xref: /openbmc/linux/drivers/scsi/csiostor/csio_wr.c (revision 8fdff1dc)
1 /*
2  * This file is part of the Chelsio FCoE driver for Linux.
3  *
4  * Copyright (c) 2008-2012 Chelsio Communications, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/kernel.h>
36 #include <linux/string.h>
37 #include <linux/compiler.h>
38 #include <linux/slab.h>
39 #include <asm/page.h>
40 #include <linux/cache.h>
41 
42 #include "csio_hw.h"
43 #include "csio_wr.h"
44 #include "csio_mb.h"
45 #include "csio_defs.h"
46 
47 int csio_intr_coalesce_cnt;		/* value:SGE_INGRESS_RX_THRESHOLD[0] */
48 static int csio_sge_thresh_reg;		/* SGE_INGRESS_RX_THRESHOLD[0] */
49 
50 int csio_intr_coalesce_time = 10;	/* value:SGE_TIMER_VALUE_1 */
51 static int csio_sge_timer_reg = 1;
52 
53 #define CSIO_SET_FLBUF_SIZE(_hw, _reg, _val)				\
54 	csio_wr_reg32((_hw), (_val), SGE_FL_BUFFER_SIZE##_reg)
55 
56 static void
57 csio_get_flbuf_size(struct csio_hw *hw, struct csio_sge *sge, uint32_t reg)
58 {
59 	sge->sge_fl_buf_size[reg] = csio_rd_reg32(hw, SGE_FL_BUFFER_SIZE0 +
60 							reg * sizeof(uint32_t));
61 }
62 
63 /* Free list buffer size */
64 static inline uint32_t
65 csio_wr_fl_bufsz(struct csio_sge *sge, struct csio_dma_buf *buf)
66 {
67 	return sge->sge_fl_buf_size[buf->paddr & 0xF];
68 }
69 
70 /* Size of the egress queue status page */
71 static inline uint32_t
72 csio_wr_qstat_pgsz(struct csio_hw *hw)
73 {
74 	return (hw->wrm.sge.sge_control & EGRSTATUSPAGESIZE(1)) ?  128 : 64;
75 }
76 
77 /* Ring freelist doorbell */
78 static inline void
79 csio_wr_ring_fldb(struct csio_hw *hw, struct csio_q *flq)
80 {
81 	/*
82 	 * Ring the doorbell only when we have atleast CSIO_QCREDIT_SZ
83 	 * number of bytes in the freelist queue. This translates to atleast
84 	 * 8 freelist buffer pointers (since each pointer is 8 bytes).
85 	 */
86 	if (flq->inc_idx >= 8) {
87 		csio_wr_reg32(hw, DBPRIO(1) | QID(flq->un.fl.flid) |
88 			      PIDX(flq->inc_idx / 8),
89 			      MYPF_REG(SGE_PF_KDOORBELL));
90 		flq->inc_idx &= 7;
91 	}
92 }
93 
94 /* Write a 0 cidx increment value to enable SGE interrupts for this queue */
95 static void
96 csio_wr_sge_intr_enable(struct csio_hw *hw, uint16_t iqid)
97 {
98 	csio_wr_reg32(hw, CIDXINC(0)		|
99 			  INGRESSQID(iqid)	|
100 			  TIMERREG(X_TIMERREG_RESTART_COUNTER),
101 			  MYPF_REG(SGE_PF_GTS));
102 }
103 
104 /*
105  * csio_wr_fill_fl - Populate the FL buffers of a FL queue.
106  * @hw: HW module.
107  * @flq: Freelist queue.
108  *
109  * Fill up freelist buffer entries with buffers of size specified
110  * in the size register.
111  *
112  */
113 static int
114 csio_wr_fill_fl(struct csio_hw *hw, struct csio_q *flq)
115 {
116 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
117 	struct csio_sge *sge = &wrm->sge;
118 	__be64 *d = (__be64 *)(flq->vstart);
119 	struct csio_dma_buf *buf = &flq->un.fl.bufs[0];
120 	uint64_t paddr;
121 	int sreg = flq->un.fl.sreg;
122 	int n = flq->credits;
123 
124 	while (n--) {
125 		buf->len = sge->sge_fl_buf_size[sreg];
126 		buf->vaddr = pci_alloc_consistent(hw->pdev, buf->len,
127 						  &buf->paddr);
128 		if (!buf->vaddr) {
129 			csio_err(hw, "Could only fill %d buffers!\n", n + 1);
130 			return -ENOMEM;
131 		}
132 
133 		paddr = buf->paddr | (sreg & 0xF);
134 
135 		*d++ = cpu_to_be64(paddr);
136 		buf++;
137 	}
138 
139 	return 0;
140 }
141 
142 /*
143  * csio_wr_update_fl -
144  * @hw: HW module.
145  * @flq: Freelist queue.
146  *
147  *
148  */
149 static inline void
150 csio_wr_update_fl(struct csio_hw *hw, struct csio_q *flq, uint16_t n)
151 {
152 
153 	flq->inc_idx += n;
154 	flq->pidx += n;
155 	if (unlikely(flq->pidx >= flq->credits))
156 		flq->pidx -= (uint16_t)flq->credits;
157 
158 	CSIO_INC_STATS(flq, n_flq_refill);
159 }
160 
161 /*
162  * csio_wr_alloc_q - Allocate a WR queue and initialize it.
163  * @hw: HW module
164  * @qsize: Size of the queue in bytes
165  * @wrsize: Since of WR in this queue, if fixed.
166  * @type: Type of queue (Ingress/Egress/Freelist)
167  * @owner: Module that owns this queue.
168  * @nflb: Number of freelist buffers for FL.
169  * @sreg: What is the FL buffer size register?
170  * @iq_int_handler: Ingress queue handler in INTx mode.
171  *
172  * This function allocates and sets up a queue for the caller
173  * of size qsize, aligned at the required boundary. This is subject to
174  * be free entries being available in the queue array. If one is found,
175  * it is initialized with the allocated queue, marked as being used (owner),
176  * and a handle returned to the caller in form of the queue's index
177  * into the q_arr array.
178  * If user has indicated a freelist (by specifying nflb > 0), create
179  * another queue (with its own index into q_arr) for the freelist. Allocate
180  * memory for DMA buffer metadata (vaddr, len etc). Save off the freelist
181  * idx in the ingress queue's flq.idx. This is how a Freelist is associated
182  * with its owning ingress queue.
183  */
184 int
185 csio_wr_alloc_q(struct csio_hw *hw, uint32_t qsize, uint32_t wrsize,
186 		uint16_t type, void *owner, uint32_t nflb, int sreg,
187 		iq_handler_t iq_intx_handler)
188 {
189 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
190 	struct csio_q	*q, *flq;
191 	int		free_idx = wrm->free_qidx;
192 	int		ret_idx = free_idx;
193 	uint32_t	qsz;
194 	int flq_idx;
195 
196 	if (free_idx >= wrm->num_q) {
197 		csio_err(hw, "No more free queues.\n");
198 		return -1;
199 	}
200 
201 	switch (type) {
202 	case CSIO_EGRESS:
203 		qsz = ALIGN(qsize, CSIO_QCREDIT_SZ) + csio_wr_qstat_pgsz(hw);
204 		break;
205 	case CSIO_INGRESS:
206 		switch (wrsize) {
207 		case 16:
208 		case 32:
209 		case 64:
210 		case 128:
211 			break;
212 		default:
213 			csio_err(hw, "Invalid Ingress queue WR size:%d\n",
214 				    wrsize);
215 			return -1;
216 		}
217 
218 		/*
219 		 * Number of elements must be a multiple of 16
220 		 * So this includes status page size
221 		 */
222 		qsz = ALIGN(qsize/wrsize, 16) * wrsize;
223 
224 		break;
225 	case CSIO_FREELIST:
226 		qsz = ALIGN(qsize/wrsize, 8) * wrsize + csio_wr_qstat_pgsz(hw);
227 		break;
228 	default:
229 		csio_err(hw, "Invalid queue type: 0x%x\n", type);
230 		return -1;
231 	}
232 
233 	q = wrm->q_arr[free_idx];
234 
235 	q->vstart = pci_alloc_consistent(hw->pdev, qsz, &q->pstart);
236 	if (!q->vstart) {
237 		csio_err(hw,
238 			 "Failed to allocate DMA memory for "
239 			 "queue at id: %d size: %d\n", free_idx, qsize);
240 		return -1;
241 	}
242 
243 	/*
244 	 * We need to zero out the contents, importantly for ingress,
245 	 * since we start with a generatiom bit of 1 for ingress.
