /* for now, leave its cppi IRQ enabled (we won't trigger it) */ | |
346 | c->hw_ep = NULL; | |
347 | channel->status = MUSB_DMA_STATUS_UNKNOWN; | |
348 | } | |
349 | ||
350 | /* Context: controller irqlocked */ | |
351 | static void | |
352 | cppi_dump_rx(int level, struct cppi_channel *c, const char *tag) | |
353 | { | |
354 | void __iomem *base = c->controller->mregs; | |
355 | struct cppi_rx_stateram __iomem *rx = c->state_ram; | |
356 | ||
357 | musb_ep_select(base, c->index + 1); | |
358 | ||
359 | DBG(level, "RX DMA%d%s: %d left, csr %04x, " | |
360 | "%08x H%08x S%08x C%08x, " | |
361 | "B%08x L%08x %08x .. %08x" | |
362 | "\n", | |
363 | c->index, tag, | |
364 | musb_readl(c->controller->tibase, | |
365 | DAVINCI_RXCPPI_BUFCNT0_REG + 4 * c->index), | |
366 | musb_readw(c->hw_ep->regs, MUSB_RXCSR), | |
367 | ||
368 | musb_readl(&rx->rx_skipbytes, 0), | |
369 | musb_readl(&rx->rx_head, 0), | |
370 | musb_readl(&rx->rx_sop, 0), | |
371 | musb_readl(&rx->rx_current, 0), | |
372 | ||
373 | musb_readl(&rx->rx_buf_current, 0), | |
374 | musb_readl(&rx->rx_len_len, 0), | |
375 | musb_readl(&rx->rx_cnt_cnt, 0), | |
376 | musb_readl(&rx->rx_complete, 0) | |
377 | ); | |
378 | } | |
379 | ||
380 | /* Context: controller irqlocked */ | |
381 | static void | |
382 | cppi_dump_tx(int level, struct cppi_channel *c, const char *tag) | |
383 | { | |
384 | void __iomem *base = c->controller->mregs; | |
385 | struct cppi_tx_stateram __iomem *tx = c->state_ram; | |
386 | ||
387 | musb_ep_select(base, c->index + 1); | |
388 | ||
389 | DBG(level, "TX DMA%d%s: csr %04x, " | |
390 | "H%08x S%08x C%08x %08x, " | |
391 | "F%08x L%08x .. %08x" | |
392 | "\n", | |
393 | c->index, tag, | |
394 | musb_readw(c->hw_ep->regs, MUSB_TXCSR), | |
395 | ||
396 | musb_readl(&tx->tx_head, 0), | |
397 | musb_readl(&tx->tx_buf, 0), | |
398 | musb_readl(&tx->tx_current, 0), | |
399 | musb_readl(&tx->tx_buf_current, 0), | |
400 | ||
401 | musb_readl(&tx->tx_info, 0), | |
402 | musb_readl(&tx->tx_rem_len, 0), | |
403 | /* dummy/unused word 6 */ | |
404 | musb_readl(&tx->tx_complete, 0) | |
405 | ); | |
406 | } | |
407 | ||
408 | /* Context: controller irqlocked */ | |
409 | static inline void | |
410 | cppi_rndis_update(struct cppi_channel *c, int is_rx, | |
411 | void __iomem *tibase, int is_rndis) | |
412 | { | |
413 | /* we may need to change the rndis flag for this cppi channel */ | |
414 | if (c->is_rndis != is_rndis) { | |
415 | u32 value = musb_readl(tibase, DAVINCI_RNDIS_REG); | |
416 | u32 temp = 1 << (c->index); | |
417 | ||
418 | if (is_rx) | |
419 | temp <<= 16; | |
420 | if (is_rndis) | |
421 | value |= temp; | |
422 | else | |
423 | value &= ~temp; | |
424 | musb_writel(tibase, DAVINCI_RNDIS_REG, value); | |
425 | c->is_rndis = is_rndis; | |
426 | } | |
427 | } | |
428 | ||
704a1485 | 429 | #ifdef CONFIG_USB_MUSB_DEBUG |
550a7375 FB |
430 | static void cppi_dump_rxbd(const char *tag, struct cppi_descriptor *bd) |
431 | { | |
432 | pr_debug("RXBD/%s %08x: " | |
433 | "nxt %08x buf %08x off.blen %08x opt.plen %08x\n", | |
434 | tag, bd->dma, | |
435 | bd->hw_next, bd->hw_bufp, bd->hw_off_len, | |
436 | bd->hw_options); | |
437 | } | |
704a1485 | 438 | #endif |
550a7375 FB |
439 | |
440 | static void cppi_dump_rxq(int level, const char *tag, struct cppi_channel *rx) | |
441 | { | |
704a1485 | 442 | #ifdef CONFIG_USB_MUSB_DEBUG |
550a7375 FB |
443 | struct cppi_descriptor *bd; |
444 | ||
445 | if (!_dbg_level(level)) | |
446 | return; | |
447 | cppi_dump_rx(level, rx, tag); | |
448 | if (rx->last_processed) | |
449 | cppi_dump_rxbd("last", rx->last_processed); | |
450 | for (bd = rx->head; bd; bd = bd->next) | |
451 | cppi_dump_rxbd("active", bd); | |
452 | #endif | |
453 | } | |
454 | ||
455 | ||
456 | /* NOTE: DaVinci autoreq is ignored except for host side "RNDIS" mode RX; | |
457 | * so we won't ever use it (see "CPPI RX Woes" below). | |
458 | */ | |
459 | static inline int cppi_autoreq_update(struct cppi_channel *rx, | |
460 | void __iomem *tibase, int onepacket, unsigned n_bds) | |
461 | { | |
462 | u32 val; | |
463 | ||
464 | #ifdef RNDIS_RX_IS_USABLE | |
465 | u32 tmp; | |
466 | /* assert(is_host_active(musb)) */ | |
467 | ||
468 | /* start from "AutoReq never" */ | |
469 | tmp = musb_readl(tibase, DAVINCI_AUTOREQ_REG); | |
470 | val = tmp & ~((0x3) << (rx->index * 2)); | |
471 | ||
472 | /* HCD arranged reqpkt for packet #1. we arrange int | |
473 | * for all but the last one, maybe in two segments. | |
474 | */ | |
475 | if (!onepacket) { | |
476 | #if 0 | |
477 | /* use two segments, autoreq "all" then the last "never" */ | |
478 | val |= ((0x3) << (rx->index * 2)); | |
479 | n_bds--; | |
480 | #else | |
481 | /* one segment, autoreq "all-but-last" */ | |
482 | val |= ((0x1) << (rx->index * 2)); | |
483 | #endif | |
484 | } | |
485 | ||
486 | if (val != tmp) { | |
487 | int n = 100; | |
488 | ||
489 | /* make sure that autoreq is updated before continuing */ | |
490 | musb_writel(tibase, DAVINCI_AUTOREQ_REG, val); | |
491 | do { | |
492 | tmp = musb_readl(tibase, DAVINCI_AUTOREQ_REG); | |
493 | if (tmp == val) | |
494 | break; | |
495 | cpu_relax(); | |
496 | } while (n-- > 0); | |
497 | } | |
498 | #endif | |
499 | ||
500 | /* REQPKT is turned off after each segment */ | |
501 | if (n_bds && rx->channel.actual_len) { | |
502 | void __iomem *regs = rx->hw_ep->regs; | |
503 | ||
504 | val = musb_readw(regs, MUSB_RXCSR); | |
505 | if (!(val & MUSB_RXCSR_H_REQPKT)) { | |
506 | val |= MUSB_RXCSR_H_REQPKT | MUSB_RXCSR_H_WZC_BITS; | |
507 | musb_writew(regs, MUSB_RXCSR, val); | |
508 | /* flush writebufer */ | |
509 | val = musb_readw(regs, MUSB_RXCSR); | |
510 | } | |
511 | } | |
