[go: nahoru, domu]

1/*
2 * linux/drivers/misc/xillybus_core.c
3 *
4 * Copyright 2011 Xillybus Ltd, http://xillybus.com
5 *
6 * Driver for the Xillybus FPGA/host framework.
7 *
8 * This driver interfaces with a special IP core in an FPGA, setting up
9 * a pipe between a hardware FIFO in the programmable logic and a device
10 * file in the host. The number of such pipes and their attributes are
11 * set up on the logic. This driver detects these automatically and
12 * creates the device files accordingly.
13 *
14 * This program is free software; you can redistribute it and/or modify
15 * it under the smems of the GNU General Public License as published by
16 * the Free Software Foundation; version 2 of the License.
17 */
18
19#include <linux/list.h>
20#include <linux/device.h>
21#include <linux/module.h>
22#include <linux/io.h>
23#include <linux/dma-mapping.h>
24#include <linux/interrupt.h>
25#include <linux/sched.h>
26#include <linux/fs.h>
27#include <linux/cdev.h>
28#include <linux/spinlock.h>
29#include <linux/mutex.h>
30#include <linux/crc32.h>
31#include <linux/poll.h>
32#include <linux/delay.h>
33#include <linux/slab.h>
34#include <linux/workqueue.h>
35#include "xillybus.h"
36
37MODULE_DESCRIPTION("Xillybus core functions");
38MODULE_AUTHOR("Eli Billauer, Xillybus Ltd.");
39MODULE_VERSION("1.07");
40MODULE_ALIAS("xillybus_core");
41MODULE_LICENSE("GPL v2");
42
43/* General timeout is 100 ms, rx timeout is 10 ms */
44#define XILLY_RX_TIMEOUT (10*HZ/1000)
45#define XILLY_TIMEOUT (100*HZ/1000)
46
47#define fpga_msg_ctrl_reg              0x0008
48#define fpga_dma_control_reg           0x0020
49#define fpga_dma_bufno_reg             0x0024
50#define fpga_dma_bufaddr_lowaddr_reg   0x0028
51#define fpga_dma_bufaddr_highaddr_reg  0x002c
52#define fpga_buf_ctrl_reg              0x0030
53#define fpga_buf_offset_reg            0x0034
54#define fpga_endian_reg                0x0040
55
56#define XILLYMSG_OPCODE_RELEASEBUF 1
57#define XILLYMSG_OPCODE_QUIESCEACK 2
58#define XILLYMSG_OPCODE_FIFOEOF 3
59#define XILLYMSG_OPCODE_FATAL_ERROR 4
60#define XILLYMSG_OPCODE_NONEMPTY 5
61
62static const char xillyname[] = "xillybus";
63
64static struct class *xillybus_class;
65
66/*
67 * ep_list_lock is the last lock to be taken; No other lock requests are
68 * allowed while holding it. It merely protects list_of_endpoints, and not
69 * the endpoints listed in it.
70 */
71
72static LIST_HEAD(list_of_endpoints);
73static struct mutex ep_list_lock;
74static struct workqueue_struct *xillybus_wq;
75
76/*
77 * Locking scheme: Mutexes protect invocations of character device methods.
78 * If both locks are taken, wr_mutex is taken first, rd_mutex second.
79 *
80 * wr_spinlock protects wr_*_buf_idx, wr_empty, wr_sleepy, wr_ready and the
81 * buffers' end_offset fields against changes made by IRQ handler (and in
82 * theory, other file request handlers, but the mutex handles that). Nothing
83 * else.
84 * They are held for short direct memory manipulations. Needless to say,
85 * no mutex locking is allowed when a spinlock is held.
86 *
87 * rd_spinlock does the same with rd_*_buf_idx, rd_empty and end_offset.
88 *
89 * register_mutex is endpoint-specific, and is held when non-atomic
90 * register operations are performed. wr_mutex and rd_mutex may be
91 * held when register_mutex is taken, but none of the spinlocks. Note that
92 * register_mutex doesn't protect against sporadic buf_ctrl_reg writes
93 * which are unrelated to buf_offset_reg, since they are harmless.
94 *
95 * Blocking on the wait queues is allowed with mutexes held, but not with
96 * spinlocks.
97 *
98 * Only interruptible blocking is allowed on mutexes and wait queues.
99 *
100 * All in all, the locking order goes (with skips allowed, of course):
101 * wr_mutex -> rd_mutex -> register_mutex -> wr_spinlock -> rd_spinlock
102 */
103
104static void malformed_message(struct xilly_endpoint *endpoint, u32 *buf)
105{
106	int opcode;
107	int msg_channel, msg_bufno, msg_data, msg_dir;
108
109	opcode = (buf[0] >> 24) & 0xff;
110	msg_dir = buf[0] & 1;
111	msg_channel = (buf[0] >> 1) & 0x7ff;
112	msg_bufno = (buf[0] >> 12) & 0x3ff;
113	msg_data = buf[1] & 0xfffffff;
114
115	dev_warn(endpoint->dev,
116		 "Malformed message (skipping): opcode=%d, channel=%03x, dir=%d, bufno=%03x, data=%07x\n",
117		 opcode, msg_channel, msg_dir, msg_bufno, msg_data);
118}
119
120/*
121 * xillybus_isr assumes the interrupt is allocated exclusively to it,
122 * which is the natural case MSI and several other hardware-oriented
123 * interrupts. Sharing is not allowed.
124 */
125
126irqreturn_t xillybus_isr(int irq, void *data)
127{
128	struct xilly_endpoint *ep = data;
129	u32 *buf;
130	unsigned int buf_size;
131	int i;
132	int opcode;
133	unsigned int msg_channel, msg_bufno, msg_data, msg_dir;
134	struct xilly_channel *channel;
135
136	buf = ep->msgbuf_addr;
137	buf_size = ep->msg_buf_size/sizeof(u32);
138
139	ep->ephw->hw_sync_sgl_for_cpu(ep,
140				      ep->msgbuf_dma_addr,
141				      ep->msg_buf_size,
142				      DMA_FROM_DEVICE);
143
144	for (i = 0; i < buf_size; i += 2) {
145		if (((buf[i+1] >> 28) & 0xf) != ep->msg_counter) {
146			malformed_message(ep, &buf[i]);
147			dev_warn(ep->dev,
148				 "Sending a NACK on counter %x (instead of %x) on entry %d\n",
149				 ((buf[i+1] >> 28) & 0xf),
150				 ep->msg_counter,
151				 i/2);
152
153			if (++ep->failed_messages > 10) {
154				dev_err(ep->dev,
155					"Lost sync with interrupt messages. Stopping.\n");
156			} else {
157				ep->ephw->hw_sync_sgl_for_device(
158					ep,
159					ep->msgbuf_dma_addr,
160					ep->msg_buf_size,
161					DMA_FROM_DEVICE);
162
163				iowrite32(0x01,  /* Message NACK */
164					  ep->registers + fpga_msg_ctrl_reg);
165			}
166			return IRQ_HANDLED;
167		} else if (buf[i] & (1 << 22)) /* Last message */
168			break;
169	}
170
171	if (i >= buf_size) {
172		dev_err(ep->dev, "Bad interrupt message. Stopping.\n");
173		return IRQ_HANDLED;
174	}
175
176	buf_size = i + 2;
177
178	for (i = 0; i < buf_size; i += 2) { /* Scan through messages */
179		opcode = (buf[i] >> 24) & 0xff;
180
181		msg_dir = buf[i] & 1;
182		msg_channel = (buf[i] >> 1) & 0x7ff;
183		msg_bufno = (buf[i] >> 12) & 0x3ff;
184		msg_data = buf[i+1] & 0xfffffff;
185
186		switch (opcode) {
187		case XILLYMSG_OPCODE_RELEASEBUF:
188			if ((msg_channel > ep->num_channels) ||
189			    (msg_channel == 0)) {
190				malformed_message(ep, &buf[i]);
191				break;
192			}
193
194			channel = ep->channels[msg_channel];
195
196			if (msg_dir) { /* Write channel */
197				if (msg_bufno >= channel->num_wr_buffers) {
198					malformed_message(ep, &buf[i]);
199					break;
200				}
201				spin_lock(&channel->wr_spinlock);
202				channel->wr_buffers[msg_bufno]->end_offset =
203					msg_data;
204				channel->wr_fpga_buf_idx = msg_bufno;
205				channel->wr_empty = 0;
206				channel->wr_sleepy = 0;
207				spin_unlock(&channel->wr_spinlock);
208
209				wake_up_interruptible(&channel->wr_wait);
210
211			} else {
212				/* Read channel */
213
214				if (msg_bufno >= channel->num_rd_buffers) {
215					malformed_message(ep, &buf[i]);
216					break;
217				}
218
219				spin_lock(&channel->rd_spinlock);
220				channel->rd_fpga_buf_idx = msg_bufno;
221				channel->rd_full = 0;
222				spin_unlock(&channel->rd_spinlock);
223
224				wake_up_interruptible(&channel->rd_wait);
225				if (!channel->rd_synchronous)
226					queue_delayed_work(
227						xillybus_wq,
228						&channel->rd_workitem,
229						XILLY_RX_TIMEOUT);
230			}
231
232			break;
233		case XILLYMSG_OPCODE_NONEMPTY:
234			if ((msg_channel > ep->num_channels) ||
235			    (msg_channel == 0) || (!msg_dir) ||
236			    !ep->channels[msg_channel]->wr_supports_nonempty) {
237				malformed_message(ep, &buf[i]);
238				break;
239			}
240
241			channel = ep->channels[msg_channel];
242
243			if (msg_bufno >= channel->num_wr_buffers) {
