xref: /openbmc/linux/drivers/scsi/csiostor/csio_wr.c (revision ba61bb17)
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##_A)
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_A +
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_F) ?  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_F | QID_V(flq->un.fl.flid) |
88 				  PIDX_T5_V(flq->inc_idx / 8) | DBTYPE_F,
89 				  MYPF_REG(SGE_PF_KDOORBELL_A));
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_V(0)		|
99 			  INGRESSQID_V(iqid)	|
100 			  TIMERREG_V(X_TIMERREG_RESTART_COUNTER),
101 			  MYPF_REG(SGE_PF_GTS_A));
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_zalloc_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 	q->type		= type;
244 	q->owner	= owner;
245 	q->pidx		= q->cidx = q->inc_idx = 0;
246 	q->size		= qsz;
247 	q->wr_sz	= wrsize;	/* If using fixed size WRs */
248 
249 	wrm->free_qidx++;
250 
251 	if (type == CSIO_INGRESS) {
252 		/* Since queue area is set to zero */
253 		q->un.iq.genbit	= 1;
254 
255 		/*
256 		 * Ingress queue status page size is always the size of
257 		 * the ingress queue entry.
258 		 */
259 		q->credits	= (qsz - q->wr_sz) / q->wr_sz;
260 		q->vwrap	= (void *)((uintptr_t)(q->vstart) + qsz
261 							- q->wr_sz);
262 
263 		/* Allocate memory for FL if requested */
264 		if (nflb > 0) {
265 			flq_idx = csio_wr_alloc_q(hw, nflb * sizeof(__be64),
266 						  sizeof(__be64), CSIO_FREELIST,
267 						  owner, 0, sreg, NULL);
268 			if (flq_idx == -1) {
269 				csio_err(hw,
270 					 "Failed to allocate FL queue"
271 					 " for IQ idx:%d\n", free_idx);
272 				return -1;
273 			}
274 
275 			/* Associate the new FL with the Ingress quue */
276 			q->un.iq.flq_idx = flq_idx;
277 
278 			flq = wrm->q_arr[q->un.iq.flq_idx];
279 			flq->un.fl.bufs = kcalloc(flq->credits,
280 						  sizeof(struct csio_dma_buf),
281 						  GFP_KERNEL);
282 			if (!flq->un.fl.bufs) {
283 				csio_err(hw,
284 					 "Failed to allocate FL queue bufs"
285 					 " for IQ idx:%d\n", free_idx);
286 				return -1;
287 			}
288 
289 			flq->un.fl.packen = 0;
290 			flq->un.fl.offset = 0;
291 			flq->un.fl.sreg = sreg;
292 
293 			/* Fill up the free list buffers */
294 			if (csio_wr_fill_fl(hw, flq))
295 				return -1;
296 
297 			/*
298 			 * Make sure in a FLQ, atleast 1 credit (8 FL buffers)
299 			 * remains unpopulated,otherwise HW thinks
300 			 * FLQ is empty.
301 			 */
302 			flq->pidx = flq->inc_idx = flq->credits - 8;
303 		} else {
304 			q->un.iq.flq_idx = -1;
305 		}
306 
307 		/* Associate the IQ INTx handler. */
308 		q->un.iq.iq_intx_handler = iq_intx_handler;
309 
310 		csio_q_iqid(hw, ret_idx) = CSIO_MAX_QID;
311 
312 	} else if (type == CSIO_EGRESS) {
313 		q->credits = (qsz - csio_wr_qstat_pgsz(hw)) / CSIO_QCREDIT_SZ;
314 		q->vwrap   = (void *)((uintptr_t)(q->vstart) + qsz
315 						- csio_wr_qstat_pgsz(hw));
316 		csio_q_eqid(hw, ret_idx) = CSIO_MAX_QID;
317 	} else { /* Freelist */
318 		q->credits = (qsz - csio_wr_qstat_pgsz(hw)) / sizeof(__be64);
319 		q->vwrap   = (void *)((uintptr_t)(q->vstart) + qsz
320 						- csio_wr_qstat_pgsz(hw));
321 		csio_q_flid(hw, ret_idx) = CSIO_MAX_QID;
322 	}
323 
324 	return ret_idx;
325 }
326 
327 /*
328  * csio_wr_iq_create_rsp - Response handler for IQ creation.
329  * @hw: The HW module.
330  * @mbp: Mailbox.
331  * @iq_idx: Ingress queue that got created.
332  *
333  * Handle FW_IQ_CMD mailbox completion. Save off the assigned IQ/FL ids.
334  */
335 static int
336 csio_wr_iq_create_rsp(struct csio_hw *hw, struct csio_mb *mbp, int iq_idx)
337 {
338 	struct csio_iq_params iqp;
339 	enum fw_retval retval;
340 	uint32_t iq_id;
341 	int flq_idx;
342 
343 	memset(&iqp, 0, sizeof(struct csio_iq_params));
344 
345 	csio_mb_iq_alloc_write_rsp(hw, mbp, &retval, &iqp);
346 
347 	if (retval != FW_SUCCESS) {
348 		csio_err(hw, "IQ cmd returned 0x%x!\n", retval);
349 		mempool_free(mbp, hw->mb_mempool);
350 		return -EINVAL;
351 	}
352 
353 	csio_q_iqid(hw, iq_idx)		= iqp.iqid;
354 	csio_q_physiqid(hw, iq_idx)	= iqp.physiqid;
355 	csio_q_pidx(hw, iq_idx)		= csio_q_cidx(hw, iq_idx) = 0;
356 	csio_q_inc_idx(hw, iq_idx)	= 0;
357 
358 	/* Actual iq-id. */
359 	iq_id = iqp.iqid - hw->wrm.fw_iq_start;
360 
361 	/* Set the iq-id to iq map table. */
362 	if (iq_id >= CSIO_MAX_IQ) {
363 		csio_err(hw,
364 			 "Exceeding MAX_IQ(%d) supported!"
365 			 " iqid:%d rel_iqid:%d FW iq_start:%d\n",
366 			 CSIO_MAX_IQ, iq_id, iqp.iqid, hw->wrm.fw_iq_start);
367 		mempool_free(mbp, hw->mb_mempool);
368 		return -EINVAL;
369 	}
370 	csio_q_set_intr_map(hw, iq_idx, iq_id);
371 
372 	/*
373 	 * During FW_IQ_CMD, FW sets interrupt_sent bit to 1 in the SGE
374 	 * ingress context of this queue. This will block interrupts to
375 	 * this queue until the next GTS write. Therefore, we do a
376 	 * 0-cidx increment GTS write for this queue just to clear the
377 	 * interrupt_sent bit. This will re-enable interrupts to this
378 	 * queue.
379 	 */
380 	csio_wr_sge_intr_enable(hw, iqp.physiqid);
381 
382 	flq_idx = csio_q_iq_flq_idx(hw, iq_idx);
383 	if (flq_idx != -1) {
384 		struct csio_q *flq = hw->wrm.q_arr[flq_idx];
385 
386 		csio_q_flid(hw, flq_idx) = iqp.fl0id;
387 		csio_q_cidx(hw, flq_idx) = 0;
388 		csio_q_pidx(hw, flq_idx)    = csio_q_credits(hw, flq_idx) - 8;
389 		csio_q_inc_idx(hw, flq_idx) = csio_q_credits(hw, flq_idx) - 8;
390 
391 		/* Now update SGE about the buffers allocated during init */
392 		csio_wr_ring_fldb(hw, flq);
393 	}
394 
395 	mempool_free(mbp, hw->mb_mempool);
396 
397 	return 0;
398 }
399 
400 /*
401  * csio_wr_iq_create - Configure an Ingress queue with FW.
402  * @hw: The HW module.
403  * @priv: Private data object.
404  * @iq_idx: Ingress queue index in the WR module.
405  * @vec: MSIX vector.