512 | return n_bds; | |
513 | } | |
514 | ||
515 | ||
516 | /* Buffer enqueuing Logic: | |
517 | * | |
518 | * - RX builds new queues each time, to help handle routine "early | |
519 | * termination" cases (faults, including errors and short reads) | |
520 | * more correctly. | |
521 | * | |
522 | * - for now, TX reuses the same queue of BDs every time | |
523 | * | |
524 | * REVISIT long term, we want a normal dynamic model. | |
525 | * ... the goal will be to append to the | |
526 | * existing queue, processing completed "dma buffers" (segments) on the fly. | |
527 | * | |
528 | * Otherwise we force an IRQ latency between requests, which slows us a lot | |
529 | * (especially in "transparent" dma). Unfortunately that model seems to be | |
530 | * inherent in the DMA model from the Mentor code, except in the rare case | |
531 | * of transfers big enough (~128+ KB) that we could append "middle" segments | |
532 | * in the TX paths. (RX can't do this, see below.) | |
533 | * | |
534 | * That's true even in the CPPI- friendly iso case, where most urbs have | |
535 | * several small segments provided in a group and where the "packet at a time" | |
536 | * "transparent" DMA model is always correct, even on the RX side. | |
537 | */ | |
538 | ||
539 | /* | |
540 | * CPPI TX: | |
541 | * ======== | |
542 | * TX is a lot more reasonable than RX; it doesn't need to run in | |
543 | * irq-per-packet mode very often. RNDIS mode seems to behave too | |
544 | * (except how it handles the exactly-N-packets case). Building a | |
545 | * txdma queue with multiple requests (urb or usb_request) looks | |
546 | * like it would work ... but fault handling would need much testing. | |
547 | * | |
548 | * The main issue with TX mode RNDIS relates to transfer lengths that | |
549 | * are an exact multiple of the packet length. It appears that there's | |
550 | * a hiccup in that case (maybe the DMA completes before the ZLP gets | |
551 | * written?) boiling down to not being able to rely on CPPI writing any | |
552 | * terminating zero length packet before the next transfer is written. | |
553 | * So that's punted to PIO; better yet, gadget drivers can avoid it. | |
554 | * | |
555 | * Plus, there's allegedly an undocumented constraint that rndis transfer | |
556 | * length be a multiple of 64 bytes ... but the chip doesn't act that | |
557 | * way, and we really don't _want_ that behavior anyway. | |
558 | * | |
559 | * On TX, "transparent" mode works ... although experiments have shown | |
560 | * problems trying to use the SOP/EOP bits in different USB packets. | |
561 | * | |
562 | * REVISIT try to handle terminating zero length packets using CPPI | |
563 | * instead of doing it by PIO after an IRQ. (Meanwhile, make Ethernet | |
564 | * links avoid that issue by forcing them to avoid zlps.) | |
565 | */ | |
566 | static void | |
567 | cppi_next_tx_segment(struct musb *musb, struct cppi_channel *tx) | |
568 | { | |
569 | unsigned maxpacket = tx->maxpacket; | |
570 | dma_addr_t addr = tx->buf_dma + tx->offset; | |
571 | size_t length = tx->buf_len - tx->offset; | |
572 | struct cppi_descriptor *bd; | |
573 | unsigned n_bds; | |
574 | unsigned i; | |
575 | struct cppi_tx_stateram __iomem *tx_ram = tx->state_ram; | |
576 | int rndis; | |
577 | ||
578 | /* TX can use the CPPI "rndis" mode, where we can probably fit this | |
579 | * transfer in one BD and one IRQ. The only time we would NOT want | |
580 | * to use it is when hardware constraints prevent it, or if we'd | |
581 | * trigger the "send a ZLP?" confusion. | |
582 | */ | |
583 | rndis = (maxpacket & 0x3f) == 0 | |
6b6e9710 | 584 | && length > maxpacket |
550a7375 FB |
585 | && length < 0xffff |
586 | && (length % maxpacket) != 0; | |
587 | ||
588 | if (rndis) { | |
589 | maxpacket = length; | |
590 | n_bds = 1; | |
591 | } else { | |
592 | n_bds = length / maxpacket; | |
593 | if (!length || (length % maxpacket)) | |
594 | n_bds++; | |
595 | n_bds = min(n_bds, (unsigned) NUM_TXCHAN_BD); | |
596 | length = min(n_bds * maxpacket, length); | |
597 | } | |
598 | ||
599 | DBG(4, "TX DMA%d, pktSz %d %s bds %d dma 0x%x len %u\n", | |
600 | tx->index, | |
601 | maxpacket, | |
602 | rndis ? "rndis" : "transparent", | |
603 | n_bds, | |
604 | addr, length); | |
605 | ||
606 | cppi_rndis_update(tx, 0, musb->ctrl_base, rndis); | |
607 | ||
608 | /* assuming here that channel_program is called during | |
609 | * transfer initiation ... current code maintains state | |
610 | * for one outstanding request only (no queues, not even | |
611 | * the implicit ones of an iso urb). | |
612 | */ | |
613 | ||
614 | bd = tx->freelist; | |
615 | tx->head = bd; | |
616 | tx->last_processed = NULL; | |
617 | ||
618 | /* FIXME use BD pool like RX side does, and just queue | |
619 | * the minimum number for this request. | |
620 | */ | |
621 | ||
622 | /* Prepare queue of BDs first, then hand it to hardware. | |
623 | * All BDs except maybe the last should be of full packet | |
624 | * size; for RNDIS there _is_ only that last packet. | |
625 | */ | |
626 | for (i = 0; i < n_bds; ) { | |
627 | if (++i < n_bds && bd->next) | |
628 | bd->hw_next = bd->next->dma; | |
629 | else | |
630 | bd->hw_next = 0; | |
631 | ||
632 | bd->hw_bufp = tx->buf_dma + tx->offset; | |
633 | ||
634 | /* FIXME set EOP only on the last packet, | |
635 | * SOP only on the first ... avoid IRQs | |
636 | */ | |
637 | if ((tx->offset + maxpacket) <= tx->buf_len) { | |
638 | tx->offset += maxpacket; | |
639 | bd->hw_off_len = maxpacket; | |
640 | bd->hw_options = CPPI_SOP_SET | CPPI_EOP_SET | |
641 | | CPPI_OWN_SET | maxpacket; | |
642 | } else { | |
643 | /* only this one may be a partial USB Packet */ | |