244				malformed_message(ep, &buf[i]);
245				break;
246			}
247			spin_lock(&channel->wr_spinlock);
248			if (msg_bufno == channel->wr_host_buf_idx)
249				channel->wr_ready = 1;
250			spin_unlock(&channel->wr_spinlock);
251
252			wake_up_interruptible(&channel->wr_ready_wait);
253
254			break;
255		case XILLYMSG_OPCODE_QUIESCEACK:
256			ep->idtlen = msg_data;
257			wake_up_interruptible(&ep->ep_wait);
258
259			break;
260		case XILLYMSG_OPCODE_FIFOEOF:
261			if ((msg_channel > ep->num_channels) ||
262			    (msg_channel == 0) || (!msg_dir) ||
263			    !ep->channels[msg_channel]->num_wr_buffers) {
264				malformed_message(ep, &buf[i]);
265				break;
266			}
267			channel = ep->channels[msg_channel];
268			spin_lock(&channel->wr_spinlock);
269			channel->wr_eof = msg_bufno;
270			channel->wr_sleepy = 0;
271
272			channel->wr_hangup = channel->wr_empty &&
273				(channel->wr_host_buf_idx == msg_bufno);
274
275			spin_unlock(&channel->wr_spinlock);
276
277			wake_up_interruptible(&channel->wr_wait);
278
279			break;
280		case XILLYMSG_OPCODE_FATAL_ERROR:
281			ep->fatal_error = 1;
282			wake_up_interruptible(&ep->ep_wait); /* For select() */
283			dev_err(ep->dev,
284				"FPGA reported a fatal error. This means that the low-level communication with the device has failed. This hardware problem is most likely unrelated to Xillybus (neither kernel module nor FPGA core), but reports are still welcome. All I/O is aborted.\n");
285			break;
286		default:
287			malformed_message(ep, &buf[i]);
288			break;
289		}
290	}
291
292	ep->ephw->hw_sync_sgl_for_device(ep,
293					 ep->msgbuf_dma_addr,
294					 ep->msg_buf_size,
295					 DMA_FROM_DEVICE);
296
297	ep->msg_counter = (ep->msg_counter + 1) & 0xf;
298	ep->failed_messages = 0;
299	iowrite32(0x03, ep->registers + fpga_msg_ctrl_reg); /* Message ACK */
300
301	return IRQ_HANDLED;
302}
303EXPORT_SYMBOL(xillybus_isr);
304
305/*
306 * A few trivial memory management functions.
307 * NOTE: These functions are used only on probe and remove, and therefore
308 * no locks are applied!
309 */
310
311static void xillybus_autoflush(struct work_struct *work);
312
313struct xilly_alloc_state {
314	void *salami;
315	int left_of_salami;
316	int nbuffer;
317	enum dma_data_direction direction;
318	u32 regdirection;
319};
320
321static int xilly_get_dma_buffers(struct xilly_endpoint *ep,
322				 struct xilly_alloc_state *s,
323				 struct xilly_buffer **buffers,
324				 int bufnum, int bytebufsize)
325{
326	int i, rc;
327	dma_addr_t dma_addr;
328	struct device *dev = ep->dev;
329	struct xilly_buffer *this_buffer = NULL; /* Init to silence warning */
330
331	if (buffers) { /* Not the message buffer */
332		this_buffer = devm_kcalloc(dev, bufnum,
333					   sizeof(struct xilly_buffer),
334					   GFP_KERNEL);
335		if (!this_buffer)
336			return -ENOMEM;
337	}
338
339	for (i = 0; i < bufnum; i++) {
340		/*
341		 * Buffers are expected in descending size order, so there
342		 * is either enough space for this buffer or none at all.
343		 */
344
345		if ((s->left_of_salami < bytebufsize) &&
346		    (s->left_of_salami > 0)) {
347			dev_err(ep->dev,
348				"Corrupt buffer allocation in IDT. Aborting.\n");
349			return -ENODEV;
350		}
351
352		if (s->left_of_salami == 0) {
353			int allocorder, allocsize;
354
355			allocsize = PAGE_SIZE;
356			allocorder = 0;
357			while (bytebufsize > allocsize) {
358				allocsize *= 2;
359				allocorder++;
360			}
361
362			s->salami = (void *) devm_get_free_pages(
363				dev,
364				GFP_KERNEL | __GFP_DMA32 | __GFP_ZERO,
365				allocorder);
366			if (!s->salami)
367				return -ENOMEM;
368
369			s->left_of_salami = allocsize;
370		}
371
372		rc = ep->ephw->map_single(ep, s->salami,
373					  bytebufsize, s->direction,
374					  &dma_addr);
375		if (rc)
376			return rc;
377
378		iowrite32((u32) (dma_addr & 0xffffffff),
379			  ep->registers + fpga_dma_bufaddr_lowaddr_reg);
380		iowrite32(((u32) ((((u64) dma_addr) >> 32) & 0xffffffff)),
381			  ep->registers + fpga_dma_bufaddr_highaddr_reg);
382
383		if (buffers) { /* Not the message buffer */
384			this_buffer->addr = s->salami;
385			this_buffer->dma_addr = dma_addr;
386			buffers[i] = this_buffer++;
387
388			iowrite32(s->regdirection | s->nbuffer++,
389				  ep->registers + fpga_dma_bufno_reg);
390		} else {
391			ep->msgbuf_addr = s->salami;
392			ep->msgbuf_dma_addr = dma_addr;
393			ep->msg_buf_size = bytebufsize;
394
395			iowrite32(s->regdirection,
396				  ep->registers + fpga_dma_bufno_reg);
397		}
398
399		s->left_of_salami -= bytebufsize;
400		s->salami += bytebufsize;
401	}
402	return 0;
403}
404
405static int xilly_setupchannels(struct xilly_endpoint *ep,
406			       unsigned char *chandesc,
407			       int entries)
408{
409	struct device *dev = ep->dev;
410	int i, entry, rc;
411	struct xilly_channel *channel;
412	int channelnum, bufnum, bufsize, format, is_writebuf;
413	int bytebufsize;
414	int synchronous, allowpartial, exclusive_open, seekable;
415	int supports_nonempty;
416	int msg_buf_done = 0;
417
418	struct xilly_alloc_state rd_alloc = {
419		.salami = NULL,
420		.left_of_salami = 0,
421		.nbuffer = 1,
422		.direction = DMA_TO_DEVICE,
423		.regdirection = 0,
424	};
425
426	struct xilly_alloc_state wr_alloc = {
427		.salami = NULL,
428		.left_of_salami = 0,
429		.nbuffer = 1,
430		.direction = DMA_FROM_DEVICE,
431		.regdirection = 0x80000000,
432	};
433
434	channel = devm_kcalloc(dev, ep->num_channels,
435			       sizeof(struct xilly_channel), GFP_KERNEL);
436	if (!channel)
437		return -ENOMEM;
438
439	ep->channels = devm_kcalloc(dev, ep->num_channels + 1,
440				    sizeof(struct xilly_channel *),
441				    GFP_KERNEL);
442	if (!ep->channels)
443		return -ENOMEM;
444
445	ep->channels[0] = NULL; /* Channel 0 is message buf. */
446
447	/* Initialize all channels with defaults */
448
449	for (i = 1; i <= ep->num_channels; i++) {
450		channel->wr_buffers = NULL;
451		channel->rd_buffers = NULL;
452		channel->num_wr_buffers = 0;
453		channel->num_rd_buffers = 0;
454		channel->wr_fpga_buf_idx = -1;
455		channel->wr_host_buf_idx = 0;
456		channel->wr_host_buf_pos = 0;
457		channel->wr_empty = 1;
458		channel->wr_ready = 0;
459		channel->wr_sleepy = 1;
460		channel->rd_fpga_buf_idx = 0;
461		channel->rd_host_buf_idx = 0;
462		channel->rd_host_buf_pos = 0;
463		channel->rd_full = 0;
464		channel->wr_ref_count = 0;
465		channel->rd_ref_count = 0;
466
467		spin_lock_init(&channel->wr_spinlock);
468		spin_lock_init(&channel->rd_spinlock);
469		mutex_init(&channel->wr_mutex);
470		mutex_init(&channel->rd_mutex);
471		init_waitqueue_head(&channel->rd_wait);
472		init_waitqueue_head(&channel->wr_wait);
473		init_waitqueue_head(&channel->wr_ready_wait);
474
475		INIT_DELAYED_WORK(&channel->rd_workitem, xillybus_autoflush);
476
477		channel->endpoint = ep;
478		channel->chan_num = i;
479
480		channel->log2_element_size = 0;
481
482		ep->channels[i] = channel++;
483	}
484
485	for (entry = 0; entry < entries; entry++, chandesc += 4) {
486		struct xilly_buffer **buffers = NULL;
487
488		is_writebuf = chandesc[0] & 0x01;
489		channelnum = (chandesc[0] >> 1) | ((chandesc[1] & 0x0f) << 7);
490		format = (chandesc[1] >> 4) & 0x03;
491		allowpartial = (chandesc[1] >> 6) & 0x01;
492		synchronous = (chandesc[1] >> 7) & 0x01;
493		bufsize = 1 << (chandesc[2] & 0x1f);
494		bufnum = 1 << (chandesc[3] & 0x0f);
495		exclusive_open = (chandesc[2] >> 7) & 0x01;
496		seekable = (chandesc[2] >> 6) & 0x01;
497		supports_nonempty = (chandesc[2] >> 5) & 0x01;
498
499		if ((channelnum > ep->num_channels) ||
500		    ((channelnum == 0) && !is_writebuf)) {
501			dev_err(ep->dev,
502				"IDT requests channel out of range. Aborting.\n");
503			return -ENODEV;
504		}
505
506		channel = ep->channels[channelnum]; /* NULL for msg channel */
507
508		if (!is_writebuf || channelnum > 0) {
509			channel->log2_element_size = ((format > 2) ?
