1 /******************************************************************************
2  *
3  * This file is provided under a dual BSD/GPLv2 license.  When using or
4  * redistributing this file, you may do so under either license.
5  *
6  * GPL LICENSE SUMMARY
7  *
8  * Copyright(c) 2003 - 2014 Intel Corporation. All rights reserved.
9  * Copyright(c) 2013 - 2015 Intel Mobile Communications GmbH
10  * Copyright(c) 2016 - 2017 Intel Deutschland GmbH
11  * Copyright(c) 2018 - 2020 Intel Corporation
12  *
13  * This program is free software; you can redistribute it and/or modify it
14  * under the terms of version 2 of the GNU General Public License as
15  * published by the Free Software Foundation.
16  *
17  * This program is distributed in the hope that it will be useful, but WITHOUT
18  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
19  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
20  * more details.
21  *
22  * The full GNU General Public License is included in this distribution in the
23  * file called COPYING.
24  *
25  * Contact Information:
26  *  Intel Linux Wireless <linuxwifi@intel.com>
27  * Intel Corporation, 5200 N.E. Elam Young Parkway, Hillsboro, OR 97124-6497
28  *
29  * BSD LICENSE
30  *
31  * Copyright(c) 2003 - 2014 Intel Corporation. All rights reserved.
32  * Copyright(c) 2013 - 2015 Intel Mobile Communications GmbH
33  * Copyright(c) 2016 - 2017 Intel Deutschland GmbH
34  * Copyright(c) 2018 - 2020 Intel Corporation
35  * All rights reserved.
36  *
37  * Redistribution and use in source and binary forms, with or without
38  * modification, are permitted provided that the following conditions
39  * are met:
40  *
41  *  * Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  *  * Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in
45  *    the documentation and/or other materials provided with the
46  *    distribution.
47  *  * Neither the name Intel Corporation nor the names of its
48  *    contributors may be used to endorse or promote products derived
49  *    from this software without specific prior written permission.
50  *
51  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
52  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
53  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
54  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
55  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
56  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
57  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
58  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
59  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
60  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
61  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
62  *
63  *****************************************************************************/
64 #include <linux/etherdevice.h>
65 #include <linux/ieee80211.h>
66 #include <linux/slab.h>
67 #include <linux/sched.h>
68 #include <net/ip6_checksum.h>
69 #include <net/tso.h>
70 
71 #include "iwl-debug.h"
72 #include "iwl-csr.h"
73 #include "iwl-prph.h"
74 #include "iwl-io.h"
75 #include "iwl-scd.h"
76 #include "iwl-op-mode.h"
77 #include "internal.h"
78 #include "fw/api/tx.h"
79 
80 #define IWL_TX_CRC_SIZE 4
81 #define IWL_TX_DELIMITER_SIZE 4
82 
83 /*************** DMA-QUEUE-GENERAL-FUNCTIONS  *****
84  * DMA services
85  *
86  * Theory of operation
87  *
88  * A Tx or Rx queue resides in host DRAM, and is comprised of a circular buffer
89  * of buffer descriptors, each of which points to one or more data buffers for
90  * the device to read from or fill.  Driver and device exchange status of each
91  * queue via "read" and "write" pointers.  Driver keeps minimum of 2 empty
92  * entries in each circular buffer, to protect against confusing empty and full
93  * queue states.
94  *
95  * The device reads or writes the data in the queues via the device's several
96  * DMA/FIFO channels.  Each queue is mapped to a single DMA channel.
97  *
98  * For Tx queue, there are low mark and high mark limits. If, after queuing
99  * the packet for Tx, free space become < low mark, Tx queue stopped. When
100  * reclaiming packets (on 'tx done IRQ), if free space become > high mark,
101  * Tx queue resumed.
102  *
103  ***************************************************/
104 
105 int iwl_queue_space(struct iwl_trans *trans, const struct iwl_txq *q)
106 {
107 	unsigned int max;
108 	unsigned int used;
109 
110 	/*
111 	 * To avoid ambiguity between empty and completely full queues, there
112 	 * should always be less than max_tfd_queue_size elements in the queue.
113 	 * If q->n_window is smaller than max_tfd_queue_size, there is no need
114 	 * to reserve any queue entries for this purpose.
115 	 */
116 	if (q->n_window < trans->trans_cfg->base_params->max_tfd_queue_size)
117 		max = q->n_window;
118 	else
119 		max = trans->trans_cfg->base_params->max_tfd_queue_size - 1;
120 
121 	/*
122 	 * max_tfd_queue_size is a power of 2, so the following is equivalent to
123 	 * modulo by max_tfd_queue_size and is well defined.