644 | u32 partial_len; | |
645 | ||
646 | partial_len = tx->buf_len - tx->offset; | |
647 | tx->offset = tx->buf_len; | |
648 | bd->hw_off_len = partial_len; | |
649 | ||
650 | bd->hw_options = CPPI_SOP_SET | CPPI_EOP_SET | |
651 | | CPPI_OWN_SET | partial_len; | |
652 | if (partial_len == 0) | |
653 | bd->hw_options |= CPPI_ZERO_SET; | |
654 | } | |
655 | ||
656 | DBG(5, "TXBD %p: nxt %08x buf %08x len %04x opt %08x\n", | |
657 | bd, bd->hw_next, bd->hw_bufp, | |
658 | bd->hw_off_len, bd->hw_options); | |
659 | ||
660 | /* update the last BD enqueued to the list */ | |
661 | tx->tail = bd; | |
662 | bd = bd->next; | |
663 | } | |
664 | ||
665 | /* BDs live in DMA-coherent memory, but writes might be pending */ | |
666 | cpu_drain_writebuffer(); | |
667 | ||
668 | /* Write to the HeadPtr in state RAM to trigger */ | |
669 | musb_writel(&tx_ram->tx_head, 0, (u32)tx->freelist->dma); | |
670 | ||
671 | cppi_dump_tx(5, tx, "/S"); | |
672 | } | |
673 | ||
674 | /* | |
675 | * CPPI RX Woes: | |
676 | * ============= | |
677 | * Consider a 1KB bulk RX buffer in two scenarios: (a) it's fed two 300 byte | |
678 | * packets back-to-back, and (b) it's fed two 512 byte packets back-to-back. | |
679 | * (Full speed transfers have similar scenarios.) | |
680 | * | |
681 | * The correct behavior for Linux is that (a) fills the buffer with 300 bytes, | |
682 | * and the next packet goes into a buffer that's queued later; while (b) fills | |
683 | * the buffer with 1024 bytes. How to do that with CPPI? | |
684 | * | |
685 | * - RX queues in "rndis" mode -- one single BD -- handle (a) correctly, but | |
686 | * (b) loses **BADLY** because nothing (!) happens when that second packet | |
687 | * fills the buffer, much less when a third one arrives. (Which makes this | |
688 | * not a "true" RNDIS mode. In the RNDIS protocol short-packet termination | |
689 | * is optional, and it's fine if peripherals -- not hosts! -- pad messages | |
690 | * out to end-of-buffer. Standard PCI host controller DMA descriptors | |
691 | * implement that mode by default ... which is no accident.) | |
692 | * | |
693 | * - RX queues in "transparent" mode -- two BDs with 512 bytes each -- have | |
694 | * converse problems: (b) is handled right, but (a) loses badly. CPPI RX | |
695 | * ignores SOP/EOP markings and processes both of those BDs; so both packets | |
696 | * are loaded into the buffer (with a 212 byte gap between them), and the next | |
697 | * buffer queued will NOT get its 300 bytes of data. (It seems like SOP/EOP | |
698 | * are intended as outputs for RX queues, not inputs...) | |
699 | * | |
700 | * - A variant of "transparent" mode -- one BD at a time -- is the only way to | |
701 | * reliably make both cases work, with software handling both cases correctly | |
702 | * and at the significant penalty of needing an IRQ per packet. (The lack of | |
703 | * I/O overlap can be slightly ameliorated by enabling double buffering.) | |
704 | * | |
705 | * So how to get rid of IRQ-per-packet? The transparent multi-BD case could | |
706 | * be used in special cases like mass storage, which sets URB_SHORT_NOT_OK | |
707 | * (or maybe its peripheral side counterpart) to flag (a) scenarios as errors | |
708 | * with guaranteed driver level fault recovery and scrubbing out what's left | |
709 | * of that garbaged datastream. | |
710 | * | |
711 | * But there seems to be no way to identify the cases where CPPI RNDIS mode | |
712 | * is appropriate -- which do NOT include RNDIS host drivers, but do include | |
713 | * the CDC Ethernet driver! -- and the documentation is incomplete/wrong. | |
714 | * So we can't _ever_ use RX RNDIS mode ... except by using a heuristic | |
715 | * that applies best on the peripheral side (and which could fail rudely). | |
716 | * | |
717 | * Leaving only "transparent" mode; we avoid multi-bd modes in almost all | |
718 | * cases other than mass storage class. Otherwise we're correct but slow, | |
719 | * since CPPI penalizes our need for a "true RNDIS" default mode. | |
720 | */ | |
721 | ||
722 | ||
723 | /* Heuristic, intended to kick in for ethernet/rndis peripheral ONLY | |
724 | * | |
725 | * IFF | |
726 | * (a) peripheral mode ... since rndis peripherals could pad their | |
727 | * writes to hosts, causing i/o failure; or we'd have to cope with | |
728 | * a largely unknowable variety of host side protocol variants | |
729 | * (b) and short reads are NOT errors ... since full reads would | |
730 | * cause those same i/o failures | |
731 | * (c) and read length is | |
732 | * - less than 64KB (max per cppi descriptor) | |
733 | * - not a multiple of 4096 (g_zero default, full reads typical) | |
734 | * - N (>1) packets long, ditto (full reads not EXPECTED) | |
735 | * THEN | |
736 | * try rx rndis mode | |
737 | * | |
738 | * Cost of heuristic failing: RXDMA wedges at the end of transfers that | |
739 | * fill out the whole buffer. Buggy host side usb network drivers could | |
740 | * trigger that, but "in the field" such bugs seem to be all but unknown. | |
741 | * | |
742 | * So this module parameter lets the heuristic be disabled. When using | |
743 | * gadgetfs, the heuristic will probably need to be disabled. | |
744 | */ | |
745 | static int cppi_rx_rndis = 1; | |
746 | ||
747 | module_param(cppi_rx_rndis, bool, 0); | |
748 | MODULE_PARM_DESC(cppi_rx_rndis, "enable/disable RX RNDIS heuristic"); | |
749 | ||
750 | ||
751 | /** | |
752 | * cppi_next_rx_segment - dma read for the next chunk of a buffer | |
753 | * @musb: the controller | |
754 | * @rx: dma channel | |
755 | * @onepacket: true unless caller treats short reads as errors, and | |