510						      2 : format);
511
512			bytebufsize = channel->rd_buf_size = bufsize *
513				(1 << channel->log2_element_size);
514
515			buffers = devm_kcalloc(dev, bufnum,
516					       sizeof(struct xilly_buffer *),
517					       GFP_KERNEL);
518			if (!buffers)
519				return -ENOMEM;
520		} else {
521			bytebufsize = bufsize << 2;
522		}
523
524		if (!is_writebuf) {
525			channel->num_rd_buffers = bufnum;
526			channel->rd_allow_partial = allowpartial;
527			channel->rd_synchronous = synchronous;
528			channel->rd_exclusive_open = exclusive_open;
529			channel->seekable = seekable;
530
531			channel->rd_buffers = buffers;
532			rc = xilly_get_dma_buffers(ep, &rd_alloc, buffers,
533						   bufnum, bytebufsize);
534		} else if (channelnum > 0) {
535			channel->num_wr_buffers = bufnum;
536
537			channel->seekable = seekable;
538			channel->wr_supports_nonempty = supports_nonempty;
539
540			channel->wr_allow_partial = allowpartial;
541			channel->wr_synchronous = synchronous;
542			channel->wr_exclusive_open = exclusive_open;
543
544			channel->wr_buffers = buffers;
545			rc = xilly_get_dma_buffers(ep, &wr_alloc, buffers,
546						   bufnum, bytebufsize);
547		} else {
548			rc = xilly_get_dma_buffers(ep, &wr_alloc, NULL,
549						   bufnum, bytebufsize);
550			msg_buf_done++;
551		}
552
553		if (rc)
554			return -ENOMEM;
555	}
556
557	if (!msg_buf_done) {
558		dev_err(ep->dev,
559			"Corrupt IDT: No message buffer. Aborting.\n");
560		return -ENODEV;
561	}
562	return 0;
563}
564
565static int xilly_scan_idt(struct xilly_endpoint *endpoint,
566			  struct xilly_idt_handle *idt_handle)
567{
568	int count = 0;
569	unsigned char *idt = endpoint->channels[1]->wr_buffers[0]->addr;
570	unsigned char *end_of_idt = idt + endpoint->idtlen - 4;
571	unsigned char *scan;
572	int len;
573
574	scan = idt;
575	idt_handle->idt = idt;
576
577	scan++; /* Skip version number */
578
579	while ((scan <= end_of_idt) && *scan) {
580		while ((scan <= end_of_idt) && *scan++)
581			/* Do nothing, just scan thru string */;
582		count++;
583	}
584
585	scan++;
586
587	if (scan > end_of_idt) {
588		dev_err(endpoint->dev,
589			"IDT device name list overflow. Aborting.\n");
590		return -ENODEV;
591	}
592	idt_handle->chandesc = scan;
593
594	len = endpoint->idtlen - (3 + ((int) (scan - idt)));
595
596	if (len & 0x03) {
597		dev_err(endpoint->dev,
598			"Corrupt IDT device name list. Aborting.\n");
599		return -ENODEV;
600	}
601
602	idt_handle->entries = len >> 2;
603	endpoint->num_channels = count;
604
605	return 0;
606}
607
608static int xilly_obtain_idt(struct xilly_endpoint *endpoint)
609{
610	struct xilly_channel *channel;
611	unsigned char *version;
612	long t;
613
614	channel = endpoint->channels[1]; /* This should be generated ad-hoc */
615
616	channel->wr_sleepy = 1;
617
618	iowrite32(1 |
619		  (3 << 24), /* Opcode 3 for channel 0 = Send IDT */
620		  endpoint->registers + fpga_buf_ctrl_reg);
621
622	t = wait_event_interruptible_timeout(channel->wr_wait,
623					     (!channel->wr_sleepy),
624					     XILLY_TIMEOUT);
625
626	if (t <= 0) {
627		dev_err(endpoint->dev, "Failed to obtain IDT. Aborting.\n");
628
629		if (endpoint->fatal_error)
630			return -EIO;
631
632		return -ENODEV;
633	}
634
635	endpoint->ephw->hw_sync_sgl_for_cpu(
636		channel->endpoint,
637		channel->wr_buffers[0]->dma_addr,
638		channel->wr_buf_size,
639		DMA_FROM_DEVICE);
640
641	if (channel->wr_buffers[0]->end_offset != endpoint->idtlen) {
642		dev_err(endpoint->dev,
643			"IDT length mismatch (%d != %d). Aborting.\n",
644			channel->wr_buffers[0]->end_offset, endpoint->idtlen);
645		return -ENODEV;
646	}
647
648	if (crc32_le(~0, channel->wr_buffers[0]->addr,
649		     endpoint->idtlen+1) != 0) {
650		dev_err(endpoint->dev, "IDT failed CRC check. Aborting.\n");
651		return -ENODEV;
652	}
653
654	version = channel->wr_buffers[0]->addr;
655
656	/* Check version number. Accept anything below 0x82 for now. */
657	if (*version > 0x82) {
658		dev_err(endpoint->dev,
659			"No support for IDT version 0x%02x. Maybe the xillybus driver needs an upgarde. Aborting.\n",
660			*version);
661		return -ENODEV;
662	}
663
664	return 0;
665}
666
667static ssize_t xillybus_read(struct file *filp, char __user *userbuf,
668			     size_t count, loff_t *f_pos)
669{
670	ssize_t rc;
671	unsigned long flags;
672	int bytes_done = 0;
673	int no_time_left = 0;
674	long deadline, left_to_sleep;
675	struct xilly_channel *channel = filp->private_data;
676
677	int empty, reached_eof, exhausted, ready;
678	/* Initializations are there only to silence warnings */
679
680	int howmany = 0, bufpos = 0, bufidx = 0, bufferdone = 0;
681	int waiting_bufidx;
682
683	if (channel->endpoint->fatal_error)
684		return -EIO;
685
686	deadline = jiffies + 1 + XILLY_RX_TIMEOUT;
687
688	rc = mutex_lock_interruptible(&channel->wr_mutex);
689	if (rc)
690		return rc;
691
692	while (1) { /* Note that we may drop mutex within this loop */
693		int bytes_to_do = count - bytes_done;
694
695		spin_lock_irqsave(&channel->wr_spinlock, flags);
696
697		empty = channel->wr_empty;
698		ready = !empty || channel->wr_ready;
699
700		if (!empty) {
701			bufidx = channel->wr_host_buf_idx;
702			bufpos = channel->wr_host_buf_pos;
703			howmany = ((channel->wr_buffers[bufidx]->end_offset
704				    + 1) << channel->log2_element_size)
705				- bufpos;
706
707			/* Update wr_host_* to its post-operation state */
708			if (howmany > bytes_to_do) {
709				bufferdone = 0;
710
711				howmany = bytes_to_do;
712				channel->wr_host_buf_pos += howmany;
713			} else {
714				bufferdone = 1;
715
716				channel->wr_host_buf_pos = 0;
717
718				if (bufidx == channel->wr_fpga_buf_idx) {
719					channel->wr_empty = 1;
720					channel->wr_sleepy = 1;
721					channel->wr_ready = 0;
722				}
723
724				if (bufidx >= (channel->num_wr_buffers - 1))
725					channel->wr_host_buf_idx = 0;
726				else
727					channel->wr_host_buf_idx++;
728			}
729		}
730
731		/*
732		 * Marking our situation after the possible changes above,
733		 * for use after releasing the spinlock.
734		 *
735		 * empty = empty before change
736		 * exhasted = empty after possible change
737		 */
738
739		reached_eof = channel->wr_empty &&
740			(channel->wr_host_buf_idx == channel->wr_eof);
741		channel->wr_hangup = reached_eof;
742		exhausted = channel->wr_empty;
743		waiting_bufidx = channel->wr_host_buf_idx;
744
745		spin_unlock_irqrestore(&channel->wr_spinlock, flags);
746
747		if (!empty) { /* Go on, now without the spinlock */
748
749			if (bufpos == 0) /* Position zero means it's virgin */
750				channel->endpoint->ephw->hw_sync_sgl_for_cpu(
751					channel->endpoint,
752					channel->wr_buffers[bufidx]->dma_addr,
753					channel->wr_buf_size,
754					DMA_FROM_DEVICE);
755
756			if (copy_to_user(
757				    userbuf,
758				    channel->wr_buffers[bufidx]->addr
759				    + bufpos, howmany))
760				rc = -EFAULT;
761
762			userbuf += howmany;
763			bytes_done += howmany;
764
765			if (bufferdone) {
766				channel->endpoint->ephw->hw_sync_sgl_for_device(
767					channel->endpoint,
768					channel->wr_buffers[bufidx]->dma_addr,
769					channel->wr_buf_size,
770					DMA_FROM_DEVICE);
771
772				/*
773				 * Tell FPGA the buffer is done with. It's an
774				 * atomic operation to the FPGA, so what
775				 * happens with other channels doesn't matter,
776				 * and the certain channel is protected with
777				 * the channel-specific mutex.
