xref: /openbmc/linux/drivers/misc/genwqe/card_ddcb.c (revision 8a10bc9d)
1 /**
2  * IBM Accelerator Family 'GenWQE'
3  *
4  * (C) Copyright IBM Corp. 2013
5  *
6  * Author: Frank Haverkamp <haver@linux.vnet.ibm.com>
7  * Author: Joerg-Stephan Vogt <jsvogt@de.ibm.com>
8  * Author: Michael Jung <mijung@de.ibm.com>
9  * Author: Michael Ruettger <michael@ibmra.de>
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License (version 2 only)
13  * as published by the Free Software Foundation.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18  * GNU General Public License for more details.
19  */
20 
21 /*
22  * Device Driver Control Block (DDCB) queue support. Definition of
23  * interrupt handlers for queue support as well as triggering the
24  * health monitor code in case of problems. The current hardware uses
25  * an MSI interrupt which is shared between error handling and
26  * functional code.
27  */
28 
29 #include <linux/types.h>
30 #include <linux/module.h>
31 #include <linux/sched.h>
32 #include <linux/wait.h>
33 #include <linux/pci.h>
34 #include <linux/string.h>
35 #include <linux/dma-mapping.h>
36 #include <linux/delay.h>
37 #include <linux/module.h>
38 #include <linux/interrupt.h>
39 #include <linux/crc-itu-t.h>
40 
41 #include "card_base.h"
42 #include "card_ddcb.h"
43 
44 /*
45  * N: next DDCB, this is where the next DDCB will be put.
46  * A: active DDCB, this is where the code will look for the next completion.
47  * x: DDCB is enqueued, we are waiting for its completion.
48 
49  * Situation (1): Empty queue
50  *  +---+---+---+---+---+---+---+---+
51  *  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
52  *  |   |   |   |   |   |   |   |   |
53  *  +---+---+---+---+---+---+---+---+
54  *           A/N
55  *  enqueued_ddcbs = A - N = 2 - 2 = 0
56  *
57  * Situation (2): Wrapped, N > A
58  *  +---+---+---+---+---+---+---+---+
59  *  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
60  *  |   |   | x | x |   |   |   |   |
61  *  +---+---+---+---+---+---+---+---+
62  *            A       N
63  *  enqueued_ddcbs = N - A = 4 - 2 = 2
64  *
65  * Situation (3): Queue wrapped, A > N
66  *  +---+---+---+---+---+---+---+---+
67  *  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
68  *  | x | x |   |   | x | x | x | x |
69  *  +---+---+---+---+---+---+---+---+
70  *            N       A
71  *  enqueued_ddcbs = queue_max  - (A - N) = 8 - (4 - 2) = 6
72  *
73  * Situation (4a): Queue full N > A
74  *  +---+---+---+---+---+---+---+---+
75  *  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
76  *  | x | x | x | x | x | x | x |   |
77  *  +---+---+---+---+---+---+---+---+
78  *    A                           N
79  *
80  *  enqueued_ddcbs = N - A = 7 - 0 = 7
81  *
82  * Situation (4a): Queue full A > N
83  *  +---+---+---+---+---+---+---+---+
84  *  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
85  *  | x | x | x |   | x | x | x | x |
86  *  +---+---+---+---+---+---+---+---+
87  *                N   A
88  *  enqueued_ddcbs = queue_max - (A - N) = 8 - (4 - 3) = 7
89  */
90 
91 static int queue_empty(struct ddcb_queue *queue)
92 {
93 	return queue->ddcb_next == queue->ddcb_act;
94 }
95 
96 static int queue_enqueued_ddcbs(struct ddcb_queue *queue)
97 {
98 	if (queue->ddcb_next >= queue->ddcb_act)
99 		return queue->ddcb_next - queue->ddcb_act;
100 
101 	return queue->ddcb_max - (queue->ddcb_act - queue->ddcb_next);
102 }
103 
104 static int queue_free_ddcbs(struct ddcb_queue *queue)
105 {
106 	int free_ddcbs = queue->ddcb_max - queue_enqueued_ddcbs(queue) - 1;
107 
108 	if (WARN_ON_ONCE(free_ddcbs < 0)) { /* must never ever happen! */
109 		return 0;
110 	}
111 	return free_ddcbs;
112 }
113 
114 /*
115  * Use of the PRIV field in the DDCB for queue debugging:
116  *
117  * (1) Trying to get rid of a DDCB which saw a timeout:
118  *     pddcb->priv[6] = 0xcc;   # cleared
119  *
120  * (2) Append a DDCB via NEXT bit:
121  *     pddcb->priv[7] = 0xaa;	# appended
122  *
123  * (3) DDCB needed tapping:
124  *     pddcb->priv[7] = 0xbb;   # tapped
125  *
126  * (4) DDCB marked as correctly finished:
127  *     pddcb->priv[6] = 0xff;	# finished
128  */
129 
130 static inline void ddcb_mark_tapped(struct ddcb *pddcb)
131 {
132 	pddcb->priv[7] = 0xbb;  /* tapped */
133 }
134 
135 static inline void ddcb_mark_appended(struct ddcb *pddcb)
136 {
137 	pddcb->priv[7] = 0xaa;	/* appended */
138 }
139 
140 static inline void ddcb_mark_cleared(struct ddcb *pddcb)
141 {
142 	pddcb->priv[6] = 0xcc; /* cleared */
143 }
144 
145 static inline void ddcb_mark_finished(struct ddcb *pddcb)
146 {
147 	pddcb->priv[6] = 0xff;	/* finished */
148 }
149 
150 static inline void ddcb_mark_unused(struct ddcb *pddcb)
151 {
152 	pddcb->priv_64 = cpu_to_be64(0); /* not tapped */
153 }
154 
155 /**
156  * genwqe_crc16() - Generate 16-bit crc as required for DDCBs
157  * @buff:       pointer to data buffer
158  * @len:        length of data for calculation
159  * @init:       initial crc (0xffff at start)
160  *
161  * Polynomial = x^16 + x^12 + x^5 + 1   (0x1021)
162  * Example: 4 bytes 0x01 0x02 0x03 0x04 with init = 0xffff
163  *          should result in a crc16 of 0x89c3
164  *
165  * Return: crc16 checksum in big endian format !