406  * @portid: PCIE Channel to be associated with this queue.
407  * @async: Is this a FW asynchronous message handling queue?
408  * @cbfn: Completion callback.
409  *
410  * This API configures an ingress queue with FW by issuing a FW_IQ_CMD mailbox
411  * with alloc/write bits set.
412  */
413 int
414 csio_wr_iq_create(struct csio_hw *hw, void *priv, int iq_idx,
415 		  uint32_t vec, uint8_t portid, bool async,
416 		  void (*cbfn) (struct csio_hw *, struct csio_mb *))
417 {
418 	struct csio_mb  *mbp;
419 	struct csio_iq_params iqp;
420 	int flq_idx;
421 
422 	memset(&iqp, 0, sizeof(struct csio_iq_params));
423 	csio_q_portid(hw, iq_idx) = portid;
424 
425 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
426 	if (!mbp) {
427 		csio_err(hw, "IQ command out of memory!\n");
428 		return -ENOMEM;
429 	}
430 
431 	switch (hw->intr_mode) {
432 	case CSIO_IM_INTX:
433 	case CSIO_IM_MSI:
434 		/* For interrupt forwarding queue only */
435 		if (hw->intr_iq_idx == iq_idx)
436 			iqp.iqandst	= X_INTERRUPTDESTINATION_PCIE;
437 		else
438 			iqp.iqandst	= X_INTERRUPTDESTINATION_IQ;
439 		iqp.iqandstindex	=
440 			csio_q_physiqid(hw, hw->intr_iq_idx);
441 		break;
442 	case CSIO_IM_MSIX:
443 		iqp.iqandst		= X_INTERRUPTDESTINATION_PCIE;
444 		iqp.iqandstindex	= (uint16_t)vec;
445 		break;
446 	case CSIO_IM_NONE:
447 		mempool_free(mbp, hw->mb_mempool);
448 		return -EINVAL;
449 	}
450 
451 	/* Pass in the ingress queue cmd parameters */
452 	iqp.pfn			= hw->pfn;
453 	iqp.vfn			= 0;
454 	iqp.iq_start		= 1;
455 	iqp.viid		= 0;
456 	iqp.type		= FW_IQ_TYPE_FL_INT_CAP;
457 	iqp.iqasynch		= async;
458 	if (csio_intr_coalesce_cnt)
459 		iqp.iqanus	= X_UPDATESCHEDULING_COUNTER_OPTTIMER;
460 	else
461 		iqp.iqanus	= X_UPDATESCHEDULING_TIMER;
462 	iqp.iqanud		= X_UPDATEDELIVERY_INTERRUPT;
463 	iqp.iqpciech		= portid;
464 	iqp.iqintcntthresh	= (uint8_t)csio_sge_thresh_reg;
465 
466 	switch (csio_q_wr_sz(hw, iq_idx)) {
467 	case 16:
468 		iqp.iqesize = 0; break;
469 	case 32:
470 		iqp.iqesize = 1; break;
471 	case 64:
472 		iqp.iqesize = 2; break;
473 	case 128:
474 		iqp.iqesize = 3; break;
475 	}
476 
477 	iqp.iqsize		= csio_q_size(hw, iq_idx) /
478 						csio_q_wr_sz(hw, iq_idx);
479 	iqp.iqaddr		= csio_q_pstart(hw, iq_idx);
480 
481 	flq_idx = csio_q_iq_flq_idx(hw, iq_idx);
482 	if (flq_idx != -1) {
483 		enum chip_type chip = CHELSIO_CHIP_VERSION(hw->chip_id);
484 		struct csio_q *flq = hw->wrm.q_arr[flq_idx];
485 
486 		iqp.fl0paden	= 1;
487 		iqp.fl0packen	= flq->un.fl.packen ? 1 : 0;
488 		iqp.fl0fbmin	= X_FETCHBURSTMIN_64B;
489 		iqp.fl0fbmax	= ((chip == CHELSIO_T5) ?
490 				  X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B);
491 		iqp.fl0size	= csio_q_size(hw, flq_idx) / CSIO_QCREDIT_SZ;
492 		iqp.fl0addr	= csio_q_pstart(hw, flq_idx);
493 	}
494 
495 	csio_mb_iq_alloc_write(hw, mbp, priv, CSIO_MB_DEFAULT_TMO, &iqp, cbfn);
496 
497 	if (csio_mb_issue(hw, mbp)) {
498 		csio_err(hw, "Issue of IQ cmd failed!\n");
499 		mempool_free(mbp, hw->mb_mempool);
500 		return -EINVAL;
501 	}
502 
503 	if (cbfn != NULL)
504 		return 0;
505 
506 	return csio_wr_iq_create_rsp(hw, mbp, iq_idx);
507 }
508 
509 /*
510  * csio_wr_eq_create_rsp - Response handler for EQ creation.
511  * @hw: The HW module.
512  * @mbp: Mailbox.
513  * @eq_idx: Egress queue that got created.
514  *
515  * Handle FW_EQ_OFLD_CMD mailbox completion. Save off the assigned EQ ids.
516  */
517 static int
518 csio_wr_eq_cfg_rsp(struct csio_hw *hw, struct csio_mb *mbp, int eq_idx)
519 {
520 	struct csio_eq_params eqp;
521 	enum fw_retval retval;
522 
523 	memset(&eqp, 0, sizeof(struct csio_eq_params));
524 
525 	csio_mb_eq_ofld_alloc_write_rsp(hw, mbp, &retval, &eqp);
526 
527 	if (retval != FW_SUCCESS) {
528 		csio_err(hw, "EQ OFLD cmd returned 0x%x!\n", retval);
529 		mempool_free(mbp, hw->mb_mempool);
530 		return -EINVAL;
531 	}
532 
533 	csio_q_eqid(hw, eq_idx)	= (uint16_t)eqp.eqid;
534 	csio_q_physeqid(hw, eq_idx) = (uint16_t)eqp.physeqid;
535 	csio_q_pidx(hw, eq_idx)	= csio_q_cidx(hw, eq_idx) = 0;
536 	csio_q_inc_idx(hw, eq_idx) = 0;
537 
538 	mempool_free(mbp, hw->mb_mempool);
539 
540 	return 0;
541 }
542 
543 /*
544  * csio_wr_eq_create - Configure an Egress queue with FW.
545  * @hw: HW module.
546  * @priv: Private data.
547  * @eq_idx: Egress queue index in the WR module.
548  * @iq_idx: Associated ingress queue index.
549  * @cbfn: Completion callback.
550  *
551  * This API configures a offload egress queue with FW by issuing a
552  * FW_EQ_OFLD_CMD  (with alloc + write ) mailbox.
553  */
554 int
555 csio_wr_eq_create(struct csio_hw *hw, void *priv, int eq_idx,
556 		  int iq_idx, uint8_t portid,
557 		  void (*cbfn) (struct csio_hw *, struct csio_mb *))
558 {
559 	struct csio_mb  *mbp;
560 	struct csio_eq_params eqp;
561 
562 	memset(&eqp, 0, sizeof(struct csio_eq_params));
563 
564 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
565 	if (!mbp) {
566 		csio_err(hw, "EQ command out of memory!\n");
567 		return -ENOMEM;
568 	}
569 
570 	eqp.pfn			= hw->pfn;
571 	eqp.vfn			= 0;
572 	eqp.eqstart		= 1;
573 	eqp.hostfcmode		= X_HOSTFCMODE_STATUS_PAGE;
574 	eqp.iqid		= csio_q_iqid(hw, iq_idx);
575 	eqp.fbmin		= X_FETCHBURSTMIN_64B;
576 	eqp.fbmax		= X_FETCHBURSTMAX_512B;
577 	eqp.cidxfthresh		= 0;
578 	eqp.pciechn		= portid;
579 	eqp.eqsize		= csio_q_size(hw, eq_idx) / CSIO_QCREDIT_SZ;
580 	eqp.eqaddr		= csio_q_pstart(hw, eq_idx);
581 
582 	csio_mb_eq_ofld_alloc_write(hw, mbp, priv, CSIO_MB_DEFAULT_TMO,
583 				    &eqp, cbfn);
584 
585 	if (csio_mb_issue(hw, mbp)) {
586 		csio_err(hw, "Issue of EQ OFLD cmd failed!\n");
587 		mempool_free(mbp, hw->mb_mempool);
588 		return -EINVAL;
589 	}
590 
591 	if (cbfn != NULL)
592 		return 0;
593 
594 	return csio_wr_eq_cfg_rsp(hw, mbp, eq_idx);
595 }
596 
597 /*
598  * csio_wr_iq_destroy_rsp - Response handler for IQ removal.