778				 */
779
780				iowrite32(1 | (channel->chan_num << 1) |
781					  (bufidx << 12),
782					  channel->endpoint->registers +
783					  fpga_buf_ctrl_reg);
784			}
785
786			if (rc) {
787				mutex_unlock(&channel->wr_mutex);
788				return rc;
789			}
790		}
791
792		/* This includes a zero-count return = EOF */
793		if ((bytes_done >= count) || reached_eof)
794			break;
795
796		if (!exhausted)
797			continue; /* More in RAM buffer(s)? Just go on. */
798
799		if ((bytes_done > 0) &&
800		    (no_time_left ||
801		     (channel->wr_synchronous && channel->wr_allow_partial)))
802			break;
803
804		/*
805		 * Nonblocking read: The "ready" flag tells us that the FPGA
806		 * has data to send. In non-blocking mode, if it isn't on,
807		 * just return. But if there is, we jump directly to the point
808		 * where we ask for the FPGA to send all it has, and wait
809		 * until that data arrives. So in a sense, we *do* block in
810		 * nonblocking mode, but only for a very short time.
811		 */
812
813		if (!no_time_left && (filp->f_flags & O_NONBLOCK)) {
814			if (bytes_done > 0)
815				break;
816
817			if (ready)
818				goto desperate;
819
820			rc = -EAGAIN;
821			break;
822		}
823
824		if (!no_time_left || (bytes_done > 0)) {
825			/*
826			 * Note that in case of an element-misaligned read
827			 * request, offsetlimit will include the last element,
828			 * which will be partially read from.
829			 */
830			int offsetlimit = ((count - bytes_done) - 1) >>
831				channel->log2_element_size;
832			int buf_elements = channel->wr_buf_size >>
833				channel->log2_element_size;
834
835			/*
836			 * In synchronous mode, always send an offset limit.
837			 * Just don't send a value too big.
838			 */
839
840			if (channel->wr_synchronous) {
841				/* Don't request more than one buffer */
842				if (channel->wr_allow_partial &&
843				    (offsetlimit >= buf_elements))
844					offsetlimit = buf_elements - 1;
845
846				/* Don't request more than all buffers */
847				if (!channel->wr_allow_partial &&
848				    (offsetlimit >=
849				     (buf_elements * channel->num_wr_buffers)))
850					offsetlimit = buf_elements *
851						channel->num_wr_buffers - 1;
852			}
853
854			/*
855			 * In asynchronous mode, force early flush of a buffer
856			 * only if that will allow returning a full count. The
857			 * "offsetlimit < ( ... )" rather than "<=" excludes
858			 * requesting a full buffer, which would obviously
859			 * cause a buffer transmission anyhow
860			 */
861
862			if (channel->wr_synchronous ||
863			    (offsetlimit < (buf_elements - 1))) {
864				mutex_lock(&channel->endpoint->register_mutex);
865
866				iowrite32(offsetlimit,
867					  channel->endpoint->registers +
868					  fpga_buf_offset_reg);
869
870				iowrite32(1 | (channel->chan_num << 1) |
871					  (2 << 24) |  /* 2 = offset limit */
872					  (waiting_bufidx << 12),
873					  channel->endpoint->registers +
874					  fpga_buf_ctrl_reg);
875
876				mutex_unlock(&channel->endpoint->
877					     register_mutex);
878			}
879		}
880
881		/*
882		 * If partial completion is disallowed, there is no point in
883		 * timeout sleeping. Neither if no_time_left is set and
884		 * there's no data.
885		 */
886
887		if (!channel->wr_allow_partial ||
888		    (no_time_left && (bytes_done == 0))) {
889			/*
890			 * This do-loop will run more than once if another
891			 * thread reasserted wr_sleepy before we got the mutex
892			 * back, so we try again.
893			 */
894
895			do {
896				mutex_unlock(&channel->wr_mutex);
897
898				if (wait_event_interruptible(
899					    channel->wr_wait,
900					    (!channel->wr_sleepy)))
901					goto interrupted;
902
903				if (mutex_lock_interruptible(
904					    &channel->wr_mutex))
905					goto interrupted;
906			} while (channel->wr_sleepy);
907
908			continue;
909
910interrupted: /* Mutex is not held if got here */
911			if (channel->endpoint->fatal_error)
912				return -EIO;
913			if (bytes_done)
914				return bytes_done;
915			if (filp->f_flags & O_NONBLOCK)
916				return -EAGAIN; /* Don't admit snoozing */
917			return -EINTR;
918		}
919
920		left_to_sleep = deadline - ((long) jiffies);
921
922		/*
923		 * If our time is out, skip the waiting. We may miss wr_sleepy
924		 * being deasserted but hey, almost missing the train is like
925		 * missing it.
926		 */
927
928		if (left_to_sleep > 0) {
929			left_to_sleep =
930				wait_event_interruptible_timeout(
931					channel->wr_wait,
932					(!channel->wr_sleepy),
933					left_to_sleep);
934
935			if (left_to_sleep > 0) /* wr_sleepy deasserted */
936				continue;
937
938			if (left_to_sleep < 0) { /* Interrupt */
939				mutex_unlock(&channel->wr_mutex);
940				if (channel->endpoint->fatal_error)
941					return -EIO;
942				if (bytes_done)
943					return bytes_done;
944				return -EINTR;
945			}
946		}
947
948desperate:
949		no_time_left = 1; /* We're out of sleeping time. Desperate! */
950
951		if (bytes_done == 0) {
952			/*
953			 * Reaching here means that we allow partial return,
954			 * that we've run out of time, and that we have
955			 * nothing to return.
956			 * So tell the FPGA to send anything it has or gets.
957			 */
958
959			iowrite32(1 | (channel->chan_num << 1) |
960				  (3 << 24) |  /* Opcode 3, flush it all! */
961				  (waiting_bufidx << 12),
962				  channel->endpoint->registers +
963				  fpga_buf_ctrl_reg);
964		}
965
966		/*
967		 * Reaching here means that we *do* have data in the buffer,
968		 * but the "partial" flag disallows returning less than
969		 * required. And we don't have as much. So loop again,
970		 * which is likely to end up blocking indefinitely until
971		 * enough data has arrived.
972		 */
973	}
974
975	mutex_unlock(&channel->wr_mutex);
976
977	if (channel->endpoint->fatal_error)
978		return -EIO;
979
980	if (rc)
981		return rc;
982
983	return bytes_done;
984}
985
986/*
987 * The timeout argument takes values as follows:
988 *  >0 : Flush with timeout
989 * ==0 : Flush, and wait idefinitely for the flush to complete
990 *  <0 : Autoflush: Flush only if there's a single buffer occupied
991 */
992
993static int xillybus_myflush(struct xilly_channel *channel, long timeout)
994{
995	int rc;
996	unsigned long flags;
997
998	int end_offset_plus1;
999	int bufidx, bufidx_minus1;
1000	int i;
1001	int empty;
1002	int new_rd_host_buf_pos;
1003
1004	if (channel->endpoint->fatal_error)
1005		return -EIO;
1006	rc = mutex_lock_interruptible(&channel->rd_mutex);
1007	if (rc)
1008		return rc;
1009
1010	/*
1011	 * Don't flush a closed channel. This can happen when the work queued
1012	 * autoflush thread fires off after the file has closed. This is not
1013	 * an error, just something to dismiss.
1014	 */
1015
1016	if (!channel->rd_ref_count)
1017		goto done;
1018
1019	bufidx = channel->rd_host_buf_idx;
1020
1021	bufidx_minus1 = (bufidx == 0) ?
1022		channel->num_rd_buffers - 1 :
1023		bufidx - 1;
1024
1025	end_offset_plus1 = channel->rd_host_buf_pos >>
1026		channel->log2_element_size;
1027
1028	new_rd_host_buf_pos = channel->rd_host_buf_pos -
1029		(end_offset_plus1 << channel->log2_element_size);
1030
1031	/* Submit the current buffer if it's nonempty */
1032	if (end_offset_plus1) {
1033		unsigned char *tail = channel->rd_buffers[bufidx]->addr +
1034			(end_offset_plus1 << channel->log2_element_size);
1035
1036		/* Copy  unflushed data, so we can put it in next buffer */
1037		for (i = 0; i < new_rd_host_buf_pos; i++)
1038			channel->rd_leftovers[i] = *tail++;
1039
1040		spin_lock_irqsave(&channel->rd_spinlock, flags);
1041
1042		/* Autoflush only if a single buffer is occupied */
1043
1044		if ((timeout < 0) &&
1045		    (channel->rd_full ||
1046		     (bufidx_minus1 != channel->rd_fpga_buf_idx))) {
1047			spin_unlock_irqrestore(&channel->rd_spinlock, flags);
1048			/*
1049			 * A new work item may be queued by the ISR exactly
1050			 * now, since the execution of a work item allows the
1051			 * queuing of a new one while it's running.