166  */
167 static inline u16 genwqe_crc16(const u8 *buff, size_t len, u16 init)
168 {
169 	return crc_itu_t(init, buff, len);
170 }
171 
172 static void print_ddcb_info(struct genwqe_dev *cd, struct ddcb_queue *queue)
173 {
174 	int i;
175 	struct ddcb *pddcb;
176 	unsigned long flags;
177 	struct pci_dev *pci_dev = cd->pci_dev;
178 
179 	spin_lock_irqsave(&cd->print_lock, flags);
180 
181 	dev_info(&pci_dev->dev,
182 		 "DDCB list for card #%d (ddcb_act=%d / ddcb_next=%d):\n",
183 		 cd->card_idx, queue->ddcb_act, queue->ddcb_next);
184 
185 	pddcb = queue->ddcb_vaddr;
186 	for (i = 0; i < queue->ddcb_max; i++) {
187 		dev_err(&pci_dev->dev,
188 			"  %c %-3d: RETC=%03x SEQ=%04x "
189 			"HSI=%02X SHI=%02x PRIV=%06llx CMD=%03x\n",
190 			i == queue->ddcb_act ? '>' : ' ',
191 			i,
192 			be16_to_cpu(pddcb->retc_16),
193 			be16_to_cpu(pddcb->seqnum_16),
194 			pddcb->hsi,
195 			pddcb->shi,
196 			be64_to_cpu(pddcb->priv_64),
197 			pddcb->cmd);
198 		pddcb++;
199 	}
200 	spin_unlock_irqrestore(&cd->print_lock, flags);
201 }
202 
203 struct genwqe_ddcb_cmd *ddcb_requ_alloc(void)
204 {
205 	struct ddcb_requ *req;
206 
207 	req = kzalloc(sizeof(*req), GFP_ATOMIC);
208 	if (!req)
209 		return NULL;
210 
211 	return &req->cmd;
212 }
213 
214 void ddcb_requ_free(struct genwqe_ddcb_cmd *cmd)
215 {
216 	struct ddcb_requ *req = container_of(cmd, struct ddcb_requ, cmd);
217 	kfree(req);
218 }
219 
220 static inline enum genwqe_requ_state ddcb_requ_get_state(struct ddcb_requ *req)
221 {
222 	return req->req_state;
223 }
224 
225 static inline void ddcb_requ_set_state(struct ddcb_requ *req,
226 				       enum genwqe_requ_state new_state)
227 {
228 	req->req_state = new_state;
229 }
230 
231 static inline int ddcb_requ_collect_debug_data(struct ddcb_requ *req)
232 {
233 	return req->cmd.ddata_addr != 0x0;
234 }
235 
236 /**
237  * ddcb_requ_finished() - Returns the hardware state of the associated DDCB
238  * @cd:          pointer to genwqe device descriptor
239  * @req:         DDCB work request
240  *
241  * Status of ddcb_requ mirrors this hardware state, but is copied in
242  * the ddcb_requ on interrupt/polling function. The lowlevel code
243  * should check the hardware state directly, the higher level code
244  * should check the copy.
245  *
246  * This function will also return true if the state of the queue is
247  * not GENWQE_CARD_USED. This enables us to purge all DDCBs in the
248  * shutdown case.
249  */
250 static int ddcb_requ_finished(struct genwqe_dev *cd, struct ddcb_requ *req)
251 {
252 	return (ddcb_requ_get_state(req) == GENWQE_REQU_FINISHED) ||
253 		(cd->card_state != GENWQE_CARD_USED);
254 }
255 
256 /**
257  * enqueue_ddcb() - Enqueue a DDCB
258  * @cd:         pointer to genwqe device descriptor
259  * @queue:	queue this operation should be done on
260  * @ddcb_no:    pointer to ddcb number being tapped
261  *
262  * Start execution of DDCB by tapping or append to queue via NEXT
263  * bit. This is done by an atomic 'compare and swap' instruction and
264  * checking SHI and HSI of the previous DDCB.
265  *
266  * This function must only be called with ddcb_lock held.
267  *
268  * Return: 1 if new DDCB is appended to previous
269  *         2 if DDCB queue is tapped via register/simulation
270  */
271 #define RET_DDCB_APPENDED 1
272 #define RET_DDCB_TAPPED   2
273 
274 static int enqueue_ddcb(struct genwqe_dev *cd, struct ddcb_queue *queue,
275 			struct ddcb *pddcb, int ddcb_no)
276 {
277 	unsigned int try;
278 	int prev_no;
279 	struct ddcb *prev_ddcb;
280 	__be32 old, new, icrc_hsi_shi;
281 	u64 num;
282 
283 	/*
284 	 * For performance checks a Dispatch Timestamp can be put into
285 	 * DDCB It is supposed to use the SLU's free running counter,
286 	 * but this requires PCIe cycles.
287 	 */
288 	ddcb_mark_unused(pddcb);
289 
290 	/* check previous DDCB if already fetched */
291 	prev_no = (ddcb_no == 0) ? queue->ddcb_max - 1 : ddcb_no - 1;
292 	prev_ddcb = &queue->ddcb_vaddr[prev_no];
293 
294 	/*
295 	 * It might have happened that the HSI.FETCHED bit is
296 	 * set. Retry in this case. Therefore I expect maximum 2 times
297 	 * trying.
298 	 */
299 	ddcb_mark_appended(pddcb);
300 	for (try = 0; try < 2; try++) {
301 		old = prev_ddcb->icrc_hsi_shi_32; /* read SHI/HSI in BE32 */
302 
303 		/* try to append via NEXT bit if prev DDCB is not completed */
304 		if ((old & DDCB_COMPLETED_BE32) != 0x00000000)
305 			break;
306 
307 		new = (old | DDCB_NEXT_BE32);
308 		icrc_hsi_shi = cmpxchg(&prev_ddcb->icrc_hsi_shi_32, old, new);
309 
310 		if (icrc_hsi_shi == old)
311 			return RET_DDCB_APPENDED; /* appended to queue */
312 	}
313 
314 	/* Queue must be re-started by updating QUEUE_OFFSET */
315 	ddcb_mark_tapped(pddcb);
316 	num = (u64)ddcb_no << 8;
317 	__genwqe_writeq(cd, queue->IO_QUEUE_OFFSET, num); /* start queue */
318 
319 	return RET_DDCB_TAPPED;
320 }
321 
322 /**
323  * copy_ddcb_results() - Copy output state from real DDCB to request
324  *
325  * Copy DDCB ASV to request struct. There is no endian
326  * conversion made, since data structure in ASV is still
327  * unknown here.
328  *
329  * This is needed by:
330  *   - genwqe_purge_ddcb()
331  *   - genwqe_check_ddcb_queue()
332  */
333 static void copy_ddcb_results(struct ddcb_requ *req, int ddcb_no)
334 {
335 	struct ddcb_queue *queue = req->queue;
336 	struct ddcb *pddcb = &queue->ddcb_vaddr[req->num];
337 
338 	memcpy(&req->cmd.asv[0], &pddcb->asv[0], DDCB_ASV_LENGTH);
339 
340 	/* copy status flags of the variant part */
341 	req->cmd.vcrc     = be16_to_cpu(pddcb->vcrc_16);
342 	req->cmd.deque_ts = be64_to_cpu(pddcb->deque_ts_64);
343 	req->cmd.cmplt_ts = be64_to_cpu(pddcb->cmplt_ts_64);
344 
345 	req->cmd.attn     = be16_to_cpu(pddcb->attn_16);
346 	req->cmd.progress = be32_to_cpu(pddcb->progress_32);
347 	req->cmd.retc     = be16_to_cpu(pddcb->retc_16);
348 
349 	if (ddcb_requ_collect_debug_data(req)) {
350 		int prev_no = (ddcb_no == 0) ?
351 			queue->ddcb_max - 1 : ddcb_no - 1;
352 		struct ddcb *prev_pddcb = &queue->ddcb_vaddr[prev_no];
353 
354 		memcpy(&req->debug_data.ddcb_finished, pddcb,
355 		       sizeof(req->debug_data.ddcb_finished));
356 		memcpy(&req->debug_data.ddcb_prev, prev_pddcb,
357 		       sizeof(req->debug_data.ddcb_prev));
358 	}
359 }
360 
361 /**
362  * genwqe_check_ddcb_queue() - Checks DDCB queue for completed work equests.