756 | * performs fault recovery above usbcore. | |
757 | * Context: controller irqlocked | |
758 | * | |
759 | * See above notes about why we can't use multi-BD RX queues except in | |
760 | * rare cases (mass storage class), and can never use the hardware "rndis" | |
761 | * mode (since it's not a "true" RNDIS mode) with complete safety.. | |
762 | * | |
763 | * It's ESSENTIAL that callers specify "onepacket" mode unless they kick in | |
764 | * code to recover from corrupted datastreams after each short transfer. | |
765 | */ | |
766 | static void | |
767 | cppi_next_rx_segment(struct musb *musb, struct cppi_channel *rx, int onepacket) | |
768 | { | |
769 | unsigned maxpacket = rx->maxpacket; | |
770 | dma_addr_t addr = rx->buf_dma + rx->offset; | |
771 | size_t length = rx->buf_len - rx->offset; | |
772 | struct cppi_descriptor *bd, *tail; | |
773 | unsigned n_bds; | |
774 | unsigned i; | |
775 | void __iomem *tibase = musb->ctrl_base; | |
776 | int is_rndis = 0; | |
777 | struct cppi_rx_stateram __iomem *rx_ram = rx->state_ram; | |
778 | ||
779 | if (onepacket) { | |
780 | /* almost every USB driver, host or peripheral side */ | |
781 | n_bds = 1; | |
782 | ||
783 | /* maybe apply the heuristic above */ | |
784 | if (cppi_rx_rndis | |
785 | && is_peripheral_active(musb) | |
786 | && length > maxpacket | |
787 | && (length & ~0xffff) == 0 | |
788 | && (length & 0x0fff) != 0 | |
789 | && (length & (maxpacket - 1)) == 0) { | |
790 | maxpacket = length; | |
791 | is_rndis = 1; | |
792 | } | |
793 | } else { | |
794 | /* virtually nothing except mass storage class */ | |
795 | if (length > 0xffff) { | |
796 | n_bds = 0xffff / maxpacket; | |
797 | length = n_bds * maxpacket; | |
798 | } else { | |
799 | n_bds = length / maxpacket; | |
800 | if (length % maxpacket) | |
801 | n_bds++; | |
802 | } | |
803 | if (n_bds == 1) | |
804 | onepacket = 1; | |
805 | else | |
806 | n_bds = min(n_bds, (unsigned) NUM_RXCHAN_BD); | |
807 | } | |
808 | ||
809 | /* In host mode, autorequest logic can generate some IN tokens; it's | |
810 | * tricky since we can't leave REQPKT set in RXCSR after the transfer | |
811 | * finishes. So: multipacket transfers involve two or more segments. | |
812 | * And always at least two IRQs ... RNDIS mode is not an option. | |
813 | */ | |
814 | if (is_host_active(musb)) | |
815 | n_bds = cppi_autoreq_update(rx, tibase, onepacket, n_bds); | |
816 | ||
817 | cppi_rndis_update(rx, 1, musb->ctrl_base, is_rndis); | |
818 | ||
819 | length = min(n_bds * maxpacket, length); | |
820 | ||
821 | DBG(4, "RX DMA%d seg, maxp %d %s bds %d (cnt %d) " | |
822 | "dma 0x%x len %u %u/%u\n", | |
823 | rx->index, maxpacket, | |
824 | onepacket | |
825 | ? (is_rndis ? "rndis" : "onepacket") | |
826 | : "multipacket", | |
827 | n_bds, | |
828 | musb_readl(tibase, | |
829 | DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4)) | |
830 | & 0xffff, | |
831 | addr, length, rx->channel.actual_len, rx->buf_len); | |
832 | ||
833 | /* only queue one segment at a time, since the hardware prevents | |
834 | * correct queue shutdown after unexpected short packets | |
835 | */ | |
836 | bd = cppi_bd_alloc(rx); | |
837 | rx->head = bd; | |
838 | ||
839 | /* Build BDs for all packets in this segment */ | |
840 | for (i = 0, tail = NULL; bd && i < n_bds; i++, tail = bd) { | |
841 | u32 bd_len; | |
842 | ||
843 | if (i) { | |
844 | bd = cppi_bd_alloc(rx); | |
845 | if (!bd) | |
846 | break; | |
847 | tail->next = bd; | |
848 | tail->hw_next = bd->dma; | |
849 | } | |
850 | bd->hw_next = 0; | |
851 | ||
852 | /* all but the last packet will be maxpacket size */ | |
853 | if (maxpacket < length) | |
854 | bd_len = maxpacket; | |
855 | else | |
856 | bd_len = length; | |
857 | ||
858 | bd->hw_bufp = addr; | |
859 | addr += bd_len; | |
860 | rx->offset += bd_len; | |
861 | ||
862 | bd->hw_off_len = (0 /*offset*/ << 16) + bd_len; | |
863 | bd->buflen = bd_len; | |
864 | ||
865 | bd->hw_options = CPPI_OWN_SET | (i == 0 ? length : 0); | |
866 | length -= bd_len; | |
867 | } | |
868 | ||
869 | /* we always expect at least one reusable BD! */ | |
870 | if (!tail) { | |
871 | WARNING("rx dma%d -- no BDs? need %d\n", rx->index, n_bds); | |
872 | return; | |
873 | } else if (i < n_bds) | |
874 | WARNING("rx dma%d -- only %d of %d BDs\n", rx->index, i, n_bds); | |
875 | ||
876 | tail->next = NULL; | |
877 | tail->hw_next = 0; | |
878 | ||
879 | bd = rx->head; | |
880 | rx->tail = tail; | |
881 | ||
882 | /* short reads and other faults should terminate this entire | |
883 | * dma segment. we want one "dma packet" per dma segment, not | |
884 | * one per USB packet, terminating the whole queue at once... | |
885 | * NOTE that current hardware seems to ignore SOP and EOP. | |
886 | */ | |
887 | bd->hw_options |= CPPI_SOP_SET; | |
888 | tail->hw_options |= CPPI_EOP_SET; | |
889 | ||
704a1485 HV |
890 | #ifdef CONFIG_USB_MUSB_DEBUG |
891 | if (_dbg_level(5)) { | |
550a7375 FB |
892 | struct cppi_descriptor *d; |
893 | ||
894 | for (d = rx->head; d; d = d->next) | |
895 | cppi_dump_rxbd("S", d); | |
896 | } | |
704a1485 | 897 | #endif |
550a7375 FB |
898 | |
899 | /* in case the preceding transfer left some state... */ | |
900 | tail = rx->last_processed; | |
901 | if (tail) { | |
902 | tail->next = bd; | |
903 | tail->hw_next = bd->dma; | |
904 | } | |
905 | ||
906 | core_rxirq_enable(tibase, rx->index + 1); | |
907 | ||
908 | /* BDs live in DMA-coherent memory, but writes might be pending */ | |
909 | cpu_drain_writebuffer(); | |
910 | ||
911 | /* REVISIT specs say to write this AFTER the BUFCNT register | |
912 | * below ... but that loses badly. | |
913 | */ | |
914 | musb_writel(&rx_ram->rx_head, 0, bd->dma); | |
915 | ||