1052			 */
1053			goto done;
1054		}
1055
1056		/* The 4th element is never needed for data, so it's a flag */
1057		channel->rd_leftovers[3] = (new_rd_host_buf_pos != 0);
1058
1059		/* Set up rd_full to reflect a certain moment's state */
1060
1061		if (bufidx == channel->rd_fpga_buf_idx)
1062			channel->rd_full = 1;
1063		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
1064
1065		if (bufidx >= (channel->num_rd_buffers - 1))
1066			channel->rd_host_buf_idx = 0;
1067		else
1068			channel->rd_host_buf_idx++;
1069
1070		channel->endpoint->ephw->hw_sync_sgl_for_device(
1071			channel->endpoint,
1072			channel->rd_buffers[bufidx]->dma_addr,
1073			channel->rd_buf_size,
1074			DMA_TO_DEVICE);
1075
1076		mutex_lock(&channel->endpoint->register_mutex);
1077
1078		iowrite32(end_offset_plus1 - 1,
1079			  channel->endpoint->registers + fpga_buf_offset_reg);
1080
1081		iowrite32((channel->chan_num << 1) | /* Channel ID */
1082			  (2 << 24) |  /* Opcode 2, submit buffer */
1083			  (bufidx << 12),
1084			  channel->endpoint->registers + fpga_buf_ctrl_reg);
1085
1086		mutex_unlock(&channel->endpoint->register_mutex);
1087	} else if (bufidx == 0) {
1088		bufidx = channel->num_rd_buffers - 1;
1089	} else {
1090		bufidx--;
1091	}
1092
1093	channel->rd_host_buf_pos = new_rd_host_buf_pos;
1094
1095	if (timeout < 0)
1096		goto done; /* Autoflush */
1097
1098	/*
1099	 * bufidx is now the last buffer written to (or equal to
1100	 * rd_fpga_buf_idx if buffer was never written to), and
1101	 * channel->rd_host_buf_idx the one after it.
1102	 *
1103	 * If bufidx == channel->rd_fpga_buf_idx we're either empty or full.
1104	 */
1105
1106	while (1) { /* Loop waiting for draining of buffers */
1107		spin_lock_irqsave(&channel->rd_spinlock, flags);
1108
1109		if (bufidx != channel->rd_fpga_buf_idx)
1110			channel->rd_full = 1; /*
1111					       * Not really full,
1112					       * but needs waiting.
1113					       */
1114
1115		empty = !channel->rd_full;
1116
1117		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
1118
1119		if (empty)
1120			break;
1121
1122		/*
1123		 * Indefinite sleep with mutex taken. With data waiting for
1124		 * flushing user should not be surprised if open() for write
1125		 * sleeps.
1126		 */
1127		if (timeout == 0)
1128			wait_event_interruptible(channel->rd_wait,
1129						 (!channel->rd_full));
1130
1131		else if (wait_event_interruptible_timeout(
1132				 channel->rd_wait,
1133				 (!channel->rd_full),
1134				 timeout) == 0) {
1135			dev_warn(channel->endpoint->dev,
1136				 "Timed out while flushing. Output data may be lost.\n");
1137
1138			rc = -ETIMEDOUT;
1139			break;
1140		}
1141
1142		if (channel->rd_full) {
1143			rc = -EINTR;
1144			break;
1145		}
1146	}
1147
1148done:
1149	mutex_unlock(&channel->rd_mutex);
1150
1151	if (channel->endpoint->fatal_error)
1152		return -EIO;
1153
1154	return rc;
1155}
1156
1157static int xillybus_flush(struct file *filp, fl_owner_t id)
1158{
1159	if (!(filp->f_mode & FMODE_WRITE))
1160		return 0;
1161
1162	return xillybus_myflush(filp->private_data, HZ); /* 1 second timeout */
1163}
1164
1165static void xillybus_autoflush(struct work_struct *work)
1166{
1167	struct delayed_work *workitem = container_of(
1168		work, struct delayed_work, work);
1169	struct xilly_channel *channel = container_of(
1170		workitem, struct xilly_channel, rd_workitem);
1171	int rc;
1172
1173	rc = xillybus_myflush(channel, -1);
1174	if (rc == -EINTR)
1175		dev_warn(channel->endpoint->dev,
1176			 "Autoflush failed because work queue thread got a signal.\n");
1177	else if (rc)
1178		dev_err(channel->endpoint->dev,
1179			"Autoflush failed under weird circumstances.\n");
1180}
1181
1182static ssize_t xillybus_write(struct file *filp, const char __user *userbuf,
1183			      size_t count, loff_t *f_pos)
1184{
1185	ssize_t rc;
1186	unsigned long flags;
1187	int bytes_done = 0;
1188	struct xilly_channel *channel = filp->private_data;
1189
1190	int full, exhausted;
1191	/* Initializations are there only to silence warnings */
1192
1193	int howmany = 0, bufpos = 0, bufidx = 0, bufferdone = 0;
1194	int end_offset_plus1 = 0;
1195
1196	if (channel->endpoint->fatal_error)
1197		return -EIO;
1198
1199	rc = mutex_lock_interruptible(&channel->rd_mutex);
1200	if (rc)
1201		return rc;
1202
1203	while (1) {
1204		int bytes_to_do = count - bytes_done;
1205
1206		spin_lock_irqsave(&channel->rd_spinlock, flags);
1207
1208		full = channel->rd_full;
1209
1210		if (!full) {
1211			bufidx = channel->rd_host_buf_idx;
1212			bufpos = channel->rd_host_buf_pos;
1213			howmany = channel->rd_buf_size - bufpos;
1214
1215			/*
1216			 * Update rd_host_* to its state after this operation.
1217			 * count=0 means committing the buffer immediately,
1218			 * which is like flushing, but not necessarily block.
1219			 */
1220
1221			if ((howmany > bytes_to_do) &&
1222			    (count ||
1223			     ((bufpos >> channel->log2_element_size) == 0))) {
1224				bufferdone = 0;
1225
1226				howmany = bytes_to_do;
1227				channel->rd_host_buf_pos += howmany;
1228			} else {
1229				bufferdone = 1;
1230
1231				if (count) {
1232					end_offset_plus1 =
1233						channel->rd_buf_size >>
1234						channel->log2_element_size;
1235					channel->rd_host_buf_pos = 0;
1236				} else {
1237					unsigned char *tail;
1238					int i;
1239
1240					end_offset_plus1 = bufpos >>
1241						channel->log2_element_size;
1242
1243					channel->rd_host_buf_pos -=
1244						end_offset_plus1 <<
1245						channel->log2_element_size;
1246
1247					tail = channel->
1248						rd_buffers[bufidx]->addr +
1249						(end_offset_plus1 <<
1250						 channel->log2_element_size);
1251
1252					for (i = 0;
1253					     i < channel->rd_host_buf_pos;
1254					     i++)
1255						channel->rd_leftovers[i] =
1256							*tail++;
1257				}
1258
1259				if (bufidx == channel->rd_fpga_buf_idx)
1260					channel->rd_full = 1;
1261
1262				if (bufidx >= (channel->num_rd_buffers - 1))
1263					channel->rd_host_buf_idx = 0;
1264				else
1265					channel->rd_host_buf_idx++;
1266			}
1267		}
1268
1269		/*
1270		 * Marking our situation after the possible changes above,
1271		 * for use  after releasing the spinlock.
1272		 *
1273		 * full = full before change
1274		 * exhasted = full after possible change
1275		 */
1276
1277		exhausted = channel->rd_full;
1278
1279		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
1280
1281		if (!full) { /* Go on, now without the spinlock */
1282			unsigned char *head =
1283				channel->rd_buffers[bufidx]->addr;
1284			int i;
1285
1286			if ((bufpos == 0) || /* Zero means it's virgin */
1287			    (channel->rd_leftovers[3] != 0)) {
1288				channel->endpoint->ephw->hw_sync_sgl_for_cpu(
1289					channel->endpoint,
1290					channel->rd_buffers[bufidx]->dma_addr,
1291					channel->rd_buf_size,
1292					DMA_TO_DEVICE);
1293
1294				/* Virgin, but leftovers are due */
1295				for (i = 0; i < bufpos; i++)
1296					*head++ = channel->rd_leftovers[i];
1297
1298				channel->rd_leftovers[3] = 0; /* Clear flag */
1299			}
1300
1301			if (copy_from_user(
1302				    channel->rd_buffers[bufidx]->addr + bufpos,
1303				    userbuf, howmany))
1304				rc = -EFAULT;
1305
1306			userbuf += howmany;
1307			bytes_done += howmany;
1308
1309			if (bufferdone) {
1310				channel->endpoint->ephw->hw_sync_sgl_for_device(
1311					channel->endpoint,
1312					channel->rd_buffers[bufidx]->dma_addr,
1313					channel->rd_buf_size,
1314					DMA_TO_DEVICE);
1315
1316				mutex_lock(&channel->endpoint->register_mutex);
1317
1318				iowrite32(end_offset_plus1 - 1,
1319					  channel->endpoint->registers +
1320					  fpga_buf_offset_reg);
1321
1322				iowrite32((channel->chan_num << 1) |
1323					  (2 << 24) |  /* 2 = submit buffer */
1324					  (bufidx << 12),
1325					  channel->endpoint->registers +
1326					  fpga_buf_ctrl_reg);
1327
1328				mutex_unlock(&channel->endpoint->
1329					     register_mutex);
1330
1331				channel->rd_leftovers[3] =
1332					(channel->rd_host_buf_pos != 0);
1333			}
1334
1335			if (rc) {
1336				mutex_unlock(&channel->rd_mutex);
1337
1338				if (channel->endpoint->fatal_error)
1339					return -EIO;
1340
1341				if (!channel->rd_synchronous)
1342					queue_delayed_work(
1343						xillybus_wq,
1344						&channel->rd_workitem,
1345						XILLY_RX_TIMEOUT);
1346
1347				return rc;
1348			}
1349		}
1350
1351		if (bytes_done >= count)
1352			break;
1353
1354		if (!exhausted)
1355			continue; /* If there's more space, just go on */
1356
1357		if ((bytes_done > 0) && channel->rd_allow_partial)
1358			break;
1359
1360		/*
1361		 * Indefinite sleep with mutex taken. With data waiting for
1362		 * flushing, user should not be surprised if open() for write
1363		 * sleeps.
