xref: /openbmc/linux/drivers/net/wireless/intel/iwlwifi/pcie/tx.c (revision 2eb3ed33e55d003d721d4d1a5e72fe323c12b4c0)
1 /******************************************************************************
2  *
3  * Copyright(c) 2003 - 2014 Intel Corporation. All rights reserved.
4  * Copyright(c) 2013 - 2015 Intel Mobile Communications GmbH
5  * Copyright(c) 2016 - 2017 Intel Deutschland GmbH
6  *
7  * Portions of this file are derived from the ipw3945 project, as well
8  * as portions of the ieee80211 subsystem header files.
9  *
10  * This program is free software; you can redistribute it and/or modify it
11  * under the terms of version 2 of the GNU General Public License as
12  * published by the Free Software Foundation.
13  *
14  * This program is distributed in the hope that it will be useful, but WITHOUT
15  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
16  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
17  * more details.
18  *
19  * You should have received a copy of the GNU General Public License along with
20  * this program; if not, write to the Free Software Foundation, Inc.,
21  * 51 Franklin Street, Fifth Floor, Boston, MA 02110, USA
22  *
23  * The full GNU General Public License is included in this distribution in the
24  * file called LICENSE.
25  *
26  * Contact Information:
27  *  Intel Linux Wireless <linuxwifi@intel.com>
28  * Intel Corporation, 5200 N.E. Elam Young Parkway, Hillsboro, OR 97124-6497
29  *
30  *****************************************************************************/
31 #include <linux/etherdevice.h>
32 #include <linux/ieee80211.h>
33 #include <linux/slab.h>
34 #include <linux/sched.h>
35 #include <linux/pm_runtime.h>
36 #include <net/ip6_checksum.h>
37 #include <net/tso.h>
38 
39 #include "iwl-debug.h"
40 #include "iwl-csr.h"
41 #include "iwl-prph.h"
42 #include "iwl-io.h"
43 #include "iwl-scd.h"
44 #include "iwl-op-mode.h"
45 #include "internal.h"
46 #include "fw/api/tx.h"
47 
48 #define IWL_TX_CRC_SIZE 4
49 #define IWL_TX_DELIMITER_SIZE 4
50 
51 /*************** DMA-QUEUE-GENERAL-FUNCTIONS  *****
52  * DMA services
53  *
54  * Theory of operation
55  *
56  * A Tx or Rx queue resides in host DRAM, and is comprised of a circular buffer
57  * of buffer descriptors, each of which points to one or more data buffers for
58  * the device to read from or fill.  Driver and device exchange status of each
59  * queue via "read" and "write" pointers.  Driver keeps minimum of 2 empty
60  * entries in each circular buffer, to protect against confusing empty and full
61  * queue states.
62  *
63  * The device reads or writes the data in the queues via the device's several
64  * DMA/FIFO channels.  Each queue is mapped to a single DMA channel.
65  *
66  * For Tx queue, there are low mark and high mark limits. If, after queuing
67  * the packet for Tx, free space become < low mark, Tx queue stopped. When
68  * reclaiming packets (on 'tx done IRQ), if free space become > high mark,
69  * Tx queue resumed.
70  *
71  ***************************************************/
72 
73 int iwl_queue_space(const struct iwl_txq *q)
74 {
75 	unsigned int max;
76 	unsigned int used;
77 
78 	/*
79 	 * To avoid ambiguity between empty and completely full queues, there
80 	 * should always be less than TFD_QUEUE_SIZE_MAX elements in the queue.
81 	 * If q->n_window is smaller than TFD_QUEUE_SIZE_MAX, there is no need
82 	 * to reserve any queue entries for this purpose.
83 	 */
84 	if (q->n_window < TFD_QUEUE_SIZE_MAX)
85 		max = q->n_window;
86 	else
87 		max = TFD_QUEUE_SIZE_MAX - 1;
88 
89 	/*
90 	 * TFD_QUEUE_SIZE_MAX is a power of 2, so the following is equivalent to
91 	 * modulo by TFD_QUEUE_SIZE_MAX and is well defined.
92 	 */
93 	used = (q->write_ptr - q->read_ptr) & (TFD_QUEUE_SIZE_MAX - 1);
94 
95 	if (WARN_ON(used > max))
96 		return 0;
97 
98 	return max - used;
99 }
100 
101 /*
102  * iwl_queue_init - Initialize queue's high/low-water and read/write indexes
103  */
104 static int iwl_queue_init(struct iwl_txq *q, int slots_num)
105 {
106 	q->n_window = slots_num;
107 
108 	/* slots_num must be power-of-two size, otherwise
109 	 * iwl_pcie_get_cmd_index is broken. */
110 	if (WARN_ON(!is_power_of_2(slots_num)))
111 		return -EINVAL;
112 
113 	q->low_mark = q->n_window / 4;
114 	if (q->low_mark < 4)
115 		q->low_mark = 4;
116 
117 	q->high_mark = q->n_window / 8;
118 	if (q->high_mark < 2)
119 		q->high_mark = 2;
120 
121 	q->write_ptr = 0;
122 	q->read_ptr = 0;
123 
124 	return 0;
125 }
126 
127 int iwl_pcie_alloc_dma_ptr(struct iwl_trans *trans,
128 			   struct iwl_dma_ptr *ptr, size_t size)
129 {
130 	if (WARN_ON(ptr->addr))
131 		return -EINVAL;
132 
133 	ptr->addr = dma_alloc_coherent(trans->dev, size,
134 				       &ptr->dma, GFP_KERNEL);
135 	if (!ptr->addr)
136 		return -ENOMEM;
137 	ptr->size = size;
138 	return 0;
139 }
140 
141 void iwl_pcie_free_dma_ptr(struct iwl_trans *trans, struct iwl_dma_ptr *ptr)
142 {
143 	if (unlikely(!ptr->addr))
144 		return;
145 
146 	dma_free_coherent(trans->dev, ptr->size, ptr->addr, ptr->dma);
147 	memset(ptr, 0, sizeof(*ptr));
148 }
149 
150 static void iwl_pcie_txq_stuck_timer(unsigned long data)
151 {
152 	struct iwl_txq *txq = (void *)data;
153 	struct iwl_trans_pcie *trans_pcie = txq->trans_pcie;
154 	struct iwl_trans *trans = iwl_trans_pcie_get_trans(trans_pcie);
155 
156 	spin_lock(&txq->lock);
157 	/* check if triggered erroneously */
158 	if (txq->read_ptr == txq->write_ptr) {
159 		spin_unlock(&txq->lock);
160 		return;
161 	}
162 	spin_unlock(&txq->lock);
163 
164 	iwl_trans_pcie_log_scd_error(trans, txq);
165 
166 	iwl_force_nmi(trans);
167 }
168 
169 /*
170  * iwl_pcie_txq_update_byte_cnt_tbl - Set up entry in Tx byte-count array
171  */
172 static void iwl_pcie_txq_update_byte_cnt_tbl(struct iwl_trans *trans,
173 					     struct iwl_txq *txq, u16 byte_cnt,
174 					     int num_tbs)
175 {
176 	struct iwlagn_scd_bc_tbl *scd_bc_tbl;
177 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
178 	int write_ptr = txq->write_ptr;
179 	int txq_id = txq->id;
180 	u8 sec_ctl = 0;
181 	u16 len = byte_cnt + IWL_TX_CRC_SIZE + IWL_TX_DELIMITER_SIZE;
182 	__le16 bc_ent;
183 	struct iwl_tx_cmd *tx_cmd =
184 		(void *)txq->entries[txq->write_ptr].cmd->payload;
185 	u8 sta_id = tx_cmd->sta_id;
186 
187 	scd_bc_tbl = trans_pcie->scd_bc_tbls.addr;
188 
189 	sec_ctl = tx_cmd->sec_ctl;
190 
191 	switch (sec_ctl & TX_CMD_SEC_MSK) {
192 	case TX_CMD_SEC_CCM:
193 		len += IEEE80211_CCMP_MIC_LEN;
194 		break;
195 	case TX_CMD_SEC_TKIP:
196 		len += IEEE80211_TKIP_ICV_LEN;
197 		break;
198 	case TX_CMD_SEC_WEP:
199 		len += IEEE80211_WEP_IV_LEN + IEEE80211_WEP_ICV_LEN;
200 		break;
201 	}
202 	if (trans_pcie->bc_table_dword)
203 		len = DIV_ROUND_UP(len, 4);
204 
205 	if (WARN_ON(len > 0xFFF || write_ptr >= TFD_QUEUE_SIZE_MAX))
206 		return;
207 
208 	bc_ent = cpu_to_le16(len | (sta_id << 12));
209 
210 	scd_bc_tbl[txq_id].tfd_offset[write_ptr] = bc_ent;
211 
212 	if (write_ptr < TFD_QUEUE_SIZE_BC_DUP)
213 		scd_bc_tbl[txq_id].
214 			tfd_offset[TFD_QUEUE_SIZE_MAX + write_ptr] = bc_ent;
215 }
216 
217 static void iwl_pcie_txq_inval_byte_cnt_tbl(struct iwl_trans *trans,
218 					    struct iwl_txq *txq)
219 {
220 	struct iwl_trans_pcie *trans_pcie =
221 		IWL_TRANS_GET_PCIE_TRANS(trans);
222 	struct iwlagn_scd_bc_tbl *scd_bc_tbl = trans_pcie->scd_bc_tbls.addr;
223 	int txq_id = txq->id;
224 	int read_ptr = txq->read_ptr;
225 	u8 sta_id = 0;
226 	__le16 bc_ent;
227 	struct iwl_tx_cmd *tx_cmd =
228 		(void *)txq->entries[read_ptr].cmd->payload;
229 
230 	WARN_ON(read_ptr >= TFD_QUEUE_SIZE_MAX);
231 
232 	if (txq_id != trans_pcie->cmd_queue)
233 		sta_id = tx_cmd->sta_id;
234 
235 	bc_ent = cpu_to_le16(1 | (sta_id << 12));
236 
237 	scd_bc_tbl[txq_id].tfd_offset[read_ptr] = bc_ent;
238 
239 	if (read_ptr < TFD_QUEUE_SIZE_BC_DUP)
240 		scd_bc_tbl[txq_id].
241 			tfd_offset[TFD_QUEUE_SIZE_MAX + read_ptr] = bc_ent;
242 }
243 
244 /*
245  * iwl_pcie_txq_inc_wr_ptr - Send new write index to hardware
246  */
247 static void iwl_pcie_txq_inc_wr_ptr(struct iwl_trans *trans,
248 				    struct iwl_txq *txq)
249 {
250 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
251 	u32 reg = 0;
252 	int txq_id = txq->id;
253 
254 	lockdep_assert_held(&txq->lock);
255 
256 	/*
257 	 * explicitly wake up the NIC if:
258 	 * 1. shadow registers aren't enabled
259 	 * 2. NIC is woken up for CMD regardless of shadow outside this function
260 	 * 3. there is a chance that the NIC is asleep
261 	 */
262 	if (!trans->cfg->base_params->shadow_reg_enable &&
263 	    txq_id != trans_pcie->cmd_queue &&
264 	    test_bit(STATUS_TPOWER_PMI, &trans->status)) {
265 		/*
266 		 * wake up nic if it's powered down ...
267 		 * uCode will wake up, and interrupt us again, so next
268 		 * time we'll skip this part.
269 		 */
270 		reg = iwl_read32(trans, CSR_UCODE_DRV_GP1);
271 
272 		if (reg & CSR_UCODE_DRV_GP1_BIT_MAC_SLEEP) {
273 			IWL_DEBUG_INFO(trans, "Tx queue %d requesting wakeup, GP1 = 0x%x\n",
274 				       txq_id, reg);
275 			iwl_set_bit(trans, CSR_GP_CNTRL,
276 				    CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
277 			txq->need_update = true;
278 			return;
279 		}
280 	}
281 
282 	/*
283 	 * if not in power-save mode, uCode will never sleep when we're
284 	 * trying to tx (during RFKILL, we're not trying to tx).