246 	 */
247 	memset(q->vstart, 0, qsz);
248 
249 	q->type		= type;
250 	q->owner	= owner;
251 	q->pidx		= q->cidx = q->inc_idx = 0;
252 	q->size		= qsz;
253 	q->wr_sz	= wrsize;	/* If using fixed size WRs */
254 
255 	wrm->free_qidx++;
256 
257 	if (type == CSIO_INGRESS) {
258 		/* Since queue area is set to zero */
259 		q->un.iq.genbit	= 1;
260 
261 		/*
262 		 * Ingress queue status page size is always the size of
263 		 * the ingress queue entry.
264 		 */
265 		q->credits	= (qsz - q->wr_sz) / q->wr_sz;
266 		q->vwrap	= (void *)((uintptr_t)(q->vstart) + qsz
267 							- q->wr_sz);
268 
269 		/* Allocate memory for FL if requested */
270 		if (nflb > 0) {
271 			flq_idx = csio_wr_alloc_q(hw, nflb * sizeof(__be64),
272 						  sizeof(__be64), CSIO_FREELIST,
273 						  owner, 0, sreg, NULL);
274 			if (flq_idx == -1) {
275 				csio_err(hw,
276 					 "Failed to allocate FL queue"
277 					 " for IQ idx:%d\n", free_idx);
278 				return -1;
279 			}
280 
281 			/* Associate the new FL with the Ingress quue */
282 			q->un.iq.flq_idx = flq_idx;
283 
284 			flq = wrm->q_arr[q->un.iq.flq_idx];
285 			flq->un.fl.bufs = kzalloc(flq->credits *
286 						  sizeof(struct csio_dma_buf),
287 						  GFP_KERNEL);
288 			if (!flq->un.fl.bufs) {
289 				csio_err(hw,
290 					 "Failed to allocate FL queue bufs"
291 					 " for IQ idx:%d\n", free_idx);
292 				return -1;
293 			}
294 
295 			flq->un.fl.packen = 0;
296 			flq->un.fl.offset = 0;
297 			flq->un.fl.sreg = sreg;
298 
299 			/* Fill up the free list buffers */
300 			if (csio_wr_fill_fl(hw, flq))
301 				return -1;
302 
303 			/*
304 			 * Make sure in a FLQ, atleast 1 credit (8 FL buffers)
305 			 * remains unpopulated,otherwise HW thinks
306 			 * FLQ is empty.
307 			 */
308 			flq->pidx = flq->inc_idx = flq->credits - 8;
309 		} else {
310 			q->un.iq.flq_idx = -1;
311 		}
312 
313 		/* Associate the IQ INTx handler. */
314 		q->un.iq.iq_intx_handler = iq_intx_handler;
315 
316 		csio_q_iqid(hw, ret_idx) = CSIO_MAX_QID;
317 
318 	} else if (type == CSIO_EGRESS) {
319 		q->credits = (qsz - csio_wr_qstat_pgsz(hw)) / CSIO_QCREDIT_SZ;
320 		q->vwrap   = (void *)((uintptr_t)(q->vstart) + qsz
321 						- csio_wr_qstat_pgsz(hw));
322 		csio_q_eqid(hw, ret_idx) = CSIO_MAX_QID;
323 	} else { /* Freelist */
324 		q->credits = (qsz - csio_wr_qstat_pgsz(hw)) / sizeof(__be64);
325 		q->vwrap   = (void *)((uintptr_t)(q->vstart) + qsz
326 						- csio_wr_qstat_pgsz(hw));
327 		csio_q_flid(hw, ret_idx) = CSIO_MAX_QID;
328 	}
329 
330 	return ret_idx;
331 }
332 
333 /*
334  * csio_wr_iq_create_rsp - Response handler for IQ creation.
335  * @hw: The HW module.
336  * @mbp: Mailbox.
337  * @iq_idx: Ingress queue that got created.
338  *
339  * Handle FW_IQ_CMD mailbox completion. Save off the assigned IQ/FL ids.
340  */
341 static int
342 csio_wr_iq_create_rsp(struct csio_hw *hw, struct csio_mb *mbp, int iq_idx)
343 {
344 	struct csio_iq_params iqp;
345 	enum fw_retval retval;
346 	uint32_t iq_id;
347 	int flq_idx;
348 
349 	memset(&iqp, 0, sizeof(struct csio_iq_params));
350 
351 	csio_mb_iq_alloc_write_rsp(hw, mbp, &retval, &iqp);
352 
353 	if (retval != FW_SUCCESS) {
354 		csio_err(hw, "IQ cmd returned 0x%x!\n", retval);
355 		mempool_free(mbp, hw->mb_mempool);
356 		return -EINVAL;
357 	}
358 
359 	csio_q_iqid(hw, iq_idx)		= iqp.iqid;
360 	csio_q_physiqid(hw, iq_idx)	= iqp.physiqid;
361 	csio_q_pidx(hw, iq_idx)		= csio_q_cidx(hw, iq_idx) = 0;
362 	csio_q_inc_idx(hw, iq_idx)	= 0;
363 
364 	/* Actual iq-id. */
365 	iq_id = iqp.iqid - hw->wrm.fw_iq_start;
366 
367 	/* Set the iq-id to iq map table. */
368 	if (iq_id >= CSIO_MAX_IQ) {
369 		csio_err(hw,
370 			 "Exceeding MAX_IQ(%d) supported!"
371 			 " iqid:%d rel_iqid:%d FW iq_start:%d\n",
372 			 CSIO_MAX_IQ, iq_id, iqp.iqid, hw->wrm.fw_iq_start);
373 		mempool_free(mbp, hw->mb_mempool);
374 		return -EINVAL;
375 	}
376 	csio_q_set_intr_map(hw, iq_idx, iq_id);
377 
378 	/*
379 	 * During FW_IQ_CMD, FW sets interrupt_sent bit to 1 in the SGE
380 	 * ingress context of this queue. This will block interrupts to
381 	 * this queue until the next GTS write. Therefore, we do a
382 	 * 0-cidx increment GTS write for this queue just to clear the
383 	 * interrupt_sent bit. This will re-enable interrupts to this
384 	 * queue.
385 	 */
386 	csio_wr_sge_intr_enable(hw, iqp.physiqid);
387 
388 	flq_idx = csio_q_iq_flq_idx(hw, iq_idx);
389 	if (flq_idx != -1) {
390 		struct csio_q *flq = hw->wrm.q_arr[flq_idx];
391 
392 		csio_q_flid(hw, flq_idx) = iqp.fl0id;
393 		csio_q_cidx(hw, flq_idx) = 0;
394 		csio_q_pidx(hw, flq_idx)    = csio_q_credits(hw, flq_idx) - 8;
395 		csio_q_inc_idx(hw, flq_idx) = csio_q_credits(hw, flq_idx) - 8;
396 
397 		/* Now update SGE about the buffers allocated during init */
398 		csio_wr_ring_fldb(hw, flq);
399 	}
400 
401 	mempool_free(mbp, hw->mb_mempool);
402 
403 	return 0;
404 }
405 
406 /*
407  * csio_wr_iq_create - Configure an Ingress queue with FW.
408  * @hw: The HW module.
409  * @priv: Private data object.
410  * @iq_idx: Ingress queue index in the WR module.
411  * @vec: MSIX vector.
412  * @portid: PCIE Channel to be associated with this queue.
413  * @async: Is this a FW asynchronous message handling queue?
414  * @cbfn: Completion callback.
415  *
416  * This API configures an ingress queue with FW by issuing a FW_IQ_CMD mailbox
417  * with alloc/write bits set.