1364		 */
1365
1366		if (filp->f_flags & O_NONBLOCK) {
1367			rc = -EAGAIN;
1368			break;
1369		}
1370
1371		if (wait_event_interruptible(channel->rd_wait,
1372					     (!channel->rd_full))) {
1373			mutex_unlock(&channel->rd_mutex);
1374
1375			if (channel->endpoint->fatal_error)
1376				return -EIO;
1377
1378			if (bytes_done)
1379				return bytes_done;
1380			return -EINTR;
1381		}
1382	}
1383
1384	mutex_unlock(&channel->rd_mutex);
1385
1386	if (!channel->rd_synchronous)
1387		queue_delayed_work(xillybus_wq,
1388				   &channel->rd_workitem,
1389				   XILLY_RX_TIMEOUT);
1390
1391	if (channel->endpoint->fatal_error)
1392		return -EIO;
1393
1394	if (rc)
1395		return rc;
1396
1397	if ((channel->rd_synchronous) && (bytes_done > 0)) {
1398		rc = xillybus_myflush(filp->private_data, 0); /* No timeout */
1399
1400		if (rc && (rc != -EINTR))
1401			return rc;
1402	}
1403
1404	return bytes_done;
1405}
1406
1407static int xillybus_open(struct inode *inode, struct file *filp)
1408{
1409	int rc = 0;
1410	unsigned long flags;
1411	int minor = iminor(inode);
1412	int major = imajor(inode);
1413	struct xilly_endpoint *ep_iter, *endpoint = NULL;
1414	struct xilly_channel *channel;
1415
1416	mutex_lock(&ep_list_lock);
1417
1418	list_for_each_entry(ep_iter, &list_of_endpoints, ep_list) {
1419		if ((ep_iter->major == major) &&
1420		    (minor >= ep_iter->lowest_minor) &&
1421		    (minor < (ep_iter->lowest_minor +
1422			      ep_iter->num_channels))) {
1423			endpoint = ep_iter;
1424			break;
1425		}
1426	}
1427	mutex_unlock(&ep_list_lock);
1428
1429	if (!endpoint) {
1430		pr_err("xillybus: open() failed to find a device for major=%d and minor=%d\n",
1431		       major, minor);
1432		return -ENODEV;
1433	}
1434
1435	if (endpoint->fatal_error)
1436		return -EIO;
1437
1438	channel = endpoint->channels[1 + minor - endpoint->lowest_minor];
1439	filp->private_data = channel;
1440
1441	/*
1442	 * It gets complicated because:
1443	 * 1. We don't want to take a mutex we don't have to
1444	 * 2. We don't want to open one direction if the other will fail.
1445	 */
1446
1447	if ((filp->f_mode & FMODE_READ) && (!channel->num_wr_buffers))
1448		return -ENODEV;
1449
1450	if ((filp->f_mode & FMODE_WRITE) && (!channel->num_rd_buffers))
1451		return -ENODEV;
1452
1453	if ((filp->f_mode & FMODE_READ) && (filp->f_flags & O_NONBLOCK) &&
1454	    (channel->wr_synchronous || !channel->wr_allow_partial ||
1455	     !channel->wr_supports_nonempty)) {
1456		dev_err(endpoint->dev,
1457			"open() failed: O_NONBLOCK not allowed for read on this device\n");
1458		return -ENODEV;
1459	}
1460
1461	if ((filp->f_mode & FMODE_WRITE) && (filp->f_flags & O_NONBLOCK) &&
1462	    (channel->rd_synchronous || !channel->rd_allow_partial)) {
1463		dev_err(endpoint->dev,
1464			"open() failed: O_NONBLOCK not allowed for write on this device\n");
1465		return -ENODEV;
1466	}
1467
1468	/*
1469	 * Note: open() may block on getting mutexes despite O_NONBLOCK.
1470	 * This shouldn't occur normally, since multiple open of the same
1471	 * file descriptor is almost always prohibited anyhow
1472	 * (*_exclusive_open is normally set in real-life systems).
1473	 */
1474
1475	if (filp->f_mode & FMODE_READ) {
1476		rc = mutex_lock_interruptible(&channel->wr_mutex);
1477		if (rc)
1478			return rc;
1479	}
1480
1481	if (filp->f_mode & FMODE_WRITE) {
1482		rc = mutex_lock_interruptible(&channel->rd_mutex);
1483		if (rc)
1484			goto unlock_wr;
1485	}
1486
1487	if ((filp->f_mode & FMODE_READ) &&
1488	    (channel->wr_ref_count != 0) &&
1489	    (channel->wr_exclusive_open)) {
1490		rc = -EBUSY;
1491		goto unlock;
1492	}
1493
1494	if ((filp->f_mode & FMODE_WRITE) &&
1495	    (channel->rd_ref_count != 0) &&
1496	    (channel->rd_exclusive_open)) {
1497		rc = -EBUSY;
1498		goto unlock;
1499	}
1500
1501	if (filp->f_mode & FMODE_READ) {
1502		if (channel->wr_ref_count == 0) { /* First open of file */
1503			/* Move the host to first buffer */
1504			spin_lock_irqsave(&channel->wr_spinlock, flags);
1505			channel->wr_host_buf_idx = 0;
1506			channel->wr_host_buf_pos = 0;
1507			channel->wr_fpga_buf_idx = -1;
1508			channel->wr_empty = 1;
1509			channel->wr_ready = 0;
1510			channel->wr_sleepy = 1;
1511			channel->wr_eof = -1;
1512			channel->wr_hangup = 0;
1513
1514			spin_unlock_irqrestore(&channel->wr_spinlock, flags);
1515
1516			iowrite32(1 | (channel->chan_num << 1) |
1517				  (4 << 24) |  /* Opcode 4, open channel */
1518				  ((channel->wr_synchronous & 1) << 23),
1519				  channel->endpoint->registers +
1520				  fpga_buf_ctrl_reg);
1521		}
1522
1523		channel->wr_ref_count++;
1524	}
1525
1526	if (filp->f_mode & FMODE_WRITE) {
1527		if (channel->rd_ref_count == 0) { /* First open of file */
1528			/* Move the host to first buffer */
1529			spin_lock_irqsave(&channel->rd_spinlock, flags);
1530			channel->rd_host_buf_idx = 0;
1531			channel->rd_host_buf_pos = 0;
1532			channel->rd_leftovers[3] = 0; /* No leftovers. */
1533			channel->rd_fpga_buf_idx = channel->num_rd_buffers - 1;
1534			channel->rd_full = 0;
1535
1536			spin_unlock_irqrestore(&channel->rd_spinlock, flags);
1537
1538			iowrite32((channel->chan_num << 1) |
1539				  (4 << 24),   /* Opcode 4, open channel */
1540				  channel->endpoint->registers +
1541				  fpga_buf_ctrl_reg);
1542		}
1543
1544		channel->rd_ref_count++;
1545	}
1546
1547unlock:
1548	if (filp->f_mode & FMODE_WRITE)
1549		mutex_unlock(&channel->rd_mutex);
1550unlock_wr:
1551	if (filp->f_mode & FMODE_READ)
1552		mutex_unlock(&channel->wr_mutex);
1553
1554	if (!rc && (!channel->seekable))
1555		return nonseekable_open(inode, filp);
1556
1557	return rc;
1558}
1559
1560static int xillybus_release(struct inode *inode, struct file *filp)
1561{
1562	unsigned long flags;
1563	struct xilly_channel *channel = filp->private_data;
1564
1565	int buf_idx;
1566	int eof;
1567
1568	if (channel->endpoint->fatal_error)
1569		return -EIO;
1570
1571	if (filp->f_mode & FMODE_WRITE) {
1572		mutex_lock(&channel->rd_mutex);
1573
1574		channel->rd_ref_count--;
1575
1576		if (channel->rd_ref_count == 0) {
1577			/*
1578			 * We rely on the kernel calling flush()
1579			 * before we get here.