363  * @cd:         pointer to genwqe device descriptor
364  *
365  * Return: Number of DDCBs which were finished
366  */
367 static int genwqe_check_ddcb_queue(struct genwqe_dev *cd,
368 				   struct ddcb_queue *queue)
369 {
370 	unsigned long flags;
371 	int ddcbs_finished = 0;
372 	struct pci_dev *pci_dev = cd->pci_dev;
373 
374 	spin_lock_irqsave(&queue->ddcb_lock, flags);
375 
376 	/* FIXME avoid soft locking CPU */
377 	while (!queue_empty(queue) && (ddcbs_finished < queue->ddcb_max)) {
378 
379 		struct ddcb *pddcb;
380 		struct ddcb_requ *req;
381 		u16 vcrc, vcrc_16, retc_16;
382 
383 		pddcb = &queue->ddcb_vaddr[queue->ddcb_act];
384 
385 		if ((pddcb->icrc_hsi_shi_32 & DDCB_COMPLETED_BE32) ==
386 		    0x00000000)
387 			goto go_home; /* not completed, continue waiting */
388 
389 		/* Note: DDCB could be purged */
390 
391 		req = queue->ddcb_req[queue->ddcb_act];
392 		if (req == NULL) {
393 			/* this occurs if DDCB is purged, not an error */
394 			/* Move active DDCB further; Nothing to do anymore. */
395 			goto pick_next_one;
396 		}
397 
398 		/*
399 		 * HSI=0x44 (fetched and completed), but RETC is
400 		 * 0x101, or even worse 0x000.
401 		 *
402 		 * In case of seeing the queue in inconsistent state
403 		 * we read the errcnts and the queue status to provide
404 		 * a trigger for our PCIe analyzer stop capturing.
405 		 */
406 		retc_16 = be16_to_cpu(pddcb->retc_16);
407 		if ((pddcb->hsi == 0x44) && (retc_16 <= 0x101)) {
408 			u64 errcnts, status;
409 			u64 ddcb_offs = (u64)pddcb - (u64)queue->ddcb_vaddr;
410 
411 			errcnts = __genwqe_readq(cd, queue->IO_QUEUE_ERRCNTS);
412 			status  = __genwqe_readq(cd, queue->IO_QUEUE_STATUS);
413 
414 			dev_err(&pci_dev->dev,
415 				"[%s] SEQN=%04x HSI=%02x RETC=%03x "
416 				" Q_ERRCNTS=%016llx Q_STATUS=%016llx\n"
417 				" DDCB_DMA_ADDR=%016llx\n",
418 				__func__, be16_to_cpu(pddcb->seqnum_16),
419 				pddcb->hsi, retc_16, errcnts, status,
420 				queue->ddcb_daddr + ddcb_offs);
421 		}
422 
423 		copy_ddcb_results(req, queue->ddcb_act);
424 		queue->ddcb_req[queue->ddcb_act] = NULL; /* take from queue */
425 
426 		dev_dbg(&pci_dev->dev, "FINISHED DDCB#%d\n", req->num);
427 		genwqe_hexdump(pci_dev, pddcb, sizeof(*pddcb));
428 
429 		ddcb_mark_finished(pddcb);
430 
431 		/* calculate CRC_16 to see if VCRC is correct */
432 		vcrc = genwqe_crc16(pddcb->asv,
433 				   VCRC_LENGTH(req->cmd.asv_length),
434 				   0xffff);
435 		vcrc_16 = be16_to_cpu(pddcb->vcrc_16);
436 		if (vcrc != vcrc_16) {
437 			printk_ratelimited(KERN_ERR
438 				"%s %s: err: wrong VCRC pre=%02x vcrc_len=%d "
439 				"bytes vcrc_data=%04x is not vcrc_card=%04x\n",
440 				GENWQE_DEVNAME, dev_name(&pci_dev->dev),
441 				pddcb->pre, VCRC_LENGTH(req->cmd.asv_length),
442 				vcrc, vcrc_16);
443 		}
444 
445 		ddcb_requ_set_state(req, GENWQE_REQU_FINISHED);
446 		queue->ddcbs_completed++;
447 		queue->ddcbs_in_flight--;
448 
449 		/* wake up process waiting for this DDCB */
450 		wake_up_interruptible(&queue->ddcb_waitqs[queue->ddcb_act]);
451 
452 pick_next_one:
453 		queue->ddcb_act = (queue->ddcb_act + 1) % queue->ddcb_max;
454 		ddcbs_finished++;
455 	}
456 
457  go_home:
458 	spin_unlock_irqrestore(&queue->ddcb_lock, flags);
459 	return ddcbs_finished;
460 }
461 
462 /**
463  * __genwqe_wait_ddcb(): Waits until DDCB is completed
464  * @cd:         pointer to genwqe device descriptor
465  * @req:        pointer to requsted DDCB parameters
466  *
467  * The Service Layer will update the RETC in DDCB when processing is
468  * pending or done.
469  *
470  * Return: > 0 remaining jiffies, DDCB completed
471  *           -ETIMEDOUT	when timeout
472  *           -ERESTARTSYS when ^C
473  *           -EINVAL when unknown error condition
474  *
475  * When an error is returned the called needs to ensure that
476  * purge_ddcb() is being called to get the &req removed from the
477  * queue.
478  */
479 int __genwqe_wait_ddcb(struct genwqe_dev *cd, struct ddcb_requ *req)
480 {
481 	int rc;
482 	unsigned int ddcb_no;
483 	struct ddcb_queue *queue;
484 	struct pci_dev *pci_dev = cd->pci_dev;
485 
486 	if (req == NULL)
487 		return -EINVAL;
488 
489 	queue = req->queue;
490 	if (queue == NULL)
491 		return -EINVAL;
492 
493 	ddcb_no = req->num;
494 	if (ddcb_no >= queue->ddcb_max)
495 		return -EINVAL;
496 
497 	rc = wait_event_interruptible_timeout(queue->ddcb_waitqs[ddcb_no],
498 				ddcb_requ_finished(cd, req),
499 				genwqe_ddcb_software_timeout * HZ);
500 
501 	/*
502 	 * We need to distinguish 3 cases here:
503 	 *   1. rc == 0              timeout occured
504 	 *   2. rc == -ERESTARTSYS   signal received
505 	 *   3. rc > 0               remaining jiffies condition is true
506 	 */
507 	if (rc == 0) {
508 		struct ddcb_queue *queue = req->queue;
509 		struct ddcb *pddcb;
510 
511 		/*
512 		 * Timeout may be caused by long task switching time.
513 		 * When timeout happens, check if the request has
514 		 * meanwhile completed.