285 	 */
286 	IWL_DEBUG_TX(trans, "Q:%d WR: 0x%x\n", txq_id, txq->write_ptr);
287 	if (!txq->block)
288 		iwl_write32(trans, HBUS_TARG_WRPTR,
289 			    txq->write_ptr | (txq_id << 8));
290 }
291 
292 void iwl_pcie_txq_check_wrptrs(struct iwl_trans *trans)
293 {
294 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
295 	int i;
296 
297 	for (i = 0; i < trans->cfg->base_params->num_of_queues; i++) {
298 		struct iwl_txq *txq = trans_pcie->txq[i];
299 
300 		if (!test_bit(i, trans_pcie->queue_used))
301 			continue;
302 
303 		spin_lock_bh(&txq->lock);
304 		if (txq->need_update) {
305 			iwl_pcie_txq_inc_wr_ptr(trans, txq);
306 			txq->need_update = false;
307 		}
308 		spin_unlock_bh(&txq->lock);
309 	}
310 }
311 
312 static inline dma_addr_t iwl_pcie_tfd_tb_get_addr(struct iwl_trans *trans,
313 						  void *_tfd, u8 idx)
314 {
315 
316 	if (trans->cfg->use_tfh) {
317 		struct iwl_tfh_tfd *tfd = _tfd;
318 		struct iwl_tfh_tb *tb = &tfd->tbs[idx];
319 
320 		return (dma_addr_t)(le64_to_cpu(tb->addr));
321 	} else {
322 		struct iwl_tfd *tfd = _tfd;
323 		struct iwl_tfd_tb *tb = &tfd->tbs[idx];
324 		dma_addr_t addr = get_unaligned_le32(&tb->lo);
325 		dma_addr_t hi_len;
326 
327 		if (sizeof(dma_addr_t) <= sizeof(u32))
328 			return addr;
329 
330 		hi_len = le16_to_cpu(tb->hi_n_len) & 0xF;
331 
332 		/*
333 		 * shift by 16 twice to avoid warnings on 32-bit
334 		 * (where this code never runs anyway due to the
335 		 * if statement above)
336 		 */
337 		return addr | ((hi_len << 16) << 16);
338 	}
339 }
340 
341 static inline void iwl_pcie_tfd_set_tb(struct iwl_trans *trans, void *tfd,
342 				       u8 idx, dma_addr_t addr, u16 len)
343 {
344 	struct iwl_tfd *tfd_fh = (void *)tfd;
345 	struct iwl_tfd_tb *tb = &tfd_fh->tbs[idx];
346 
347 	u16 hi_n_len = len << 4;
348 
349 	put_unaligned_le32(addr, &tb->lo);
350 	hi_n_len |= iwl_get_dma_hi_addr(addr);
351 
352 	tb->hi_n_len = cpu_to_le16(hi_n_len);
353 
354 	tfd_fh->num_tbs = idx + 1;
355 }
356 
357 static inline u8 iwl_pcie_tfd_get_num_tbs(struct iwl_trans *trans, void *_tfd)
358 {
359 	if (trans->cfg->use_tfh) {
360 		struct iwl_tfh_tfd *tfd = _tfd;
361 
362 		return le16_to_cpu(tfd->num_tbs) & 0x1f;
363 	} else {
364 		struct iwl_tfd *tfd = _tfd;
365 
366 		return tfd->num_tbs & 0x1f;
367 	}
368 }
369 
370 static void iwl_pcie_tfd_unmap(struct iwl_trans *trans,
371 			       struct iwl_cmd_meta *meta,
372 			       struct iwl_txq *txq, int index)
373 {
374 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
375 	int i, num_tbs;
376 	void *tfd = iwl_pcie_get_tfd(trans_pcie, txq, index);
377 
378 	/* Sanity check on number of chunks */
379 	num_tbs = iwl_pcie_tfd_get_num_tbs(trans, tfd);
380 
381 	if (num_tbs >= trans_pcie->max_tbs) {
382 		IWL_ERR(trans, "Too many chunks: %i\n", num_tbs);
383 		/* @todo issue fatal error, it is quite serious situation */
384 		return;
385 	}
386 
387 	/* first TB is never freed - it's the bidirectional DMA data */
388 
389 	for (i = 1; i < num_tbs; i++) {
390 		if (meta->tbs & BIT(i))
391 			dma_unmap_page(trans->dev,
392 				       iwl_pcie_tfd_tb_get_addr(trans, tfd, i),
393 				       iwl_pcie_tfd_tb_get_len(trans, tfd, i),
394 				       DMA_TO_DEVICE);
395 		else
396 			dma_unmap_single(trans->dev,
397 					 iwl_pcie_tfd_tb_get_addr(trans, tfd,
398 								  i),
399 					 iwl_pcie_tfd_tb_get_len(trans, tfd,
400 								 i),
401 					 DMA_TO_DEVICE);
402 	}
403 
404 	if (trans->cfg->use_tfh) {
405 		struct iwl_tfh_tfd *tfd_fh = (void *)tfd;
406 
407 		tfd_fh->num_tbs = 0;
408 	} else {
409 		struct iwl_tfd *tfd_fh = (void *)tfd;
410 
411 		tfd_fh->num_tbs = 0;
412 	}
413 
414 }
415 
416 /*
417  * iwl_pcie_txq_free_tfd - Free all chunks referenced by TFD [txq->q.read_ptr]
418  * @trans - transport private data
419  * @txq - tx queue
420  * @dma_dir - the direction of the DMA mapping
421  *
422  * Does NOT advance any TFD circular buffer read/write indexes
423  * Does NOT free the TFD itself (which is within circular buffer)
424  */
425 void iwl_pcie_txq_free_tfd(struct iwl_trans *trans, struct iwl_txq *txq)
426 {
427 	/* rd_ptr is bounded by TFD_QUEUE_SIZE_MAX and
428 	 * idx is bounded by n_window
429 	 */
430 	int rd_ptr = txq->read_ptr;
431 	int idx = iwl_pcie_get_cmd_index(txq, rd_ptr);
432 
433 	lockdep_assert_held(&txq->lock);
434 
435 	/* We have only q->n_window txq->entries, but we use
436 	 * TFD_QUEUE_SIZE_MAX tfds
437 	 */
438 	iwl_pcie_tfd_unmap(trans, &txq->entries[idx].meta, txq, rd_ptr);
439 
440 	/* free SKB */
441 	if (txq->entries) {
442 		struct sk_buff *skb;
443 
444 		skb = txq->entries[idx].skb;
445 
446 		/* Can be called from irqs-disabled context
447 		 * If skb is not NULL, it means that the whole queue is being
448 		 * freed and that the queue is not empty - free the skb
449 		 */
450 		if (skb) {
451 			iwl_op_mode_free_skb(trans->op_mode, skb);
452 			txq->entries[idx].skb = NULL;
453 		}
454 	}
455 }
456 
457 static int iwl_pcie_txq_build_tfd(struct iwl_trans *trans, struct iwl_txq *txq,
458 				  dma_addr_t addr, u16 len, bool reset)
459 {
460 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
461 	void *tfd;
462 	u32 num_tbs;
463 
464 	tfd = txq->tfds + trans_pcie->tfd_size * txq->write_ptr;
465 
466 	if (reset)
467 		memset(tfd, 0, trans_pcie->tfd_size);
468 
469 	num_tbs = iwl_pcie_tfd_get_num_tbs(trans, tfd);
470 
471 	/* Each TFD can point to a maximum max_tbs Tx buffers */
472 	if (num_tbs >= trans_pcie->max_tbs) {
473 		IWL_ERR(trans, "Error can not send more than %d chunks\n",
474 			trans_pcie->max_tbs);
475 		return -EINVAL;
476 	}
477 
478 	if (WARN(addr & ~IWL_TX_DMA_MASK,
479 		 "Unaligned address = %llx\n", (unsigned long long)addr))
480 		return -EINVAL;
481 
482 	iwl_pcie_tfd_set_tb(trans, tfd, num_tbs, addr, len);
483 
484 	return num_tbs;
485 }
486 
487 int iwl_pcie_txq_alloc(struct iwl_trans *trans, struct iwl_txq *txq,
488 		       int slots_num, bool cmd_queue)
489 {
490 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
491 	size_t tfd_sz = trans_pcie->tfd_size * TFD_QUEUE_SIZE_MAX;
492 	size_t tb0_buf_sz;
493 	int i;
494 
495 	if (WARN_ON(txq->entries || txq->tfds))
496 		return -EINVAL;
497 
498 	setup_timer(&txq->stuck_timer, iwl_pcie_txq_stuck_timer,
499 		    (unsigned long)txq);
500 	txq->trans_pcie = trans_pcie;
501 
502 	txq->n_window = slots_num;
503 
504 	txq->entries = kcalloc(slots_num,
505 			       sizeof(struct iwl_pcie_txq_entry),
506 			       GFP_KERNEL);
507 
508 	if (!txq->entries)
509 		goto error;
510 
511 	if (cmd_queue)
512 		for (i = 0; i < slots_num; i++) {
513 			txq->entries[i].cmd =
514 				kmalloc(sizeof(struct iwl_device_cmd),
515 					GFP_KERNEL);
516 			if (!txq->entries[i].cmd)
517 				goto error;
518 		}
519 
520 	/* Circular buffer of transmit frame descriptors (TFDs),
521 	 * shared with device */
522 	txq->tfds = dma_alloc_coherent(trans->dev, tfd_sz,
523 				       &txq->dma_addr, GFP_KERNEL);
524 	if (!txq->tfds)
525 		goto error;
526 
527 	BUILD_BUG_ON(IWL_FIRST_TB_SIZE_ALIGN != sizeof(*txq->first_tb_bufs));
528 
529 	tb0_buf_sz = sizeof(*txq->first_tb_bufs) * slots_num;
530 
531 	txq->first_tb_bufs = dma_alloc_coherent(trans->dev, tb0_buf_sz,
532 					      &txq->first_tb_dma,
533 					      GFP_KERNEL);
534 	if (!txq->first_tb_bufs)
535 		goto err_free_tfds;
536 
537 	return 0;
538 err_free_tfds:
539 	dma_free_coherent(trans->dev, tfd_sz, txq->tfds, txq->dma_addr);
540 error:
541 	if (txq->entries && cmd_queue)
542 		for (i = 0; i < slots_num; i++)
543 			kfree(txq->entries[i].cmd);
544 	kfree(txq->entries);
545 	txq->entries = NULL;
546 
547 	return -ENOMEM;
548 
549 }
550 
551 int iwl_pcie_txq_init(struct iwl_trans *trans, struct iwl_txq *txq,
552 		      int slots_num, bool cmd_queue)
553 {
554 	int ret;
555 
556 	txq->need_update = false;
557 
558 	/* TFD_QUEUE_SIZE_MAX must be power-of-two size, otherwise
559 	 * iwl_queue_inc_wrap and iwl_queue_dec_wrap are broken. */
560 	BUILD_BUG_ON(TFD_QUEUE_SIZE_MAX & (TFD_QUEUE_SIZE_MAX - 1));
561 
562 	/* Initialize queue's high/low-water marks, and head/tail indexes */
563 	ret = iwl_queue_init(txq, slots_num);
564 	if (ret)
565 		return ret;
566 
567 	spin_lock_init(&txq->lock);
568 
569 	if (cmd_queue) {
570 		static struct lock_class_key iwl_pcie_cmd_queue_lock_class;
571 
572 		lockdep_set_class(&txq->lock, &iwl_pcie_cmd_queue_lock_class);
573 	}
574 
575 	__skb_queue_head_init(&txq->overflow_q);
576 
577 	return 0;
578 }
579 
580 void iwl_pcie_free_tso_page(struct iwl_trans_pcie *trans_pcie,
581 			    struct sk_buff *skb)
582 {
583 	struct page **page_ptr;
584 
585 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
586 
587 	if (*page_ptr) {
588 		__free_page(*page_ptr);
589 		*page_ptr = NULL;
590 	}
591 }
592 
593 static void iwl_pcie_clear_cmd_in_flight(struct iwl_trans *trans)
594 {
595 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
596 
597 	lockdep_assert_held(&trans_pcie->reg_lock);
598 
599 	if (trans_pcie->ref_cmd_in_flight) {
600 		trans_pcie->ref_cmd_in_flight = false;
601 		IWL_DEBUG_RPM(trans, "clear ref_cmd_in_flight - unref\n");
602 		iwl_trans_unref(trans);
603 	}
604 
605 	if (!trans->cfg->base_params->apmg_wake_up_wa)
606 		return;
607 	if (WARN_ON(!trans_pcie->cmd_hold_nic_awake))
608 		return;
609 
610 	trans_pcie->cmd_hold_nic_awake = false;
611 	__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
612 				   CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
613 }
614 
615 /*
616  * iwl_pcie_txq_unmap -  Unmap any remaining DMA mappings and free skb's
617  */
618 static void iwl_pcie_txq_unmap(struct iwl_trans *trans, int txq_id)
619 {
620 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
621 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
622 
623 	spin_lock_bh(&txq->lock);
624 	while (txq->write_ptr != txq->read_ptr) {
625 		IWL_DEBUG_TX_REPLY(trans, "Q %d Free %d\n",
626 				   txq_id, txq->read_ptr);
627 
628 		if (txq_id != trans_pcie->cmd_queue) {
629 			struct sk_buff *skb = txq->entries[txq->read_ptr].skb;
630 
631 			if (WARN_ON_ONCE(!skb))
632 				continue;
633 
634 			iwl_pcie_free_tso_page(trans_pcie, skb);
635 		}
636 		iwl_pcie_txq_free_tfd(trans, txq);
637 		txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr);
638 
639 		if (txq->read_ptr == txq->write_ptr) {
640 			unsigned long flags;
641 
642 			spin_lock_irqsave(&trans_pcie->reg_lock, flags);
643 			if (txq_id != trans_pcie->cmd_queue) {
644 				IWL_DEBUG_RPM(trans, "Q %d - last tx freed\n",
645 					      txq->id);
646 				iwl_trans_unref(trans);
647 			} else {
648 				iwl_pcie_clear_cmd_in_flight(trans);
649 			}
650 			spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
651 		}
652 	}
653 
654 	while (!skb_queue_empty(&txq->overflow_q)) {
655 		struct sk_buff *skb = __skb_dequeue(&txq->overflow_q);
656 
657 		iwl_op_mode_free_skb(trans->op_mode, skb);
658 	}
659 
660 	spin_unlock_bh(&txq->lock);
661 
662 	/* just in case - this queue may have been stopped */
663 	iwl_wake_queue(trans, txq);
664 }
665 
666 /*
667  * iwl_pcie_txq_free - Deallocate DMA queue.