418  */
419 int
420 csio_wr_iq_create(struct csio_hw *hw, void *priv, int iq_idx,
421 		  uint32_t vec, uint8_t portid, bool async,
422 		  void (*cbfn) (struct csio_hw *, struct csio_mb *))
423 {
424 	struct csio_mb  *mbp;
425 	struct csio_iq_params iqp;
426 	int flq_idx;
427 
428 	memset(&iqp, 0, sizeof(struct csio_iq_params));
429 	csio_q_portid(hw, iq_idx) = portid;
430 
431 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
432 	if (!mbp) {
433 		csio_err(hw, "IQ command out of memory!\n");
434 		return -ENOMEM;
435 	}
436 
437 	switch (hw->intr_mode) {
438 	case CSIO_IM_INTX:
439 	case CSIO_IM_MSI:
440 		/* For interrupt forwarding queue only */
441 		if (hw->intr_iq_idx == iq_idx)
442 			iqp.iqandst	= X_INTERRUPTDESTINATION_PCIE;
443 		else
444 			iqp.iqandst	= X_INTERRUPTDESTINATION_IQ;
445 		iqp.iqandstindex	=
446 			csio_q_physiqid(hw, hw->intr_iq_idx);
447 		break;
448 	case CSIO_IM_MSIX:
449 		iqp.iqandst		= X_INTERRUPTDESTINATION_PCIE;
450 		iqp.iqandstindex	= (uint16_t)vec;
451 		break;
452 	case CSIO_IM_NONE:
453 		mempool_free(mbp, hw->mb_mempool);
454 		return -EINVAL;
455 	}
456 
457 	/* Pass in the ingress queue cmd parameters */
458 	iqp.pfn			= hw->pfn;
459 	iqp.vfn			= 0;
460 	iqp.iq_start		= 1;
461 	iqp.viid		= 0;
462 	iqp.type		= FW_IQ_TYPE_FL_INT_CAP;
463 	iqp.iqasynch		= async;
464 	if (csio_intr_coalesce_cnt)
465 		iqp.iqanus	= X_UPDATESCHEDULING_COUNTER_OPTTIMER;
466 	else
467 		iqp.iqanus	= X_UPDATESCHEDULING_TIMER;
468 	iqp.iqanud		= X_UPDATEDELIVERY_INTERRUPT;
469 	iqp.iqpciech		= portid;
470 	iqp.iqintcntthresh	= (uint8_t)csio_sge_thresh_reg;
471 
472 	switch (csio_q_wr_sz(hw, iq_idx)) {
473 	case 16:
474 		iqp.iqesize = 0; break;
475 	case 32:
476 		iqp.iqesize = 1; break;
477 	case 64:
478 		iqp.iqesize = 2; break;
479 	case 128:
480 		iqp.iqesize = 3; break;
481 	}
482 
483 	iqp.iqsize		= csio_q_size(hw, iq_idx) /
484 						csio_q_wr_sz(hw, iq_idx);
485 	iqp.iqaddr		= csio_q_pstart(hw, iq_idx);
486 
487 	flq_idx = csio_q_iq_flq_idx(hw, iq_idx);
488 	if (flq_idx != -1) {
489 		struct csio_q *flq = hw->wrm.q_arr[flq_idx];
490 
491 		iqp.fl0paden	= 1;
492 		iqp.fl0packen	= flq->un.fl.packen ? 1 : 0;
493 		iqp.fl0fbmin	= X_FETCHBURSTMIN_64B;
494 		iqp.fl0fbmax	= X_FETCHBURSTMAX_512B;
495 		iqp.fl0size	= csio_q_size(hw, flq_idx) / CSIO_QCREDIT_SZ;
496 		iqp.fl0addr	= csio_q_pstart(hw, flq_idx);
497 	}
498 
499 	csio_mb_iq_alloc_write(hw, mbp, priv, CSIO_MB_DEFAULT_TMO, &iqp, cbfn);
500 
501 	if (csio_mb_issue(hw, mbp)) {
502 		csio_err(hw, "Issue of IQ cmd failed!\n");
503 		mempool_free(mbp, hw->mb_mempool);
504 		return -EINVAL;
505 	}
506 
507 	if (cbfn != NULL)
508 		return 0;
509 
510 	return csio_wr_iq_create_rsp(hw, mbp, iq_idx);
511 }
512 
513 /*
514  * csio_wr_eq_create_rsp - Response handler for EQ creation.
515  * @hw: The HW module.
516  * @mbp: Mailbox.
517  * @eq_idx: Egress queue that got created.
518  *
519  * Handle FW_EQ_OFLD_CMD mailbox completion. Save off the assigned EQ ids.
520  */
521 static int
522 csio_wr_eq_cfg_rsp(struct csio_hw *hw, struct csio_mb *mbp, int eq_idx)
523 {
524 	struct csio_eq_params eqp;
525 	enum fw_retval retval;
526 
527 	memset(&eqp, 0, sizeof(struct csio_eq_params));
528 
529 	csio_mb_eq_ofld_alloc_write_rsp(hw, mbp, &retval, &eqp);
530 
531 	if (retval != FW_SUCCESS) {
532 		csio_err(hw, "EQ OFLD cmd returned 0x%x!\n", retval);
533 		mempool_free(mbp, hw->mb_mempool);
534 		return -EINVAL;
535 	}
536 
537 	csio_q_eqid(hw, eq_idx)	= (uint16_t)eqp.eqid;
538 	csio_q_physeqid(hw, eq_idx) = (uint16_t)eqp.physeqid;
539 	csio_q_pidx(hw, eq_idx)	= csio_q_cidx(hw, eq_idx) = 0;
540 	csio_q_inc_idx(hw, eq_idx) = 0;
541 
542 	mempool_free(mbp, hw->mb_mempool);
543 
544 	return 0;
545 }
546 
547 /*
548  * csio_wr_eq_create - Configure an Egress queue with FW.
549  * @hw: HW module.
550  * @priv: Private data.
551  * @eq_idx: Egress queue index in the WR module.
552  * @iq_idx: Associated ingress queue index.
553  * @cbfn: Completion callback.
554  *
555  * This API configures a offload egress queue with FW by issuing a
556  * FW_EQ_OFLD_CMD  (with alloc + write ) mailbox.
557  */
558 int
559 csio_wr_eq_create(struct csio_hw *hw, void *priv, int eq_idx,
560 		  int iq_idx, uint8_t portid,
561 		  void (*cbfn) (struct csio_hw *, struct csio_mb *))
562 {
563 	struct csio_mb  *mbp;
564 	struct csio_eq_params eqp;
565 
566 	memset(&eqp, 0, sizeof(struct csio_eq_params));
567 
568 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
569 	if (!mbp) {
570 		csio_err(hw, "EQ command out of memory!\n");
571 		return -ENOMEM;
572 	}
573 
574 	eqp.pfn			= hw->pfn;
575 	eqp.vfn			= 0;
576 	eqp.eqstart		= 1;
577 	eqp.hostfcmode		= X_HOSTFCMODE_STATUS_PAGE;
578 	eqp.iqid		= csio_q_iqid(hw, iq_idx);
579 	eqp.fbmin		= X_FETCHBURSTMIN_64B;
580 	eqp.fbmax		= X_FETCHBURSTMAX_512B;
581 	eqp.cidxfthresh		= 0;
582 	eqp.pciechn		= portid;
583 	eqp.eqsize		= csio_q_size(hw, eq_idx) / CSIO_QCREDIT_SZ;
584 	eqp.eqaddr		= csio_q_pstart(hw, eq_idx);
585 
586 	csio_mb_eq_ofld_alloc_write(hw, mbp, priv, CSIO_MB_DEFAULT_TMO,
587 				    &eqp, cbfn);
588 
589 	if (csio_mb_issue(hw, mbp)) {
590 		csio_err(hw, "Issue of EQ OFLD cmd failed!\n");
591 		mempool_free(mbp, hw->mb_mempool);
592 		return -EINVAL;
593 	}
594 
595 	if (cbfn != NULL)
596 		return 0;
597 
598 	return csio_wr_eq_cfg_rsp(hw, mbp, eq_idx);
599 }
600 
601 /*
602  * csio_wr_iq_destroy_rsp - Response handler for IQ removal.
603  * @hw: The HW module.
604  * @mbp: Mailbox.
605  * @iq_idx: Ingress queue that was freed.
606  *
607  * Handle FW_IQ_CMD (free) mailbox completion.
608  */
609 static int
610 csio_wr_iq_destroy_rsp(struct csio_hw *hw, struct csio_mb *mbp, int iq_idx)
611 {
612 	enum fw_retval retval = csio_mb_fw_retval(mbp);
613 	int rv = 0;
614 
615 	if (retval != FW_SUCCESS)
616 		rv = -EINVAL;
617 
618 	mempool_free(mbp, hw->mb_mempool);
619 
620 	return rv;
621 }
622 
623 /*
624  * csio_wr_iq_destroy - Free an ingress queue.
625  * @hw: The HW module.
626  * @priv: Private data object.
627  * @iq_idx: Ingress queue index to destroy
628  * @cbfn: Completion callback.
629  *
630  * This API frees an ingress queue by issuing the FW_IQ_CMD
631  * with the free bit set.