599  * @hw: The HW module.
600  * @mbp: Mailbox.
601  * @iq_idx: Ingress queue that was freed.
602  *
603  * Handle FW_IQ_CMD (free) mailbox completion.
604  */
605 static int
606 csio_wr_iq_destroy_rsp(struct csio_hw *hw, struct csio_mb *mbp, int iq_idx)
607 {
608 	enum fw_retval retval = csio_mb_fw_retval(mbp);
609 	int rv = 0;
610 
611 	if (retval != FW_SUCCESS)
612 		rv = -EINVAL;
613 
614 	mempool_free(mbp, hw->mb_mempool);
615 
616 	return rv;
617 }
618 
619 /*
620  * csio_wr_iq_destroy - Free an ingress queue.
621  * @hw: The HW module.
622  * @priv: Private data object.
623  * @iq_idx: Ingress queue index to destroy
624  * @cbfn: Completion callback.
625  *
626  * This API frees an ingress queue by issuing the FW_IQ_CMD
627  * with the free bit set.
628  */
629 static int
630 csio_wr_iq_destroy(struct csio_hw *hw, void *priv, int iq_idx,
631 		   void (*cbfn)(struct csio_hw *, struct csio_mb *))
632 {
633 	int rv = 0;
634 	struct csio_mb  *mbp;
635 	struct csio_iq_params iqp;
636 	int flq_idx;
637 
638 	memset(&iqp, 0, sizeof(struct csio_iq_params));
639 
640 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
641 	if (!mbp)
642 		return -ENOMEM;
643 
644 	iqp.pfn		= hw->pfn;
645 	iqp.vfn		= 0;
646 	iqp.iqid	= csio_q_iqid(hw, iq_idx);
647 	iqp.type	= FW_IQ_TYPE_FL_INT_CAP;
648 
649 	flq_idx = csio_q_iq_flq_idx(hw, iq_idx);
650 	if (flq_idx != -1)
651 		iqp.fl0id = csio_q_flid(hw, flq_idx);
652 	else
653 		iqp.fl0id = 0xFFFF;
654 
655 	iqp.fl1id = 0xFFFF;
656 
657 	csio_mb_iq_free(hw, mbp, priv, CSIO_MB_DEFAULT_TMO, &iqp, cbfn);
658 
659 	rv = csio_mb_issue(hw, mbp);
660 	if (rv != 0) {
661 		mempool_free(mbp, hw->mb_mempool);
662 		return rv;
663 	}
664 
665 	if (cbfn != NULL)
666 		return 0;
667 
668 	return csio_wr_iq_destroy_rsp(hw, mbp, iq_idx);
669 }
670 
671 /*
672  * csio_wr_eq_destroy_rsp - Response handler for OFLD EQ creation.
673  * @hw: The HW module.
674  * @mbp: Mailbox.
675  * @eq_idx: Egress queue that was freed.
676  *
677  * Handle FW_OFLD_EQ_CMD (free) mailbox completion.
678  */
679 static int
680 csio_wr_eq_destroy_rsp(struct csio_hw *hw, struct csio_mb *mbp, int eq_idx)
681 {
682 	enum fw_retval retval = csio_mb_fw_retval(mbp);
683 	int rv = 0;
684 
685 	if (retval != FW_SUCCESS)
686 		rv = -EINVAL;
687 
688 	mempool_free(mbp, hw->mb_mempool);
689 
690 	return rv;
691 }
692 
693 /*
694  * csio_wr_eq_destroy - Free an Egress queue.
695  * @hw: The HW module.
696  * @priv: Private data object.
697  * @eq_idx: Egress queue index to destroy
698  * @cbfn: Completion callback.
699  *
700  * This API frees an Egress queue by issuing the FW_EQ_OFLD_CMD
701  * with the free bit set.
702  */
703 static int
704 csio_wr_eq_destroy(struct csio_hw *hw, void *priv, int eq_idx,
705 		   void (*cbfn) (struct csio_hw *, struct csio_mb *))
706 {
707 	int rv = 0;
708 	struct csio_mb  *mbp;
709 	struct csio_eq_params eqp;
710 
711 	memset(&eqp, 0, sizeof(struct csio_eq_params));
712 
713 	mbp = mempool_alloc(hw->mb_mempool, GFP_ATOMIC);
714 	if (!mbp)
715 		return -ENOMEM;
716 
717 	eqp.pfn		= hw->pfn;
718 	eqp.vfn		= 0;
719 	eqp.eqid	= csio_q_eqid(hw, eq_idx);
720 
721 	csio_mb_eq_ofld_free(hw, mbp, priv, CSIO_MB_DEFAULT_TMO, &eqp, cbfn);
722 
723 	rv = csio_mb_issue(hw, mbp);
724 	if (rv != 0) {
725 		mempool_free(mbp, hw->mb_mempool);
726 		return rv;
727 	}
728 
729 	if (cbfn != NULL)
730 		return 0;
731 
732 	return csio_wr_eq_destroy_rsp(hw, mbp, eq_idx);
733 }
734 
735 /*
736  * csio_wr_cleanup_eq_stpg - Cleanup Egress queue status page
737  * @hw: HW module
738  * @qidx: Egress queue index
739  *
740  * Cleanup the Egress queue status page.
741  */
742 static void
743 csio_wr_cleanup_eq_stpg(struct csio_hw *hw, int qidx)
744 {
745 	struct csio_q	*q = csio_hw_to_wrm(hw)->q_arr[qidx];
746 	struct csio_qstatus_page *stp = (struct csio_qstatus_page *)q->vwrap;
747 
748 	memset(stp, 0, sizeof(*stp));
749 }
750 
751 /*
752  * csio_wr_cleanup_iq_ftr - Cleanup Footer entries in IQ
753  * @hw: HW module
754  * @qidx: Ingress queue index
755  *
756  * Cleanup the footer entries in the given ingress queue,
757  * set to 1 the internal copy of genbit.