916 | /* bufferCount must be at least 3, and zeroes on completion | |
917 | * unless it underflows below zero, or stops at two, or keeps | |
918 | * growing ... grr. | |
919 | */ | |
920 | i = musb_readl(tibase, | |
921 | DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4)) | |
922 | & 0xffff; | |
923 | ||
924 | if (!i) | |
925 | musb_writel(tibase, | |
926 | DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4), | |
927 | n_bds + 2); | |
928 | else if (n_bds > (i - 3)) | |
929 | musb_writel(tibase, | |
930 | DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4), | |
931 | n_bds - (i - 3)); | |
932 | ||
933 | i = musb_readl(tibase, | |
934 | DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4)) | |
935 | & 0xffff; | |
936 | if (i < (2 + n_bds)) { | |
937 | DBG(2, "bufcnt%d underrun - %d (for %d)\n", | |
938 | rx->index, i, n_bds); | |
939 | musb_writel(tibase, | |
940 | DAVINCI_RXCPPI_BUFCNT0_REG + (rx->index * 4), | |
941 | n_bds + 2); | |
942 | } | |
943 | ||
944 | cppi_dump_rx(4, rx, "/S"); | |
945 | } | |
946 | ||
947 | /** | |
948 | * cppi_channel_program - program channel for data transfer | |
949 | * @ch: the channel | |
950 | * @maxpacket: max packet size | |
951 | * @mode: For RX, 1 unless the usb protocol driver promised to treat | |
952 | * all short reads as errors and kick in high level fault recovery. | |
953 | * For TX, ignored because of RNDIS mode races/glitches. | |
954 | * @dma_addr: dma address of buffer | |
955 | * @len: length of buffer | |
956 | * Context: controller irqlocked | |
957 | */ | |
958 | static int cppi_channel_program(struct dma_channel *ch, | |
959 | u16 maxpacket, u8 mode, | |
960 | dma_addr_t dma_addr, u32 len) | |
961 | { | |
962 | struct cppi_channel *cppi_ch; | |
963 | struct cppi *controller; | |
964 | struct musb *musb; | |
965 | ||
966 | cppi_ch = container_of(ch, struct cppi_channel, channel); | |
967 | controller = cppi_ch->controller; | |
968 | musb = controller->musb; | |
969 | ||
970 | switch (ch->status) { | |
971 | case MUSB_DMA_STATUS_BUS_ABORT: | |
972 | case MUSB_DMA_STATUS_CORE_ABORT: | |
973 | /* fault irq handler should have handled cleanup */ | |
974 | WARNING("%cX DMA%d not cleaned up after abort!\n", | |
975 | cppi_ch->transmit ? 'T' : 'R', | |
976 | cppi_ch->index); | |
977 | /* WARN_ON(1); */ | |
978 | break; | |
979 | case MUSB_DMA_STATUS_BUSY: | |
980 | WARNING("program active channel? %cX DMA%d\n", | |
981 | cppi_ch->transmit ? 'T' : 'R', | |
982 | cppi_ch->index); | |
983 | /* WARN_ON(1); */ | |
984 | break; | |
985 | case MUSB_DMA_STATUS_UNKNOWN: | |
986 | DBG(1, "%cX DMA%d not allocated!\n", | |
987 | cppi_ch->transmit ? 'T' : 'R', | |
988 | cppi_ch->index); | |
989 | /* FALLTHROUGH */ | |
990 | case MUSB_DMA_STATUS_FREE: | |
991 | break; | |
992 | } | |
993 | ||
994 | ch->status = MUSB_DMA_STATUS_BUSY; | |
995 | ||
996 | /* set transfer parameters, then queue up its first segment */ | |
997 | cppi_ch->buf_dma = dma_addr; | |
998 | cppi_ch->offset = 0; | |
999 | cppi_ch->maxpacket = maxpacket; | |
1000 | cppi_ch->buf_len = len; | |
191b7766 | 1001 | cppi_ch->channel.actual_len = 0; |
550a7375 FB |
1002 | |
1003 | /* TX channel? or RX? */ | |
1004 | if (cppi_ch->transmit) | |
1005 | cppi_next_tx_segment(musb, cppi_ch); | |
1006 | else | |
1007 | cppi_next_rx_segment(musb, cppi_ch, mode); | |
1008 | ||
1009 | return true; | |
1010 | } | |
1011 | ||
1012 | static bool cppi_rx_scan(struct cppi *cppi, unsigned ch) | |
1013 | { | |
1014 | struct cppi_channel *rx = &cppi->rx[ch]; | |
1015 | struct cppi_rx_stateram __iomem *state = rx->state_ram; | |
1016 | struct cppi_descriptor *bd; | |
1017 | struct cppi_descriptor *last = rx->last_processed; | |
1018 | bool completed = false; | |
1019 | bool acked = false; | |
1020 | int i; | |
1021 | dma_addr_t safe2ack; | |
1022 | void __iomem *regs = rx->hw_ep->regs; | |
1023 | ||
1024 | cppi_dump_rx(6, rx, "/K"); | |
1025 | ||
1026 | bd = last ? last->next : rx->head; | |
1027 | if (!bd) | |
1028 | return false; | |
1029 | ||
1030 | /* run through all completed BDs */ | |
1031 | for (i = 0, safe2ack = musb_readl(&state->rx_complete, 0); | |
1032 | (safe2ack || completed) && bd && i < NUM_RXCHAN_BD; | |
1033 | i++, bd = bd->next) { | |
1034 | u16 len; | |
1035 | ||
1036 | /* catch latest BD writes from CPPI */ | |
1037 | rmb(); | |
1038 | if (!completed && (bd->hw_options & CPPI_OWN_SET)) | |
1039 | break; | |
1040 | ||
1041 | DBG(5, "C/RXBD %08x: nxt %08x buf %08x " | |
1042 | "off.len %08x opt.len %08x (%d)\n", | |
1043 | bd->dma, bd->hw_next, bd->hw_bufp, | |
1044 | bd->hw_off_len, bd->hw_options, | |
1045 | rx->channel.actual_len); | |
1046 | ||
1047 | /* actual packet received length */ | |
1048 | if ((bd->hw_options & CPPI_SOP_SET) && !completed) | |
1049 | len = bd->hw_off_len & CPPI_RECV_PKTLEN_MASK; | |
1050 | else | |
1051 | len = 0; | |
1052 | ||
1053 | if (bd->hw_options & CPPI_EOQ_MASK) | |
1054 | completed = true; | |
1055 | ||
1056 | if (!completed && len < bd->buflen) { | |
1057 | /* NOTE: when we get a short packet, RXCSR_H_REQPKT | |
1058 | * must have been cleared, and no more DMA packets may | |
1059 | * active be in the queue... TI docs didn't say, but | |
1060 | * CPPI ignores those BDs even though OWN is still set. | |
1061 | */ | |
1062 | completed = true; | |
1063 | DBG(3, "rx short %d/%d (%d)\n", | |
1064 | len, bd->buflen, | |
1065 | rx->channel.actual_len); | |
1066 | } | |
1067 | ||
1068 | /* If we got here, we expect to ack at least one BD; meanwhile | |
1069 | * CPPI may completing other BDs while we scan this list... | |
1070 | * | |
1071 | * RACE: we can notice OWN cleared before CPPI raises the | |