668  * @txq: Transmit queue to deallocate.
669  *
670  * Empty queue by removing and destroying all BD's.
671  * Free all buffers.
672  * 0-fill, but do not free "txq" descriptor structure.
673  */
674 static void iwl_pcie_txq_free(struct iwl_trans *trans, int txq_id)
675 {
676 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
677 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
678 	struct device *dev = trans->dev;
679 	int i;
680 
681 	if (WARN_ON(!txq))
682 		return;
683 
684 	iwl_pcie_txq_unmap(trans, txq_id);
685 
686 	/* De-alloc array of command/tx buffers */
687 	if (txq_id == trans_pcie->cmd_queue)
688 		for (i = 0; i < txq->n_window; i++) {
689 			kzfree(txq->entries[i].cmd);
690 			kzfree(txq->entries[i].free_buf);
691 		}
692 
693 	/* De-alloc circular buffer of TFDs */
694 	if (txq->tfds) {
695 		dma_free_coherent(dev,
696 				  trans_pcie->tfd_size * TFD_QUEUE_SIZE_MAX,
697 				  txq->tfds, txq->dma_addr);
698 		txq->dma_addr = 0;
699 		txq->tfds = NULL;
700 
701 		dma_free_coherent(dev,
702 				  sizeof(*txq->first_tb_bufs) * txq->n_window,
703 				  txq->first_tb_bufs, txq->first_tb_dma);
704 	}
705 
706 	kfree(txq->entries);
707 	txq->entries = NULL;
708 
709 	del_timer_sync(&txq->stuck_timer);
710 
711 	/* 0-fill queue descriptor structure */
712 	memset(txq, 0, sizeof(*txq));
713 }
714 
715 void iwl_pcie_tx_start(struct iwl_trans *trans, u32 scd_base_addr)
716 {
717 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
718 	int nq = trans->cfg->base_params->num_of_queues;
719 	int chan;
720 	u32 reg_val;
721 	int clear_dwords = (SCD_TRANS_TBL_OFFSET_QUEUE(nq) -
722 				SCD_CONTEXT_MEM_LOWER_BOUND) / sizeof(u32);
723 
724 	/* make sure all queue are not stopped/used */
725 	memset(trans_pcie->queue_stopped, 0, sizeof(trans_pcie->queue_stopped));
726 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
727 
728 	trans_pcie->scd_base_addr =
729 		iwl_read_prph(trans, SCD_SRAM_BASE_ADDR);
730 
731 	WARN_ON(scd_base_addr != 0 &&
732 		scd_base_addr != trans_pcie->scd_base_addr);
733 
734 	/* reset context data, TX status and translation data */
735 	iwl_trans_write_mem(trans, trans_pcie->scd_base_addr +
736 				   SCD_CONTEXT_MEM_LOWER_BOUND,
737 			    NULL, clear_dwords);
738 
739 	iwl_write_prph(trans, SCD_DRAM_BASE_ADDR,
740 		       trans_pcie->scd_bc_tbls.dma >> 10);
741 
742 	/* The chain extension of the SCD doesn't work well. This feature is
743 	 * enabled by default by the HW, so we need to disable it manually.
744 	 */
745 	if (trans->cfg->base_params->scd_chain_ext_wa)
746 		iwl_write_prph(trans, SCD_CHAINEXT_EN, 0);
747 
748 	iwl_trans_ac_txq_enable(trans, trans_pcie->cmd_queue,
749 				trans_pcie->cmd_fifo,
750 				trans_pcie->cmd_q_wdg_timeout);
751 
752 	/* Activate all Tx DMA/FIFO channels */
753 	iwl_scd_activate_fifos(trans);
754 
755 	/* Enable DMA channel */
756 	for (chan = 0; chan < FH_TCSR_CHNL_NUM; chan++)
757 		iwl_write_direct32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(chan),
758 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CHNL_ENABLE |
759 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CREDIT_ENABLE);
760 
761 	/* Update FH chicken bits */
762 	reg_val = iwl_read_direct32(trans, FH_TX_CHICKEN_BITS_REG);
763 	iwl_write_direct32(trans, FH_TX_CHICKEN_BITS_REG,
764 			   reg_val | FH_TX_CHICKEN_BITS_SCD_AUTO_RETRY_EN);
765 
766 	/* Enable L1-Active */
767 	if (trans->cfg->device_family < IWL_DEVICE_FAMILY_8000)
768 		iwl_clear_bits_prph(trans, APMG_PCIDEV_STT_REG,
769 				    APMG_PCIDEV_STT_VAL_L1_ACT_DIS);
770 }
771 
772 void iwl_trans_pcie_tx_reset(struct iwl_trans *trans)
773 {
774 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
775 	int txq_id;
776 
777 	/*
778 	 * we should never get here in gen2 trans mode return early to avoid
779 	 * having invalid accesses
780 	 */
781 	if (WARN_ON_ONCE(trans->cfg->gen2))
782 		return;
783 
784 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
785 	     txq_id++) {
786 		struct iwl_txq *txq = trans_pcie->txq[txq_id];
787 		if (trans->cfg->use_tfh)
788 			iwl_write_direct64(trans,
789 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
790 					   txq->dma_addr);
791 		else
792 			iwl_write_direct32(trans,
793 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
794 					   txq->dma_addr >> 8);
795 		iwl_pcie_txq_unmap(trans, txq_id);
796 		txq->read_ptr = 0;
797 		txq->write_ptr = 0;
798 	}
799 
800 	/* Tell NIC where to find the "keep warm" buffer */
801 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
802 			   trans_pcie->kw.dma >> 4);
803 
804 	/*
805 	 * Send 0 as the scd_base_addr since the device may have be reset
806 	 * while we were in WoWLAN in which case SCD_SRAM_BASE_ADDR will
807 	 * contain garbage.
808 	 */
809 	iwl_pcie_tx_start(trans, 0);
810 }
811 
812 static void iwl_pcie_tx_stop_fh(struct iwl_trans *trans)
813 {
814 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
815 	unsigned long flags;
816 	int ch, ret;
817 	u32 mask = 0;
818 
819 	spin_lock(&trans_pcie->irq_lock);
820 
821 	if (!iwl_trans_grab_nic_access(trans, &flags))
822 		goto out;
823 
824 	/* Stop each Tx DMA channel */
825 	for (ch = 0; ch < FH_TCSR_CHNL_NUM; ch++) {
826 		iwl_write32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(ch), 0x0);
827 		mask |= FH_TSSR_TX_STATUS_REG_MSK_CHNL_IDLE(ch);
828 	}
829 
830 	/* Wait for DMA channels to be idle */
831 	ret = iwl_poll_bit(trans, FH_TSSR_TX_STATUS_REG, mask, mask, 5000);
832 	if (ret < 0)
833 		IWL_ERR(trans,
834 			"Failing on timeout while stopping DMA channel %d [0x%08x]\n",
835 			ch, iwl_read32(trans, FH_TSSR_TX_STATUS_REG));
836 
837 	iwl_trans_release_nic_access(trans, &flags);
838 
839 out:
840 	spin_unlock(&trans_pcie->irq_lock);
841 }
842 
843 /*
844  * iwl_pcie_tx_stop - Stop all Tx DMA channels
845  */
846 int iwl_pcie_tx_stop(struct iwl_trans *trans)
847 {
848 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
849 	int txq_id;
850 
851 	/* Turn off all Tx DMA fifos */
852 	iwl_scd_deactivate_fifos(trans);
853 
854 	/* Turn off all Tx DMA channels */
855 	iwl_pcie_tx_stop_fh(trans);
856 
857 	/*
858 	 * This function can be called before the op_mode disabled the
859 	 * queues. This happens when we have an rfkill interrupt.
860 	 * Since we stop Tx altogether - mark the queues as stopped.
861 	 */
862 	memset(trans_pcie->queue_stopped, 0, sizeof(trans_pcie->queue_stopped));
863 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
864 
865 	/* This can happen: start_hw, stop_device */
866 	if (!trans_pcie->txq_memory)
867 		return 0;
868 
869 	/* Unmap DMA from host system and free skb's */
870 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
871 	     txq_id++)
872 		iwl_pcie_txq_unmap(trans, txq_id);
873 
874 	return 0;
875 }
876 
877 /*
878  * iwl_trans_tx_free - Free TXQ Context
879  *
880  * Destroy all TX DMA queues and structures
881  */
882 void iwl_pcie_tx_free(struct iwl_trans *trans)
883 {
884 	int txq_id;
885 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
886 
887 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
888 
889 	/* Tx queues */
890 	if (trans_pcie->txq_memory) {
891 		for (txq_id = 0;
892 		     txq_id < trans->cfg->base_params->num_of_queues;
893 		     txq_id++) {
894 			iwl_pcie_txq_free(trans, txq_id);
895 			trans_pcie->txq[txq_id] = NULL;
896 		}
897 	}
898 
899 	kfree(trans_pcie->txq_memory);
900 	trans_pcie->txq_memory = NULL;
901 
902 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->kw);
903 
904 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->scd_bc_tbls);
905 }
906 
907 /*
908  * iwl_pcie_tx_alloc - allocate TX context
909  * Allocate all Tx DMA structures and initialize them
910  */
911 static int iwl_pcie_tx_alloc(struct iwl_trans *trans)
912 {
913 	int ret;
914 	int txq_id, slots_num;
915 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
916 
917 	u16 scd_bc_tbls_size = trans->cfg->base_params->num_of_queues *
918 			sizeof(struct iwlagn_scd_bc_tbl);
919 
920 	/*It is not allowed to alloc twice, so warn when this happens.
921 	 * We cannot rely on the previous allocation, so free and fail */
922 	if (WARN_ON(trans_pcie->txq_memory)) {
923 		ret = -EINVAL;
924 		goto error;
925 	}
926 
927 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->scd_bc_tbls,
928 				   scd_bc_tbls_size);
929 	if (ret) {
930 		IWL_ERR(trans, "Scheduler BC Table allocation failed\n");
931 		goto error;
932 	}
933 
934 	/* Alloc keep-warm buffer */
935 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->kw, IWL_KW_SIZE);
936 	if (ret) {
937 		IWL_ERR(trans, "Keep Warm allocation failed\n");
938 		goto error;
939 	}
940 
941 	trans_pcie->txq_memory = kcalloc(trans->cfg->base_params->num_of_queues,
942 					 sizeof(struct iwl_txq), GFP_KERNEL);
943 	if (!trans_pcie->txq_memory) {
944 		IWL_ERR(trans, "Not enough memory for txq\n");
945 		ret = -ENOMEM;
946 		goto error;
947 	}
948 
949 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
950 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
951 	     txq_id++) {
952 		bool cmd_queue = (txq_id == trans_pcie->cmd_queue);
953 
954 		slots_num = cmd_queue ? trans_pcie->tx_cmd_queue_size :
955 			TFD_TX_CMD_SLOTS;
956 		trans_pcie->txq[txq_id] = &trans_pcie->txq_memory[txq_id];
957 		ret = iwl_pcie_txq_alloc(trans, trans_pcie->txq[txq_id],
958 					 slots_num, cmd_queue);
959 		if (ret) {
960 			IWL_ERR(trans, "Tx %d queue alloc failed\n", txq_id);
961 			goto error;
962 		}
963 		trans_pcie->txq[txq_id]->id = txq_id;
964 	}
965 
966 	return 0;
967 
968 error:
969 	iwl_pcie_tx_free(trans);
970 
971 	return ret;
972 }
973 
974 void iwl_pcie_set_tx_cmd_queue_size(struct iwl_trans *trans)
975 {
976 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
977 	int queue_size = TFD_CMD_SLOTS;
978 
979 	if (trans->cfg->tx_cmd_queue_size)
980 		queue_size = trans->cfg->tx_cmd_queue_size;
981 
982 	if (WARN_ON(!(is_power_of_2(queue_size) &&
983 		      TFD_QUEUE_CB_SIZE(queue_size) > 0)))
984 		trans_pcie->tx_cmd_queue_size = TFD_CMD_SLOTS;
985 	else
986 		trans_pcie->tx_cmd_queue_size = queue_size;
987 }
988 
989 int iwl_pcie_tx_init(struct iwl_trans *trans)
990 {
991 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
992 	int ret;
993 	int txq_id, slots_num;
994 	bool alloc = false;
995 
996 	iwl_pcie_set_tx_cmd_queue_size(trans);
997 
998 	if (!trans_pcie->txq_memory) {
999 		ret = iwl_pcie_tx_alloc(trans);
1000 		if (ret)
1001 			goto error;
1002 		alloc = true;
1003 	}
1004 
1005 	spin_lock(&trans_pcie->irq_lock);
1006 
1007 	/* Turn off all Tx DMA fifos */
1008 	iwl_scd_deactivate_fifos(trans);
1009 
1010 	/* Tell NIC where to find the "keep warm" buffer */
1011 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
1012 			   trans_pcie->kw.dma >> 4);
1013 
1014 	spin_unlock(&trans_pcie->irq_lock);
1015 
1016 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
1017 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
1018 	     txq_id++) {
1019 		bool cmd_queue = (txq_id == trans_pcie->cmd_queue);
1020 
1021 		slots_num = cmd_queue ? trans_pcie->tx_cmd_queue_size :
1022 			TFD_TX_CMD_SLOTS;
1023 		ret = iwl_pcie_txq_init(trans, trans_pcie->txq[txq_id],
1024 					slots_num, cmd_queue);
1025 		if (ret) {
1026 			IWL_ERR(trans, "Tx %d queue init failed\n", txq_id);
1027 			goto error;
1028 		}
1029 
1030 		/*
1031 		 * Tell nic where to find circular buffer of TFDs for a
1032 		 * given Tx queue, and enable the DMA channel used for that
1033 		 * queue.