124 	 */
125 	used = (q->write_ptr - q->read_ptr) &
126 		(trans->trans_cfg->base_params->max_tfd_queue_size - 1);
127 
128 	if (WARN_ON(used > max))
129 		return 0;
130 
131 	return max - used;
132 }
133 
134 /*
135  * iwl_queue_init - Initialize queue's high/low-water and read/write indexes
136  */
137 static int iwl_queue_init(struct iwl_txq *q, int slots_num)
138 {
139 	q->n_window = slots_num;
140 
141 	/* slots_num must be power-of-two size, otherwise
142 	 * iwl_pcie_get_cmd_index is broken. */
143 	if (WARN_ON(!is_power_of_2(slots_num)))
144 		return -EINVAL;
145 
146 	q->low_mark = q->n_window / 4;
147 	if (q->low_mark < 4)
148 		q->low_mark = 4;
149 
150 	q->high_mark = q->n_window / 8;
151 	if (q->high_mark < 2)
152 		q->high_mark = 2;
153 
154 	q->write_ptr = 0;
155 	q->read_ptr = 0;
156 
157 	return 0;
158 }
159 
160 int iwl_pcie_alloc_dma_ptr(struct iwl_trans *trans,
161 			   struct iwl_dma_ptr *ptr, size_t size)
162 {
163 	if (WARN_ON(ptr->addr))
164 		return -EINVAL;
165 
166 	ptr->addr = dma_alloc_coherent(trans->dev, size,
167 				       &ptr->dma, GFP_KERNEL);
168 	if (!ptr->addr)
169 		return -ENOMEM;
170 	ptr->size = size;
171 	return 0;
172 }
173 
174 void iwl_pcie_free_dma_ptr(struct iwl_trans *trans, struct iwl_dma_ptr *ptr)
175 {
176 	if (unlikely(!ptr->addr))
177 		return;
178 
179 	dma_free_coherent(trans->dev, ptr->size, ptr->addr, ptr->dma);
180 	memset(ptr, 0, sizeof(*ptr));
181 }
182 
183 static void iwl_pcie_txq_stuck_timer(struct timer_list *t)
184 {
185 	struct iwl_txq *txq = from_timer(txq, t, stuck_timer);
186 	struct iwl_trans *trans = txq->trans;
187 
188 	spin_lock(&txq->lock);
189 	/* check if triggered erroneously */
190 	if (txq->read_ptr == txq->write_ptr) {
191 		spin_unlock(&txq->lock);
192 		return;
193 	}
194 	spin_unlock(&txq->lock);
195 
196 	iwl_trans_pcie_log_scd_error(trans, txq);
197 
198 	iwl_force_nmi(trans);
199 }
200 
201 /*
202  * iwl_pcie_txq_update_byte_cnt_tbl - Set up entry in Tx byte-count array
203  */
204 static void iwl_pcie_txq_update_byte_cnt_tbl(struct iwl_trans *trans,
205 					     struct iwl_txq *txq, u16 byte_cnt,
206 					     int num_tbs)
207 {
208 	struct iwlagn_scd_bc_tbl *scd_bc_tbl;
209 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
210 	int write_ptr = txq->write_ptr;
211 	int txq_id = txq->id;
212 	u8 sec_ctl = 0;
213 	u16 len = byte_cnt + IWL_TX_CRC_SIZE + IWL_TX_DELIMITER_SIZE;
214 	__le16 bc_ent;
215 	struct iwl_device_tx_cmd *dev_cmd = txq->entries[txq->write_ptr].cmd;
216 	struct iwl_tx_cmd *tx_cmd = (void *)dev_cmd->payload;
217 	u8 sta_id = tx_cmd->sta_id;
218 
219 	scd_bc_tbl = trans_pcie->scd_bc_tbls.addr;
220 
221 	sec_ctl = tx_cmd->sec_ctl;
222 
223 	switch (sec_ctl & TX_CMD_SEC_MSK) {
224 	case TX_CMD_SEC_CCM:
225 		len += IEEE80211_CCMP_MIC_LEN;
226 		break;
227 	case TX_CMD_SEC_TKIP:
228 		len += IEEE80211_TKIP_ICV_LEN;
229 		break;
230 	case TX_CMD_SEC_WEP:
231 		len += IEEE80211_WEP_IV_LEN + IEEE80211_WEP_ICV_LEN;
232 		break;
233 	}
234 	if (trans_pcie->bc_table_dword)
235 		len = DIV_ROUND_UP(len, 4);
236 
237 	if (WARN_ON(len > 0xFFF || write_ptr >= TFD_QUEUE_SIZE_MAX))
238 		return;
239 
240 	bc_ent = cpu_to_le16(len | (sta_id << 12));
241 
242 	scd_bc_tbl[txq_id].tfd_offset[write_ptr] = bc_ent;
243 
244 	if (write_ptr < TFD_QUEUE_SIZE_BC_DUP)
245 		scd_bc_tbl[txq_id].