1072 | * matching irq by writing that BD as the completion pointer. | |
1073 | * In such cases, stop scanning and wait for the irq, avoiding | |
1074 | * lost acks and states where BD ownership is unclear. | |
1075 | */ | |
1076 | if (bd->dma == safe2ack) { | |
1077 | musb_writel(&state->rx_complete, 0, safe2ack); | |
1078 | safe2ack = musb_readl(&state->rx_complete, 0); | |
1079 | acked = true; | |
1080 | if (bd->dma == safe2ack) | |
1081 | safe2ack = 0; | |
1082 | } | |
1083 | ||
1084 | rx->channel.actual_len += len; | |
1085 | ||
1086 | cppi_bd_free(rx, last); | |
1087 | last = bd; | |
1088 | ||
1089 | /* stop scanning on end-of-segment */ | |
1090 | if (bd->hw_next == 0) | |
1091 | completed = true; | |
1092 | } | |
1093 | rx->last_processed = last; | |
1094 | ||
1095 | /* dma abort, lost ack, or ... */ | |
1096 | if (!acked && last) { | |
1097 | int csr; | |
1098 | ||
1099 | if (safe2ack == 0 || safe2ack == rx->last_processed->dma) | |
1100 | musb_writel(&state->rx_complete, 0, safe2ack); | |
1101 | if (safe2ack == 0) { | |
1102 | cppi_bd_free(rx, last); | |
1103 | rx->last_processed = NULL; | |
1104 | ||
1105 | /* if we land here on the host side, H_REQPKT will | |
1106 | * be clear and we need to restart the queue... | |
1107 | */ | |
1108 | WARN_ON(rx->head); | |
1109 | } | |
1110 | musb_ep_select(cppi->mregs, rx->index + 1); | |
1111 | csr = musb_readw(regs, MUSB_RXCSR); | |
1112 | if (csr & MUSB_RXCSR_DMAENAB) { | |
1113 | DBG(4, "list%d %p/%p, last %08x%s, csr %04x\n", | |
1114 | rx->index, | |
1115 | rx->head, rx->tail, | |
1116 | rx->last_processed | |
1117 | ? rx->last_processed->dma | |
1118 | : 0, | |
1119 | completed ? ", completed" : "", | |
1120 | csr); | |
1121 | cppi_dump_rxq(4, "/what?", rx); | |
1122 | } | |
1123 | } | |
1124 | if (!completed) { | |
1125 | int csr; | |
1126 | ||
1127 | rx->head = bd; | |
1128 | ||
1129 | /* REVISIT seems like "autoreq all but EOP" doesn't... | |
1130 | * setting it here "should" be racey, but seems to work | |
1131 | */ | |
1132 | csr = musb_readw(rx->hw_ep->regs, MUSB_RXCSR); | |
1133 | if (is_host_active(cppi->musb) | |
1134 | && bd | |
1135 | && !(csr & MUSB_RXCSR_H_REQPKT)) { | |
1136 | csr |= MUSB_RXCSR_H_REQPKT; | |
1137 | musb_writew(regs, MUSB_RXCSR, | |
1138 | MUSB_RXCSR_H_WZC_BITS | csr); | |
1139 | csr = musb_readw(rx->hw_ep->regs, MUSB_RXCSR); | |
1140 | } | |
1141 | } else { | |
1142 | rx->head = NULL; | |
1143 | rx->tail = NULL; | |
1144 | } | |
1145 | ||
1146 | cppi_dump_rx(6, rx, completed ? "/completed" : "/cleaned"); | |
1147 | return completed; | |
1148 | } | |
1149 | ||
91e9c4fe | 1150 | irqreturn_t cppi_interrupt(int irq, void *dev_id) |
550a7375 | 1151 | { |
91e9c4fe | 1152 | struct musb *musb = dev_id; |
550a7375 | 1153 | struct cppi *cppi; |
91e9c4fe | 1154 | void __iomem *tibase; |
550a7375 | 1155 | struct musb_hw_ep *hw_ep = NULL; |
91e9c4fe SS |
1156 | u32 rx, tx; |
1157 | int i, index; | |
93aa3dab | 1158 | unsigned long flags; |
550a7375 FB |
1159 | |
1160 | cppi = container_of(musb->dma_controller, struct cppi, controller); | |
93aa3dab S |
1161 | if (cppi->irq) |
1162 | spin_lock_irqsave(&musb->lock, flags); | |
550a7375 FB |
1163 | |
1164 | tibase = musb->ctrl_base; | |
1165 | ||
91e9c4fe SS |
1166 | tx = musb_readl(tibase, DAVINCI_TXCPPI_MASKED_REG); |
1167 | rx = musb_readl(tibase, DAVINCI_RXCPPI_MASKED_REG); | |
1168 | ||
1169 | if (!tx && !rx) | |
1170 | return IRQ_NONE; | |
1171 | ||
1172 | DBG(4, "CPPI IRQ Tx%x Rx%x\n", tx, rx); | |
1173 | ||
550a7375 FB |
1174 | /* process TX channels */ |
1175 | for (index = 0; tx; tx = tx >> 1, index++) { | |
1176 | struct cppi_channel *tx_ch; | |
1177 | struct cppi_tx_stateram __iomem *tx_ram; | |
1178 | bool completed = false; | |
1179 | struct cppi_descriptor *bd; | |
1180 | ||
1181 | if (!(tx & 1)) | |
1182 | continue; | |
1183 | ||
1184 | tx_ch = cppi->tx + index; | |
1185 | tx_ram = tx_ch->state_ram; | |
1186 | ||
1187 | /* FIXME need a cppi_tx_scan() routine, which | |
1188 | * can also be called from abort code | |
1189 | */ | |
1190 | ||
1191 | cppi_dump_tx(5, tx_ch, "/E"); | |
1192 | ||
1193 | bd = tx_ch->head; | |
1194 | ||
56596923 S |
1195 | /* |
1196 | * If Head is null then this could mean that a abort interrupt | |
1197 | * that needs to be acknowledged. | |
1198 | */ | |
550a7375 FB |
1199 | if (NULL == bd) { |
1200 | DBG(1, "null BD\n"); | |
56596923 | 1201 | tx_ram->tx_complete = 0; |
550a7375 FB |
1202 | continue; |
1203 | } | |
1204 | ||
1205 | /* run through all completed BDs */ | |
1206 | for (i = 0; !completed && bd && i < NUM_TXCHAN_BD; | |
1207 | i++, bd = bd->next) { | |
1208 | u16 len; | |
1209 | ||
1210 | /* catch latest BD writes from CPPI */ | |
1211 | rmb(); | |
1212 | if (bd->hw_options & CPPI_OWN_SET) | |
1213 | break; | |
1214 | ||
1215 | DBG(5, "C/TXBD %p n %x b %x off %x opt %x\n", | |
1216 | bd, bd->hw_next, bd->hw_bufp, | |
1217 | bd->hw_off_len, bd->hw_options); | |
1218 | ||
1219 | len = bd->hw_off_len & CPPI_BUFFER_LEN_MASK; | |
1220 | tx_ch->channel.actual_len += len; | |
1221 | ||
1222 | tx_ch->last_processed = bd; | |
1223 | ||
1224 | /* write completion register to acknowledge | |
1225 | * processing of completed BDs, and possibly | |
1226 | * release the IRQ; EOQ might not be set ... | |
1227 | * | |
1228 | * REVISIT use the same ack strategy as rx | |
1229 | * | |
1230 | * REVISIT have observed bit 18 set; huh?? | |
1231 | */ | |
1232 | /* if ((bd->hw_options & CPPI_EOQ_MASK)) */ | |
1233 | musb_writel(&tx_ram->tx_complete, 0, bd->dma); | |
1234 | ||
1235 | /* stop scanning on end-of-segment */ | |
1236 | if (bd->hw_next == 0) | |
1237 | completed = true; | |
1238 | } | |
1239 | ||
1240 | /* on end of segment, maybe go to next one */ | |