758  */
759 static void
760 csio_wr_cleanup_iq_ftr(struct csio_hw *hw, int qidx)
761 {
762 	struct csio_wrm *wrm	= csio_hw_to_wrm(hw);
763 	struct csio_q	*q	= wrm->q_arr[qidx];
764 	void *wr;
765 	struct csio_iqwr_footer *ftr;
766 	uint32_t i = 0;
767 
768 	/* set to 1 since we are just about zero out genbit */
769 	q->un.iq.genbit = 1;
770 
771 	for (i = 0; i < q->credits; i++) {
772 		/* Get the WR */
773 		wr = (void *)((uintptr_t)q->vstart +
774 					   (i * q->wr_sz));
775 		/* Get the footer */
776 		ftr = (struct csio_iqwr_footer *)((uintptr_t)wr +
777 					  (q->wr_sz - sizeof(*ftr)));
778 		/* Zero out footer */
779 		memset(ftr, 0, sizeof(*ftr));
780 	}
781 }
782 
783 int
784 csio_wr_destroy_queues(struct csio_hw *hw, bool cmd)
785 {
786 	int i, flq_idx;
787 	struct csio_q *q;
788 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
789 	int rv;
790 
791 	for (i = 0; i < wrm->free_qidx; i++) {
792 		q = wrm->q_arr[i];
793 
794 		switch (q->type) {
795 		case CSIO_EGRESS:
796 			if (csio_q_eqid(hw, i) != CSIO_MAX_QID) {
797 				csio_wr_cleanup_eq_stpg(hw, i);
798 				if (!cmd) {
799 					csio_q_eqid(hw, i) = CSIO_MAX_QID;
800 					continue;
801 				}
802 
803 				rv = csio_wr_eq_destroy(hw, NULL, i, NULL);
804 				if ((rv == -EBUSY) || (rv == -ETIMEDOUT))
805 					cmd = false;
806 
807 				csio_q_eqid(hw, i) = CSIO_MAX_QID;
808 			}
809 		case CSIO_INGRESS:
810 			if (csio_q_iqid(hw, i) != CSIO_MAX_QID) {
811 				csio_wr_cleanup_iq_ftr(hw, i);
812 				if (!cmd) {
813 					csio_q_iqid(hw, i) = CSIO_MAX_QID;
814 					flq_idx = csio_q_iq_flq_idx(hw, i);
815 					if (flq_idx != -1)
816 						csio_q_flid(hw, flq_idx) =
817 								CSIO_MAX_QID;
818 					continue;
819 				}
820 
821 				rv = csio_wr_iq_destroy(hw, NULL, i, NULL);
822 				if ((rv == -EBUSY) || (rv == -ETIMEDOUT))
823 					cmd = false;
824 
825 				csio_q_iqid(hw, i) = CSIO_MAX_QID;
826 				flq_idx = csio_q_iq_flq_idx(hw, i);
827 				if (flq_idx != -1)
828 					csio_q_flid(hw, flq_idx) = CSIO_MAX_QID;
829 			}
830 		default:
831 			break;
832 		}
833 	}
834 
835 	hw->flags &= ~CSIO_HWF_Q_FW_ALLOCED;
836 
837 	return 0;
838 }
839 
840 /*
841  * csio_wr_get - Get requested size of WR entry/entries from queue.
842  * @hw: HW module.
843  * @qidx: Index of queue.
844  * @size: Cumulative size of Work request(s).
845  * @wrp: Work request pair.
846  *
847  * If requested credits are available, return the start address of the
848  * work request in the work request pair. Set pidx accordingly and
849  * return.
850  *
851  * NOTE about WR pair:
852  * ==================
853  * A WR can start towards the end of a queue, and then continue at the
854  * beginning, since the queue is considered to be circular. This will
855  * require a pair of address/size to be passed back to the caller -
856  * hence Work request pair format.
857  */
858 int
859 csio_wr_get(struct csio_hw *hw, int qidx, uint32_t size,
860 	    struct csio_wr_pair *wrp)
861 {
862 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
863 	struct csio_q *q = wrm->q_arr[qidx];
864 	void *cwr = (void *)((uintptr_t)(q->vstart) +
865 						(q->pidx * CSIO_QCREDIT_SZ));
866 	struct csio_qstatus_page *stp = (struct csio_qstatus_page *)q->vwrap;
867 	uint16_t cidx = q->cidx = ntohs(stp->cidx);
868 	uint16_t pidx = q->pidx;
869 	uint32_t req_sz	= ALIGN(size, CSIO_QCREDIT_SZ);
870 	int req_credits	= req_sz / CSIO_QCREDIT_SZ;
871 	int credits;
872 
873 	CSIO_DB_ASSERT(q->owner != NULL);
874 	CSIO_DB_ASSERT((qidx >= 0) && (qidx < wrm->free_qidx));
875 	CSIO_DB_ASSERT(cidx <= q->credits);
876 
877 	/* Calculate credits */
878 	if (pidx > cidx) {
879 		credits = q->credits - (pidx - cidx) - 1;
880 	} else if (cidx > pidx) {
881 		credits = cidx - pidx - 1;
882 	} else {
883 		/* cidx == pidx, empty queue */
884 		credits = q->credits;
885 		CSIO_INC_STATS(q, n_qempty);
886 	}
887 
888 	/*
889 	 * Check if we have enough credits.
890 	 * credits = 1 implies queue is full.
891 	 */
892 	if (!credits || (req_credits > credits)) {
893 		CSIO_INC_STATS(q, n_qfull);
894 		return -EBUSY;
895 	}
896 
897 	/*
898 	 * If we are here, we have enough credits to satisfy the
899 	 * request. Check if we are near the end of q, and if WR spills over.
900 	 * If it does, use the first addr/size to cover the queue until
901 	 * the end. Fit the remainder portion of the request at the top
902 	 * of queue and return it in the second addr/len. Set pidx
903 	 * accordingly.
904 	 */
905 	if (unlikely(((uintptr_t)cwr + req_sz) > (uintptr_t)(q->vwrap))) {
906 		wrp->addr1 = cwr;
907 		wrp->size1 = (uint32_t)((uintptr_t)q->vwrap - (uintptr_t)cwr);
908 		wrp->addr2 = q->vstart;
909 		wrp->size2 = req_sz - wrp->size1;
910 		q->pidx	= (uint16_t)(ALIGN(wrp->size2, CSIO_QCREDIT_SZ) /
911 							CSIO_QCREDIT_SZ);
912 		CSIO_INC_STATS(q, n_qwrap);
913 		CSIO_INC_STATS(q, n_eq_wr_split);
914 	} else {
915 		wrp->addr1 = cwr;
916 		wrp->size1 = req_sz;
917 		wrp->addr2 = NULL;
918 		wrp->size2 = 0;
919 		q->pidx	+= (uint16_t)req_credits;
920 
921 		/* We are the end of queue, roll back pidx to top of queue */
922 		if (unlikely(q->pidx == q->credits)) {
923 			q->pidx = 0;
924 			CSIO_INC_STATS(q, n_qwrap);
925 		}
926 	}
927 
928 	q->inc_idx = (uint16_t)req_credits;
929 
930 	CSIO_INC_STATS(q, n_tot_reqs);
931 
932 	return 0;
933 }
934 
935 /*
936  * csio_wr_copy_to_wrp - Copies given data into WR.
937  * @data_buf - Data buffer
938  * @wrp - Work request pair.
939  * @wr_off - Work request offset.
940  * @data_len - Data length.
941  *
942  * Copies the given data in Work Request. Work request pair(wrp) specifies
943  * address information of Work request.
944  * Returns: none
945  */
946 void
947 csio_wr_copy_to_wrp(void *data_buf, struct csio_wr_pair *wrp,
948 		   uint32_t wr_off, uint32_t data_len)
949 {
950 	uint32_t nbytes;
951 
952 	/* Number of space available in buffer addr1 of WRP */
953 	nbytes = ((wrp->size1 - wr_off) >= data_len) ?
954 					data_len : (wrp->size1 - wr_off);
955 
956 	memcpy((uint8_t *) wrp->addr1 + wr_off, data_buf, nbytes);
957 	data_len -= nbytes;
958 
959 	/* Write the remaining data from the begining of circular buffer */
960 	if (data_len) {
961 		CSIO_DB_ASSERT(data_len <= wrp->size2);
962 		CSIO_DB_ASSERT(wrp->addr2 != NULL);
963 		memcpy(wrp->addr2, (uint8_t *) data_buf + nbytes, data_len);
964 	}
965 }
966 
967 /*
968  * csio_wr_issue - Notify chip of Work request.
969  * @hw: HW module.
970  * @qidx: Index of queue.
971  * @prio: 0: Low priority, 1: High priority
972  *
973  * Rings the SGE Doorbell by writing the current producer index of the passed
974  * in queue into the register.