632  */
633 static int
634 csio_wr_iq_destroy(struct csio_hw *hw, void *priv, int iq_idx,
635 		   void (*cbfn)(struct csio_hw *, struct csio_mb *))
636 {
637 	int rv = 0;
638 	struct csio_mb  *mbp;
639 	struct csio_iq_params iqp;
640 	int flq_idx;
641 
642 	memset(&iqp, 0, sizeof(struct csio_iq_params));
643 
644 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
645 	if (!mbp)
646 		return -ENOMEM;
647 
648 	iqp.pfn		= hw->pfn;
649 	iqp.vfn		= 0;
650 	iqp.iqid	= csio_q_iqid(hw, iq_idx);
651 	iqp.type	= FW_IQ_TYPE_FL_INT_CAP;
652 
653 	flq_idx = csio_q_iq_flq_idx(hw, iq_idx);
654 	if (flq_idx != -1)
655 		iqp.fl0id = csio_q_flid(hw, flq_idx);
656 	else
657 		iqp.fl0id = 0xFFFF;
658 
659 	iqp.fl1id = 0xFFFF;
660 
661 	csio_mb_iq_free(hw, mbp, priv, CSIO_MB_DEFAULT_TMO, &iqp, cbfn);
662 
663 	rv = csio_mb_issue(hw, mbp);
664 	if (rv != 0) {
665 		mempool_free(mbp, hw->mb_mempool);
666 		return rv;
667 	}
668 
669 	if (cbfn != NULL)
670 		return 0;
671 
672 	return csio_wr_iq_destroy_rsp(hw, mbp, iq_idx);
673 }
674 
675 /*
676  * csio_wr_eq_destroy_rsp - Response handler for OFLD EQ creation.
677  * @hw: The HW module.
678  * @mbp: Mailbox.
679  * @eq_idx: Egress queue that was freed.
680  *
681  * Handle FW_OFLD_EQ_CMD (free) mailbox completion.
682  */
683 static int
684 csio_wr_eq_destroy_rsp(struct csio_hw *hw, struct csio_mb *mbp, int eq_idx)
685 {
686 	enum fw_retval retval = csio_mb_fw_retval(mbp);
687 	int rv = 0;
688 
689 	if (retval != FW_SUCCESS)
690 		rv = -EINVAL;
691 
692 	mempool_free(mbp, hw->mb_mempool);
693 
694 	return rv;
695 }
696 
697 /*
698  * csio_wr_eq_destroy - Free an Egress queue.
699  * @hw: The HW module.
700  * @priv: Private data object.
701  * @eq_idx: Egress queue index to destroy
702  * @cbfn: Completion callback.
703  *
704  * This API frees an Egress queue by issuing the FW_EQ_OFLD_CMD
705  * with the free bit set.
706  */
707 static int
708 csio_wr_eq_destroy(struct csio_hw *hw, void *priv, int eq_idx,
709 		   void (*cbfn) (struct csio_hw *, struct csio_mb *))
710 {
711 	int rv = 0;
712 	struct csio_mb  *mbp;
713 	struct csio_eq_params eqp;
714 
715 	memset(&eqp, 0, sizeof(struct csio_eq_params));
716 
717 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
718 	if (!mbp)
719 		return -ENOMEM;
720 
721 	eqp.pfn		= hw->pfn;
722 	eqp.vfn		= 0;
723 	eqp.eqid	= csio_q_eqid(hw, eq_idx);
724 
725 	csio_mb_eq_ofld_free(hw, mbp, priv, CSIO_MB_DEFAULT_TMO, &eqp, cbfn);
726 
727 	rv = csio_mb_issue(hw, mbp);
728 	if (rv != 0) {
729 		mempool_free(mbp, hw->mb_mempool);
730 		return rv;
731 	}
732 
733 	if (cbfn != NULL)
734 		return 0;
735 
736 	return csio_wr_eq_destroy_rsp(hw, mbp, eq_idx);
737 }
738 
739 /*
740  * csio_wr_cleanup_eq_stpg - Cleanup Egress queue status page
741  * @hw: HW module
742  * @qidx: Egress queue index
743  *
744  * Cleanup the Egress queue status page.
745  */
746 static void
747 csio_wr_cleanup_eq_stpg(struct csio_hw *hw, int qidx)
748 {
749 	struct csio_q	*q = csio_hw_to_wrm(hw)->q_arr[qidx];
750 	struct csio_qstatus_page *stp = (struct csio_qstatus_page *)q->vwrap;
751 
752 	memset(stp, 0, sizeof(*stp));
753 }
754 
755 /*
756  * csio_wr_cleanup_iq_ftr - Cleanup Footer entries in IQ
757  * @hw: HW module
758  * @qidx: Ingress queue index
759  *
760  * Cleanup the footer entries in the given ingress queue,
761  * set to 1 the internal copy of genbit.
762  */
763 static void
764 csio_wr_cleanup_iq_ftr(struct csio_hw *hw, int qidx)
765 {
766 	struct csio_wrm *wrm	= csio_hw_to_wrm(hw);
767 	struct csio_q	*q	= wrm->q_arr[qidx];
768 	void *wr;
769 	struct csio_iqwr_footer *ftr;
770 	uint32_t i = 0;
771 
772 	/* set to 1 since we are just about zero out genbit */
773 	q->un.iq.genbit = 1;
774 
775 	for (i = 0; i < q->credits; i++) {
776 		/* Get the WR */
777 		wr = (void *)((uintptr_t)q->vstart +
778 					   (i * q->wr_sz));
779 		/* Get the footer */
780 		ftr = (struct csio_iqwr_footer *)((uintptr_t)wr +
781 					  (q->wr_sz - sizeof(*ftr)));
782 		/* Zero out footer */
783 		memset(ftr, 0, sizeof(*ftr));
784 	}
785 }
786 
787 int
788 csio_wr_destroy_queues(struct csio_hw *hw, bool cmd)
789 {
790 	int i, flq_idx;
791 	struct csio_q *q;
792 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
793 	int rv;
794 
795 	for (i = 0; i < wrm->free_qidx; i++) {
796 		q = wrm->q_arr[i];
797 
798 		switch (q->type) {
799 		case CSIO_EGRESS:
800 			if (csio_q_eqid(hw, i) != CSIO_MAX_QID) {
801 				csio_wr_cleanup_eq_stpg(hw, i);
802 				if (!cmd) {
803 					csio_q_eqid(hw, i) = CSIO_MAX_QID;
804 					continue;
805 				}
806 
807 				rv = csio_wr_eq_destroy(hw, NULL, i, NULL);
808 				if ((rv == -EBUSY) || (rv == -ETIMEDOUT))
809 					cmd = false;
810 
811 				csio_q_eqid(hw, i) = CSIO_MAX_QID;
812 			}
813 		case CSIO_INGRESS:
814 			if (csio_q_iqid(hw, i) != CSIO_MAX_QID) {
815 				csio_wr_cleanup_iq_ftr(hw, i);
816 				if (!cmd) {
817 					csio_q_iqid(hw, i) = CSIO_MAX_QID;
818 					flq_idx = csio_q_iq_flq_idx(hw, i);
819 					if (flq_idx != -1)
820 						csio_q_flid(hw, flq_idx) =
821 								CSIO_MAX_QID;
822 					continue;
823 				}
824 
825 				rv = csio_wr_iq_destroy(hw, NULL, i, NULL);
826 				if ((rv == -EBUSY) || (rv == -ETIMEDOUT))
827 					cmd = false;
828 
829 				csio_q_iqid(hw, i) = CSIO_MAX_QID;
830 				flq_idx = csio_q_iq_flq_idx(hw, i);
831 				if (flq_idx != -1)
832 					csio_q_flid(hw, flq_idx) = CSIO_MAX_QID;
833 			}
834 		default:
835 			break;
836 		}
837 	}
838 
839 	hw->flags &= ~CSIO_HWF_Q_FW_ALLOCED;
840 
841 	return 0;
842 }
843 
844 /*
845  * csio_wr_get - Get requested size of WR entry/entries from queue.
846  * @hw: HW module.
847  * @qidx: Index of queue.
848  * @size: Cumulative size of Work request(s).
849  * @wrp: Work request pair.
850  *
851  * If requested credits are available, return the start address of the
852  * work request in the work request pair. Set pidx accordingly and
853  * return.
854  *
855  * NOTE about WR pair:
856  * ==================
857  * A WR can start towards the end of a queue, and then continue at the
858  * beginning, since the queue is considered to be circular. This will
859  * require a pair of address/size to be passed back to the caller -
860  * hence Work request pair format.