1034 		 * Circular buffer (TFD queue in DRAM) physical base address
1035 		 */
1036 		iwl_write_direct32(trans, FH_MEM_CBBC_QUEUE(trans, txq_id),
1037 				   trans_pcie->txq[txq_id]->dma_addr >> 8);
1038 	}
1039 
1040 	iwl_set_bits_prph(trans, SCD_GP_CTRL, SCD_GP_CTRL_AUTO_ACTIVE_MODE);
1041 	if (trans->cfg->base_params->num_of_queues > 20)
1042 		iwl_set_bits_prph(trans, SCD_GP_CTRL,
1043 				  SCD_GP_CTRL_ENABLE_31_QUEUES);
1044 
1045 	return 0;
1046 error:
1047 	/*Upon error, free only if we allocated something */
1048 	if (alloc)
1049 		iwl_pcie_tx_free(trans);
1050 	return ret;
1051 }
1052 
1053 static inline void iwl_pcie_txq_progress(struct iwl_txq *txq)
1054 {
1055 	lockdep_assert_held(&txq->lock);
1056 
1057 	if (!txq->wd_timeout)
1058 		return;
1059 
1060 	/*
1061 	 * station is asleep and we send data - that must
1062 	 * be uAPSD or PS-Poll. Don't rearm the timer.
1063 	 */
1064 	if (txq->frozen)
1065 		return;
1066 
1067 	/*
1068 	 * if empty delete timer, otherwise move timer forward
1069 	 * since we're making progress on this queue
1070 	 */
1071 	if (txq->read_ptr == txq->write_ptr)
1072 		del_timer(&txq->stuck_timer);
1073 	else
1074 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1075 }
1076 
1077 /* Frees buffers until index _not_ inclusive */
1078 void iwl_trans_pcie_reclaim(struct iwl_trans *trans, int txq_id, int ssn,
1079 			    struct sk_buff_head *skbs)
1080 {
1081 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1082 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1083 	int tfd_num = ssn & (TFD_QUEUE_SIZE_MAX - 1);
1084 	int last_to_free;
1085 
1086 	/* This function is not meant to release cmd queue*/
1087 	if (WARN_ON(txq_id == trans_pcie->cmd_queue))
1088 		return;
1089 
1090 	spin_lock_bh(&txq->lock);
1091 
1092 	if (!test_bit(txq_id, trans_pcie->queue_used)) {
1093 		IWL_DEBUG_TX_QUEUES(trans, "Q %d inactive - ignoring idx %d\n",
1094 				    txq_id, ssn);
1095 		goto out;
1096 	}
1097 
1098 	if (txq->read_ptr == tfd_num)
1099 		goto out;
1100 
1101 	IWL_DEBUG_TX_REPLY(trans, "[Q %d] %d -> %d (%d)\n",
1102 			   txq_id, txq->read_ptr, tfd_num, ssn);
1103 
1104 	/*Since we free until index _not_ inclusive, the one before index is
1105 	 * the last we will free. This one must be used */
1106 	last_to_free = iwl_queue_dec_wrap(tfd_num);
1107 
1108 	if (!iwl_queue_used(txq, last_to_free)) {
1109 		IWL_ERR(trans,
1110 			"%s: Read index for DMA queue txq id (%d), last_to_free %d is out of range [0-%d] %d %d.\n",
1111 			__func__, txq_id, last_to_free, TFD_QUEUE_SIZE_MAX,
1112 			txq->write_ptr, txq->read_ptr);
1113 		goto out;
1114 	}
1115 
1116 	if (WARN_ON(!skb_queue_empty(skbs)))
1117 		goto out;
1118 
1119 	for (;
1120 	     txq->read_ptr != tfd_num;
1121 	     txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr)) {
1122 		int idx = iwl_pcie_get_cmd_index(txq, txq->read_ptr);
1123 		struct sk_buff *skb = txq->entries[idx].skb;
1124 
1125 		if (WARN_ON_ONCE(!skb))
1126 			continue;
1127 
1128 		iwl_pcie_free_tso_page(trans_pcie, skb);
1129 
1130 		__skb_queue_tail(skbs, skb);
1131 
1132 		txq->entries[idx].skb = NULL;
1133 
1134 		if (!trans->cfg->use_tfh)
1135 			iwl_pcie_txq_inval_byte_cnt_tbl(trans, txq);
1136 
1137 		iwl_pcie_txq_free_tfd(trans, txq);
1138 	}
1139 
1140 	iwl_pcie_txq_progress(txq);
1141 
1142 	if (iwl_queue_space(txq) > txq->low_mark &&
1143 	    test_bit(txq_id, trans_pcie->queue_stopped)) {
1144 		struct sk_buff_head overflow_skbs;
1145 
1146 		__skb_queue_head_init(&overflow_skbs);
1147 		skb_queue_splice_init(&txq->overflow_q, &overflow_skbs);
1148 
1149 		/*
1150 		 * This is tricky: we are in reclaim path which is non
1151 		 * re-entrant, so noone will try to take the access the
1152 		 * txq data from that path. We stopped tx, so we can't
1153 		 * have tx as well. Bottom line, we can unlock and re-lock
1154 		 * later.
1155 		 */
1156 		spin_unlock_bh(&txq->lock);
1157 
1158 		while (!skb_queue_empty(&overflow_skbs)) {
1159 			struct sk_buff *skb = __skb_dequeue(&overflow_skbs);
1160 			struct iwl_device_cmd *dev_cmd_ptr;
1161 
1162 			dev_cmd_ptr = *(void **)((u8 *)skb->cb +
1163 						 trans_pcie->dev_cmd_offs);
1164 
1165 			/*
1166 			 * Note that we can very well be overflowing again.
1167 			 * In that case, iwl_queue_space will be small again
1168 			 * and we won't wake mac80211's queue.
1169 			 */
1170 			iwl_trans_pcie_tx(trans, skb, dev_cmd_ptr, txq_id);
1171 		}
1172 		spin_lock_bh(&txq->lock);
1173 
1174 		if (iwl_queue_space(txq) > txq->low_mark)
1175 			iwl_wake_queue(trans, txq);
1176 	}
1177 
1178 	if (txq->read_ptr == txq->write_ptr) {
1179 		IWL_DEBUG_RPM(trans, "Q %d - last tx reclaimed\n", txq->id);
1180 		iwl_trans_unref(trans);
1181 	}
1182 
1183 out:
1184 	spin_unlock_bh(&txq->lock);
1185 }
1186 
1187 static int iwl_pcie_set_cmd_in_flight(struct iwl_trans *trans,
1188 				      const struct iwl_host_cmd *cmd)
1189 {
1190 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1191 	int ret;
1192 
1193 	lockdep_assert_held(&trans_pcie->reg_lock);
1194 
1195 	if (!(cmd->flags & CMD_SEND_IN_IDLE) &&
1196 	    !trans_pcie->ref_cmd_in_flight) {
1197 		trans_pcie->ref_cmd_in_flight = true;
1198 		IWL_DEBUG_RPM(trans, "set ref_cmd_in_flight - ref\n");
1199 		iwl_trans_ref(trans);
1200 	}
1201 
1202 	/*
1203 	 * wake up the NIC to make sure that the firmware will see the host
1204 	 * command - we will let the NIC sleep once all the host commands
1205 	 * returned. This needs to be done only on NICs that have
1206 	 * apmg_wake_up_wa set.
1207 	 */
1208 	if (trans->cfg->base_params->apmg_wake_up_wa &&
1209 	    !trans_pcie->cmd_hold_nic_awake) {
1210 		__iwl_trans_pcie_set_bit(trans, CSR_GP_CNTRL,
1211 					 CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1212 
1213 		ret = iwl_poll_bit(trans, CSR_GP_CNTRL,
1214 				   CSR_GP_CNTRL_REG_VAL_MAC_ACCESS_EN,
1215 				   (CSR_GP_CNTRL_REG_FLAG_MAC_CLOCK_READY |
1216 				    CSR_GP_CNTRL_REG_FLAG_GOING_TO_SLEEP),
1217 				   15000);
1218 		if (ret < 0) {
1219 			__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
1220 					CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1221 			IWL_ERR(trans, "Failed to wake NIC for hcmd\n");
1222 			return -EIO;
1223 		}
1224 		trans_pcie->cmd_hold_nic_awake = true;
1225 	}
1226 
1227 	return 0;
1228 }
1229 
1230 /*
1231  * iwl_pcie_cmdq_reclaim - Reclaim TX command queue entries already Tx'd
1232  *
1233  * When FW advances 'R' index, all entries between old and new 'R' index
1234  * need to be reclaimed. As result, some free space forms.  If there is
1235  * enough free space (> low mark), wake the stack that feeds us.
1236  */
1237 static void iwl_pcie_cmdq_reclaim(struct iwl_trans *trans, int txq_id, int idx)
1238 {
1239 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1240 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1241 	unsigned long flags;
1242 	int nfreed = 0;
1243 
1244 	lockdep_assert_held(&txq->lock);
1245 
1246 	if ((idx >= TFD_QUEUE_SIZE_MAX) || (!iwl_queue_used(txq, idx))) {
1247 		IWL_ERR(trans,
1248 			"%s: Read index for DMA queue txq id (%d), index %d is out of range [0-%d] %d %d.\n",
1249 			__func__, txq_id, idx, TFD_QUEUE_SIZE_MAX,
1250 			txq->write_ptr, txq->read_ptr);
1251 		return;
1252 	}
1253 
1254 	for (idx = iwl_queue_inc_wrap(idx); txq->read_ptr != idx;
1255 	     txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr)) {
1256 
1257 		if (nfreed++ > 0) {
1258 			IWL_ERR(trans, "HCMD skipped: index (%d) %d %d\n",
1259 				idx, txq->write_ptr, txq->read_ptr);
1260 			iwl_force_nmi(trans);
1261 		}
1262 	}
1263 
1264 	if (txq->read_ptr == txq->write_ptr) {
1265 		spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1266 		iwl_pcie_clear_cmd_in_flight(trans);
1267 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1268 	}
1269 
1270 	iwl_pcie_txq_progress(txq);
1271 }
1272 
1273 static int iwl_pcie_txq_set_ratid_map(struct iwl_trans *trans, u16 ra_tid,
1274 				 u16 txq_id)
1275 {
1276 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1277 	u32 tbl_dw_addr;
1278 	u32 tbl_dw;
1279 	u16 scd_q2ratid;
1280 
1281 	scd_q2ratid = ra_tid & SCD_QUEUE_RA_TID_MAP_RATID_MSK;
1282 
1283 	tbl_dw_addr = trans_pcie->scd_base_addr +
1284 			SCD_TRANS_TBL_OFFSET_QUEUE(txq_id);
1285 
1286 	tbl_dw = iwl_trans_read_mem32(trans, tbl_dw_addr);
1287 
1288 	if (txq_id & 0x1)
1289 		tbl_dw = (scd_q2ratid << 16) | (tbl_dw & 0x0000FFFF);
1290 	else
1291 		tbl_dw = scd_q2ratid | (tbl_dw & 0xFFFF0000);
1292 
1293 	iwl_trans_write_mem32(trans, tbl_dw_addr, tbl_dw);
1294 
1295 	return 0;
1296 }
1297 
1298 /* Receiver address (actually, Rx station's index into station table),
1299  * combined with Traffic ID (QOS priority), in format used by Tx Scheduler */
1300 #define BUILD_RAxTID(sta_id, tid)	(((sta_id) << 4) + (tid))
1301 
1302 bool iwl_trans_pcie_txq_enable(struct iwl_trans *trans, int txq_id, u16 ssn,
1303 			       const struct iwl_trans_txq_scd_cfg *cfg,
1304 			       unsigned int wdg_timeout)
1305 {
1306 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1307 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1308 	int fifo = -1;
1309 	bool scd_bug = false;
1310 
1311 	if (test_and_set_bit(txq_id, trans_pcie->queue_used))
1312 		WARN_ONCE(1, "queue %d already used - expect issues", txq_id);
1313 
1314 	txq->wd_timeout = msecs_to_jiffies(wdg_timeout);
1315 
1316 	if (cfg) {
1317 		fifo = cfg->fifo;
1318 
1319 		/* Disable the scheduler prior configuring the cmd queue */
1320 		if (txq_id == trans_pcie->cmd_queue &&
1321 		    trans_pcie->scd_set_active)
1322 			iwl_scd_enable_set_active(trans, 0);
1323 
1324 		/* Stop this Tx queue before configuring it */
1325 		iwl_scd_txq_set_inactive(trans, txq_id);
1326 
1327 		/* Set this queue as a chain-building queue unless it is CMD */
1328 		if (txq_id != trans_pcie->cmd_queue)
1329 			iwl_scd_txq_set_chain(trans, txq_id);
1330 
1331 		if (cfg->aggregate) {
1332 			u16 ra_tid = BUILD_RAxTID(cfg->sta_id, cfg->tid);
1333 
1334 			/* Map receiver-address / traffic-ID to this queue */
1335 			iwl_pcie_txq_set_ratid_map(trans, ra_tid, txq_id);
1336 
1337 			/* enable aggregations for the queue */
1338 			iwl_scd_txq_enable_agg(trans, txq_id);
1339 			txq->ampdu = true;
1340 		} else {
1341 			/*
1342 			 * disable aggregations for the queue, this will also
1343 			 * make the ra_tid mapping configuration irrelevant
1344 			 * since it is now a non-AGG queue.