1241 | if (completed) { | |
1242 | /* cppi_dump_tx(4, tx_ch, "/complete"); */ | |
1243 | ||
1244 | /* transfer more, or report completion */ | |
1245 | if (tx_ch->offset >= tx_ch->buf_len) { | |
1246 | tx_ch->head = NULL; | |
1247 | tx_ch->tail = NULL; | |
1248 | tx_ch->channel.status = MUSB_DMA_STATUS_FREE; | |
1249 | ||
1250 | hw_ep = tx_ch->hw_ep; | |
1251 | ||
c7bbc056 | 1252 | musb_dma_completion(musb, index + 1, 1); |
550a7375 FB |
1253 | |
1254 | } else { | |
1255 | /* Bigger transfer than we could fit in | |
1256 | * that first batch of descriptors... | |
1257 | */ | |
1258 | cppi_next_tx_segment(musb, tx_ch); | |
1259 | } | |
1260 | } else | |
1261 | tx_ch->head = bd; | |
1262 | } | |
1263 | ||
1264 | /* Start processing the RX block */ | |
1265 | for (index = 0; rx; rx = rx >> 1, index++) { | |
1266 | ||
1267 | if (rx & 1) { | |
1268 | struct cppi_channel *rx_ch; | |
1269 | ||
1270 | rx_ch = cppi->rx + index; | |
1271 | ||
1272 | /* let incomplete dma segments finish */ | |
1273 | if (!cppi_rx_scan(cppi, index)) | |
1274 | continue; | |
1275 | ||
1276 | /* start another dma segment if needed */ | |
1277 | if (rx_ch->channel.actual_len != rx_ch->buf_len | |
1278 | && rx_ch->channel.actual_len | |
1279 | == rx_ch->offset) { | |
1280 | cppi_next_rx_segment(musb, rx_ch, 1); | |
1281 | continue; | |
1282 | } | |
1283 | ||
1284 | /* all segments completed! */ | |
1285 | rx_ch->channel.status = MUSB_DMA_STATUS_FREE; | |
1286 | ||
1287 | hw_ep = rx_ch->hw_ep; | |
1288 | ||
1289 | core_rxirq_disable(tibase, index + 1); | |
1290 | musb_dma_completion(musb, index + 1, 0); | |
1291 | } | |
1292 | } | |
1293 | ||
1294 | /* write to CPPI EOI register to re-enable interrupts */ | |
1295 | musb_writel(tibase, DAVINCI_CPPI_EOI_REG, 0); | |
91e9c4fe | 1296 | |
93aa3dab S |
1297 | if (cppi->irq) |
1298 | spin_unlock_irqrestore(&musb->lock, flags); | |
1299 | ||
91e9c4fe | 1300 | return IRQ_HANDLED; |
550a7375 FB |
1301 | } |
1302 | ||
1303 | /* Instantiate a software object representing a DMA controller. */ | |
1304 | struct dma_controller *__init | |
1305 | dma_controller_create(struct musb *musb, void __iomem *mregs) | |
1306 | { | |
1307 | struct cppi *controller; | |
91e9c4fe SS |
1308 | struct device *dev = musb->controller; |
1309 | struct platform_device *pdev = to_platform_device(dev); | |
1310 | int irq = platform_get_irq(pdev, 1); | |
550a7375 FB |
1311 | |
1312 | controller = kzalloc(sizeof *controller, GFP_KERNEL); | |
1313 | if (!controller) | |
1314 | return NULL; | |
1315 | ||
1316 | controller->mregs = mregs; | |
1317 | controller->tibase = mregs - DAVINCI_BASE_OFFSET; | |
1318 | ||
1319 | controller->musb = musb; | |
1320 | controller->controller.start = cppi_controller_start; | |
1321 | controller->controller.stop = cppi_controller_stop; | |
1322 | controller->controller.channel_alloc = cppi_channel_allocate; | |
1323 | controller->controller.channel_release = cppi_channel_release; | |
1324 | controller->controller.channel_program = cppi_channel_program; | |
1325 | controller->controller.channel_abort = cppi_channel_abort; | |
1326 | ||
1327 | /* NOTE: allocating from on-chip SRAM would give the least | |
1328 | * contention for memory access, if that ever matters here. | |
1329 | */ | |
1330 | ||
1331 | /* setup BufferPool */ | |
1332 | controller->pool = dma_pool_create("cppi", | |
1333 | controller->musb->controller, | |
1334 | sizeof(struct cppi_descriptor), | |
1335 | CPPI_DESCRIPTOR_ALIGN, 0); | |
1336 | if (!controller->pool) { | |
1337 | kfree(controller); | |
1338 | return NULL; | |
1339 | } | |
1340 | ||
91e9c4fe SS |
1341 | if (irq > 0) { |
1342 | if (request_irq(irq, cppi_interrupt, 0, "cppi-dma", musb)) { | |
1343 | dev_err(dev, "request_irq %d failed!\n", irq); | |
1344 | dma_controller_destroy(&controller->controller); | |
1345 | return NULL; | |
1346 | } | |
1347 | controller->irq = irq; | |
1348 | } | |
1349 | ||
550a7375 FB |
1350 | return &controller->controller; |
1351 | } | |
1352 | ||
1353 | /* | |
1354 | * Destroy a previously-instantiated DMA controller. | |
1355 | */ | |
1356 | void dma_controller_destroy(struct dma_controller *c) | |
1357 | { | |
1358 | struct cppi *cppi; | |
1359 | ||
1360 | cppi = container_of(c, struct cppi, controller); | |
1361 | ||
91e9c4fe SS |
1362 | if (cppi->irq) |
1363 | free_irq(cppi->irq, cppi->musb); | |
1364 | ||
550a7375 FB |
1365 | /* assert: caller stopped the controller first */ |
1366 | dma_pool_destroy(cppi->pool); | |
1367 | ||
1368 | kfree(cppi); | |
1369 | } | |
1370 | ||
1371 | /* | |
1372 | * Context: controller irqlocked, endpoint selected | |
1373 | */ | |
1374 | static int cppi_channel_abort(struct dma_channel *channel) | |
1375 | { | |
1376 | struct cppi_channel *cppi_ch; | |
1377 | struct cppi *controller; | |
1378 | void __iomem *mbase; | |
1379 | void __iomem *tibase; | |
1380 | void __iomem *regs; | |
1381 | u32 value; | |
1382 | struct cppi_descriptor *queue; | |
1383 | ||
1384 | cppi_ch = container_of(channel, struct cppi_channel, channel); | |
1385 | ||
1386 | controller = cppi_ch->controller; | |
1387 | ||
1388 | switch (channel->status) { | |
1389 | case MUSB_DMA_STATUS_BUS_ABORT: | |
1390 | case MUSB_DMA_STATUS_CORE_ABORT: | |
1391 | /* from RX or TX fault irq handler */ | |
1392 | case MUSB_DMA_STATUS_BUSY: | |
1393 | /* the hardware needs shutting down */ | |
1394 | regs = cppi_ch->hw_ep->regs; | |
1395 | break; | |
1396 | case MUSB_DMA_STATUS_UNKNOWN: | |
1397 | case MUSB_DMA_STATUS_FREE: | |
1398 | return 0; | |
1399 | default: | |
1400 | return -EINVAL; | |
1401 | } | |
1402 | ||
1403 | if (!cppi_ch->transmit && cppi_ch->head) | |