861  */
862 int
863 csio_wr_get(struct csio_hw *hw, int qidx, uint32_t size,
864 	    struct csio_wr_pair *wrp)
865 {
866 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
867 	struct csio_q *q = wrm->q_arr[qidx];
868 	void *cwr = (void *)((uintptr_t)(q->vstart) +
869 						(q->pidx * CSIO_QCREDIT_SZ));
870 	struct csio_qstatus_page *stp = (struct csio_qstatus_page *)q->vwrap;
871 	uint16_t cidx = q->cidx = ntohs(stp->cidx);
872 	uint16_t pidx = q->pidx;
873 	uint32_t req_sz	= ALIGN(size, CSIO_QCREDIT_SZ);
874 	int req_credits	= req_sz / CSIO_QCREDIT_SZ;
875 	int credits;
876 
877 	CSIO_DB_ASSERT(q->owner != NULL);
878 	CSIO_DB_ASSERT((qidx >= 0) && (qidx < wrm->free_qidx));
879 	CSIO_DB_ASSERT(cidx <= q->credits);
880 
881 	/* Calculate credits */
882 	if (pidx > cidx) {
883 		credits = q->credits - (pidx - cidx) - 1;
884 	} else if (cidx > pidx) {
885 		credits = cidx - pidx - 1;
886 	} else {
887 		/* cidx == pidx, empty queue */
888 		credits = q->credits;
889 		CSIO_INC_STATS(q, n_qempty);
890 	}
891 
892 	/*
893 	 * Check if we have enough credits.
894 	 * credits = 1 implies queue is full.
895 	 */
896 	if (!credits || (req_credits > credits)) {
897 		CSIO_INC_STATS(q, n_qfull);
898 		return -EBUSY;
899 	}
900 
901 	/*
902 	 * If we are here, we have enough credits to satisfy the
903 	 * request. Check if we are near the end of q, and if WR spills over.
904 	 * If it does, use the first addr/size to cover the queue until
905 	 * the end. Fit the remainder portion of the request at the top
906 	 * of queue and return it in the second addr/len. Set pidx
907 	 * accordingly.
908 	 */
909 	if (unlikely(((uintptr_t)cwr + req_sz) > (uintptr_t)(q->vwrap))) {
910 		wrp->addr1 = cwr;
911 		wrp->size1 = (uint32_t)((uintptr_t)q->vwrap - (uintptr_t)cwr);
912 		wrp->addr2 = q->vstart;
913 		wrp->size2 = req_sz - wrp->size1;
914 		q->pidx	= (uint16_t)(ALIGN(wrp->size2, CSIO_QCREDIT_SZ) /
915 							CSIO_QCREDIT_SZ);
916 		CSIO_INC_STATS(q, n_qwrap);
917 		CSIO_INC_STATS(q, n_eq_wr_split);
918 	} else {
919 		wrp->addr1 = cwr;
920 		wrp->size1 = req_sz;
921 		wrp->addr2 = NULL;
922 		wrp->size2 = 0;
923 		q->pidx	+= (uint16_t)req_credits;
924 
925 		/* We are the end of queue, roll back pidx to top of queue */
926 		if (unlikely(q->pidx == q->credits)) {
927 			q->pidx = 0;
928 			CSIO_INC_STATS(q, n_qwrap);
929 		}
930 	}
931 
932 	q->inc_idx = (uint16_t)req_credits;
933 
934 	CSIO_INC_STATS(q, n_tot_reqs);
935 
936 	return 0;
937 }
938 
939 /*
940  * csio_wr_copy_to_wrp - Copies given data into WR.
941  * @data_buf - Data buffer
942  * @wrp - Work request pair.
943  * @wr_off - Work request offset.
944  * @data_len - Data length.
945  *
946  * Copies the given data in Work Request. Work request pair(wrp) specifies
947  * address information of Work request.
948  * Returns: none
949  */
950 void
951 csio_wr_copy_to_wrp(void *data_buf, struct csio_wr_pair *wrp,
952 		   uint32_t wr_off, uint32_t data_len)
953 {
954 	uint32_t nbytes;
955 
956 	/* Number of space available in buffer addr1 of WRP */
957 	nbytes = ((wrp->size1 - wr_off) >= data_len) ?
958 					data_len : (wrp->size1 - wr_off);
959 
960 	memcpy((uint8_t *) wrp->addr1 + wr_off, data_buf, nbytes);
961 	data_len -= nbytes;
962 
963 	/* Write the remaining data from the begining of circular buffer */
964 	if (data_len) {
965 		CSIO_DB_ASSERT(data_len <= wrp->size2);
966 		CSIO_DB_ASSERT(wrp->addr2 != NULL);
967 		memcpy(wrp->addr2, (uint8_t *) data_buf + nbytes, data_len);
968 	}
969 }
970 
971 /*
972  * csio_wr_issue - Notify chip of Work request.
973  * @hw: HW module.
974  * @qidx: Index of queue.
975  * @prio: 0: Low priority, 1: High priority
976  *
977  * Rings the SGE Doorbell by writing the current producer index of the passed
978  * in queue into the register.
979  *
980  */
981 int
982 csio_wr_issue(struct csio_hw *hw, int qidx, bool prio)
983 {
984 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
985 	struct csio_q *q = wrm->q_arr[qidx];
986 
987 	CSIO_DB_ASSERT((qidx >= 0) && (qidx < wrm->free_qidx));
988 
989 	wmb();
990 	/* Ring SGE Doorbell writing q->pidx into it */
991 	csio_wr_reg32(hw, DBPRIO(prio) | QID(q->un.eq.physeqid) |
992 		      PIDX(q->inc_idx), MYPF_REG(SGE_PF_KDOORBELL));
993 	q->inc_idx = 0;
994 
995 	return 0;
996 }
997 
998 static inline uint32_t
999 csio_wr_avail_qcredits(struct csio_q *q)
1000 {
1001 	if (q->pidx > q->cidx)
1002 		return q->pidx - q->cidx;
1003 	else if (q->cidx > q->pidx)
1004 		return q->credits - (q->cidx - q->pidx);
1005 	else
1006 		return 0;	/* cidx == pidx, empty queue */
1007 }
1008 
1009 /*
1010  * csio_wr_inval_flq_buf - Invalidate a free list buffer entry.
1011  * @hw: HW module.
1012  * @flq: The freelist queue.
1013  *
1014  * Invalidate the driver's version of a freelist buffer entry,
1015  * without freeing the associated the DMA memory. The entry
1016  * to be invalidated is picked up from the current Free list
1017  * queue cidx.
1018  *
1019  */
1020 static inline void
1021 csio_wr_inval_flq_buf(struct csio_hw *hw, struct csio_q *flq)
1022 {
1023 	flq->cidx++;
1024 	if (flq->cidx == flq->credits) {
1025 		flq->cidx = 0;
1026 		CSIO_INC_STATS(flq, n_qwrap);
1027 	}
1028 }
1029 
1030 /*
1031  * csio_wr_process_fl - Process a freelist completion.
1032  * @hw: HW module.
1033  * @q: The ingress queue attached to the Freelist.
1034  * @wr: The freelist completion WR in the ingress queue.
1035  * @len_to_qid: The lower 32-bits of the first flit of the RSP footer
1036  * @iq_handler: Caller's handler for this completion.
1037  * @priv: Private pointer of caller
1038  *
1039  */
1040 static inline void
1041 csio_wr_process_fl(struct csio_hw *hw, struct csio_q *q,
1042 		   void *wr, uint32_t len_to_qid,
1043 		   void (*iq_handler)(struct csio_hw *, void *,
1044 				      uint32_t, struct csio_fl_dma_buf *,
1045 				      void *),
1046 		   void *priv)
1047 {
1048 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1049 	struct csio_sge *sge = &wrm->sge;
1050 	struct csio_fl_dma_buf flb;
1051 	struct csio_dma_buf *buf, *fbuf;
1052 	uint32_t bufsz, len, lastlen = 0;
1053 	int n;
1054 	struct csio_q *flq = hw->wrm.q_arr[q->un.iq.flq_idx];
1055 
1056 	CSIO_DB_ASSERT(flq != NULL);
1057 
1058 	len = len_to_qid;
1059 
1060 	if (len & IQWRF_NEWBUF) {
1061 		if (flq->un.fl.offset > 0) {
1062 			csio_wr_inval_flq_buf(hw, flq);
1063 			flq->un.fl.offset = 0;
1064 		}
1065 		len = IQWRF_LEN_GET(len);
1066 	}
1067 
1068 	CSIO_DB_ASSERT(len != 0);
1069 
1070 	flb.totlen = len;
1071 
1072 	/* Consume all freelist buffers used for len bytes */
1073 	for (n = 0, fbuf = flb.flbufs; ; n++, fbuf++) {
1074 		buf = &flq->un.fl.bufs[flq->cidx];
1075 		bufsz = csio_wr_fl_bufsz(sge, buf);
1076 
1077 		fbuf->paddr	= buf->paddr;
1078 		fbuf->vaddr	= buf->vaddr;
1079 
1080 		flb.offset	= flq->un.fl.offset;
1081 		lastlen		= min(bufsz, len);
1082 		fbuf->len	= lastlen;
1083 
1084 		len -= lastlen;
1085 		if (!len)
1086 			break;
1087 		csio_wr_inval_flq_buf(hw, flq);
1088 	}
1089 
1090 	flb.defer_free = flq->un.fl.packen ? 0 : 1;
1091 
1092 	iq_handler(hw, wr, q->wr_sz - sizeof(struct csio_iqwr_footer),
1093 		   &flb, priv);
1094 
1095 	if (flq->un.fl.packen)
1096 		flq->un.fl.offset += ALIGN(lastlen, sge->csio_fl_align);
1097 	else
1098 		csio_wr_inval_flq_buf(hw, flq);
1099 
1100 }
1101 
1102 /*
1103  * csio_is_new_iqwr - Is this a new Ingress queue entry ?