1345 			 */
1346 			iwl_scd_txq_disable_agg(trans, txq_id);
1347 
1348 			ssn = txq->read_ptr;
1349 		}
1350 	} else {
1351 		/*
1352 		 * If we need to move the SCD write pointer by steps of
1353 		 * 0x40, 0x80 or 0xc0, it gets stuck. Avoids this and let
1354 		 * the op_mode know by returning true later.
1355 		 * Do this only in case cfg is NULL since this trick can
1356 		 * be done only if we have DQA enabled which is true for mvm
1357 		 * only. And mvm never sets a cfg pointer.
1358 		 * This is really ugly, but this is the easiest way out for
1359 		 * this sad hardware issue.
1360 		 * This bug has been fixed on devices 9000 and up.
1361 		 */
1362 		scd_bug = !trans->cfg->mq_rx_supported &&
1363 			!((ssn - txq->write_ptr) & 0x3f) &&
1364 			(ssn != txq->write_ptr);
1365 		if (scd_bug)
1366 			ssn++;
1367 	}
1368 
1369 	/* Place first TFD at index corresponding to start sequence number.
1370 	 * Assumes that ssn_idx is valid (!= 0xFFF) */
1371 	txq->read_ptr = (ssn & 0xff);
1372 	txq->write_ptr = (ssn & 0xff);
1373 	iwl_write_direct32(trans, HBUS_TARG_WRPTR,
1374 			   (ssn & 0xff) | (txq_id << 8));
1375 
1376 	if (cfg) {
1377 		u8 frame_limit = cfg->frame_limit;
1378 
1379 		iwl_write_prph(trans, SCD_QUEUE_RDPTR(txq_id), ssn);
1380 
1381 		/* Set up Tx window size and frame limit for this queue */
1382 		iwl_trans_write_mem32(trans, trans_pcie->scd_base_addr +
1383 				SCD_CONTEXT_QUEUE_OFFSET(txq_id), 0);
1384 		iwl_trans_write_mem32(trans,
1385 			trans_pcie->scd_base_addr +
1386 			SCD_CONTEXT_QUEUE_OFFSET(txq_id) + sizeof(u32),
1387 			SCD_QUEUE_CTX_REG2_VAL(WIN_SIZE, frame_limit) |
1388 			SCD_QUEUE_CTX_REG2_VAL(FRAME_LIMIT, frame_limit));
1389 
1390 		/* Set up status area in SRAM, map to Tx DMA/FIFO, activate */
1391 		iwl_write_prph(trans, SCD_QUEUE_STATUS_BITS(txq_id),
1392 			       (1 << SCD_QUEUE_STTS_REG_POS_ACTIVE) |
1393 			       (cfg->fifo << SCD_QUEUE_STTS_REG_POS_TXF) |
1394 			       (1 << SCD_QUEUE_STTS_REG_POS_WSL) |
1395 			       SCD_QUEUE_STTS_REG_MSK);
1396 
1397 		/* enable the scheduler for this queue (only) */
1398 		if (txq_id == trans_pcie->cmd_queue &&
1399 		    trans_pcie->scd_set_active)
1400 			iwl_scd_enable_set_active(trans, BIT(txq_id));
1401 
1402 		IWL_DEBUG_TX_QUEUES(trans,
1403 				    "Activate queue %d on FIFO %d WrPtr: %d\n",
1404 				    txq_id, fifo, ssn & 0xff);
1405 	} else {
1406 		IWL_DEBUG_TX_QUEUES(trans,
1407 				    "Activate queue %d WrPtr: %d\n",
1408 				    txq_id, ssn & 0xff);
1409 	}
1410 
1411 	return scd_bug;
1412 }
1413 
1414 void iwl_trans_pcie_txq_set_shared_mode(struct iwl_trans *trans, u32 txq_id,
1415 					bool shared_mode)
1416 {
1417 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1418 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1419 
1420 	txq->ampdu = !shared_mode;
1421 }
1422 
1423 void iwl_trans_pcie_txq_disable(struct iwl_trans *trans, int txq_id,
1424 				bool configure_scd)
1425 {
1426 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1427 	u32 stts_addr = trans_pcie->scd_base_addr +
1428 			SCD_TX_STTS_QUEUE_OFFSET(txq_id);
1429 	static const u32 zero_val[4] = {};
1430 
1431 	trans_pcie->txq[txq_id]->frozen_expiry_remainder = 0;
1432 	trans_pcie->txq[txq_id]->frozen = false;
1433 
1434 	/*
1435 	 * Upon HW Rfkill - we stop the device, and then stop the queues
1436 	 * in the op_mode. Just for the sake of the simplicity of the op_mode,
1437 	 * allow the op_mode to call txq_disable after it already called
1438 	 * stop_device.
1439 	 */
1440 	if (!test_and_clear_bit(txq_id, trans_pcie->queue_used)) {
1441 		WARN_ONCE(test_bit(STATUS_DEVICE_ENABLED, &trans->status),
1442 			  "queue %d not used", txq_id);
1443 		return;
1444 	}
1445 
1446 	if (configure_scd) {
1447 		iwl_scd_txq_set_inactive(trans, txq_id);
1448 
1449 		iwl_trans_write_mem(trans, stts_addr, (void *)zero_val,
1450 				    ARRAY_SIZE(zero_val));
1451 	}
1452 
1453 	iwl_pcie_txq_unmap(trans, txq_id);
1454 	trans_pcie->txq[txq_id]->ampdu = false;
1455 
1456 	IWL_DEBUG_TX_QUEUES(trans, "Deactivate queue %d\n", txq_id);
1457 }
1458 
1459 /*************** HOST COMMAND QUEUE FUNCTIONS   *****/
1460 
1461 /*
1462  * iwl_pcie_enqueue_hcmd - enqueue a uCode command
1463  * @priv: device private data point
1464  * @cmd: a pointer to the ucode command structure
1465  *
1466  * The function returns < 0 values to indicate the operation
1467  * failed. On success, it returns the index (>= 0) of command in the
1468  * command queue.
1469  */
1470 static int iwl_pcie_enqueue_hcmd(struct iwl_trans *trans,
1471 				 struct iwl_host_cmd *cmd)
1472 {
1473 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1474 	struct iwl_txq *txq = trans_pcie->txq[trans_pcie->cmd_queue];
1475 	struct iwl_device_cmd *out_cmd;
1476 	struct iwl_cmd_meta *out_meta;
1477 	unsigned long flags;
1478 	void *dup_buf = NULL;
1479 	dma_addr_t phys_addr;
1480 	int idx;
1481 	u16 copy_size, cmd_size, tb0_size;
1482 	bool had_nocopy = false;
1483 	u8 group_id = iwl_cmd_groupid(cmd->id);
1484 	int i, ret;
1485 	u32 cmd_pos;
1486 	const u8 *cmddata[IWL_MAX_CMD_TBS_PER_TFD];
1487 	u16 cmdlen[IWL_MAX_CMD_TBS_PER_TFD];
1488 
1489 	if (WARN(!trans->wide_cmd_header &&
1490 		 group_id > IWL_ALWAYS_LONG_GROUP,
1491 		 "unsupported wide command %#x\n", cmd->id))
1492 		return -EINVAL;
1493 
1494 	if (group_id != 0) {
1495 		copy_size = sizeof(struct iwl_cmd_header_wide);
1496 		cmd_size = sizeof(struct iwl_cmd_header_wide);
1497 	} else {
1498 		copy_size = sizeof(struct iwl_cmd_header);
1499 		cmd_size = sizeof(struct iwl_cmd_header);
1500 	}
1501 
1502 	/* need one for the header if the first is NOCOPY */
1503 	BUILD_BUG_ON(IWL_MAX_CMD_TBS_PER_TFD > IWL_NUM_OF_TBS - 1);
1504 
1505 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1506 		cmddata[i] = cmd->data[i];
1507 		cmdlen[i] = cmd->len[i];
1508 
1509 		if (!cmd->len[i])
1510 			continue;
1511 
1512 		/* need at least IWL_FIRST_TB_SIZE copied */
1513 		if (copy_size < IWL_FIRST_TB_SIZE) {
1514 			int copy = IWL_FIRST_TB_SIZE - copy_size;
1515 
1516 			if (copy > cmdlen[i])
1517 				copy = cmdlen[i];
1518 			cmdlen[i] -= copy;
1519 			cmddata[i] += copy;
1520 			copy_size += copy;
1521 		}
1522 
1523 		if (cmd->dataflags[i] & IWL_HCMD_DFL_NOCOPY) {
1524 			had_nocopy = true;
1525 			if (WARN_ON(cmd->dataflags[i] & IWL_HCMD_DFL_DUP)) {
1526 				idx = -EINVAL;
1527 				goto free_dup_buf;
1528 			}
1529 		} else if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP) {
1530 			/*
1531 			 * This is also a chunk that isn't copied
1532 			 * to the static buffer so set had_nocopy.
1533 			 */
1534 			had_nocopy = true;
1535 
1536 			/* only allowed once */
1537 			if (WARN_ON(dup_buf)) {
1538 				idx = -EINVAL;
1539 				goto free_dup_buf;
1540 			}
1541 
1542 			dup_buf = kmemdup(cmddata[i], cmdlen[i],
1543 					  GFP_ATOMIC);
1544 			if (!dup_buf)
1545 				return -ENOMEM;
1546 		} else {
1547 			/* NOCOPY must not be followed by normal! */
1548 			if (WARN_ON(had_nocopy)) {
1549 				idx = -EINVAL;
1550 				goto free_dup_buf;
1551 			}
1552 			copy_size += cmdlen[i];
1553 		}
1554 		cmd_size += cmd->len[i];
1555 	}
1556 
1557 	/*
1558 	 * If any of the command structures end up being larger than
1559 	 * the TFD_MAX_PAYLOAD_SIZE and they aren't dynamically
1560 	 * allocated into separate TFDs, then we will need to
1561 	 * increase the size of the buffers.
1562 	 */
1563 	if (WARN(copy_size > TFD_MAX_PAYLOAD_SIZE,
1564 		 "Command %s (%#x) is too large (%d bytes)\n",
1565 		 iwl_get_cmd_string(trans, cmd->id),
1566 		 cmd->id, copy_size)) {
1567 		idx = -EINVAL;
1568 		goto free_dup_buf;
1569 	}
1570 
1571 	spin_lock_bh(&txq->lock);
1572 
1573 	if (iwl_queue_space(txq) < ((cmd->flags & CMD_ASYNC) ? 2 : 1)) {
1574 		spin_unlock_bh(&txq->lock);
1575 
1576 		IWL_ERR(trans, "No space in command queue\n");
1577 		iwl_op_mode_cmd_queue_full(trans->op_mode);
1578 		idx = -ENOSPC;
1579 		goto free_dup_buf;
1580 	}
1581 
1582 	idx = iwl_pcie_get_cmd_index(txq, txq->write_ptr);
1583 	out_cmd = txq->entries[idx].cmd;
1584 	out_meta = &txq->entries[idx].meta;
1585 
1586 	memset(out_meta, 0, sizeof(*out_meta));	/* re-initialize to NULL */
1587 	if (cmd->flags & CMD_WANT_SKB)
1588 		out_meta->source = cmd;
1589 
1590 	/* set up the header */
1591 	if (group_id != 0) {
1592 		out_cmd->hdr_wide.cmd = iwl_cmd_opcode(cmd->id);
1593 		out_cmd->hdr_wide.group_id = group_id;
1594 		out_cmd->hdr_wide.version = iwl_cmd_version(cmd->id);
1595 		out_cmd->hdr_wide.length =
1596 			cpu_to_le16(cmd_size -
1597 				    sizeof(struct iwl_cmd_header_wide));
1598 		out_cmd->hdr_wide.reserved = 0;
1599 		out_cmd->hdr_wide.sequence =
1600 			cpu_to_le16(QUEUE_TO_SEQ(trans_pcie->cmd_queue) |
1601 						 INDEX_TO_SEQ(txq->write_ptr));
1602 
1603 		cmd_pos = sizeof(struct iwl_cmd_header_wide);
1604 		copy_size = sizeof(struct iwl_cmd_header_wide);
1605 	} else {
1606 		out_cmd->hdr.cmd = iwl_cmd_opcode(cmd->id);
1607 		out_cmd->hdr.sequence =
1608 			cpu_to_le16(QUEUE_TO_SEQ(trans_pcie->cmd_queue) |
1609 						 INDEX_TO_SEQ(txq->write_ptr));
1610 		out_cmd->hdr.group_id = 0;
1611 
1612 		cmd_pos = sizeof(struct iwl_cmd_header);
1613 		copy_size = sizeof(struct iwl_cmd_header);
1614 	}
1615 
1616 	/* and copy the data that needs to be copied */
1617 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1618 		int copy;
1619 
1620 		if (!cmd->len[i])
1621 			continue;
1622 
1623 		/* copy everything if not nocopy/dup */
1624 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1625 					   IWL_HCMD_DFL_DUP))) {
1626 			copy = cmd->len[i];
1627 
1628 			memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1629 			cmd_pos += copy;
1630 			copy_size += copy;
1631 			continue;
1632 		}
1633 
1634 		/*
1635 		 * Otherwise we need at least IWL_FIRST_TB_SIZE copied
1636 		 * in total (for bi-directional DMA), but copy up to what
1637 		 * we can fit into the payload for debug dump purposes.