1404 | cppi_dump_rxq(3, "/abort", cppi_ch); | |
1405 | ||
1406 | mbase = controller->mregs; | |
1407 | tibase = controller->tibase; | |
1408 | ||
1409 | queue = cppi_ch->head; | |
1410 | cppi_ch->head = NULL; | |
1411 | cppi_ch->tail = NULL; | |
1412 | ||
1413 | /* REVISIT should rely on caller having done this, | |
1414 | * and caller should rely on us not changing it. | |
1415 | * peripheral code is safe ... check host too. | |
1416 | */ | |
1417 | musb_ep_select(mbase, cppi_ch->index + 1); | |
1418 | ||
1419 | if (cppi_ch->transmit) { | |
1420 | struct cppi_tx_stateram __iomem *tx_ram; | |
550a7375 FB |
1421 | /* REVISIT put timeouts on these controller handshakes */ |
1422 | ||
1423 | cppi_dump_tx(6, cppi_ch, " (teardown)"); | |
1424 | ||
1425 | /* teardown DMA engine then usb core */ | |
1426 | do { | |
1427 | value = musb_readl(tibase, DAVINCI_TXCPPI_TEAR_REG); | |
1428 | } while (!(value & CPPI_TEAR_READY)); | |
1429 | musb_writel(tibase, DAVINCI_TXCPPI_TEAR_REG, cppi_ch->index); | |
1430 | ||
1431 | tx_ram = cppi_ch->state_ram; | |
1432 | do { | |
1433 | value = musb_readl(&tx_ram->tx_complete, 0); | |
1434 | } while (0xFFFFFFFC != value); | |
550a7375 FB |
1435 | |
1436 | /* FIXME clean up the transfer state ... here? | |
1437 | * the completion routine should get called with | |
1438 | * an appropriate status code. | |
1439 | */ | |
1440 | ||
1441 | value = musb_readw(regs, MUSB_TXCSR); | |
1442 | value &= ~MUSB_TXCSR_DMAENAB; | |
1443 | value |= MUSB_TXCSR_FLUSHFIFO; | |
1444 | musb_writew(regs, MUSB_TXCSR, value); | |
1445 | musb_writew(regs, MUSB_TXCSR, value); | |
1446 | ||
56596923 | 1447 | /* |
550a7375 FB |
1448 | * 1. Write to completion Ptr value 0x1(bit 0 set) |
1449 | * (write back mode) | |
56596923 S |
1450 | * 2. Wait for abort interrupt and then put the channel in |
1451 | * compare mode by writing 1 to the tx_complete register. | |
550a7375 FB |
1452 | */ |
1453 | cppi_reset_tx(tx_ram, 1); | |
56596923 S |
1454 | cppi_ch->head = 0; |
1455 | musb_writel(&tx_ram->tx_complete, 0, 1); | |
550a7375 FB |
1456 | cppi_dump_tx(5, cppi_ch, " (done teardown)"); |
1457 | ||
1458 | /* REVISIT tx side _should_ clean up the same way | |
1459 | * as the RX side ... this does no cleanup at all! | |
1460 | */ | |
1461 | ||
1462 | } else /* RX */ { | |
1463 | u16 csr; | |
1464 | ||
1465 | /* NOTE: docs don't guarantee any of this works ... we | |
1466 | * expect that if the usb core stops telling the cppi core | |
1467 | * to pull more data from it, then it'll be safe to flush | |
1468 | * current RX DMA state iff any pending fifo transfer is done. | |
1469 | */ | |
1470 | ||
1471 | core_rxirq_disable(tibase, cppi_ch->index + 1); | |
1472 | ||
1473 | /* for host, ensure ReqPkt is never set again */ | |
1474 | if (is_host_active(cppi_ch->controller->musb)) { | |
1475 | value = musb_readl(tibase, DAVINCI_AUTOREQ_REG); | |
1476 | value &= ~((0x3) << (cppi_ch->index * 2)); | |
1477 | musb_writel(tibase, DAVINCI_AUTOREQ_REG, value); | |
1478 | } | |
1479 | ||
1480 | csr = musb_readw(regs, MUSB_RXCSR); | |
1481 | ||
1482 | /* for host, clear (just) ReqPkt at end of current packet(s) */ | |
1483 | if (is_host_active(cppi_ch->controller->musb)) { | |
1484 | csr |= MUSB_RXCSR_H_WZC_BITS; | |
1485 | csr &= ~MUSB_RXCSR_H_REQPKT; | |
1486 | } else | |
1487 | csr |= MUSB_RXCSR_P_WZC_BITS; | |
1488 | ||
1489 | /* clear dma enable */ | |
1490 | csr &= ~(MUSB_RXCSR_DMAENAB); | |
1491 | musb_writew(regs, MUSB_RXCSR, csr); | |
1492 | csr = musb_readw(regs, MUSB_RXCSR); | |
1493 | ||
1494 | /* Quiesce: wait for current dma to finish (if not cleanup). | |
1495 | * We can't use bit zero of stateram->rx_sop, since that | |
1496 | * refers to an entire "DMA packet" not just emptying the | |
1497 | * current fifo. Most segments need multiple usb packets. | |
1498 | */ | |
1499 | if (channel->status == MUSB_DMA_STATUS_BUSY) | |
1500 | udelay(50); | |
1501 | ||
1502 | /* scan the current list, reporting any data that was | |
1503 | * transferred and acking any IRQ | |
1504 | */ | |
1505 | cppi_rx_scan(controller, cppi_ch->index); | |
1506 | ||
1507 | /* clobber the existing state once it's idle | |
1508 | * | |
1509 | * NOTE: arguably, we should also wait for all the other | |
1510 | * RX channels to quiesce (how??) and then temporarily | |
1511 | * disable RXCPPI_CTRL_REG ... but it seems that we can | |
1512 | * rely on the controller restarting from state ram, with | |
1513 | * only RXCPPI_BUFCNT state being bogus. BUFCNT will | |
1514 | * correct itself after the next DMA transfer though. | |
1515 | * | |
1516 | * REVISIT does using rndis mode change that? | |
1517 | */ | |
1518 | cppi_reset_rx(cppi_ch->state_ram); | |
1519 | ||
1520 | /* next DMA request _should_ load cppi head ptr */ | |
1521 | ||
1522 | /* ... we don't "free" that list, only mutate it in place. */ | |
1523 | cppi_dump_rx(5, cppi_ch, " (done abort)"); | |
1524 | ||
1525 | /* clean up previously pending bds */ | |
1526 | cppi_bd_free(cppi_ch, cppi_ch->last_processed); | |
1527 | cppi_ch->last_processed = NULL; | |
1528 | ||
1529 | while (queue) { | |
1530 | struct cppi_descriptor *tmp = queue->next; | |
1531 | ||
1532 | cppi_bd_free(cppi_ch, queue); | |
1533 | queue = tmp; | |
1534 | } | |
1535 | } | |
1536 | ||
1537 | channel->status = MUSB_DMA_STATUS_FREE; | |
1538 | cppi_ch->buf_dma = 0; | |
1539 | cppi_ch->offset = 0; | |
1540 | cppi_ch->buf_len = 0; | |
1541 | cppi_ch->maxpacket = 0; | |
1542 | return 0; | |
1543 | } | |
1544 | ||
1545 | /* TBD Queries: | |
1546 | * | |
1547 | * Power Management ... probably turn off cppi during suspend, restart; | |
1548 | * check state ram? Clocking is presumably shared with usb core. | |
1549 | */ |