1580			 */
1581
1582			iowrite32((channel->chan_num << 1) | /* Channel ID */
1583				  (5 << 24),  /* Opcode 5, close channel */
1584				  channel->endpoint->registers +
1585				  fpga_buf_ctrl_reg);
1586		}
1587		mutex_unlock(&channel->rd_mutex);
1588	}
1589
1590	if (filp->f_mode & FMODE_READ) {
1591		mutex_lock(&channel->wr_mutex);
1592
1593		channel->wr_ref_count--;
1594
1595		if (channel->wr_ref_count == 0) {
1596			iowrite32(1 | (channel->chan_num << 1) |
1597				  (5 << 24),  /* Opcode 5, close channel */
1598				  channel->endpoint->registers +
1599				  fpga_buf_ctrl_reg);
1600
1601			/*
1602			 * This is crazily cautious: We make sure that not
1603			 * only that we got an EOF (be it because we closed
1604			 * the channel or because of a user's EOF), but verify
1605			 * that it's one beyond the last buffer arrived, so
1606			 * we have no leftover buffers pending before wrapping
1607			 * up (which can only happen in asynchronous channels,
1608			 * BTW)
1609			 */
1610
1611			while (1) {
1612				spin_lock_irqsave(&channel->wr_spinlock,
1613						  flags);
1614				buf_idx = channel->wr_fpga_buf_idx;
1615				eof = channel->wr_eof;
1616				channel->wr_sleepy = 1;
1617				spin_unlock_irqrestore(&channel->wr_spinlock,
1618						       flags);
1619
1620				/*
1621				 * Check if eof points at the buffer after
1622				 * the last one the FPGA submitted. Note that
1623				 * no EOF is marked by negative eof.
1624				 */
1625
1626				buf_idx++;
1627				if (buf_idx == channel->num_wr_buffers)
1628					buf_idx = 0;
1629
1630				if (buf_idx == eof)
1631					break;
1632
1633				/*
1634				 * Steal extra 100 ms if awaken by interrupt.
1635				 * This is a simple workaround for an
1636				 * interrupt pending when entering, which would
1637				 * otherwise result in declaring the hardware
1638				 * non-responsive.
1639				 */
1640
1641				if (wait_event_interruptible(
1642					    channel->wr_wait,
1643					    (!channel->wr_sleepy)))
1644					msleep(100);
1645
1646				if (channel->wr_sleepy) {
1647					mutex_unlock(&channel->wr_mutex);
1648					dev_warn(channel->endpoint->dev,
1649						 "Hardware failed to respond to close command, therefore left in messy state.\n");
1650					return -EINTR;
1651				}
1652			}
1653		}
1654
1655		mutex_unlock(&channel->wr_mutex);
1656	}
1657
1658	return 0;
1659}
1660
1661static loff_t xillybus_llseek(struct file *filp, loff_t offset, int whence)
1662{
1663	struct xilly_channel *channel = filp->private_data;
1664	loff_t pos = filp->f_pos;
1665	int rc = 0;
1666
1667	/*
1668	 * Take both mutexes not allowing interrupts, since it seems like
1669	 * common applications don't expect an -EINTR here. Besides, multiple
1670	 * access to a single file descriptor on seekable devices is a mess
1671	 * anyhow.
1672	 */
1673
1674	if (channel->endpoint->fatal_error)
1675		return -EIO;
1676
1677	mutex_lock(&channel->wr_mutex);
1678	mutex_lock(&channel->rd_mutex);
1679
1680	switch (whence) {
1681	case SEEK_SET:
1682		pos = offset;
1683		break;
1684	case SEEK_CUR:
1685		pos += offset;
1686		break;
1687	case SEEK_END:
1688		pos = offset; /* Going to the end => to the beginning */
1689		break;
1690	default:
1691		rc = -EINVAL;
1692		goto end;
1693	}
1694
1695	/* In any case, we must finish on an element boundary */
1696	if (pos & ((1 << channel->log2_element_size) - 1)) {
1697		rc = -EINVAL;
1698		goto end;
1699	}
1700
1701	mutex_lock(&channel->endpoint->register_mutex);
1702
1703	iowrite32(pos >> channel->log2_element_size,
1704		  channel->endpoint->registers + fpga_buf_offset_reg);
1705
1706	iowrite32((channel->chan_num << 1) |
1707		  (6 << 24),  /* Opcode 6, set address */
1708		  channel->endpoint->registers + fpga_buf_ctrl_reg);
1709
1710	mutex_unlock(&channel->endpoint->register_mutex);
1711
1712end:
1713	mutex_unlock(&channel->rd_mutex);
1714	mutex_unlock(&channel->wr_mutex);
1715
1716	if (rc) /* Return error after releasing mutexes */
1717		return rc;
1718
1719	filp->f_pos = pos;
1720
1721	/*
1722	 * Since seekable devices are allowed only when the channel is
1723	 * synchronous, we assume that there is no data pending in either
1724	 * direction (which holds true as long as no concurrent access on the
1725	 * file descriptor takes place).
1726	 * The only thing we may need to throw away is leftovers from partial
1727	 * write() flush.
1728	 */
1729
1730	channel->rd_leftovers[3] = 0;
1731
1732	return pos;
1733}
1734
1735static unsigned int xillybus_poll(struct file *filp, poll_table *wait)
1736{
1737	struct xilly_channel *channel = filp->private_data;
1738	unsigned int mask = 0;
1739	unsigned long flags;
1740
1741	poll_wait(filp, &channel->endpoint->ep_wait, wait);
1742
1743	/*
1744	 * poll() won't play ball regarding read() channels which
1745	 * aren't asynchronous and support the nonempty message. Allowing
1746	 * that will create situations where data has been delivered at
1747	 * the FPGA, and users expecting select() to wake up, which it may
1748	 * not.
1749	 */
1750
1751	if (!channel->wr_synchronous && channel->wr_supports_nonempty) {
1752		poll_wait(filp, &channel->wr_wait, wait);
1753		poll_wait(filp, &channel->wr_ready_wait, wait);
1754
1755		spin_lock_irqsave(&channel->wr_spinlock, flags);
1756		if (!channel->wr_empty || channel->wr_ready)
1757			mask |= POLLIN | POLLRDNORM;
1758
1759		if (channel->wr_hangup)
1760			/*
1761			 * Not POLLHUP, because its behavior is in the
1762			 * mist, and POLLIN does what we want: Wake up
1763			 * the read file descriptor so it sees EOF.
1764			 */
1765			mask |=  POLLIN | POLLRDNORM;
1766		spin_unlock_irqrestore(&channel->wr_spinlock, flags);
1767	}
1768
1769	/*
1770	 * If partial data write is disallowed on a write() channel,
1771	 * it's pointless to ever signal OK to write, because is could
1772	 * block despite some space being available.