1638 		 */
1639 		copy = min_t(int, TFD_MAX_PAYLOAD_SIZE - cmd_pos, cmd->len[i]);
1640 
1641 		memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1642 		cmd_pos += copy;
1643 
1644 		/* However, treat copy_size the proper way, we need it below */
1645 		if (copy_size < IWL_FIRST_TB_SIZE) {
1646 			copy = IWL_FIRST_TB_SIZE - copy_size;
1647 
1648 			if (copy > cmd->len[i])
1649 				copy = cmd->len[i];
1650 			copy_size += copy;
1651 		}
1652 	}
1653 
1654 	IWL_DEBUG_HC(trans,
1655 		     "Sending command %s (%.2x.%.2x), seq: 0x%04X, %d bytes at %d[%d]:%d\n",
1656 		     iwl_get_cmd_string(trans, cmd->id),
1657 		     group_id, out_cmd->hdr.cmd,
1658 		     le16_to_cpu(out_cmd->hdr.sequence),
1659 		     cmd_size, txq->write_ptr, idx, trans_pcie->cmd_queue);
1660 
1661 	/* start the TFD with the minimum copy bytes */
1662 	tb0_size = min_t(int, copy_size, IWL_FIRST_TB_SIZE);
1663 	memcpy(&txq->first_tb_bufs[idx], &out_cmd->hdr, tb0_size);
1664 	iwl_pcie_txq_build_tfd(trans, txq,
1665 			       iwl_pcie_get_first_tb_dma(txq, idx),
1666 			       tb0_size, true);
1667 
1668 	/* map first command fragment, if any remains */
1669 	if (copy_size > tb0_size) {
1670 		phys_addr = dma_map_single(trans->dev,
1671 					   ((u8 *)&out_cmd->hdr) + tb0_size,
1672 					   copy_size - tb0_size,
1673 					   DMA_TO_DEVICE);
1674 		if (dma_mapping_error(trans->dev, phys_addr)) {
1675 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1676 					   txq->write_ptr);
1677 			idx = -ENOMEM;
1678 			goto out;
1679 		}
1680 
1681 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr,
1682 				       copy_size - tb0_size, false);
1683 	}
1684 
1685 	/* map the remaining (adjusted) nocopy/dup fragments */
1686 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1687 		const void *data = cmddata[i];
1688 
1689 		if (!cmdlen[i])
1690 			continue;
1691 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1692 					   IWL_HCMD_DFL_DUP)))
1693 			continue;
1694 		if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP)
1695 			data = dup_buf;
1696 		phys_addr = dma_map_single(trans->dev, (void *)data,
1697 					   cmdlen[i], DMA_TO_DEVICE);
1698 		if (dma_mapping_error(trans->dev, phys_addr)) {
1699 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1700 					   txq->write_ptr);
1701 			idx = -ENOMEM;
1702 			goto out;
1703 		}
1704 
1705 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr, cmdlen[i], false);
1706 	}
1707 
1708 	BUILD_BUG_ON(IWL_TFH_NUM_TBS > sizeof(out_meta->tbs) * BITS_PER_BYTE);
1709 	out_meta->flags = cmd->flags;
1710 	if (WARN_ON_ONCE(txq->entries[idx].free_buf))
1711 		kzfree(txq->entries[idx].free_buf);
1712 	txq->entries[idx].free_buf = dup_buf;
1713 
1714 	trace_iwlwifi_dev_hcmd(trans->dev, cmd, cmd_size, &out_cmd->hdr_wide);
1715 
1716 	/* start timer if queue currently empty */
1717 	if (txq->read_ptr == txq->write_ptr && txq->wd_timeout)
1718 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1719 
1720 	spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1721 	ret = iwl_pcie_set_cmd_in_flight(trans, cmd);
1722 	if (ret < 0) {
1723 		idx = ret;
1724 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1725 		goto out;
1726 	}
1727 
1728 	/* Increment and update queue's write index */
1729 	txq->write_ptr = iwl_queue_inc_wrap(txq->write_ptr);
1730 	iwl_pcie_txq_inc_wr_ptr(trans, txq);
1731 
1732 	spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1733 
1734  out:
1735 	spin_unlock_bh(&txq->lock);
1736  free_dup_buf:
1737 	if (idx < 0)
1738 		kfree(dup_buf);
1739 	return idx;
1740 }
1741 
1742 /*
1743  * iwl_pcie_hcmd_complete - Pull unused buffers off the queue and reclaim them
1744  * @rxb: Rx buffer to reclaim
1745  */
1746 void iwl_pcie_hcmd_complete(struct iwl_trans *trans,
1747 			    struct iwl_rx_cmd_buffer *rxb)
1748 {
1749 	struct iwl_rx_packet *pkt = rxb_addr(rxb);
1750 	u16 sequence = le16_to_cpu(pkt->hdr.sequence);
1751 	u8 group_id;
1752 	u32 cmd_id;
1753 	int txq_id = SEQ_TO_QUEUE(sequence);
1754 	int index = SEQ_TO_INDEX(sequence);
1755 	int cmd_index;
1756 	struct iwl_device_cmd *cmd;
1757 	struct iwl_cmd_meta *meta;
1758 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1759 	struct iwl_txq *txq = trans_pcie->txq[trans_pcie->cmd_queue];
1760 
1761 	/* If a Tx command is being handled and it isn't in the actual
1762 	 * command queue then there a command routing bug has been introduced
1763 	 * in the queue management code. */
1764 	if (WARN(txq_id != trans_pcie->cmd_queue,
1765 		 "wrong command queue %d (should be %d), sequence 0x%X readp=%d writep=%d\n",
1766 		 txq_id, trans_pcie->cmd_queue, sequence, txq->read_ptr,
1767 		 txq->write_ptr)) {
1768 		iwl_print_hex_error(trans, pkt, 32);
1769 		return;
1770 	}
1771 
1772 	spin_lock_bh(&txq->lock);
1773 
1774 	cmd_index = iwl_pcie_get_cmd_index(txq, index);
1775 	cmd = txq->entries[cmd_index].cmd;
1776 	meta = &txq->entries[cmd_index].meta;
1777 	group_id = cmd->hdr.group_id;
1778 	cmd_id = iwl_cmd_id(cmd->hdr.cmd, group_id, 0);
1779 
1780 	iwl_pcie_tfd_unmap(trans, meta, txq, index);
1781 
1782 	/* Input error checking is done when commands are added to queue. */
1783 	if (meta->flags & CMD_WANT_SKB) {
1784 		struct page *p = rxb_steal_page(rxb);
1785 
1786 		meta->source->resp_pkt = pkt;
1787 		meta->source->_rx_page_addr = (unsigned long)page_address(p);
1788 		meta->source->_rx_page_order = trans_pcie->rx_page_order;
1789 	}
1790 
1791 	if (meta->flags & CMD_WANT_ASYNC_CALLBACK)
1792 		iwl_op_mode_async_cb(trans->op_mode, cmd);
1793 
1794 	iwl_pcie_cmdq_reclaim(trans, txq_id, index);
1795 
1796 	if (!(meta->flags & CMD_ASYNC)) {
1797 		if (!test_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status)) {
1798 			IWL_WARN(trans,
1799 				 "HCMD_ACTIVE already clear for command %s\n",
1800 				 iwl_get_cmd_string(trans, cmd_id));
1801 		}
1802 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1803 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1804 			       iwl_get_cmd_string(trans, cmd_id));
1805 		wake_up(&trans_pcie->wait_command_queue);
1806 	}
1807 
1808 	if (meta->flags & CMD_MAKE_TRANS_IDLE) {
1809 		IWL_DEBUG_INFO(trans, "complete %s - mark trans as idle\n",
1810 			       iwl_get_cmd_string(trans, cmd->hdr.cmd));
1811 		set_bit(STATUS_TRANS_IDLE, &trans->status);
1812 		wake_up(&trans_pcie->d0i3_waitq);
1813 	}
1814 
1815 	if (meta->flags & CMD_WAKE_UP_TRANS) {
1816 		IWL_DEBUG_INFO(trans, "complete %s - clear trans idle flag\n",
1817 			       iwl_get_cmd_string(trans, cmd->hdr.cmd));
1818 		clear_bit(STATUS_TRANS_IDLE, &trans->status);
1819 		wake_up(&trans_pcie->d0i3_waitq);
1820 	}
1821 
1822 	meta->flags = 0;
1823 
1824 	spin_unlock_bh(&txq->lock);
1825 }
1826 
1827 #define HOST_COMPLETE_TIMEOUT	(2 * HZ)
1828 
1829 static int iwl_pcie_send_hcmd_async(struct iwl_trans *trans,
1830 				    struct iwl_host_cmd *cmd)
1831 {
1832 	int ret;
1833 
1834 	/* An asynchronous command can not expect an SKB to be set. */
1835 	if (WARN_ON(cmd->flags & CMD_WANT_SKB))
1836 		return -EINVAL;
1837 
1838 	ret = iwl_pcie_enqueue_hcmd(trans, cmd);
1839 	if (ret < 0) {
1840 		IWL_ERR(trans,
1841 			"Error sending %s: enqueue_hcmd failed: %d\n",
1842 			iwl_get_cmd_string(trans, cmd->id), ret);
1843 		return ret;
1844 	}
1845 	return 0;
1846 }
1847 
1848 static int iwl_pcie_send_hcmd_sync(struct iwl_trans *trans,
1849 				   struct iwl_host_cmd *cmd)
1850 {
1851 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1852 	struct iwl_txq *txq = trans_pcie->txq[trans_pcie->cmd_queue];
1853 	int cmd_idx;
1854 	int ret;
1855 
1856 	IWL_DEBUG_INFO(trans, "Attempting to send sync command %s\n",
1857 		       iwl_get_cmd_string(trans, cmd->id));
1858 
1859 	if (WARN(test_and_set_bit(STATUS_SYNC_HCMD_ACTIVE,
1860 				  &trans->status),
1861 		 "Command %s: a command is already active!\n",
1862 		 iwl_get_cmd_string(trans, cmd->id)))
1863 		return -EIO;
1864 
1865 	IWL_DEBUG_INFO(trans, "Setting HCMD_ACTIVE for command %s\n",
1866 		       iwl_get_cmd_string(trans, cmd->id));
1867 
1868 	if (pm_runtime_suspended(&trans_pcie->pci_dev->dev)) {
1869 		ret = wait_event_timeout(trans_pcie->d0i3_waitq,
1870 				 pm_runtime_active(&trans_pcie->pci_dev->dev),
1871 				 msecs_to_jiffies(IWL_TRANS_IDLE_TIMEOUT));
1872 		if (!ret) {
1873 			IWL_ERR(trans, "Timeout exiting D0i3 before hcmd\n");
1874 			return -ETIMEDOUT;
1875 		}
1876 	}
1877 
1878 	cmd_idx = iwl_pcie_enqueue_hcmd(trans, cmd);
1879 	if (cmd_idx < 0) {
1880 		ret = cmd_idx;
1881 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1882 		IWL_ERR(trans,
1883 			"Error sending %s: enqueue_hcmd failed: %d\n",
1884 			iwl_get_cmd_string(trans, cmd->id), ret);
1885 		return ret;
1886 	}
1887 
1888 	ret = wait_event_timeout(trans_pcie->wait_command_queue,
1889 				 !test_bit(STATUS_SYNC_HCMD_ACTIVE,
1890 					   &trans->status),
1891 				 HOST_COMPLETE_TIMEOUT);
1892 	if (!ret) {
1893 		IWL_ERR(trans, "Error sending %s: time out after %dms.\n",
1894 			iwl_get_cmd_string(trans, cmd->id),
1895 			jiffies_to_msecs(HOST_COMPLETE_TIMEOUT));
1896 
1897 		IWL_ERR(trans, "Current CMD queue read_ptr %d write_ptr %d\n",
1898 			txq->read_ptr, txq->write_ptr);
1899 
1900 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1901 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1902 			       iwl_get_cmd_string(trans, cmd->id));
1903 		ret = -ETIMEDOUT;
1904 
1905 		iwl_force_nmi(trans);
1906 		iwl_trans_fw_error(trans);
1907 
1908 		goto cancel;
1909 	}
1910 
1911 	if (test_bit(STATUS_FW_ERROR, &trans->status)) {
1912 		IWL_ERR(trans, "FW error in SYNC CMD %s\n",
1913 			iwl_get_cmd_string(trans, cmd->id));
1914 		dump_stack();
1915 		ret = -EIO;
1916 		goto cancel;
1917 	}
1918 
1919 	if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1920 	    test_bit(STATUS_RFKILL_OPMODE, &trans->status)) {
1921 		IWL_DEBUG_RF_KILL(trans, "RFKILL in SYNC CMD... no rsp\n");
1922 		ret = -ERFKILL;
1923 		goto cancel;
1924 	}
1925 
1926 	if ((cmd->flags & CMD_WANT_SKB) && !cmd->resp_pkt) {
1927 		IWL_ERR(trans, "Error: Response NULL in '%s'\n",
1928 			iwl_get_cmd_string(trans, cmd->id));
1929 		ret = -EIO;
1930 		goto cancel;
1931 	}
1932 
1933 	return 0;
1934 
1935 cancel:
1936 	if (cmd->flags & CMD_WANT_SKB) {
1937 		/*
1938 		 * Cancel the CMD_WANT_SKB flag for the cmd in the
1939 		 * TX cmd queue. Otherwise in case the cmd comes
1940 		 * in later, it will possibly set an invalid
1941 		 * address (cmd->meta.source).