1104  * @q: Ingress quueue.
1105  * @ftr: Ingress queue WR SGE footer.
1106  *
1107  * The entry is new if our generation bit matches the corresponding
1108  * bit in the footer of the current WR.
1109  */
1110 static inline bool
1111 csio_is_new_iqwr(struct csio_q *q, struct csio_iqwr_footer *ftr)
1112 {
1113 	return (q->un.iq.genbit == (ftr->u.type_gen >> IQWRF_GEN_SHIFT));
1114 }
1115 
1116 /*
1117  * csio_wr_process_iq - Process elements in Ingress queue.
1118  * @hw:  HW pointer
1119  * @qidx: Index of queue
1120  * @iq_handler: Handler for this queue
1121  * @priv: Caller's private pointer
1122  *
1123  * This routine walks through every entry of the ingress queue, calling
1124  * the provided iq_handler with the entry, until the generation bit
1125  * flips.
1126  */
1127 int
1128 csio_wr_process_iq(struct csio_hw *hw, struct csio_q *q,
1129 		   void (*iq_handler)(struct csio_hw *, void *,
1130 				      uint32_t, struct csio_fl_dma_buf *,
1131 				      void *),
1132 		   void *priv)
1133 {
1134 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1135 	void *wr = (void *)((uintptr_t)q->vstart + (q->cidx * q->wr_sz));
1136 	struct csio_iqwr_footer *ftr;
1137 	uint32_t wr_type, fw_qid, qid;
1138 	struct csio_q *q_completed;
1139 	struct csio_q *flq = csio_iq_has_fl(q) ?
1140 					wrm->q_arr[q->un.iq.flq_idx] : NULL;
1141 	int rv = 0;
1142 
1143 	/* Get the footer */
1144 	ftr = (struct csio_iqwr_footer *)((uintptr_t)wr +
1145 					  (q->wr_sz - sizeof(*ftr)));
1146 
1147 	/*
1148 	 * When q wrapped around last time, driver should have inverted
1149 	 * ic.genbit as well.
1150 	 */
1151 	while (csio_is_new_iqwr(q, ftr)) {
1152 
1153 		CSIO_DB_ASSERT(((uintptr_t)wr + q->wr_sz) <=
1154 						(uintptr_t)q->vwrap);
1155 		rmb();
1156 		wr_type = IQWRF_TYPE_GET(ftr->u.type_gen);
1157 
1158 		switch (wr_type) {
1159 		case X_RSPD_TYPE_CPL:
1160 			/* Subtract footer from WR len */
1161 			iq_handler(hw, wr, q->wr_sz - sizeof(*ftr), NULL, priv);
1162 			break;
1163 		case X_RSPD_TYPE_FLBUF:
1164 			csio_wr_process_fl(hw, q, wr,
1165 					   ntohl(ftr->pldbuflen_qid),
1166 					   iq_handler, priv);
1167 			break;
1168 		case X_RSPD_TYPE_INTR:
1169 			fw_qid = ntohl(ftr->pldbuflen_qid);
1170 			qid = fw_qid - wrm->fw_iq_start;
1171 			q_completed = hw->wrm.intr_map[qid];
1172 
1173 			if (unlikely(qid ==
1174 					csio_q_physiqid(hw, hw->intr_iq_idx))) {
1175 				/*
1176 				 * We are already in the Forward Interrupt
1177 				 * Interrupt Queue Service! Do-not service
1178 				 * again!
1179 				 *
1180 				 */
1181 			} else {
1182 				CSIO_DB_ASSERT(q_completed);
1183 				CSIO_DB_ASSERT(
1184 					q_completed->un.iq.iq_intx_handler);
1185 
1186 				/* Call the queue handler. */
1187 				q_completed->un.iq.iq_intx_handler(hw, NULL,
1188 						0, NULL, (void *)q_completed);
1189 			}
1190 			break;
1191 		default:
1192 			csio_warn(hw, "Unknown resp type 0x%x received\n",
1193 				 wr_type);
1194 			CSIO_INC_STATS(q, n_rsp_unknown);
1195 			break;
1196 		}
1197 
1198 		/*
1199 		 * Ingress *always* has fixed size WR entries. Therefore,
1200 		 * there should always be complete WRs towards the end of
1201 		 * queue.
1202 		 */
1203 		if (((uintptr_t)wr + q->wr_sz) == (uintptr_t)q->vwrap) {
1204 
1205 			/* Roll over to start of queue */
1206 			q->cidx = 0;
1207 			wr	= q->vstart;
1208 
1209 			/* Toggle genbit */
1210 			q->un.iq.genbit ^= 0x1;
1211 
1212 			CSIO_INC_STATS(q, n_qwrap);
1213 		} else {
1214 			q->cidx++;
1215 			wr	= (void *)((uintptr_t)(q->vstart) +
1216 					   (q->cidx * q->wr_sz));
1217 		}
1218 
1219 		ftr = (struct csio_iqwr_footer *)((uintptr_t)wr +
1220 						  (q->wr_sz - sizeof(*ftr)));
1221 		q->inc_idx++;
1222 
1223 	} /* while (q->un.iq.genbit == hdr->genbit) */
1224 
1225 	/*
1226 	 * We need to re-arm SGE interrupts in case we got a stray interrupt,
1227 	 * especially in msix mode. With INTx, this may be a common occurence.
1228 	 */
1229 	if (unlikely(!q->inc_idx)) {
1230 		CSIO_INC_STATS(q, n_stray_comp);
1231 		rv = -EINVAL;
1232 		goto restart;
1233 	}
1234 
1235 	/* Replenish free list buffers if pending falls below low water mark */
1236 	if (flq) {
1237 		uint32_t avail  = csio_wr_avail_qcredits(flq);
1238 		if (avail <= 16) {
1239 			/* Make sure in FLQ, atleast 1 credit (8 FL buffers)
1240 			 * remains unpopulated otherwise HW thinks
1241 			 * FLQ is empty.