515 		 */
516 		genwqe_check_ddcb_queue(cd, req->queue);
517 		if (ddcb_requ_finished(cd, req))
518 			return rc;
519 
520 		dev_err(&pci_dev->dev,
521 			"[%s] err: DDCB#%d timeout rc=%d state=%d req @ %p\n",
522 			__func__, req->num, rc,	ddcb_requ_get_state(req),
523 			req);
524 		dev_err(&pci_dev->dev,
525 			"[%s]      IO_QUEUE_STATUS=0x%016llx\n", __func__,
526 			__genwqe_readq(cd, queue->IO_QUEUE_STATUS));
527 
528 		pddcb = &queue->ddcb_vaddr[req->num];
529 		genwqe_hexdump(pci_dev, pddcb, sizeof(*pddcb));
530 
531 		print_ddcb_info(cd, req->queue);
532 		return -ETIMEDOUT;
533 
534 	} else if (rc == -ERESTARTSYS) {
535 		return rc;
536 		/*
537 		 * EINTR:       Stops the application
538 		 * ERESTARTSYS: Restartable systemcall; called again
539 		 */
540 
541 	} else if (rc < 0) {
542 		dev_err(&pci_dev->dev,
543 			"[%s] err: DDCB#%d unknown result (rc=%d) %d!\n",
544 			__func__, req->num, rc, ddcb_requ_get_state(req));
545 		return -EINVAL;
546 	}
547 
548 	/* Severe error occured. Driver is forced to stop operation */
549 	if (cd->card_state != GENWQE_CARD_USED) {
550 		dev_err(&pci_dev->dev,
551 			"[%s] err: DDCB#%d forced to stop (rc=%d)\n",
552 			__func__, req->num, rc);
553 		return -EIO;
554 	}
555 	return rc;
556 }
557 
558 /**
559  * get_next_ddcb() - Get next available DDCB
560  * @cd:         pointer to genwqe device descriptor
561  *
562  * DDCB's content is completely cleared but presets for PRE and
563  * SEQNUM. This function must only be called when ddcb_lock is held.
564  *
565  * Return: NULL if no empty DDCB available otherwise ptr to next DDCB.
566  */
567 static struct ddcb *get_next_ddcb(struct genwqe_dev *cd,
568 				  struct ddcb_queue *queue,
569 				  int *num)
570 {
571 	u64 *pu64;
572 	struct ddcb *pddcb;
573 
574 	if (queue_free_ddcbs(queue) == 0) /* queue is  full */
575 		return NULL;
576 
577 	/* find new ddcb */
578 	pddcb = &queue->ddcb_vaddr[queue->ddcb_next];
579 
580 	/* if it is not completed, we are not allowed to use it */
581 	/* barrier(); */
582 	if ((pddcb->icrc_hsi_shi_32 & DDCB_COMPLETED_BE32) == 0x00000000)
583 		return NULL;
584 
585 	*num = queue->ddcb_next;	/* internal DDCB number */
586 	queue->ddcb_next = (queue->ddcb_next + 1) % queue->ddcb_max;
587 
588 	/* clear important DDCB fields */
589 	pu64 = (u64 *)pddcb;
590 	pu64[0] = 0ULL;		/* offs 0x00 (ICRC,HSI,SHI,...) */
591 	pu64[1] = 0ULL;		/* offs 0x01 (ACFUNC,CMD...) */
592 
593 	/* destroy previous results in ASV */
594 	pu64[0x80/8] = 0ULL;	/* offs 0x80 (ASV + 0) */
595 	pu64[0x88/8] = 0ULL;	/* offs 0x88 (ASV + 0x08) */
596 	pu64[0x90/8] = 0ULL;	/* offs 0x90 (ASV + 0x10) */
597 	pu64[0x98/8] = 0ULL;	/* offs 0x98 (ASV + 0x18) */
598 	pu64[0xd0/8] = 0ULL;	/* offs 0xd0 (RETC,ATTN...) */
599 
600 	pddcb->pre = DDCB_PRESET_PRE; /* 128 */
601 	pddcb->seqnum_16 = cpu_to_be16(queue->ddcb_seq++);
602 	return pddcb;
603 }
604 
605 /**
606  * __genwqe_purge_ddcb() - Remove a DDCB from the workqueue
607  * @cd:         genwqe device descriptor
608  * @req:        DDCB request
609  *
610  * This will fail when the request was already FETCHED. In this case
611  * we need to wait until it is finished. Else the DDCB can be
612  * reused. This function also ensures that the request data structure
613  * is removed from ddcb_req[].
614  *
615  * Do not forget to call this function when genwqe_wait_ddcb() fails,
616  * such that the request gets really removed from ddcb_req[].
617  *
618  * Return: 0 success
619  */
620 int __genwqe_purge_ddcb(struct genwqe_dev *cd, struct ddcb_requ *req)
621 {
622 	struct ddcb *pddcb = NULL;
623 	unsigned int t;
624 	unsigned long flags;
625 	struct ddcb_queue *queue = req->queue;
626 	struct pci_dev *pci_dev = cd->pci_dev;
627 	u64 queue_status;
628 	__be32 icrc_hsi_shi = 0x0000;
629 	__be32 old, new;
630 
631 	/* unsigned long flags; */
632 	if (genwqe_ddcb_software_timeout <= 0) {
633 		dev_err(&pci_dev->dev,
634 			"[%s] err: software timeout is not set!\n", __func__);
635 		return -EFAULT;
636 	}
637 
638 	pddcb = &queue->ddcb_vaddr[req->num];
639 
640 	for (t = 0; t < genwqe_ddcb_software_timeout * 10; t++) {
641 
642 		spin_lock_irqsave(&queue->ddcb_lock, flags);
643 
644 		/* Check if req was meanwhile finished */
645 		if (ddcb_requ_get_state(req) == GENWQE_REQU_FINISHED)
646 			goto go_home;
647 
648 		/* try to set PURGE bit if FETCHED/COMPLETED are not set */
649 		old = pddcb->icrc_hsi_shi_32;	/* read SHI/HSI in BE32 */
650 		if ((old & DDCB_FETCHED_BE32) == 0x00000000) {
651 
652 			new = (old | DDCB_PURGE_BE32);
653 			icrc_hsi_shi = cmpxchg(&pddcb->icrc_hsi_shi_32,
654 					       old, new);
655 			if (icrc_hsi_shi == old)
656 				goto finish_ddcb;
657 		}
658 
659 		/* normal finish with HSI bit */
660 		barrier();
661 		icrc_hsi_shi = pddcb->icrc_hsi_shi_32;
662 		if (icrc_hsi_shi & DDCB_COMPLETED_BE32)
663 			goto finish_ddcb;
664 
665 		spin_unlock_irqrestore(&queue->ddcb_lock, flags);
666 
667 		/*
668 		 * Here the check_ddcb() function will most likely
669 		 * discover this DDCB to be finished some point in
670 		 * time. It will mark the req finished and free it up
671 		 * in the list.
672 		 */
673 
674 		copy_ddcb_results(req, req->num); /* for the failing case */
675 		msleep(100); /* sleep for 1/10 second and try again */
676 		continue;
677 
678 finish_ddcb:
679 		copy_ddcb_results(req, req->num);
680 		ddcb_requ_set_state(req, GENWQE_REQU_FINISHED);
681 		queue->ddcbs_in_flight--;
682 		queue->ddcb_req[req->num] = NULL; /* delete from array */
683 		ddcb_mark_cleared(pddcb);
684 
685 		/* Move active DDCB further; Nothing to do here anymore. */
686 
687 		/*
688 		 * We need to ensure that there is at least one free
689 		 * DDCB in the queue. To do that, we must update
690 		 * ddcb_act only if the COMPLETED bit is set for the
691 		 * DDCB we are working on else we treat that DDCB even
692 		 * if we PURGED it as occupied (hardware is supposed
693 		 * to set the COMPLETED bit yet!).