1773	 */
1774
1775	if (channel->rd_allow_partial) {
1776		poll_wait(filp, &channel->rd_wait, wait);
1777
1778		spin_lock_irqsave(&channel->rd_spinlock, flags);
1779		if (!channel->rd_full)
1780			mask |= POLLOUT | POLLWRNORM;
1781		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
1782	}
1783
1784	if (channel->endpoint->fatal_error)
1785		mask |= POLLERR;
1786
1787	return mask;
1788}
1789
1790static const struct file_operations xillybus_fops = {
1791	.owner      = THIS_MODULE,
1792	.read       = xillybus_read,
1793	.write      = xillybus_write,
1794	.open       = xillybus_open,
1795	.flush      = xillybus_flush,
1796	.release    = xillybus_release,
1797	.llseek     = xillybus_llseek,
1798	.poll       = xillybus_poll,
1799};
1800
1801static int xillybus_init_chrdev(struct xilly_endpoint *endpoint,
1802				const unsigned char *idt)
1803{
1804	int rc;
1805	dev_t dev;
1806	int devnum, i, minor, major;
1807	char devname[48];
1808	struct device *device;
1809
1810	rc = alloc_chrdev_region(&dev, 0, /* minor start */
1811				 endpoint->num_channels,
1812				 xillyname);
1813	if (rc) {
1814		dev_warn(endpoint->dev, "Failed to obtain major/minors");
1815		return rc;
1816	}
1817
1818	endpoint->major = major = MAJOR(dev);
1819	endpoint->lowest_minor = minor = MINOR(dev);
1820
1821	cdev_init(&endpoint->cdev, &xillybus_fops);
1822	endpoint->cdev.owner = endpoint->ephw->owner;
1823	rc = cdev_add(&endpoint->cdev, MKDEV(major, minor),
1824		      endpoint->num_channels);
1825	if (rc) {
1826		dev_warn(endpoint->dev, "Failed to add cdev. Aborting.\n");
1827		goto unregister_chrdev;
1828	}
1829
1830	idt++;
1831
1832	for (i = minor, devnum = 0;
1833	     devnum < endpoint->num_channels;
1834	     devnum++, i++) {
1835		snprintf(devname, sizeof(devname)-1, "xillybus_%s", idt);
1836
1837		devname[sizeof(devname)-1] = 0; /* Should never matter */
1838
1839		while (*idt++)
1840			/* Skip to next */;
1841
1842		device = device_create(xillybus_class,
1843				       NULL,
1844				       MKDEV(major, i),
1845				       NULL,
1846				       "%s", devname);
1847
1848		if (IS_ERR(device)) {
1849			dev_warn(endpoint->dev,
1850				 "Failed to create %s device. Aborting.\n",
1851				 devname);
1852			rc = -ENODEV;
1853			goto unroll_device_create;
1854		}
1855	}
1856
1857	dev_info(endpoint->dev, "Created %d device files.\n",
1858		 endpoint->num_channels);
1859	return 0; /* succeed */
1860
1861unroll_device_create:
1862	devnum--; i--;
1863	for (; devnum >= 0; devnum--, i--)
1864		device_destroy(xillybus_class, MKDEV(major, i));
1865
1866	cdev_del(&endpoint->cdev);
1867unregister_chrdev:
1868	unregister_chrdev_region(MKDEV(major, minor), endpoint->num_channels);
1869
1870	return rc;
1871}
1872
1873static void xillybus_cleanup_chrdev(struct xilly_endpoint *endpoint)
1874{
1875	int minor;
1876
1877	for (minor = endpoint->lowest_minor;
1878	     minor < (endpoint->lowest_minor + endpoint->num_channels);
1879	     minor++)
1880		device_destroy(xillybus_class, MKDEV(endpoint->major, minor));
1881	cdev_del(&endpoint->cdev);
1882	unregister_chrdev_region(MKDEV(endpoint->major,
1883				       endpoint->lowest_minor),
1884				 endpoint->num_channels);
1885
1886	dev_info(endpoint->dev, "Removed %d device files.\n",
1887		 endpoint->num_channels);
1888}
1889
1890struct xilly_endpoint *xillybus_init_endpoint(struct pci_dev *pdev,
1891					      struct device *dev,
1892					      struct xilly_endpoint_hardware
1893					      *ephw)
1894{
1895	struct xilly_endpoint *endpoint;
1896
1897	endpoint = devm_kzalloc(dev, sizeof(*endpoint), GFP_KERNEL);
1898	if (!endpoint)
1899		return NULL;
1900
1901	endpoint->pdev = pdev;
1902	endpoint->dev = dev;
1903	endpoint->ephw = ephw;
1904	endpoint->msg_counter = 0x0b;
1905	endpoint->failed_messages = 0;
1906	endpoint->fatal_error = 0;
1907
1908	init_waitqueue_head(&endpoint->ep_wait);
1909	mutex_init(&endpoint->register_mutex);
1910
1911	return endpoint;
1912}
1913EXPORT_SYMBOL(xillybus_init_endpoint);
1914
1915static int xilly_quiesce(struct xilly_endpoint *endpoint)
1916{
1917	long t;
1918
1919	endpoint->idtlen = -1;
1920
1921	iowrite32((u32) (endpoint->dma_using_dac & 0x0001),
1922		  endpoint->registers + fpga_dma_control_reg);
1923
1924	t = wait_event_interruptible_timeout(endpoint->ep_wait,
1925					     (endpoint->idtlen >= 0),
1926					     XILLY_TIMEOUT);
1927	if (t <= 0) {
1928		dev_err(endpoint->dev,
1929			"Failed to quiesce the device on exit.\n");
1930		return -ENODEV;
1931	}
1932	return 0;
1933}
1934
1935int xillybus_endpoint_discovery(struct xilly_endpoint *endpoint)
1936{
1937	int rc;
1938	long t;
1939
1940	void *bootstrap_resources;
1941	int idtbuffersize = (1 << PAGE_SHIFT);
1942	struct device *dev = endpoint->dev;
1943
1944	/*
1945	 * The bogus IDT is used during bootstrap for allocating the initial
1946	 * message buffer, and then the message buffer and space for the IDT
1947	 * itself. The initial message buffer is of a single page's size, but
1948	 * it's soon replaced with a more modest one (and memory is freed).
1949	 */
1950
1951	unsigned char bogus_idt[8] = { 1, 224, (PAGE_SHIFT)-2, 0,
1952				       3, 192, PAGE_SHIFT, 0 };
1953	struct xilly_idt_handle idt_handle;
1954
1955	/*
1956	 * Writing the value 0x00000001 to Endianness register signals which
1957	 * endianness this processor is using, so the FPGA can swap words as
1958	 * necessary.
1959	 */
1960
1961	iowrite32(1, endpoint->registers + fpga_endian_reg);
1962
1963	/* Bootstrap phase I: Allocate temporary message buffer */
1964
1965	bootstrap_resources = devres_open_group(dev, NULL, GFP_KERNEL);
1966	if (!bootstrap_resources)
1967		return -ENOMEM;
1968
1969	endpoint->num_channels = 0;
1970
1971	rc = xilly_setupchannels(endpoint, bogus_idt, 1);
1972	if (rc)
1973		return rc;
1974
1975	/* Clear the message subsystem (and counter in particular) */
1976	iowrite32(0x04, endpoint->registers + fpga_msg_ctrl_reg);
1977
1978	endpoint->idtlen = -1;
1979
1980	/*
1981	 * Set DMA 32/64 bit mode, quiesce the device (?!) and get IDT
1982	 * buffer size.
1983	 */
1984	iowrite32((u32) (endpoint->dma_using_dac & 0x0001),
1985		  endpoint->registers + fpga_dma_control_reg);
1986
1987	t = wait_event_interruptible_timeout(endpoint->ep_wait,
1988					     (endpoint->idtlen >= 0),
1989					     XILLY_TIMEOUT);
1990	if (t <= 0) {
1991		dev_err(endpoint->dev, "No response from FPGA. Aborting.\n");
1992		return -ENODEV;
1993	}
1994
1995	/* Enable DMA */
1996	iowrite32((u32) (0x0002 | (endpoint->dma_using_dac & 0x0001)),
1997		  endpoint->registers + fpga_dma_control_reg);
1998
1999	/* Bootstrap phase II: Allocate buffer for IDT and obtain it */
2000	while (endpoint->idtlen >= idtbuffersize) {
2001		idtbuffersize *= 2;
2002		bogus_idt[6]++;
2003	}
2004
2005	endpoint->num_channels = 1;
2006
2007	rc = xilly_setupchannels(endpoint, bogus_idt, 2);
2008	if (rc)
2009		goto failed_idt;
2010
2011	rc = xilly_obtain_idt(endpoint);
2012	if (rc)
2013		goto failed_idt;
2014
2015	rc = xilly_scan_idt(endpoint, &idt_handle);
2016	if (rc)
2017		goto failed_idt;
2018
2019	devres_close_group(dev, bootstrap_resources);
2020
2021	/* Bootstrap phase III: Allocate buffers according to IDT */
2022
2023	rc = xilly_setupchannels(endpoint,
2024				 idt_handle.chandesc,
2025				 idt_handle.entries);
2026	if (rc)
2027		goto failed_idt;
2028
2029	/*
2030	 * endpoint is now completely configured. We put it on the list
2031	 * available to open() before registering the char device(s)
2032	 */
2033
2034	mutex_lock(&ep_list_lock);
2035	list_add_tail(&endpoint->ep_list, &list_of_endpoints);
2036	mutex_unlock(&ep_list_lock);
2037
2038	rc = xillybus_init_chrdev(endpoint, idt_handle.idt);
2039	if (rc)
2040		goto failed_chrdevs;
2041
2042	devres_release_group(dev, bootstrap_resources);
2043
2044	return 0;
2045
2046failed_chrdevs:
2047	mutex_lock(&ep_list_lock);
2048	list_del(&endpoint->ep_list);
2049	mutex_unlock(&ep_list_lock);
2050
2051failed_idt:
2052	xilly_quiesce(endpoint);
2053	flush_workqueue(xillybus_wq);
2054
2055	return rc;
2056}
2057EXPORT_SYMBOL(xillybus_endpoint_discovery);
2058
2059void xillybus_endpoint_remove(struct xilly_endpoint *endpoint)
2060{
2061	xillybus_cleanup_chrdev(endpoint);
2062
2063	mutex_lock(&ep_list_lock);
2064	list_del(&endpoint->ep_list);
2065	mutex_unlock(&ep_list_lock);
2066
2067	xilly_quiesce(endpoint);
2068
2069	/*
2070	 * Flushing is done upon endpoint release to prevent access to memory
2071	 * just about to be released. This makes the quiesce complete.
2072	 */
2073	flush_workqueue(xillybus_wq);
2074}
2075EXPORT_SYMBOL(xillybus_endpoint_remove);
2076
2077static int __init xillybus_init(void)
2078{
2079	mutex_init(&ep_list_lock);
2080
2081	xillybus_class = class_create(THIS_MODULE, xillyname);
2082	if (IS_ERR(xillybus_class))
2083		return PTR_ERR(xillybus_class);
2084
2085	xillybus_wq = alloc_workqueue(xillyname, 0, 0);
2086	if (!xillybus_wq) {
2087		class_destroy(xillybus_class);
2088		return -ENOMEM;
2089	}
2090
2091	return 0;
2092}
2093
2094static void __exit xillybus_exit(void)
2095{
2096	/* flush_workqueue() was called for each endpoint released */
2097	destroy_workqueue(xillybus_wq);
2098
2099	class_destroy(xillybus_class);
2100}
2101
2102module_init(xillybus_init);
2103module_exit(xillybus_exit);
2104