1942 		 */
1943 		txq->entries[cmd_idx].meta.flags &= ~CMD_WANT_SKB;
1944 	}
1945 
1946 	if (cmd->resp_pkt) {
1947 		iwl_free_resp(cmd);
1948 		cmd->resp_pkt = NULL;
1949 	}
1950 
1951 	return ret;
1952 }
1953 
1954 int iwl_trans_pcie_send_hcmd(struct iwl_trans *trans, struct iwl_host_cmd *cmd)
1955 {
1956 	if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1957 	    test_bit(STATUS_RFKILL_OPMODE, &trans->status)) {
1958 		IWL_DEBUG_RF_KILL(trans, "Dropping CMD 0x%x: RF KILL\n",
1959 				  cmd->id);
1960 		return -ERFKILL;
1961 	}
1962 
1963 	if (cmd->flags & CMD_ASYNC)
1964 		return iwl_pcie_send_hcmd_async(trans, cmd);
1965 
1966 	/* We still can fail on RFKILL that can be asserted while we wait */
1967 	return iwl_pcie_send_hcmd_sync(trans, cmd);
1968 }
1969 
1970 static int iwl_fill_data_tbs(struct iwl_trans *trans, struct sk_buff *skb,
1971 			     struct iwl_txq *txq, u8 hdr_len,
1972 			     struct iwl_cmd_meta *out_meta,
1973 			     struct iwl_device_cmd *dev_cmd, u16 tb1_len)
1974 {
1975 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1976 	u16 tb2_len;
1977 	int i;
1978 
1979 	/*
1980 	 * Set up TFD's third entry to point directly to remainder
1981 	 * of skb's head, if any
1982 	 */
1983 	tb2_len = skb_headlen(skb) - hdr_len;
1984 
1985 	if (tb2_len > 0) {
1986 		dma_addr_t tb2_phys = dma_map_single(trans->dev,
1987 						     skb->data + hdr_len,
1988 						     tb2_len, DMA_TO_DEVICE);
1989 		if (unlikely(dma_mapping_error(trans->dev, tb2_phys))) {
1990 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1991 					   txq->write_ptr);
1992 			return -EINVAL;
1993 		}
1994 		iwl_pcie_txq_build_tfd(trans, txq, tb2_phys, tb2_len, false);
1995 	}
1996 
1997 	/* set up the remaining entries to point to the data */
1998 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1999 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2000 		dma_addr_t tb_phys;
2001 		int tb_idx;
2002 
2003 		if (!skb_frag_size(frag))
2004 			continue;
2005 
2006 		tb_phys = skb_frag_dma_map(trans->dev, frag, 0,
2007 					   skb_frag_size(frag), DMA_TO_DEVICE);
2008 
2009 		if (unlikely(dma_mapping_error(trans->dev, tb_phys))) {
2010 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
2011 					   txq->write_ptr);
2012 			return -EINVAL;
2013 		}
2014 		tb_idx = iwl_pcie_txq_build_tfd(trans, txq, tb_phys,
2015 						skb_frag_size(frag), false);
2016 
2017 		out_meta->tbs |= BIT(tb_idx);
2018 	}
2019 
2020 	trace_iwlwifi_dev_tx(trans->dev, skb,
2021 			     iwl_pcie_get_tfd(trans_pcie, txq, txq->write_ptr),
2022 			     trans_pcie->tfd_size,
2023 			     &dev_cmd->hdr, IWL_FIRST_TB_SIZE + tb1_len,
2024 			     hdr_len);
2025 	trace_iwlwifi_dev_tx_data(trans->dev, skb, hdr_len);
2026 	return 0;
2027 }
2028 
2029 #ifdef CONFIG_INET
2030 struct iwl_tso_hdr_page *get_page_hdr(struct iwl_trans *trans, size_t len)
2031 {
2032 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
2033 	struct iwl_tso_hdr_page *p = this_cpu_ptr(trans_pcie->tso_hdr_page);
2034 
2035 	if (!p->page)
2036 		goto alloc;
2037 
2038 	/* enough room on this page */
2039 	if (p->pos + len < (u8 *)page_address(p->page) + PAGE_SIZE)
2040 		return p;
2041 
2042 	/* We don't have enough room on this page, get a new one. */
2043 	__free_page(p->page);
2044 
2045 alloc:
2046 	p->page = alloc_page(GFP_ATOMIC);
2047 	if (!p->page)
2048 		return NULL;
2049 	p->pos = page_address(p->page);
2050 	return p;
2051 }
2052 
2053 static void iwl_compute_pseudo_hdr_csum(void *iph, struct tcphdr *tcph,
2054 					bool ipv6, unsigned int len)
2055 {
2056 	if (ipv6) {
2057 		struct ipv6hdr *iphv6 = iph;
2058 
2059 		tcph->check = ~csum_ipv6_magic(&iphv6->saddr, &iphv6->daddr,
2060 					       len + tcph->doff * 4,
2061 					       IPPROTO_TCP, 0);
2062 	} else {
2063 		struct iphdr *iphv4 = iph;
2064 
2065 		ip_send_check(iphv4);
2066 		tcph->check = ~csum_tcpudp_magic(iphv4->saddr, iphv4->daddr,
2067 						 len + tcph->doff * 4,
2068 						 IPPROTO_TCP, 0);
2069 	}
2070 }
2071 
2072 static int iwl_fill_data_tbs_amsdu(struct iwl_trans *trans, struct sk_buff *skb,
2073 				   struct iwl_txq *txq, u8 hdr_len,
2074 				   struct iwl_cmd_meta *out_meta,
2075 				   struct iwl_device_cmd *dev_cmd, u16 tb1_len)
2076 {
2077 	struct iwl_tx_cmd *tx_cmd = (void *)dev_cmd->payload;
2078 	struct iwl_trans_pcie *trans_pcie = txq->trans_pcie;
2079 	struct ieee80211_hdr *hdr = (void *)skb->data;
2080 	unsigned int snap_ip_tcp_hdrlen, ip_hdrlen, total_len, hdr_room;
2081 	unsigned int mss = skb_shinfo(skb)->gso_size;
2082 	u16 length, iv_len, amsdu_pad;
2083 	u8 *start_hdr;
2084 	struct iwl_tso_hdr_page *hdr_page;
2085 	struct page **page_ptr;
2086 	int ret;
2087 	struct tso_t tso;
2088 
2089 	/* if the packet is protected, then it must be CCMP or GCMP */
2090 	BUILD_BUG_ON(IEEE80211_CCMP_HDR_LEN != IEEE80211_GCMP_HDR_LEN);
2091 	iv_len = ieee80211_has_protected(hdr->frame_control) ?
2092 		IEEE80211_CCMP_HDR_LEN : 0;
2093 
2094 	trace_iwlwifi_dev_tx(trans->dev, skb,
2095 			     iwl_pcie_get_tfd(trans_pcie, txq, txq->write_ptr),
2096 			     trans_pcie->tfd_size,
2097 			     &dev_cmd->hdr, IWL_FIRST_TB_SIZE + tb1_len, 0);
2098 
2099 	ip_hdrlen = skb_transport_header(skb) - skb_network_header(skb);
2100 	snap_ip_tcp_hdrlen = 8 + ip_hdrlen + tcp_hdrlen(skb);
2101 	total_len = skb->len - snap_ip_tcp_hdrlen - hdr_len - iv_len;
2102 	amsdu_pad = 0;
2103 
2104 	/* total amount of header we may need for this A-MSDU */
2105 	hdr_room = DIV_ROUND_UP(total_len, mss) *
2106 		(3 + snap_ip_tcp_hdrlen + sizeof(struct ethhdr)) + iv_len;
2107 
2108 	/* Our device supports 9 segments at most, it will fit in 1 page */
2109 	hdr_page = get_page_hdr(trans, hdr_room);
2110 	if (!hdr_page)
2111 		return -ENOMEM;
2112 
2113 	get_page(hdr_page->page);
2114 	start_hdr = hdr_page->pos;
2115 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
2116 	*page_ptr = hdr_page->page;
2117 	memcpy(hdr_page->pos, skb->data + hdr_len, iv_len);
2118 	hdr_page->pos += iv_len;
2119 
2120 	/*
2121 	 * Pull the ieee80211 header + IV to be able to use TSO core,
2122 	 * we will restore it for the tx_status flow.
2123 	 */
2124 	skb_pull(skb, hdr_len + iv_len);
2125 
2126 	/*
2127 	 * Remove the length of all the headers that we don't actually
2128 	 * have in the MPDU by themselves, but that we duplicate into
2129 	 * all the different MSDUs inside the A-MSDU.
2130 	 */
2131 	le16_add_cpu(&tx_cmd->len, -snap_ip_tcp_hdrlen);
2132 
2133 	tso_start(skb, &tso);
2134 
2135 	while (total_len) {
2136 		/* this is the data left for this subframe */
2137 		unsigned int data_left =
2138 			min_t(unsigned int, mss, total_len);
2139 		struct sk_buff *csum_skb = NULL;
2140 		unsigned int hdr_tb_len;
2141 		dma_addr_t hdr_tb_phys;
2142 		struct tcphdr *tcph;
2143 		u8 *iph, *subf_hdrs_start = hdr_page->pos;
2144 
2145 		total_len -= data_left;
2146 
2147 		memset(hdr_page->pos, 0, amsdu_pad);
2148 		hdr_page->pos += amsdu_pad;
2149 		amsdu_pad = (4 - (sizeof(struct ethhdr) + snap_ip_tcp_hdrlen +
2150 				  data_left)) & 0x3;
2151 		ether_addr_copy(hdr_page->pos, ieee80211_get_DA(hdr));
2152 		hdr_page->pos += ETH_ALEN;
2153 		ether_addr_copy(hdr_page->pos, ieee80211_get_SA(hdr));
2154 		hdr_page->pos += ETH_ALEN;
2155 
2156 		length = snap_ip_tcp_hdrlen + data_left;
2157 		*((__be16 *)hdr_page->pos) = cpu_to_be16(length);
2158 		hdr_page->pos += sizeof(length);
2159 
2160 		/*
2161 		 * This will copy the SNAP as well which will be considered
2162 		 * as MAC header.