975  *
976  */
977 int
978 csio_wr_issue(struct csio_hw *hw, int qidx, bool prio)
979 {
980 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
981 	struct csio_q *q = wrm->q_arr[qidx];
982 
983 	CSIO_DB_ASSERT((qidx >= 0) && (qidx < wrm->free_qidx));
984 
985 	wmb();
986 	/* Ring SGE Doorbell writing q->pidx into it */
987 	csio_wr_reg32(hw, DBPRIO_V(prio) | QID_V(q->un.eq.physeqid) |
988 			  PIDX_T5_V(q->inc_idx) | DBTYPE_F,
989 			  MYPF_REG(SGE_PF_KDOORBELL_A));
990 	q->inc_idx = 0;
991 
992 	return 0;
993 }
994 
995 static inline uint32_t
996 csio_wr_avail_qcredits(struct csio_q *q)
997 {
998 	if (q->pidx > q->cidx)
999 		return q->pidx - q->cidx;
1000 	else if (q->cidx > q->pidx)
1001 		return q->credits - (q->cidx - q->pidx);
1002 	else
1003 		return 0;	/* cidx == pidx, empty queue */
1004 }
1005 
1006 /*
1007  * csio_wr_inval_flq_buf - Invalidate a free list buffer entry.
1008  * @hw: HW module.
1009  * @flq: The freelist queue.
1010  *
1011  * Invalidate the driver's version of a freelist buffer entry,
1012  * without freeing the associated the DMA memory. The entry
1013  * to be invalidated is picked up from the current Free list
1014  * queue cidx.
1015  *
1016  */
1017 static inline void
1018 csio_wr_inval_flq_buf(struct csio_hw *hw, struct csio_q *flq)
1019 {
1020 	flq->cidx++;
1021 	if (flq->cidx == flq->credits) {
1022 		flq->cidx = 0;
1023 		CSIO_INC_STATS(flq, n_qwrap);
1024 	}
1025 }
1026 
1027 /*
1028  * csio_wr_process_fl - Process a freelist completion.
1029  * @hw: HW module.
1030  * @q: The ingress queue attached to the Freelist.
1031  * @wr: The freelist completion WR in the ingress queue.
1032  * @len_to_qid: The lower 32-bits of the first flit of the RSP footer
1033  * @iq_handler: Caller's handler for this completion.
1034  * @priv: Private pointer of caller
1035  *
1036  */
1037 static inline void
1038 csio_wr_process_fl(struct csio_hw *hw, struct csio_q *q,
1039 		   void *wr, uint32_t len_to_qid,
1040 		   void (*iq_handler)(struct csio_hw *, void *,
1041 				      uint32_t, struct csio_fl_dma_buf *,
1042 				      void *),
1043 		   void *priv)
1044 {
1045 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1046 	struct csio_sge *sge = &wrm->sge;
1047 	struct csio_fl_dma_buf flb;
1048 	struct csio_dma_buf *buf, *fbuf;
1049 	uint32_t bufsz, len, lastlen = 0;
1050 	int n;
1051 	struct csio_q *flq = hw->wrm.q_arr[q->un.iq.flq_idx];
1052 
1053 	CSIO_DB_ASSERT(flq != NULL);
1054 
1055 	len = len_to_qid;
1056 
1057 	if (len & IQWRF_NEWBUF) {
1058 		if (flq->un.fl.offset > 0) {
1059 			csio_wr_inval_flq_buf(hw, flq);
1060 			flq->un.fl.offset = 0;
1061 		}
1062 		len = IQWRF_LEN_GET(len);
1063 	}
1064 
1065 	CSIO_DB_ASSERT(len != 0);
1066 
1067 	flb.totlen = len;
1068 
1069 	/* Consume all freelist buffers used for len bytes */
1070 	for (n = 0, fbuf = flb.flbufs; ; n++, fbuf++) {
1071 		buf = &flq->un.fl.bufs[flq->cidx];
1072 		bufsz = csio_wr_fl_bufsz(sge, buf);
1073 
1074 		fbuf->paddr	= buf->paddr;
1075 		fbuf->vaddr	= buf->vaddr;
1076 
1077 		flb.offset	= flq->un.fl.offset;
1078 		lastlen		= min(bufsz, len);
1079 		fbuf->len	= lastlen;
1080 
1081 		len -= lastlen;
1082 		if (!len)
1083 			break;
1084 		csio_wr_inval_flq_buf(hw, flq);
1085 	}
1086 
1087 	flb.defer_free = flq->un.fl.packen ? 0 : 1;
1088 
1089 	iq_handler(hw, wr, q->wr_sz - sizeof(struct csio_iqwr_footer),
1090 		   &flb, priv);
1091 
1092 	if (flq->un.fl.packen)
1093 		flq->un.fl.offset += ALIGN(lastlen, sge->csio_fl_align);
1094 	else
1095 		csio_wr_inval_flq_buf(hw, flq);
1096 
1097 }
1098 
1099 /*
1100  * csio_is_new_iqwr - Is this a new Ingress queue entry ?
1101  * @q: Ingress quueue.
1102  * @ftr: Ingress queue WR SGE footer.
1103  *
1104  * The entry is new if our generation bit matches the corresponding
1105  * bit in the footer of the current WR.
1106  */
1107 static inline bool
1108 csio_is_new_iqwr(struct csio_q *q, struct csio_iqwr_footer *ftr)
1109 {
1110 	return (q->un.iq.genbit == (ftr->u.type_gen >> IQWRF_GEN_SHIFT));
1111 }
1112 
1113 /*
1114  * csio_wr_process_iq - Process elements in Ingress queue.
1115  * @hw:  HW pointer
1116  * @qidx: Index of queue
1117  * @iq_handler: Handler for this queue
1118  * @priv: Caller's private pointer
1119  *
1120  * This routine walks through every entry of the ingress queue, calling
1121  * the provided iq_handler with the entry, until the generation bit
1122  * flips.
1123  */
1124 int
1125 csio_wr_process_iq(struct csio_hw *hw, struct csio_q *q,
1126 		   void (*iq_handler)(struct csio_hw *, void *,
1127 				      uint32_t, struct csio_fl_dma_buf *,
1128 				      void *),
1129 		   void *priv)
1130 {
1131 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1132 	void *wr = (void *)((uintptr_t)q->vstart + (q->cidx * q->wr_sz));
1133 	struct csio_iqwr_footer *ftr;
1134 	uint32_t wr_type, fw_qid, qid;
1135 	struct csio_q *q_completed;
1136 	struct csio_q *flq = csio_iq_has_fl(q) ?
1137 					wrm->q_arr[q->un.iq.flq_idx] : NULL;
1138 	int rv = 0;
1139 
1140 	/* Get the footer */
1141 	ftr = (struct csio_iqwr_footer *)((uintptr_t)wr +
1142 					  (q->wr_sz - sizeof(*ftr)));
1143 
1144 	/*
1145 	 * When q wrapped around last time, driver should have inverted
1146 	 * ic.genbit as well.
1147 	 */
1148 	while (csio_is_new_iqwr(q, ftr)) {
1149 
1150 		CSIO_DB_ASSERT(((uintptr_t)wr + q->wr_sz) <=
1151 						(uintptr_t)q->vwrap);
1152 		rmb();
1153 		wr_type = IQWRF_TYPE_GET(ftr->u.type_gen);
1154 
1155 		switch (wr_type) {
1156 		case X_RSPD_TYPE_CPL:
1157 			/* Subtract footer from WR len */
1158 			iq_handler(hw, wr, q->wr_sz - sizeof(*ftr), NULL, priv);
1159 			break;
1160 		case X_RSPD_TYPE_FLBUF:
1161 			csio_wr_process_fl(hw, q, wr,
1162 					   ntohl(ftr->pldbuflen_qid),
1163 					   iq_handler, priv);
1164 			break;
1165 		case X_RSPD_TYPE_INTR:
1166 			fw_qid = ntohl(ftr->pldbuflen_qid);
1167 			qid = fw_qid - wrm->fw_iq_start;
1168 			q_completed = hw->wrm.intr_map[qid];
1169 
1170 			if (unlikely(qid ==
1171 					csio_q_physiqid(hw, hw->intr_iq_idx))) {
1172 				/*
1173 				 * We are already in the Forward Interrupt
1174 				 * Interrupt Queue Service! Do-not service
1175 				 * again!