694 		 */
695 		icrc_hsi_shi = pddcb->icrc_hsi_shi_32;
696 		if ((icrc_hsi_shi & DDCB_COMPLETED_BE32) &&
697 		    (queue->ddcb_act == req->num)) {
698 			queue->ddcb_act = ((queue->ddcb_act + 1) %
699 					   queue->ddcb_max);
700 		}
701 go_home:
702 		spin_unlock_irqrestore(&queue->ddcb_lock, flags);
703 		return 0;
704 	}
705 
706 	/*
707 	 * If the card is dead and the queue is forced to stop, we
708 	 * might see this in the queue status register.
709 	 */
710 	queue_status = __genwqe_readq(cd, queue->IO_QUEUE_STATUS);
711 
712 	dev_dbg(&pci_dev->dev, "UN/FINISHED DDCB#%d\n", req->num);
713 	genwqe_hexdump(pci_dev, pddcb, sizeof(*pddcb));
714 
715 	dev_err(&pci_dev->dev,
716 		"[%s] err: DDCB#%d not purged and not completed "
717 		"after %d seconds QSTAT=%016llx!!\n",
718 		__func__, req->num, genwqe_ddcb_software_timeout,
719 		queue_status);
720 
721 	print_ddcb_info(cd, req->queue);
722 
723 	return -EFAULT;
724 }
725 
726 int genwqe_init_debug_data(struct genwqe_dev *cd, struct genwqe_debug_data *d)
727 {
728 	int len;
729 	struct pci_dev *pci_dev = cd->pci_dev;
730 
731 	if (d == NULL) {
732 		dev_err(&pci_dev->dev,
733 			"[%s] err: invalid memory for debug data!\n",
734 			__func__);
735 		return -EFAULT;
736 	}
737 
738 	len  = sizeof(d->driver_version);
739 	snprintf(d->driver_version, len, "%s", DRV_VERS_STRING);
740 	d->slu_unitcfg = cd->slu_unitcfg;
741 	d->app_unitcfg = cd->app_unitcfg;
742 	return 0;
743 }
744 
745 /**
746  * __genwqe_enqueue_ddcb() - Enqueue a DDCB
747  * @cd:          pointer to genwqe device descriptor
748  * @req:         pointer to DDCB execution request
749  *
750  * Return: 0 if enqueuing succeeded
751  *         -EIO if card is unusable/PCIe problems
752  *         -EBUSY if enqueuing failed
753  */
754 int __genwqe_enqueue_ddcb(struct genwqe_dev *cd, struct ddcb_requ *req)
755 {
756 	struct ddcb *pddcb;
757 	unsigned long flags;
758 	struct ddcb_queue *queue;
759 	struct pci_dev *pci_dev = cd->pci_dev;
760 	u16 icrc;
761 
762 	if (cd->card_state != GENWQE_CARD_USED) {
763 		printk_ratelimited(KERN_ERR
764 			"%s %s: [%s] Card is unusable/PCIe problem Req#%d\n",
765 			GENWQE_DEVNAME, dev_name(&pci_dev->dev),
766 			__func__, req->num);
767 		return -EIO;
768 	}
769 
770 	queue = req->queue = &cd->queue;
771 
772 	/* FIXME circumvention to improve performance when no irq is
773 	 * there.
774 	 */
775 	if (genwqe_polling_enabled)
776 		genwqe_check_ddcb_queue(cd, queue);
777 
778 	/*
779 	 * It must be ensured to process all DDCBs in successive
780 	 * order. Use a lock here in order to prevent nested DDCB
781 	 * enqueuing.
782 	 */
783 	spin_lock_irqsave(&queue->ddcb_lock, flags);
784 
785 	pddcb = get_next_ddcb(cd, queue, &req->num);	/* get ptr and num */
786 	if (pddcb == NULL) {
787 		spin_unlock_irqrestore(&queue->ddcb_lock, flags);
788 		queue->busy++;
789 		return -EBUSY;
790 	}
791 
792 	if (queue->ddcb_req[req->num] != NULL) {
793 		spin_unlock_irqrestore(&queue->ddcb_lock, flags);
794 
795 		dev_err(&pci_dev->dev,
796 			"[%s] picked DDCB %d with req=%p still in use!!\n",
797 			__func__, req->num, req);
798 		return -EFAULT;
799 	}
800 	ddcb_requ_set_state(req, GENWQE_REQU_ENQUEUED);
801 	queue->ddcb_req[req->num] = req;
802 
803 	pddcb->cmdopts_16 = cpu_to_be16(req->cmd.cmdopts);
804 	pddcb->cmd = req->cmd.cmd;
805 	pddcb->acfunc = req->cmd.acfunc;	/* functional unit */
806 
807 	/*
808 	 * We know that we can get retc 0x104 with CRC error, do not
809 	 * stop the queue in those cases for this command. XDIR = 1
810 	 * does not work for old SLU versions.
811 	 *
812 	 * Last bitstream with the old XDIR behavior had SLU_ID
813 	 * 0x34199.
814 	 */
815 	if ((cd->slu_unitcfg & 0xFFFF0ull) > 0x34199ull)
816 		pddcb->xdir = 0x1;
817 	else
818 		pddcb->xdir = 0x0;
819 
820 
821 	pddcb->psp = (((req->cmd.asiv_length / 8) << 4) |
822 		      ((req->cmd.asv_length  / 8)));
823 	pddcb->disp_ts_64 = cpu_to_be64(req->cmd.disp_ts);
824 
825 	/*
826 	 * If copying the whole DDCB_ASIV_LENGTH is impacting
827 	 * performance we need to change it to
828 	 * req->cmd.asiv_length. But simulation benefits from some
829 	 * non-architectured bits behind the architectured content.
830 	 *
831 	 * How much data is copied depends on the availability of the
832 	 * ATS field, which was introduced late. If the ATS field is
833 	 * supported ASIV is 8 bytes shorter than it used to be. Since
834 	 * the ATS field is copied too, the code should do exactly
835 	 * what it did before, but I wanted to make copying of the ATS
836 	 * field very explicit.
837 	 */
838 	if (genwqe_get_slu_id(cd) <= 0x2) {
839 		memcpy(&pddcb->__asiv[0],	/* destination */
840 		       &req->cmd.__asiv[0],	/* source */
841 		       DDCB_ASIV_LENGTH);	/* req->cmd.asiv_length */
842 	} else {
843 		pddcb->n.ats_64 = cpu_to_be64(req->cmd.ats);
844 		memcpy(&pddcb->n.asiv[0],	/* destination */
845 			&req->cmd.asiv[0],	/* source */
846 			DDCB_ASIV_LENGTH_ATS);	/* req->cmd.asiv_length */
847 	}
848 
849 	pddcb->icrc_hsi_shi_32 = cpu_to_be32(0x00000000); /* for crc */
850 
851 	/*
852 	 * Calculate CRC_16 for corresponding range PSP(7:4). Include
853 	 * empty 4 bytes prior to the data.