2163 		 */
2164 		tso_build_hdr(skb, hdr_page->pos, &tso, data_left, !total_len);
2165 		iph = hdr_page->pos + 8;
2166 		tcph = (void *)(iph + ip_hdrlen);
2167 
2168 		/* For testing on current hardware only */
2169 		if (trans_pcie->sw_csum_tx) {
2170 			csum_skb = alloc_skb(data_left + tcp_hdrlen(skb),
2171 					     GFP_ATOMIC);
2172 			if (!csum_skb) {
2173 				ret = -ENOMEM;
2174 				goto out_unmap;
2175 			}
2176 
2177 			iwl_compute_pseudo_hdr_csum(iph, tcph,
2178 						    skb->protocol ==
2179 							htons(ETH_P_IPV6),
2180 						    data_left);
2181 
2182 			skb_put_data(csum_skb, tcph, tcp_hdrlen(skb));
2183 			skb_reset_transport_header(csum_skb);
2184 			csum_skb->csum_start =
2185 				(unsigned char *)tcp_hdr(csum_skb) -
2186 						 csum_skb->head;
2187 		}
2188 
2189 		hdr_page->pos += snap_ip_tcp_hdrlen;
2190 
2191 		hdr_tb_len = hdr_page->pos - start_hdr;
2192 		hdr_tb_phys = dma_map_single(trans->dev, start_hdr,
2193 					     hdr_tb_len, DMA_TO_DEVICE);
2194 		if (unlikely(dma_mapping_error(trans->dev, hdr_tb_phys))) {
2195 			dev_kfree_skb(csum_skb);
2196 			ret = -EINVAL;
2197 			goto out_unmap;
2198 		}
2199 		iwl_pcie_txq_build_tfd(trans, txq, hdr_tb_phys,
2200 				       hdr_tb_len, false);
2201 		trace_iwlwifi_dev_tx_tso_chunk(trans->dev, start_hdr,
2202 					       hdr_tb_len);
2203 		/* add this subframe's headers' length to the tx_cmd */
2204 		le16_add_cpu(&tx_cmd->len, hdr_page->pos - subf_hdrs_start);
2205 
2206 		/* prepare the start_hdr for the next subframe */
2207 		start_hdr = hdr_page->pos;
2208 
2209 		/* put the payload */
2210 		while (data_left) {
2211 			unsigned int size = min_t(unsigned int, tso.size,
2212 						  data_left);
2213 			dma_addr_t tb_phys;
2214 
2215 			if (trans_pcie->sw_csum_tx)
2216 				skb_put_data(csum_skb, tso.data, size);
2217 
2218 			tb_phys = dma_map_single(trans->dev, tso.data,
2219 						 size, DMA_TO_DEVICE);
2220 			if (unlikely(dma_mapping_error(trans->dev, tb_phys))) {
2221 				dev_kfree_skb(csum_skb);
2222 				ret = -EINVAL;
2223 				goto out_unmap;
2224 			}
2225 
2226 			iwl_pcie_txq_build_tfd(trans, txq, tb_phys,
2227 					       size, false);
2228 			trace_iwlwifi_dev_tx_tso_chunk(trans->dev, tso.data,
2229 						       size);
2230 
2231 			data_left -= size;
2232 			tso_build_data(skb, &tso, size);
2233 		}
2234 
2235 		/* For testing on early hardware only */
2236 		if (trans_pcie->sw_csum_tx) {
2237 			__wsum csum;
2238 
2239 			csum = skb_checksum(csum_skb,
2240 					    skb_checksum_start_offset(csum_skb),
2241 					    csum_skb->len -
2242 					    skb_checksum_start_offset(csum_skb),
2243 					    0);
2244 			dev_kfree_skb(csum_skb);
2245 			dma_sync_single_for_cpu(trans->dev, hdr_tb_phys,
2246 						hdr_tb_len, DMA_TO_DEVICE);
2247 			tcph->check = csum_fold(csum);
2248 			dma_sync_single_for_device(trans->dev, hdr_tb_phys,
2249 						   hdr_tb_len, DMA_TO_DEVICE);
2250 		}
2251 	}
2252 
2253 	/* re -add the WiFi header and IV */
2254 	skb_push(skb, hdr_len + iv_len);
2255 
2256 	return 0;
2257 
2258 out_unmap:
2259 	iwl_pcie_tfd_unmap(trans, out_meta, txq, txq->write_ptr);
2260 	return ret;
2261 }
2262 #else /* CONFIG_INET */
2263 static int iwl_fill_data_tbs_amsdu(struct iwl_trans *trans, struct sk_buff *skb,
2264 				   struct iwl_txq *txq, u8 hdr_len,
2265 				   struct iwl_cmd_meta *out_meta,
2266 				   struct iwl_device_cmd *dev_cmd, u16 tb1_len)
2267 {
2268 	/* No A-MSDU without CONFIG_INET */
2269 	WARN_ON(1);
2270 
2271 	return -1;
2272 }
2273 #endif /* CONFIG_INET */
2274 
2275 int iwl_trans_pcie_tx(struct iwl_trans *trans, struct sk_buff *skb,
2276 		      struct iwl_device_cmd *dev_cmd, int txq_id)
2277 {
2278 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
2279 	struct ieee80211_hdr *hdr;
2280 	struct iwl_tx_cmd *tx_cmd = (struct iwl_tx_cmd *)dev_cmd->payload;
2281 	struct iwl_cmd_meta *out_meta;
2282 	struct iwl_txq *txq;
2283 	dma_addr_t tb0_phys, tb1_phys, scratch_phys;
2284 	void *tb1_addr;
2285 	void *tfd;
2286 	u16 len, tb1_len;
2287 	bool wait_write_ptr;
2288 	__le16 fc;
2289 	u8 hdr_len;
2290 	u16 wifi_seq;
2291 	bool amsdu;
2292 
2293 	txq = trans_pcie->txq[txq_id];
2294 
2295 	if (WARN_ONCE(!test_bit(txq_id, trans_pcie->queue_used),
2296 		      "TX on unused queue %d\n", txq_id))
2297 		return -EINVAL;
2298 
2299 	if (unlikely(trans_pcie->sw_csum_tx &&
2300 		     skb->ip_summed == CHECKSUM_PARTIAL)) {
2301 		int offs = skb_checksum_start_offset(skb);
2302 		int csum_offs = offs + skb->csum_offset;
2303 		__wsum csum;
2304 
2305 		if (skb_ensure_writable(skb, csum_offs + sizeof(__sum16)))
2306 			return -1;
2307 
2308 		csum = skb_checksum(skb, offs, skb->len - offs, 0);
2309 		*(__sum16 *)(skb->data + csum_offs) = csum_fold(csum);
2310 
2311 		skb->ip_summed = CHECKSUM_UNNECESSARY;
2312 	}
2313 
2314 	if (skb_is_nonlinear(skb) &&
2315 	    skb_shinfo(skb)->nr_frags > IWL_PCIE_MAX_FRAGS(trans_pcie) &&
2316 	    __skb_linearize(skb))
2317 		return -ENOMEM;
2318 
2319 	/* mac80211 always puts the full header into the SKB's head,
2320 	 * so there's no need to check if it's readable there
2321 	 */
2322 	hdr = (struct ieee80211_hdr *)skb->data;
2323 	fc = hdr->frame_control;
2324 	hdr_len = ieee80211_hdrlen(fc);
2325 
2326 	spin_lock(&txq->lock);
2327 
2328 	if (iwl_queue_space(txq) < txq->high_mark) {
2329 		iwl_stop_queue(trans, txq);
2330 
2331 		/* don't put the packet on the ring, if there is no room */
2332 		if (unlikely(iwl_queue_space(txq) < 3)) {
2333 			struct iwl_device_cmd **dev_cmd_ptr;
2334 
2335 			dev_cmd_ptr = (void *)((u8 *)skb->cb +
2336 					       trans_pcie->dev_cmd_offs);
2337 
2338 			*dev_cmd_ptr = dev_cmd;
2339 			__skb_queue_tail(&txq->overflow_q, skb);
2340 
2341 			spin_unlock(&txq->lock);
2342 			return 0;
2343 		}
2344 	}
2345 
2346 	/* In AGG mode, the index in the ring must correspond to the WiFi
2347 	 * sequence number. This is a HW requirements to help the SCD to parse
2348 	 * the BA.
2349 	 * Check here that the packets are in the right place on the ring.
2350 	 */
2351 	wifi_seq = IEEE80211_SEQ_TO_SN(le16_to_cpu(hdr->seq_ctrl));
2352 	WARN_ONCE(txq->ampdu &&
2353 		  (wifi_seq & 0xff) != txq->write_ptr,
2354 		  "Q: %d WiFi Seq %d tfdNum %d",
2355 		  txq_id, wifi_seq, txq->write_ptr);
2356 
2357 	/* Set up driver data for this TFD */
2358 	txq->entries[txq->write_ptr].skb = skb;
2359 	txq->entries[txq->write_ptr].cmd = dev_cmd;
2360 
2361 	dev_cmd->hdr.sequence =
2362 		cpu_to_le16((u16)(QUEUE_TO_SEQ(txq_id) |
2363 			    INDEX_TO_SEQ(txq->write_ptr)));
2364 
2365 	tb0_phys = iwl_pcie_get_first_tb_dma(txq, txq->write_ptr);
2366 	scratch_phys = tb0_phys + sizeof(struct iwl_cmd_header) +
2367 		       offsetof(struct iwl_tx_cmd, scratch);
2368 
2369 	tx_cmd->dram_lsb_ptr = cpu_to_le32(scratch_phys);
2370 	tx_cmd->dram_msb_ptr = iwl_get_dma_hi_addr(scratch_phys);
2371 
2372 	/* Set up first empty entry in queue's array of Tx/cmd buffers */
2373 	out_meta = &txq->entries[txq->write_ptr].meta;
2374 	out_meta->flags = 0;
2375 
2376 	/*
2377 	 * The second TB (tb1) points to the remainder of the TX command
2378 	 * and the 802.11 header - dword aligned size
2379 	 * (This calculation modifies the TX command, so do it before the
2380 	 * setup of the first TB)
2381 	 */
2382 	len = sizeof(struct iwl_tx_cmd) + sizeof(struct iwl_cmd_header) +
2383 	      hdr_len - IWL_FIRST_TB_SIZE;
2384 	/* do not align A-MSDU to dword as the subframe header aligns it */
2385 	amsdu = ieee80211_is_data_qos(fc) &&
2386 		(*ieee80211_get_qos_ctl(hdr) &
2387 		 IEEE80211_QOS_CTL_A_MSDU_PRESENT);
2388 	if (trans_pcie->sw_csum_tx || !amsdu) {
2389 		tb1_len = ALIGN(len, 4);
2390 		/* Tell NIC about any 2-byte padding after MAC header */
2391 		if (tb1_len != len)
2392 			tx_cmd->tx_flags |= cpu_to_le32(TX_CMD_FLG_MH_PAD);
2393 	} else {
2394 		tb1_len = len;
2395 	}
2396 
2397 	/*
2398 	 * The first TB points to bi-directional DMA data, we'll
2399 	 * memcpy the data into it later.
2400 	 */
2401 	iwl_pcie_txq_build_tfd(trans, txq, tb0_phys,
2402 			       IWL_FIRST_TB_SIZE, true);
2403 
2404 	/* there must be data left over for TB1 or this code must be changed */
2405 	BUILD_BUG_ON(sizeof(struct iwl_tx_cmd) < IWL_FIRST_TB_SIZE);
2406 
2407 	/* map the data for TB1 */
2408 	tb1_addr = ((u8 *)&dev_cmd->hdr) + IWL_FIRST_TB_SIZE;
2409 	tb1_phys = dma_map_single(trans->dev, tb1_addr, tb1_len, DMA_TO_DEVICE);
2410 	if (unlikely(dma_mapping_error(trans->dev, tb1_phys)))
2411 		goto out_err;
2412 	iwl_pcie_txq_build_tfd(trans, txq, tb1_phys, tb1_len, false);
2413 
2414 	if (amsdu) {
2415 		if (unlikely(iwl_fill_data_tbs_amsdu(trans, skb, txq, hdr_len,
2416 						     out_meta, dev_cmd,
2417 						     tb1_len)))
2418 			goto out_err;
2419 	} else if (unlikely(iwl_fill_data_tbs(trans, skb, txq, hdr_len,
2420 				       out_meta, dev_cmd, tb1_len))) {
2421 		goto out_err;
2422 	}
2423 
2424 	/* building the A-MSDU might have changed this data, so memcpy it now */
2425 	memcpy(&txq->first_tb_bufs[txq->write_ptr], &dev_cmd->hdr,
2426 	       IWL_FIRST_TB_SIZE);
2427 
2428 	tfd = iwl_pcie_get_tfd(trans_pcie, txq, txq->write_ptr);
2429 	/* Set up entry for this TFD in Tx byte-count array */
2430 	iwl_pcie_txq_update_byte_cnt_tbl(trans, txq, le16_to_cpu(tx_cmd->len),
2431 					 iwl_pcie_tfd_get_num_tbs(trans, tfd));
2432 
2433 	wait_write_ptr = ieee80211_has_morefrags(fc);
2434 
2435 	/* start timer if queue currently empty */
2436 	if (txq->read_ptr == txq->write_ptr) {
2437 		if (txq->wd_timeout) {
2438 			/*
2439 			 * If the TXQ is active, then set the timer, if not,
2440 			 * set the timer in remainder so that the timer will
2441 			 * be armed with the right value when the station will
2442 			 * wake up.
2443 			 */
2444 			if (!txq->frozen)
2445 				mod_timer(&txq->stuck_timer,
2446 					  jiffies + txq->wd_timeout);
2447 			else
2448 				txq->frozen_expiry_remainder = txq->wd_timeout;
2449 		}
2450 		IWL_DEBUG_RPM(trans, "Q: %d first tx - take ref\n", txq->id);
2451 		iwl_trans_ref(trans);
2452 	}
2453 
2454 	/* Tell device the write index *just past* this latest filled TFD */
2455 	txq->write_ptr = iwl_queue_inc_wrap(txq->write_ptr);
2456 	if (!wait_write_ptr)
2457 		iwl_pcie_txq_inc_wr_ptr(trans, txq);
2458 
2459 	/*
2460 	 * At this point the frame is "transmitted" successfully
2461 	 * and we will get a TX status notification eventually.
2462 	 */
2463 	spin_unlock(&txq->lock);
2464 	return 0;
2465 out_err:
2466 	spin_unlock(&txq->lock);
2467 	return -1;
2468 }
2469