1176 				 *
1177 				 */
1178 			} else {
1179 				CSIO_DB_ASSERT(q_completed);
1180 				CSIO_DB_ASSERT(
1181 					q_completed->un.iq.iq_intx_handler);
1182 
1183 				/* Call the queue handler. */
1184 				q_completed->un.iq.iq_intx_handler(hw, NULL,
1185 						0, NULL, (void *)q_completed);
1186 			}
1187 			break;
1188 		default:
1189 			csio_warn(hw, "Unknown resp type 0x%x received\n",
1190 				 wr_type);
1191 			CSIO_INC_STATS(q, n_rsp_unknown);
1192 			break;
1193 		}
1194 
1195 		/*
1196 		 * Ingress *always* has fixed size WR entries. Therefore,
1197 		 * there should always be complete WRs towards the end of
1198 		 * queue.
1199 		 */
1200 		if (((uintptr_t)wr + q->wr_sz) == (uintptr_t)q->vwrap) {
1201 
1202 			/* Roll over to start of queue */
1203 			q->cidx = 0;
1204 			wr	= q->vstart;
1205 
1206 			/* Toggle genbit */
1207 			q->un.iq.genbit ^= 0x1;
1208 
1209 			CSIO_INC_STATS(q, n_qwrap);
1210 		} else {
1211 			q->cidx++;
1212 			wr	= (void *)((uintptr_t)(q->vstart) +
1213 					   (q->cidx * q->wr_sz));
1214 		}
1215 
1216 		ftr = (struct csio_iqwr_footer *)((uintptr_t)wr +
1217 						  (q->wr_sz - sizeof(*ftr)));
1218 		q->inc_idx++;
1219 
1220 	} /* while (q->un.iq.genbit == hdr->genbit) */
1221 
1222 	/*
1223 	 * We need to re-arm SGE interrupts in case we got a stray interrupt,
1224 	 * especially in msix mode. With INTx, this may be a common occurence.
1225 	 */
1226 	if (unlikely(!q->inc_idx)) {
1227 		CSIO_INC_STATS(q, n_stray_comp);
1228 		rv = -EINVAL;
1229 		goto restart;
1230 	}
1231 
1232 	/* Replenish free list buffers if pending falls below low water mark */
1233 	if (flq) {
1234 		uint32_t avail  = csio_wr_avail_qcredits(flq);
1235 		if (avail <= 16) {
1236 			/* Make sure in FLQ, atleast 1 credit (8 FL buffers)
1237 			 * remains unpopulated otherwise HW thinks
1238 			 * FLQ is empty.
1239 			 */
1240 			csio_wr_update_fl(hw, flq, (flq->credits - 8) - avail);
1241 			csio_wr_ring_fldb(hw, flq);
1242 		}
1243 	}
1244 
1245 restart:
1246 	/* Now inform SGE about our incremental index value */
1247 	csio_wr_reg32(hw, CIDXINC_V(q->inc_idx)		|
1248 			  INGRESSQID_V(q->un.iq.physiqid)	|
1249 			  TIMERREG_V(csio_sge_timer_reg),
1250 			  MYPF_REG(SGE_PF_GTS_A));
1251 	q->stats.n_tot_rsps += q->inc_idx;
1252 
1253 	q->inc_idx = 0;
1254 
1255 	return rv;
1256 }
1257 
1258 int
1259 csio_wr_process_iq_idx(struct csio_hw *hw, int qidx,
1260 		   void (*iq_handler)(struct csio_hw *, void *,
1261 				      uint32_t, struct csio_fl_dma_buf *,
1262 				      void *),
1263 		   void *priv)
1264 {
1265 	struct csio_wrm *wrm	= csio_hw_to_wrm(hw);
1266 	struct csio_q	*iq	= wrm->q_arr[qidx];
1267 
1268 	return csio_wr_process_iq(hw, iq, iq_handler, priv);
1269 }
1270 
1271 static int
1272 csio_closest_timer(struct csio_sge *s, int time)
1273 {
1274 	int i, delta, match = 0, min_delta = INT_MAX;
1275 
1276 	for (i = 0; i < ARRAY_SIZE(s->timer_val); i++) {
1277 		delta = time - s->timer_val[i];
1278 		if (delta < 0)
1279 			delta = -delta;
1280 		if (delta < min_delta) {
1281 			min_delta = delta;
1282 			match = i;
1283 		}
1284 	}
1285 	return match;
1286 }
1287 
1288 static int
1289 csio_closest_thresh(struct csio_sge *s, int cnt)
1290 {
1291 	int i, delta, match = 0, min_delta = INT_MAX;
1292 
1293 	for (i = 0; i < ARRAY_SIZE(s->counter_val); i++) {
1294 		delta = cnt - s->counter_val[i];
1295 		if (delta < 0)
1296 			delta = -delta;
1297 		if (delta < min_delta) {
1298 			min_delta = delta;
1299 			match = i;
1300 		}
1301 	}
1302 	return match;
1303 }
1304 
1305 static void
1306 csio_wr_fixup_host_params(struct csio_hw *hw)
1307 {
1308 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1309 	struct csio_sge *sge = &wrm->sge;
1310 	uint32_t clsz = L1_CACHE_BYTES;
1311 	uint32_t s_hps = PAGE_SHIFT - 10;
1312 	uint32_t ingpad = 0;
1313 	uint32_t stat_len = clsz > 64 ? 128 : 64;
1314 
1315 	csio_wr_reg32(hw, HOSTPAGESIZEPF0_V(s_hps) | HOSTPAGESIZEPF1_V(s_hps) |
1316 		      HOSTPAGESIZEPF2_V(s_hps) | HOSTPAGESIZEPF3_V(s_hps) |
1317 		      HOSTPAGESIZEPF4_V(s_hps) | HOSTPAGESIZEPF5_V(s_hps) |
1318 		      HOSTPAGESIZEPF6_V(s_hps) | HOSTPAGESIZEPF7_V(s_hps),
1319 		      SGE_HOST_PAGE_SIZE_A);
1320 
1321 	sge->csio_fl_align = clsz < 32 ? 32 : clsz;
1322 	ingpad = ilog2(sge->csio_fl_align) - 5;
1323 
1324 	csio_set_reg_field(hw, SGE_CONTROL_A,
1325 			   INGPADBOUNDARY_V(INGPADBOUNDARY_M) |
1326 			   EGRSTATUSPAGESIZE_F,
1327 			   INGPADBOUNDARY_V(ingpad) |
1328 			   EGRSTATUSPAGESIZE_V(stat_len != 64));
1329 
1330 	/* FL BUFFER SIZE#0 is Page size i,e already aligned to cache line */
1331 	csio_wr_reg32(hw, PAGE_SIZE, SGE_FL_BUFFER_SIZE0_A);
1332 
1333 	/*
1334 	 * If using hard params, the following will get set correctly
1335 	 * in csio_wr_set_sge().