854 	 */
855 	icrc = genwqe_crc16((const u8 *)pddcb,
856 			   ICRC_LENGTH(req->cmd.asiv_length), 0xffff);
857 	pddcb->icrc_hsi_shi_32 = cpu_to_be32((u32)icrc << 16);
858 
859 	/* enable DDCB completion irq */
860 	if (!genwqe_polling_enabled)
861 		pddcb->icrc_hsi_shi_32 |= DDCB_INTR_BE32;
862 
863 	dev_dbg(&pci_dev->dev, "INPUT DDCB#%d\n", req->num);
864 	genwqe_hexdump(pci_dev, pddcb, sizeof(*pddcb));
865 
866 	if (ddcb_requ_collect_debug_data(req)) {
867 		/* use the kernel copy of debug data. copying back to
868 		   user buffer happens later */
869 
870 		genwqe_init_debug_data(cd, &req->debug_data);
871 		memcpy(&req->debug_data.ddcb_before, pddcb,
872 		       sizeof(req->debug_data.ddcb_before));
873 	}
874 
875 	enqueue_ddcb(cd, queue, pddcb, req->num);
876 	queue->ddcbs_in_flight++;
877 
878 	if (queue->ddcbs_in_flight > queue->ddcbs_max_in_flight)
879 		queue->ddcbs_max_in_flight = queue->ddcbs_in_flight;
880 
881 	ddcb_requ_set_state(req, GENWQE_REQU_TAPPED);
882 	spin_unlock_irqrestore(&queue->ddcb_lock, flags);
883 	wake_up_interruptible(&cd->queue_waitq);
884 
885 	return 0;
886 }
887 
888 /**
889  * __genwqe_execute_raw_ddcb() - Setup and execute DDCB
890  * @cd:         pointer to genwqe device descriptor
891  * @req:        user provided DDCB request
892  */
893 int __genwqe_execute_raw_ddcb(struct genwqe_dev *cd,
894 			     struct genwqe_ddcb_cmd *cmd)
895 {
896 	int rc = 0;
897 	struct pci_dev *pci_dev = cd->pci_dev;
898 	struct ddcb_requ *req = container_of(cmd, struct ddcb_requ, cmd);
899 
900 	if (cmd->asiv_length > DDCB_ASIV_LENGTH) {
901 		dev_err(&pci_dev->dev, "[%s] err: wrong asiv_length of %d\n",
902 			__func__, cmd->asiv_length);
903 		return -EINVAL;
904 	}
905 	if (cmd->asv_length > DDCB_ASV_LENGTH) {
906 		dev_err(&pci_dev->dev, "[%s] err: wrong asv_length of %d\n",
907 			__func__, cmd->asiv_length);
908 		return -EINVAL;
909 	}
910 	rc = __genwqe_enqueue_ddcb(cd, req);
911 	if (rc != 0)
912 		return rc;
913 
914 	rc = __genwqe_wait_ddcb(cd, req);
915 	if (rc < 0)		/* error or signal interrupt */
916 		goto err_exit;
917 
918 	if (ddcb_requ_collect_debug_data(req)) {
919 		if (copy_to_user((struct genwqe_debug_data __user *)
920 				 (unsigned long)cmd->ddata_addr,
921 				 &req->debug_data,
922 				 sizeof(struct genwqe_debug_data)))
923 			return -EFAULT;
924 	}
925 
926 	/*
927 	 * Higher values than 0x102 indicate completion with faults,
928 	 * lower values than 0x102 indicate processing faults. Note
929 	 * that DDCB might have been purged. E.g. Cntl+C.
930 	 */
931 	if (cmd->retc != DDCB_RETC_COMPLETE) {
932 		/* This might happen e.g. flash read, and needs to be
933 		   handled by the upper layer code. */
934 		rc = -EBADMSG;	/* not processed/error retc */
935 	}
936 
937 	return rc;
938 
939  err_exit:
940 	__genwqe_purge_ddcb(cd, req);
941 
942 	if (ddcb_requ_collect_debug_data(req)) {
943 		if (copy_to_user((struct genwqe_debug_data __user *)
944 				 (unsigned long)cmd->ddata_addr,
945 				 &req->debug_data,
946 				 sizeof(struct genwqe_debug_data)))
947 			return -EFAULT;
948 	}
949 	return rc;
950 }
951 
952 /**
953  * genwqe_next_ddcb_ready() - Figure out if the next DDCB is already finished
954  *
955  * We use this as condition for our wait-queue code.
956  */
957 static int genwqe_next_ddcb_ready(struct genwqe_dev *cd)
958 {
959 	unsigned long flags;
960 	struct ddcb *pddcb;
961 	struct ddcb_queue *queue = &cd->queue;
962 
963 	spin_lock_irqsave(&queue->ddcb_lock, flags);
964 
965 	if (queue_empty(queue)) { /* emtpy queue */
966 		spin_unlock_irqrestore(&queue->ddcb_lock, flags);
967 		return 0;
968 	}
969 
970 	pddcb = &queue->ddcb_vaddr[queue->ddcb_act];
971 	if (pddcb->icrc_hsi_shi_32 & DDCB_COMPLETED_BE32) { /* ddcb ready */
972 		spin_unlock_irqrestore(&queue->ddcb_lock, flags);
973 		return 1;
974 	}
975 
976 	spin_unlock_irqrestore(&queue->ddcb_lock, flags);
977 	return 0;
978 }
979 
980 /**
981  * genwqe_ddcbs_in_flight() - Check how many DDCBs are in flight
982  *
983  * Keep track on the number of DDCBs which ware currently in the
984  * queue. This is needed for statistics as well as conditon if we want
985  * to wait or better do polling in case of no interrupts available.