1242 			 */
1243 			csio_wr_update_fl(hw, flq, (flq->credits - 8) - avail);
1244 			csio_wr_ring_fldb(hw, flq);
1245 		}
1246 	}
1247 
1248 restart:
1249 	/* Now inform SGE about our incremental index value */
1250 	csio_wr_reg32(hw, CIDXINC(q->inc_idx)		|
1251 			  INGRESSQID(q->un.iq.physiqid)	|
1252 			  TIMERREG(csio_sge_timer_reg),
1253 			  MYPF_REG(SGE_PF_GTS));
1254 	q->stats.n_tot_rsps += q->inc_idx;
1255 
1256 	q->inc_idx = 0;
1257 
1258 	return rv;
1259 }
1260 
1261 int
1262 csio_wr_process_iq_idx(struct csio_hw *hw, int qidx,
1263 		   void (*iq_handler)(struct csio_hw *, void *,
1264 				      uint32_t, struct csio_fl_dma_buf *,
1265 				      void *),
1266 		   void *priv)
1267 {
1268 	struct csio_wrm *wrm	= csio_hw_to_wrm(hw);
1269 	struct csio_q	*iq	= wrm->q_arr[qidx];
1270 
1271 	return csio_wr_process_iq(hw, iq, iq_handler, priv);
1272 }
1273 
1274 static int
1275 csio_closest_timer(struct csio_sge *s, int time)
1276 {
1277 	int i, delta, match = 0, min_delta = INT_MAX;
1278 
1279 	for (i = 0; i < ARRAY_SIZE(s->timer_val); i++) {
1280 		delta = time - s->timer_val[i];
1281 		if (delta < 0)
1282 			delta = -delta;
1283 		if (delta < min_delta) {
1284 			min_delta = delta;
1285 			match = i;
1286 		}
1287 	}
1288 	return match;
1289 }
1290 
1291 static int
1292 csio_closest_thresh(struct csio_sge *s, int cnt)
1293 {
1294 	int i, delta, match = 0, min_delta = INT_MAX;
1295 
1296 	for (i = 0; i < ARRAY_SIZE(s->counter_val); i++) {
1297 		delta = cnt - s->counter_val[i];
1298 		if (delta < 0)
1299 			delta = -delta;
1300 		if (delta < min_delta) {
1301 			min_delta = delta;
1302 			match = i;
1303 		}
1304 	}
1305 	return match;
1306 }
1307 
1308 static void
1309 csio_wr_fixup_host_params(struct csio_hw *hw)
1310 {
1311 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1312 	struct csio_sge *sge = &wrm->sge;
1313 	uint32_t clsz = L1_CACHE_BYTES;
1314 	uint32_t s_hps = PAGE_SHIFT - 10;
1315 	uint32_t ingpad = 0;
1316 	uint32_t stat_len = clsz > 64 ? 128 : 64;
1317 
1318 	csio_wr_reg32(hw, HOSTPAGESIZEPF0(s_hps) | HOSTPAGESIZEPF1(s_hps) |
1319 		      HOSTPAGESIZEPF2(s_hps) | HOSTPAGESIZEPF3(s_hps) |
1320 		      HOSTPAGESIZEPF4(s_hps) | HOSTPAGESIZEPF5(s_hps) |
1321 		      HOSTPAGESIZEPF6(s_hps) | HOSTPAGESIZEPF7(s_hps),
1322 		      SGE_HOST_PAGE_SIZE);
1323 
1324 	sge->csio_fl_align = clsz < 32 ? 32 : clsz;
1325 	ingpad = ilog2(sge->csio_fl_align) - 5;
1326 
1327 	csio_set_reg_field(hw, SGE_CONTROL, INGPADBOUNDARY_MASK |
1328 					    EGRSTATUSPAGESIZE(1),
1329 			   INGPADBOUNDARY(ingpad) |
1330 			   EGRSTATUSPAGESIZE(stat_len != 64));
1331 
1332 	/* FL BUFFER SIZE#0 is Page size i,e already aligned to cache line */
1333 	csio_wr_reg32(hw, PAGE_SIZE, SGE_FL_BUFFER_SIZE0);
1334 	csio_wr_reg32(hw,
1335 		      (csio_rd_reg32(hw, SGE_FL_BUFFER_SIZE2) +
1336 		      sge->csio_fl_align - 1) & ~(sge->csio_fl_align - 1),
1337 		      SGE_FL_BUFFER_SIZE2);
1338 	csio_wr_reg32(hw,
1339 		      (csio_rd_reg32(hw, SGE_FL_BUFFER_SIZE3) +
1340 		      sge->csio_fl_align - 1) & ~(sge->csio_fl_align - 1),
1341 		      SGE_FL_BUFFER_SIZE3);
1342 
1343 	csio_wr_reg32(hw, HPZ0(PAGE_SHIFT - 12), ULP_RX_TDDP_PSZ);
1344 
1345 	/* default value of rx_dma_offset of the NIC driver */
1346 	csio_set_reg_field(hw, SGE_CONTROL, PKTSHIFT_MASK,
1347 			   PKTSHIFT(CSIO_SGE_RX_DMA_OFFSET));
1348 }
1349 
1350 static void
1351 csio_init_intr_coalesce_parms(struct csio_hw *hw)
1352 {
1353 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1354 	struct csio_sge *sge = &wrm->sge;
1355 
1356 	csio_sge_thresh_reg = csio_closest_thresh(sge, csio_intr_coalesce_cnt);
1357 	if (csio_intr_coalesce_cnt) {
1358 		csio_sge_thresh_reg = 0;
1359 		csio_sge_timer_reg = X_TIMERREG_RESTART_COUNTER;
1360 		return;
1361 	}
1362 
1363 	csio_sge_timer_reg = csio_closest_timer(sge, csio_intr_coalesce_time);
1364 }
1365 
1366 /*
1367  * csio_wr_get_sge - Get SGE register values.
1368  * @hw: HW module.
1369  *
1370  * Used by non-master functions and by master-functions relying on config file.
1371  */
1372 static void
1373 csio_wr_get_sge(struct csio_hw *hw)
1374 {
1375 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1376 	struct csio_sge *sge = &wrm->sge;
1377 	uint32_t ingpad;
1378 	int i;
1379 	u32 timer_value_0_and_1, timer_value_2_and_3, timer_value_4_and_5;
1380 	u32 ingress_rx_threshold;
1381 
1382 	sge->sge_control = csio_rd_reg32(hw, SGE_CONTROL);
1383 
1384 	ingpad = INGPADBOUNDARY_GET(sge->sge_control);
1385 
1386 	switch (ingpad) {
1387 	case X_INGPCIEBOUNDARY_32B:
1388 		sge->csio_fl_align = 32; break;
1389 	case X_INGPCIEBOUNDARY_64B:
1390 		sge->csio_fl_align = 64; break;
1391 	case X_INGPCIEBOUNDARY_128B:
1392 		sge->csio_fl_align = 128; break;
1393 	case X_INGPCIEBOUNDARY_256B:
1394 		sge->csio_fl_align = 256; break;
1395 	case X_INGPCIEBOUNDARY_512B:
1396 		sge->csio_fl_align = 512; break;
1397 	case X_INGPCIEBOUNDARY_1024B:
1398 		sge->csio_fl_align = 1024; break;
1399 	case X_INGPCIEBOUNDARY_2048B:
1400 		sge->csio_fl_align = 2048; break;
1401 	case X_INGPCIEBOUNDARY_4096B:
1402 		sge->csio_fl_align = 4096; break;
1403 	}
1404 
1405 	for (i = 0; i < CSIO_SGE_FL_SIZE_REGS; i++)
1406 		csio_get_flbuf_size(hw, sge, i);
1407 
1408 	timer_value_0_and_1 = csio_rd_reg32(hw, SGE_TIMER_VALUE_0_AND_1);
1409 	timer_value_2_and_3 = csio_rd_reg32(hw, SGE_TIMER_VALUE_2_AND_3);
1410 	timer_value_4_and_5 = csio_rd_reg32(hw, SGE_TIMER_VALUE_4_AND_5);
1411 
1412 	sge->timer_val[0] = (uint16_t)csio_core_ticks_to_us(hw,
1413 					TIMERVALUE0_GET(timer_value_0_and_1));
1414 	sge->timer_val[1] = (uint16_t)csio_core_ticks_to_us(hw,
1415 					TIMERVALUE1_GET(timer_value_0_and_1));
1416 	sge->timer_val[2] = (uint16_t)csio_core_ticks_to_us(hw,
1417 					TIMERVALUE2_GET(timer_value_2_and_3));
1418 	sge->timer_val[3] = (uint16_t)csio_core_ticks_to_us(hw,
1419 					TIMERVALUE3_GET(timer_value_2_and_3));
1420 	sge->timer_val[4] = (uint16_t)csio_core_ticks_to_us(hw,
1421 					TIMERVALUE4_GET(timer_value_4_and_5));
1422 	sge->timer_val[5] = (uint16_t)csio_core_ticks_to_us(hw,
1423 					TIMERVALUE5_GET(timer_value_4_and_5));
1424 
1425 	ingress_rx_threshold = csio_rd_reg32(hw, SGE_INGRESS_RX_THRESHOLD);
1426 	sge->counter_val[0] = THRESHOLD_0_GET(ingress_rx_threshold);
1427 	sge->counter_val[1] = THRESHOLD_1_GET(ingress_rx_threshold);
1428 	sge->counter_val[2] = THRESHOLD_2_GET(ingress_rx_threshold);
1429 	sge->counter_val[3] = THRESHOLD_3_GET(ingress_rx_threshold);
1430 
1431 	csio_init_intr_coalesce_parms(hw);
1432 }
1433 
1434 /*
1435  * csio_wr_set_sge - Initialize SGE registers
1436  * @hw: HW module.
1437  *
1438  * Used by Master function to initialize SGE registers in the absence
1439  * of a config file.
1440  */
1441 static void
1442 csio_wr_set_sge(struct csio_hw *hw)
1443 {
1444 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1445 	struct csio_sge *sge = &wrm->sge;
1446 	int i;
1447 
1448 	/*
1449 	 * Set up our basic SGE mode to deliver CPL messages to our Ingress
1450 	 * Queue and Packet Date to the Free List.
1451 	 */
1452 	csio_set_reg_field(hw, SGE_CONTROL, RXPKTCPLMODE(1), RXPKTCPLMODE(1));
1453 
1454 	sge->sge_control = csio_rd_reg32(hw, SGE_CONTROL);
1455 
1456 	/* sge->csio_fl_align is set up by csio_wr_fixup_host_params(). */
1457 
1458 	/*
1459 	 * Set up to drop DOORBELL writes when the DOORBELL FIFO overflows
1460 	 * and generate an interrupt when this occurs so we can recover.