1336 	 */
1337 	if (hw->flags & CSIO_HWF_USING_SOFT_PARAMS) {
1338 		csio_wr_reg32(hw,
1339 			(csio_rd_reg32(hw, SGE_FL_BUFFER_SIZE2_A) +
1340 			sge->csio_fl_align - 1) & ~(sge->csio_fl_align - 1),
1341 			SGE_FL_BUFFER_SIZE2_A);
1342 		csio_wr_reg32(hw,
1343 			(csio_rd_reg32(hw, SGE_FL_BUFFER_SIZE3_A) +
1344 			sge->csio_fl_align - 1) & ~(sge->csio_fl_align - 1),
1345 			SGE_FL_BUFFER_SIZE3_A);
1346 	}
1347 
1348 	csio_wr_reg32(hw, HPZ0_V(PAGE_SHIFT - 12), ULP_RX_TDDP_PSZ_A);
1349 
1350 	/* default value of rx_dma_offset of the NIC driver */
1351 	csio_set_reg_field(hw, SGE_CONTROL_A,
1352 			   PKTSHIFT_V(PKTSHIFT_M),
1353 			   PKTSHIFT_V(CSIO_SGE_RX_DMA_OFFSET));
1354 
1355 	csio_hw_tp_wr_bits_indirect(hw, TP_INGRESS_CONFIG_A,
1356 				    CSUM_HAS_PSEUDO_HDR_F, 0);
1357 }
1358 
1359 static void
1360 csio_init_intr_coalesce_parms(struct csio_hw *hw)
1361 {
1362 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1363 	struct csio_sge *sge = &wrm->sge;
1364 
1365 	csio_sge_thresh_reg = csio_closest_thresh(sge, csio_intr_coalesce_cnt);
1366 	if (csio_intr_coalesce_cnt) {
1367 		csio_sge_thresh_reg = 0;
1368 		csio_sge_timer_reg = X_TIMERREG_RESTART_COUNTER;
1369 		return;
1370 	}
1371 
1372 	csio_sge_timer_reg = csio_closest_timer(sge, csio_intr_coalesce_time);
1373 }
1374 
1375 /*
1376  * csio_wr_get_sge - Get SGE register values.
1377  * @hw: HW module.
1378  *
1379  * Used by non-master functions and by master-functions relying on config file.
1380  */
1381 static void
1382 csio_wr_get_sge(struct csio_hw *hw)
1383 {
1384 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1385 	struct csio_sge *sge = &wrm->sge;
1386 	uint32_t ingpad;
1387 	int i;
1388 	u32 timer_value_0_and_1, timer_value_2_and_3, timer_value_4_and_5;
1389 	u32 ingress_rx_threshold;
1390 
1391 	sge->sge_control = csio_rd_reg32(hw, SGE_CONTROL_A);
1392 
1393 	ingpad = INGPADBOUNDARY_G(sge->sge_control);
1394 
1395 	switch (ingpad) {
1396 	case X_INGPCIEBOUNDARY_32B:
1397 		sge->csio_fl_align = 32; break;
1398 	case X_INGPCIEBOUNDARY_64B:
1399 		sge->csio_fl_align = 64; break;
1400 	case X_INGPCIEBOUNDARY_128B:
1401 		sge->csio_fl_align = 128; break;
1402 	case X_INGPCIEBOUNDARY_256B:
1403 		sge->csio_fl_align = 256; break;
1404 	case X_INGPCIEBOUNDARY_512B:
1405 		sge->csio_fl_align = 512; break;
1406 	case X_INGPCIEBOUNDARY_1024B:
1407 		sge->csio_fl_align = 1024; break;
1408 	case X_INGPCIEBOUNDARY_2048B:
1409 		sge->csio_fl_align = 2048; break;
1410 	case X_INGPCIEBOUNDARY_4096B:
1411 		sge->csio_fl_align = 4096; break;
1412 	}
1413 
1414 	for (i = 0; i < CSIO_SGE_FL_SIZE_REGS; i++)
1415 		csio_get_flbuf_size(hw, sge, i);
1416 
1417 	timer_value_0_and_1 = csio_rd_reg32(hw, SGE_TIMER_VALUE_0_AND_1_A);
1418 	timer_value_2_and_3 = csio_rd_reg32(hw, SGE_TIMER_VALUE_2_AND_3_A);
1419 	timer_value_4_and_5 = csio_rd_reg32(hw, SGE_TIMER_VALUE_4_AND_5_A);
1420 
1421 	sge->timer_val[0] = (uint16_t)csio_core_ticks_to_us(hw,
1422 					TIMERVALUE0_G(timer_value_0_and_1));
1423 	sge->timer_val[1] = (uint16_t)csio_core_ticks_to_us(hw,
1424 					TIMERVALUE1_G(timer_value_0_and_1));
1425 	sge->timer_val[2] = (uint16_t)csio_core_ticks_to_us(hw,
1426 					TIMERVALUE2_G(timer_value_2_and_3));
1427 	sge->timer_val[3] = (uint16_t)csio_core_ticks_to_us(hw,
1428 					TIMERVALUE3_G(timer_value_2_and_3));
1429 	sge->timer_val[4] = (uint16_t)csio_core_ticks_to_us(hw,
1430 					TIMERVALUE4_G(timer_value_4_and_5));
1431 	sge->timer_val[5] = (uint16_t)csio_core_ticks_to_us(hw,
1432 					TIMERVALUE5_G(timer_value_4_and_5));
1433 
1434 	ingress_rx_threshold = csio_rd_reg32(hw, SGE_INGRESS_RX_THRESHOLD_A);
1435 	sge->counter_val[0] = THRESHOLD_0_G(ingress_rx_threshold);
1436 	sge->counter_val[1] = THRESHOLD_1_G(ingress_rx_threshold);
1437 	sge->counter_val[2] = THRESHOLD_2_G(ingress_rx_threshold);
1438 	sge->counter_val[3] = THRESHOLD_3_G(ingress_rx_threshold);
1439 
1440 	csio_init_intr_coalesce_parms(hw);
1441 }
1442 
1443 /*
1444  * csio_wr_set_sge - Initialize SGE registers
1445  * @hw: HW module.
1446  *
1447  * Used by Master function to initialize SGE registers in the absence
1448  * of a config file.
1449  */
1450 static void
1451 csio_wr_set_sge(struct csio_hw *hw)
1452 {
1453 	struct csio_wrm *wrm = csio_hw_to_wrm(hw);
1454 	struct csio_sge *sge = &wrm->sge;
1455 	int i;
1456 
1457 	/*
1458 	 * Set up our basic SGE mode to deliver CPL messages to our Ingress
1459 	 * Queue and Packet Date to the Free List.
1460 	 */
1461 	csio_set_reg_field(hw, SGE_CONTROL_A, RXPKTCPLMODE_F, RXPKTCPLMODE_F);
1462 
1463 	sge->sge_control = csio_rd_reg32(hw, SGE_CONTROL_A);
1464 
1465 	/* sge->csio_fl_align is set up by csio_wr_fixup_host_params(). */
1466 
1467 	/*
1468 	 * Set up to drop DOORBELL writes when the DOORBELL FIFO overflows
1469 	 * and generate an interrupt when this occurs so we can recover.