986  */
987 int genwqe_ddcbs_in_flight(struct genwqe_dev *cd)
988 {
989 	unsigned long flags;
990 	int ddcbs_in_flight = 0;
991 	struct ddcb_queue *queue = &cd->queue;
992 
993 	spin_lock_irqsave(&queue->ddcb_lock, flags);
994 	ddcbs_in_flight += queue->ddcbs_in_flight;
995 	spin_unlock_irqrestore(&queue->ddcb_lock, flags);
996 
997 	return ddcbs_in_flight;
998 }
999 
1000 static int setup_ddcb_queue(struct genwqe_dev *cd, struct ddcb_queue *queue)
1001 {
1002 	int rc, i;
1003 	struct ddcb *pddcb;
1004 	u64 val64;
1005 	unsigned int queue_size;
1006 	struct pci_dev *pci_dev = cd->pci_dev;
1007 
1008 	if (genwqe_ddcb_max < 2)
1009 		return -EINVAL;
1010 
1011 	queue_size = roundup(genwqe_ddcb_max * sizeof(struct ddcb), PAGE_SIZE);
1012 
1013 	queue->ddcbs_in_flight = 0;  /* statistics */
1014 	queue->ddcbs_max_in_flight = 0;
1015 	queue->ddcbs_completed = 0;
1016 	queue->busy = 0;
1017 
1018 	queue->ddcb_seq	  = 0x100; /* start sequence number */
1019 	queue->ddcb_max	  = genwqe_ddcb_max; /* module parameter */
1020 	queue->ddcb_vaddr = __genwqe_alloc_consistent(cd, queue_size,
1021 						&queue->ddcb_daddr);
1022 	if (queue->ddcb_vaddr == NULL) {
1023 		dev_err(&pci_dev->dev,
1024 			"[%s] **err: could not allocate DDCB **\n", __func__);
1025 		return -ENOMEM;
1026 	}
1027 	memset(queue->ddcb_vaddr, 0, queue_size);
1028 
1029 	queue->ddcb_req = kzalloc(sizeof(struct ddcb_requ *) *
1030 				  queue->ddcb_max, GFP_KERNEL);
1031 	if (!queue->ddcb_req) {
1032 		rc = -ENOMEM;
1033 		goto free_ddcbs;
1034 	}
1035 
1036 	queue->ddcb_waitqs = kzalloc(sizeof(wait_queue_head_t) *
1037 				     queue->ddcb_max, GFP_KERNEL);
1038 	if (!queue->ddcb_waitqs) {
1039 		rc = -ENOMEM;
1040 		goto free_requs;
1041 	}
1042 
1043 	for (i = 0; i < queue->ddcb_max; i++) {
1044 		pddcb = &queue->ddcb_vaddr[i];		     /* DDCBs */
1045 		pddcb->icrc_hsi_shi_32 = DDCB_COMPLETED_BE32;
1046 		pddcb->retc_16 = cpu_to_be16(0xfff);
1047 
1048 		queue->ddcb_req[i] = NULL;		     /* requests */
1049 		init_waitqueue_head(&queue->ddcb_waitqs[i]); /* waitqueues */
1050 	}
1051 
1052 	queue->ddcb_act  = 0;
1053 	queue->ddcb_next = 0;	/* queue is empty */
1054 
1055 	spin_lock_init(&queue->ddcb_lock);
1056 	init_waitqueue_head(&queue->ddcb_waitq);
1057 
1058 	val64 = ((u64)(queue->ddcb_max - 1) <<  8); /* lastptr */
1059 	__genwqe_writeq(cd, queue->IO_QUEUE_CONFIG,  0x07);  /* iCRC/vCRC */
1060 	__genwqe_writeq(cd, queue->IO_QUEUE_SEGMENT, queue->ddcb_daddr);
1061 	__genwqe_writeq(cd, queue->IO_QUEUE_INITSQN, queue->ddcb_seq);
1062 	__genwqe_writeq(cd, queue->IO_QUEUE_WRAP,    val64);
1063 	return 0;
1064 
1065  free_requs:
1066 	kfree(queue->ddcb_req);
1067 	queue->ddcb_req = NULL;
1068  free_ddcbs:
1069 	__genwqe_free_consistent(cd, queue_size, queue->ddcb_vaddr,
1070 				queue->ddcb_daddr);
1071 	queue->ddcb_vaddr = NULL;
1072 	queue->ddcb_daddr = 0ull;
1073 	return -ENODEV;
1074 
1075 }
1076 
1077 static int ddcb_queue_initialized(struct ddcb_queue *queue)
1078 {
1079 	return queue->ddcb_vaddr != NULL;
1080 }
1081 
1082 static void free_ddcb_queue(struct genwqe_dev *cd, struct ddcb_queue *queue)
1083 {
1084 	unsigned int queue_size;
1085 
1086 	queue_size = roundup(queue->ddcb_max * sizeof(struct ddcb), PAGE_SIZE);
1087 
1088 	kfree(queue->ddcb_req);
1089 	queue->ddcb_req = NULL;
1090 
1091 	if (queue->ddcb_vaddr) {
1092 		__genwqe_free_consistent(cd, queue_size, queue->ddcb_vaddr,
1093 					queue->ddcb_daddr);
1094 		queue->ddcb_vaddr = NULL;
1095 		queue->ddcb_daddr = 0ull;
1096 	}
1097 }
1098 
1099 static irqreturn_t genwqe_pf_isr(int irq, void *dev_id)
1100 {
1101 	u64 gfir;
1102 	struct genwqe_dev *cd = (struct genwqe_dev *)dev_id;
1103 	struct pci_dev *pci_dev = cd->pci_dev;
1104 
1105 	/*
1106 	 * In case of fatal FIR error the queue is stopped, such that
1107 	 * we can safely check it without risking anything.
1108 	 */
1109 	cd->irqs_processed++;
1110 	wake_up_interruptible(&cd->queue_waitq);
1111 
1112 	/*
1113 	 * Checking for errors before kicking the queue might be
1114 	 * safer, but slower for the good-case ... See above.
1115 	 */
1116 	gfir = __genwqe_readq(cd, IO_SLC_CFGREG_GFIR);
1117 	if ((gfir & GFIR_ERR_TRIGGER) != 0x0) {
1118 
1119 		wake_up_interruptible(&cd->health_waitq);
1120 
1121 		/*
1122 		 * By default GFIRs causes recovery actions. This
1123 		 * count is just for debug when recovery is masked.
1124 		 */
1125 		printk_ratelimited(KERN_ERR
1126 				   "%s %s: [%s] GFIR=%016llx\n",
1127 				   GENWQE_DEVNAME, dev_name(&pci_dev->dev),
1128 				   __func__, gfir);
1129 	}
1130 
1131 	return IRQ_HANDLED;
1132 }
1133 
1134 static irqreturn_t genwqe_vf_isr(int irq, void *dev_id)
1135 {
1136 	struct genwqe_dev *cd = (struct genwqe_dev *)dev_id;
1137 
1138 	cd->irqs_processed++;
1139 	wake_up_interruptible(&cd->queue_waitq);
1140 
1141 	return IRQ_HANDLED;
1142 }
1143 
1144 /**
1145  * genwqe_card_thread() - Work thread for the DDCB queue
1146  *
1147  * The idea is to check if there are DDCBs in processing. If there are
1148  * some finished DDCBs, we process them and wakeup the
1149  * requestors. Otherwise we give other processes time using
1150  * cond_resched().
1151  */
1152 static int genwqe_card_thread(void *data)
1153 {
1154 	int should_stop = 0, rc = 0;
1155 	struct genwqe_dev *cd = (struct genwqe_dev *)data;
1156 
1157 	while (!kthread_should_stop()) {
1158 
1159 		genwqe_check_ddcb_queue(cd, &cd->queue);
1160 
1161 		if (genwqe_polling_enabled) {
1162 			rc = wait_event_interruptible_timeout(
1163 				cd->queue_waitq,
1164 				genwqe_ddcbs_in_flight(cd) ||
1165 				(should_stop = kthread_should_stop()), 1);
1166 		} else {
1167 			rc = wait_event_interruptible_timeout(
1168 				cd->queue_waitq,
1169 				genwqe_next_ddcb_ready(cd) ||
1170 				(should_stop = kthread_should_stop()), HZ);
1171 		}
1172 		if (should_stop)
1173 			break;
1174 
1175 		/*
1176 		 * Avoid soft lockups on heavy loads; we do not want
1177 		 * to disable our interrupts.
1178 		 */
1179 		cond_resched();
1180 	}
1181 	return 0;
1182 }
1183 
1184 /**
1185  * genwqe_setup_service_layer() - Setup DDCB queue
1186  * @cd:         pointer to genwqe device descriptor
1187  *
1188  * Allocate DDCBs. Configure Service Layer Controller (SLC).