1461 	 */
1462 	csio_set_reg_field(hw, SGE_DBFIFO_STATUS,
1463 			   HP_INT_THRESH(HP_INT_THRESH_MASK) |
1464 			   LP_INT_THRESH(LP_INT_THRESH_MASK),
1465 			   HP_INT_THRESH(CSIO_SGE_DBFIFO_INT_THRESH) |
1466 			   LP_INT_THRESH(CSIO_SGE_DBFIFO_INT_THRESH));
1467 	csio_set_reg_field(hw, SGE_DOORBELL_CONTROL, ENABLE_DROP,
1468 			   ENABLE_DROP);
1469 
1470 	/* SGE_FL_BUFFER_SIZE0 is set up by csio_wr_fixup_host_params(). */
1471 
1472 	CSIO_SET_FLBUF_SIZE(hw, 1, CSIO_SGE_FLBUF_SIZE1);
1473 	CSIO_SET_FLBUF_SIZE(hw, 2, CSIO_SGE_FLBUF_SIZE2);
1474 	CSIO_SET_FLBUF_SIZE(hw, 3, CSIO_SGE_FLBUF_SIZE3);
1475 	CSIO_SET_FLBUF_SIZE(hw, 4, CSIO_SGE_FLBUF_SIZE4);
1476 	CSIO_SET_FLBUF_SIZE(hw, 5, CSIO_SGE_FLBUF_SIZE5);
1477 	CSIO_SET_FLBUF_SIZE(hw, 6, CSIO_SGE_FLBUF_SIZE6);
1478 	CSIO_SET_FLBUF_SIZE(hw, 7, CSIO_SGE_FLBUF_SIZE7);
1479 	CSIO_SET_FLBUF_SIZE(hw, 8, CSIO_SGE_FLBUF_SIZE8);
1480 
1481 	for (i = 0; i < CSIO_SGE_FL_SIZE_REGS; i++)
1482 		csio_get_flbuf_size(hw, sge, i);
1483 
1484 	/* Initialize interrupt coalescing attributes */
1485 	sge->timer_val[0] = CSIO_SGE_TIMER_VAL_0;
1486 	sge->timer_val[1] = CSIO_SGE_TIMER_VAL_1;
1487 	sge->timer_val[2] = CSIO_SGE_TIMER_VAL_2;
1488 	sge->timer_val[3] = CSIO_SGE_TIMER_VAL_3;
1489 	sge->timer_val[4] = CSIO_SGE_TIMER_VAL_4;
1490 	sge->timer_val[5] = CSIO_SGE_TIMER_VAL_5;
1491 
1492 	sge->counter_val[0] = CSIO_SGE_INT_CNT_VAL_0;
1493 	sge->counter_val[1] = CSIO_SGE_INT_CNT_VAL_1;
1494 	sge->counter_val[2] = CSIO_SGE_INT_CNT_VAL_2;
1495 	sge->counter_val[3] = CSIO_SGE_INT_CNT_VAL_3;
1496 
1497 	csio_wr_reg32(hw, THRESHOLD_0(sge->counter_val[0]) |
1498 		      THRESHOLD_1(sge->counter_val[1]) |
1499 		      THRESHOLD_2(sge->counter_val[2]) |
1500 		      THRESHOLD_3(sge->counter_val[3]),
1501 		      SGE_INGRESS_RX_THRESHOLD);
1502 
1503 	csio_wr_reg32(hw,
1504 		   TIMERVALUE0(csio_us_to_core_ticks(hw, sge->timer_val[0])) |
1505 		   TIMERVALUE1(csio_us_to_core_ticks(hw, sge->timer_val[1])),
1506 		   SGE_TIMER_VALUE_0_AND_1);
1507 
1508 	csio_wr_reg32(hw,
1509 		   TIMERVALUE2(csio_us_to_core_ticks(hw, sge->timer_val[2])) |
1510 		   TIMERVALUE3(csio_us_to_core_ticks(hw, sge->timer_val[3])),
1511 		   SGE_TIMER_VALUE_2_AND_3);
1512 
1513 	csio_wr_reg32(hw,
1514 		   TIMERVALUE4(csio_us_to_core_ticks(hw, sge->timer_val[4])) |
1515 		   TIMERVALUE5(csio_us_to_core_ticks(hw, sge->timer_val[5])),
1516 		   SGE_TIMER_VALUE_4_AND_5);
1517 
1518 	csio_init_intr_coalesce_parms(hw);
1519 }
1520 
1521 void
1522 csio_wr_sge_init(struct csio_hw *hw)
1523 {
1524 	/*
1525 	 * If we are master:
1526 	 *    - If we plan to use the config file, we need to fixup some
1527 	 *      host specific registers, and read the rest of the SGE
1528 	 *      configuration.
1529 	 *    - If we dont plan to use the config file, we need to initialize
1530 	 *      SGE entirely, including fixing the host specific registers.
1531 	 * If we arent the master, we are only allowed to read and work off of
1532 	 *      the already initialized SGE values.
1533 	 *
1534 	 * Therefore, before calling this function, we assume that the master-
1535 	 * ship of the card, and whether to use config file or not, have
1536 	 * already been decided. In other words, CSIO_HWF_USING_SOFT_PARAMS and
1537 	 * CSIO_HWF_MASTER should be set/unset.
1538 	 */
1539 	if (csio_is_hw_master(hw)) {
1540 		csio_wr_fixup_host_params(hw);
1541 
1542 		if (hw->flags & CSIO_HWF_USING_SOFT_PARAMS)
1543 			csio_wr_get_sge(hw);
1544 		else
1545 			csio_wr_set_sge(hw);
1546 	} else
1547 		csio_wr_get_sge(hw);
1548 }
1549 
1550 /*
1551  * csio_wrm_init - Initialize Work request module.
1552  * @wrm: WR module
1553  * @hw: HW pointer
1554  *
1555  * Allocates memory for an array of queue pointers starting at q_arr.
1556  */
1557 int
1558 csio_wrm_init(struct csio_wrm *wrm, struct csio_hw *hw)
1559 {
1560 	int i;
1561 
1562 	if (!wrm->num_q) {
1563 		csio_err(hw, "Num queues is not set\n");
1564 		return -EINVAL;
1565 	}
1566 
1567 	wrm->q_arr = kzalloc(sizeof(struct csio_q *) * wrm->num_q, GFP_KERNEL);
1568 	if (!wrm->q_arr)
1569 		goto err;
1570 
1571 	for (i = 0; i < wrm->num_q; i++) {
1572 		wrm->q_arr[i] = kzalloc(sizeof(struct csio_q), GFP_KERNEL);
1573 		if (!wrm->q_arr[i]) {
1574 			while (--i >= 0)
1575 				kfree(wrm->q_arr[i]);
1576 			goto err_free_arr;
1577 		}
1578 	}
1579 	wrm->free_qidx	= 0;
1580 
1581 	return 0;
1582 
1583 err_free_arr:
1584 	kfree(wrm->q_arr);
1585 err:
1586 	return -ENOMEM;
1587 }
1588 
1589 /*
1590  * csio_wrm_exit - Initialize Work request module.
1591  * @wrm: WR module
1592  * @hw: HW module
1593  *
1594  * Uninitialize WR module. Free q_arr and pointers in it.
1595  * We have the additional job of freeing the DMA memory associated
1596  * with the queues.
1597  */
1598 void
1599 csio_wrm_exit(struct csio_wrm *wrm, struct csio_hw *hw)
1600 {
1601 	int i;
1602 	uint32_t j;
1603 	struct csio_q *q;
1604 	struct csio_dma_buf *buf;
1605 
1606 	for (i = 0; i < wrm->num_q; i++) {
1607 		q = wrm->q_arr[i];
1608 
1609 		if (wrm->free_qidx && (i < wrm->free_qidx)) {
1610 			if (q->type == CSIO_FREELIST) {
1611 				if (!q->un.fl.bufs)
1612 					continue;
1613 				for (j = 0; j < q->credits; j++) {
1614 					buf = &q->un.fl.bufs[j];
1615 					if (!buf->vaddr)
1616 						continue;
1617 					pci_free_consistent(hw->pdev, buf->len,
1618 							    buf->vaddr,
1619 							    buf->paddr);
1620 				}
1621 				kfree(q->un.fl.bufs);
1622 			}
1623 			pci_free_consistent(hw->pdev, q->size,
1624 					    q->vstart, q->pstart);
1625 		}
1626 		kfree(q);
1627 	}
1628 
1629 	hw->flags &= ~CSIO_HWF_Q_MEM_ALLOCED;
1630 
1631 	kfree(wrm->q_arr);
1632 }
1633