1470 	 */
1471 	csio_set_reg_field(hw, SGE_DBFIFO_STATUS_A,
1472 			   LP_INT_THRESH_T5_V(LP_INT_THRESH_T5_M),
1473 			   LP_INT_THRESH_T5_V(CSIO_SGE_DBFIFO_INT_THRESH));
1474 	csio_set_reg_field(hw, SGE_DBFIFO_STATUS2_A,
1475 			   HP_INT_THRESH_T5_V(LP_INT_THRESH_T5_M),
1476 			   HP_INT_THRESH_T5_V(CSIO_SGE_DBFIFO_INT_THRESH));
1477 
1478 	csio_set_reg_field(hw, SGE_DOORBELL_CONTROL_A, ENABLE_DROP_F,
1479 			   ENABLE_DROP_F);
1480 
1481 	/* SGE_FL_BUFFER_SIZE0 is set up by csio_wr_fixup_host_params(). */
1482 
1483 	CSIO_SET_FLBUF_SIZE(hw, 1, CSIO_SGE_FLBUF_SIZE1);
1484 	csio_wr_reg32(hw, (CSIO_SGE_FLBUF_SIZE2 + sge->csio_fl_align - 1)
1485 		      & ~(sge->csio_fl_align - 1), SGE_FL_BUFFER_SIZE2_A);
1486 	csio_wr_reg32(hw, (CSIO_SGE_FLBUF_SIZE3 + sge->csio_fl_align - 1)
1487 		      & ~(sge->csio_fl_align - 1), SGE_FL_BUFFER_SIZE3_A);
1488 	CSIO_SET_FLBUF_SIZE(hw, 4, CSIO_SGE_FLBUF_SIZE4);
1489 	CSIO_SET_FLBUF_SIZE(hw, 5, CSIO_SGE_FLBUF_SIZE5);
1490 	CSIO_SET_FLBUF_SIZE(hw, 6, CSIO_SGE_FLBUF_SIZE6);
1491 	CSIO_SET_FLBUF_SIZE(hw, 7, CSIO_SGE_FLBUF_SIZE7);
1492 	CSIO_SET_FLBUF_SIZE(hw, 8, CSIO_SGE_FLBUF_SIZE8);
1493 
1494 	for (i = 0; i < CSIO_SGE_FL_SIZE_REGS; i++)
1495 		csio_get_flbuf_size(hw, sge, i);
1496 
1497 	/* Initialize interrupt coalescing attributes */
1498 	sge->timer_val[0] = CSIO_SGE_TIMER_VAL_0;
1499 	sge->timer_val[1] = CSIO_SGE_TIMER_VAL_1;
1500 	sge->timer_val[2] = CSIO_SGE_TIMER_VAL_2;
1501 	sge->timer_val[3] = CSIO_SGE_TIMER_VAL_3;
1502 	sge->timer_val[4] = CSIO_SGE_TIMER_VAL_4;
1503 	sge->timer_val[5] = CSIO_SGE_TIMER_VAL_5;
1504 
1505 	sge->counter_val[0] = CSIO_SGE_INT_CNT_VAL_0;
1506 	sge->counter_val[1] = CSIO_SGE_INT_CNT_VAL_1;
1507 	sge->counter_val[2] = CSIO_SGE_INT_CNT_VAL_2;
1508 	sge->counter_val[3] = CSIO_SGE_INT_CNT_VAL_3;
1509 
1510 	csio_wr_reg32(hw, THRESHOLD_0_V(sge->counter_val[0]) |
1511 		      THRESHOLD_1_V(sge->counter_val[1]) |
1512 		      THRESHOLD_2_V(sge->counter_val[2]) |
1513 		      THRESHOLD_3_V(sge->counter_val[3]),
1514 		      SGE_INGRESS_RX_THRESHOLD_A);
1515 
1516 	csio_wr_reg32(hw,
1517 		   TIMERVALUE0_V(csio_us_to_core_ticks(hw, sge->timer_val[0])) |
1518 		   TIMERVALUE1_V(csio_us_to_core_ticks(hw, sge->timer_val[1])),
1519 		   SGE_TIMER_VALUE_0_AND_1_A);
1520 
1521 	csio_wr_reg32(hw,
1522 		   TIMERVALUE2_V(csio_us_to_core_ticks(hw, sge->timer_val[2])) |
1523 		   TIMERVALUE3_V(csio_us_to_core_ticks(hw, sge->timer_val[3])),
1524 		   SGE_TIMER_VALUE_2_AND_3_A);
1525 
1526 	csio_wr_reg32(hw,
1527 		   TIMERVALUE4_V(csio_us_to_core_ticks(hw, sge->timer_val[4])) |
1528 		   TIMERVALUE5_V(csio_us_to_core_ticks(hw, sge->timer_val[5])),
1529 		   SGE_TIMER_VALUE_4_AND_5_A);
1530 
1531 	csio_init_intr_coalesce_parms(hw);
1532 }
1533 
1534 void
1535 csio_wr_sge_init(struct csio_hw *hw)
1536 {
1537 	/*
1538 	 * If we are master and chip is not initialized:
1539 	 *    - If we plan to use the config file, we need to fixup some
1540 	 *      host specific registers, and read the rest of the SGE
1541 	 *      configuration.
1542 	 *    - If we dont plan to use the config file, we need to initialize
1543 	 *      SGE entirely, including fixing the host specific registers.
1544 	 * If we are master and chip is initialized, just read and work off of
1545 	 *	the already initialized SGE values.
1546 	 * If we arent the master, we are only allowed to read and work off of
1547 	 *      the already initialized SGE values.
1548 	 *
1549 	 * Therefore, before calling this function, we assume that the master-
1550 	 * ship of the card, state and whether to use config file or not, have
1551 	 * already been decided.
1552 	 */
1553 	if (csio_is_hw_master(hw)) {
1554 		if (hw->fw_state != CSIO_DEV_STATE_INIT)
1555 			csio_wr_fixup_host_params(hw);
1556 
1557 		if (hw->flags & CSIO_HWF_USING_SOFT_PARAMS)
1558 			csio_wr_get_sge(hw);
1559 		else
1560 			csio_wr_set_sge(hw);
1561 	} else
1562 		csio_wr_get_sge(hw);
1563 }
1564 
1565 /*
1566  * csio_wrm_init - Initialize Work request module.
1567  * @wrm: WR module
1568  * @hw: HW pointer
1569  *
1570  * Allocates memory for an array of queue pointers starting at q_arr.
1571  */
1572 int
1573 csio_wrm_init(struct csio_wrm *wrm, struct csio_hw *hw)
1574 {
1575 	int i;
1576 
1577 	if (!wrm->num_q) {
1578 		csio_err(hw, "Num queues is not set\n");
1579 		return -EINVAL;
1580 	}
1581 
1582 	wrm->q_arr = kcalloc(wrm->num_q, sizeof(struct csio_q *), GFP_KERNEL);
1583 	if (!wrm->q_arr)
1584 		goto err;
1585 
1586 	for (i = 0; i < wrm->num_q; i++) {
1587 		wrm->q_arr[i] = kzalloc(sizeof(struct csio_q), GFP_KERNEL);
1588 		if (!wrm->q_arr[i]) {
1589 			while (--i >= 0)
1590 				kfree(wrm->q_arr[i]);
1591 			goto err_free_arr;
1592 		}
1593 	}
1594 	wrm->free_qidx	= 0;
1595 
1596 	return 0;
1597 
1598 err_free_arr:
1599 	kfree(wrm->q_arr);
1600 err:
1601 	return -ENOMEM;
1602 }
1603 
1604 /*
1605  * csio_wrm_exit - Initialize Work request module.
1606  * @wrm: WR module
1607  * @hw: HW module
1608  *
1609  * Uninitialize WR module. Free q_arr and pointers in it.
1610  * We have the additional job of freeing the DMA memory associated
1611  * with the queues.
1612  */
1613 void
1614 csio_wrm_exit(struct csio_wrm *wrm, struct csio_hw *hw)
1615 {
1616 	int i;
1617 	uint32_t j;
1618 	struct csio_q *q;
1619 	struct csio_dma_buf *buf;
1620 
1621 	for (i = 0; i < wrm->num_q; i++) {
1622 		q = wrm->q_arr[i];
1623 
1624 		if (wrm->free_qidx && (i < wrm->free_qidx)) {
1625 			if (q->type == CSIO_FREELIST) {
1626 				if (!q->un.fl.bufs)
1627 					continue;
1628 				for (j = 0; j < q->credits; j++) {
1629 					buf = &q->un.fl.bufs[j];
1630 					if (!buf->vaddr)
1631 						continue;
1632 					pci_free_consistent(hw->pdev, buf->len,
1633 							    buf->vaddr,
1634 							    buf->paddr);
1635 				}
1636 				kfree(q->un.fl.bufs);
1637 			}
1638 			pci_free_consistent(hw->pdev, q->size,
1639 					    q->vstart, q->pstart);
1640 		}
1641 		kfree(q);
1642 	}
1643 
1644 	hw->flags &= ~CSIO_HWF_Q_MEM_ALLOCED;
1645 
1646 	kfree(wrm->q_arr);
1647 }
1648