1189  *
1190  * Return: 0 success
1191  */
1192 int genwqe_setup_service_layer(struct genwqe_dev *cd)
1193 {
1194 	int rc;
1195 	struct ddcb_queue *queue;
1196 	struct pci_dev *pci_dev = cd->pci_dev;
1197 
1198 	if (genwqe_is_privileged(cd)) {
1199 		rc = genwqe_card_reset(cd);
1200 		if (rc < 0) {
1201 			dev_err(&pci_dev->dev,
1202 				"[%s] err: reset failed.\n", __func__);
1203 			return rc;
1204 		}
1205 		genwqe_read_softreset(cd);
1206 	}
1207 
1208 	queue = &cd->queue;
1209 	queue->IO_QUEUE_CONFIG  = IO_SLC_QUEUE_CONFIG;
1210 	queue->IO_QUEUE_STATUS  = IO_SLC_QUEUE_STATUS;
1211 	queue->IO_QUEUE_SEGMENT = IO_SLC_QUEUE_SEGMENT;
1212 	queue->IO_QUEUE_INITSQN = IO_SLC_QUEUE_INITSQN;
1213 	queue->IO_QUEUE_OFFSET  = IO_SLC_QUEUE_OFFSET;
1214 	queue->IO_QUEUE_WRAP    = IO_SLC_QUEUE_WRAP;
1215 	queue->IO_QUEUE_WTIME   = IO_SLC_QUEUE_WTIME;
1216 	queue->IO_QUEUE_ERRCNTS = IO_SLC_QUEUE_ERRCNTS;
1217 	queue->IO_QUEUE_LRW     = IO_SLC_QUEUE_LRW;
1218 
1219 	rc = setup_ddcb_queue(cd, queue);
1220 	if (rc != 0) {
1221 		rc = -ENODEV;
1222 		goto err_out;
1223 	}
1224 
1225 	init_waitqueue_head(&cd->queue_waitq);
1226 	cd->card_thread = kthread_run(genwqe_card_thread, cd,
1227 				      GENWQE_DEVNAME "%d_thread",
1228 				      cd->card_idx);
1229 	if (IS_ERR(cd->card_thread)) {
1230 		rc = PTR_ERR(cd->card_thread);
1231 		cd->card_thread = NULL;
1232 		goto stop_free_queue;
1233 	}
1234 
1235 	rc = genwqe_set_interrupt_capability(cd, GENWQE_MSI_IRQS);
1236 	if (rc > 0)
1237 		rc = genwqe_set_interrupt_capability(cd, rc);
1238 	if (rc != 0) {
1239 		rc = -ENODEV;
1240 		goto stop_kthread;
1241 	}
1242 
1243 	/*
1244 	 * We must have all wait-queues initialized when we enable the
1245 	 * interrupts. Otherwise we might crash if we get an early
1246 	 * irq.
1247 	 */
1248 	init_waitqueue_head(&cd->health_waitq);
1249 
1250 	if (genwqe_is_privileged(cd)) {
1251 		rc = request_irq(pci_dev->irq, genwqe_pf_isr, IRQF_SHARED,
1252 				 GENWQE_DEVNAME, cd);
1253 	} else {
1254 		rc = request_irq(pci_dev->irq, genwqe_vf_isr, IRQF_SHARED,
1255 				 GENWQE_DEVNAME, cd);
1256 	}
1257 	if (rc < 0) {
1258 		dev_err(&pci_dev->dev, "irq %d not free.\n", pci_dev->irq);
1259 		goto stop_irq_cap;
1260 	}
1261 
1262 	cd->card_state = GENWQE_CARD_USED;
1263 	return 0;
1264 
1265  stop_irq_cap:
1266 	genwqe_reset_interrupt_capability(cd);
1267  stop_kthread:
1268 	kthread_stop(cd->card_thread);
1269 	cd->card_thread = NULL;
1270  stop_free_queue:
1271 	free_ddcb_queue(cd, queue);
1272  err_out:
1273 	return rc;
1274 }
1275 
1276 /**
1277  * queue_wake_up_all() - Handles fatal error case
1278  *
1279  * The PCI device got unusable and we have to stop all pending
1280  * requests as fast as we can. The code after this must purge the
1281  * DDCBs in question and ensure that all mappings are freed.
1282  */
1283 static int queue_wake_up_all(struct genwqe_dev *cd)
1284 {
1285 	unsigned int i;
1286 	unsigned long flags;
1287 	struct ddcb_queue *queue = &cd->queue;
1288 
1289 	spin_lock_irqsave(&queue->ddcb_lock, flags);
1290 
1291 	for (i = 0; i < queue->ddcb_max; i++)
1292 		wake_up_interruptible(&queue->ddcb_waitqs[queue->ddcb_act]);
1293 
1294 	spin_unlock_irqrestore(&queue->ddcb_lock, flags);
1295 
1296 	return 0;
1297 }
1298 
1299 /**
1300  * genwqe_finish_queue() - Remove any genwqe devices and user-interfaces
1301  *
1302  * Relies on the pre-condition that there are no users of the card
1303  * device anymore e.g. with open file-descriptors.
1304  *
1305  * This function must be robust enough to be called twice.
1306  */
1307 int genwqe_finish_queue(struct genwqe_dev *cd)
1308 {
1309 	int i, rc, in_flight;
1310 	int waitmax = genwqe_ddcb_software_timeout;
1311 	struct pci_dev *pci_dev = cd->pci_dev;
1312 	struct ddcb_queue *queue = &cd->queue;
1313 
1314 	if (!ddcb_queue_initialized(queue))
1315 		return 0;
1316 
1317 	/* Do not wipe out the error state. */
1318 	if (cd->card_state == GENWQE_CARD_USED)
1319 		cd->card_state = GENWQE_CARD_UNUSED;
1320 
1321 	/* Wake up all requests in the DDCB queue such that they
1322 	   should be removed nicely. */
1323 	queue_wake_up_all(cd);
1324 
1325 	/* We must wait to get rid of the DDCBs in flight */
1326 	for (i = 0; i < waitmax; i++) {
1327 		in_flight = genwqe_ddcbs_in_flight(cd);
1328 
1329 		if (in_flight == 0)
1330 			break;
1331 
1332 		dev_dbg(&pci_dev->dev,
1333 			"  DEBUG [%d/%d] waiting for queue to get empty: "
1334 			"%d requests!\n", i, waitmax, in_flight);
1335 
1336 		/*
1337 		 * Severe severe error situation: The card itself has
1338 		 * 16 DDCB queues, each queue has e.g. 32 entries,
1339 		 * each DDBC has a hardware timeout of currently 250
1340 		 * msec but the PFs have a hardware timeout of 8 sec
1341 		 * ... so I take something large.
1342 		 */
1343 		msleep(1000);
1344 	}
1345 	if (i == waitmax) {
1346 		dev_err(&pci_dev->dev, "  [%s] err: queue is not empty!!\n",
1347 			__func__);
1348 		rc = -EIO;
1349 	}
1350 	return rc;
1351 }
1352 
1353 /**
1354  * genwqe_release_service_layer() - Shutdown DDCB queue
1355  * @cd:       genwqe device descriptor
1356  *
1357  * This function must be robust enough to be called twice.
1358  */
1359 int genwqe_release_service_layer(struct genwqe_dev *cd)
1360 {
1361 	struct pci_dev *pci_dev = cd->pci_dev;
1362 
1363 	if (!ddcb_queue_initialized(&cd->queue))
1364 		return 1;
1365 
1366 	free_irq(pci_dev->irq, cd);
1367 	genwqe_reset_interrupt_capability(cd);
1368 
1369 	if (cd->card_thread != NULL) {
1370 		kthread_stop(cd->card_thread);
1371 		cd->card_thread = NULL;
1372 	}
1373 
1374 	free_ddcb_queue(cd, &cd->queue);
1375 	return 0;
1376 }
1377