246 			tfd_offset[TFD_QUEUE_SIZE_MAX + write_ptr] = bc_ent;
247 }
248 
249 static void iwl_pcie_txq_inval_byte_cnt_tbl(struct iwl_trans *trans,
250 					    struct iwl_txq *txq)
251 {
252 	struct iwl_trans_pcie *trans_pcie =
253 		IWL_TRANS_GET_PCIE_TRANS(trans);
254 	struct iwlagn_scd_bc_tbl *scd_bc_tbl = trans_pcie->scd_bc_tbls.addr;
255 	int txq_id = txq->id;
256 	int read_ptr = txq->read_ptr;
257 	u8 sta_id = 0;
258 	__le16 bc_ent;
259 	struct iwl_device_tx_cmd *dev_cmd = txq->entries[read_ptr].cmd;
260 	struct iwl_tx_cmd *tx_cmd = (void *)dev_cmd->payload;
261 
262 	WARN_ON(read_ptr >= TFD_QUEUE_SIZE_MAX);
263 
264 	if (txq_id != trans->txqs.cmd.q_id)
265 		sta_id = tx_cmd->sta_id;
266 
267 	bc_ent = cpu_to_le16(1 | (sta_id << 12));
268 
269 	scd_bc_tbl[txq_id].tfd_offset[read_ptr] = bc_ent;
270 
271 	if (read_ptr < TFD_QUEUE_SIZE_BC_DUP)
272 		scd_bc_tbl[txq_id].
273 			tfd_offset[TFD_QUEUE_SIZE_MAX + read_ptr] = bc_ent;
274 }
275 
276 /*
277  * iwl_pcie_txq_inc_wr_ptr - Send new write index to hardware
278  */
279 static void iwl_pcie_txq_inc_wr_ptr(struct iwl_trans *trans,
280 				    struct iwl_txq *txq)
281 {
282 	u32 reg = 0;
283 	int txq_id = txq->id;
284 
285 	lockdep_assert_held(&txq->lock);
286 
287 	/*
288 	 * explicitly wake up the NIC if:
289 	 * 1. shadow registers aren't enabled
290 	 * 2. NIC is woken up for CMD regardless of shadow outside this function
291 	 * 3. there is a chance that the NIC is asleep
292 	 */
293 	if (!trans->trans_cfg->base_params->shadow_reg_enable &&
294 	    txq_id != trans->txqs.cmd.q_id &&
295 	    test_bit(STATUS_TPOWER_PMI, &trans->status)) {
296 		/*
297 		 * wake up nic if it's powered down ...
298 		 * uCode will wake up, and interrupt us again, so next
299 		 * time we'll skip this part.
300 		 */
301 		reg = iwl_read32(trans, CSR_UCODE_DRV_GP1);
302 
303 		if (reg & CSR_UCODE_DRV_GP1_BIT_MAC_SLEEP) {
304 			IWL_DEBUG_INFO(trans, "Tx queue %d requesting wakeup, GP1 = 0x%x\n",
305 				       txq_id, reg);
306 			iwl_set_bit(trans, CSR_GP_CNTRL,
307 				    CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
308 			txq->need_update = true;
309 			return;
310 		}
311 	}
312 
313 	/*
314 	 * if not in power-save mode, uCode will never sleep when we're
315 	 * trying to tx (during RFKILL, we're not trying to tx).
316 	 */
317 	IWL_DEBUG_TX(trans, "Q:%d WR: 0x%x\n", txq_id, txq->write_ptr);
318 	if (!txq->block)
319 		iwl_write32(trans, HBUS_TARG_WRPTR,
320 			    txq->write_ptr | (txq_id << 8));
321 }
322 
323 void iwl_pcie_txq_check_wrptrs(struct iwl_trans *trans)
324 {
325 	int i;
326 
327 	for (i = 0; i < trans->trans_cfg->base_params->num_of_queues; i++) {
328 		struct iwl_txq *txq = trans->txqs.txq[i];
329 
330 		if (!test_bit(i, trans->txqs.queue_used))
331 			continue;
332 
333 		spin_lock_bh(&txq->lock);
334 		if (txq->need_update) {
335 			iwl_pcie_txq_inc_wr_ptr(trans, txq);
336 			txq->need_update = false;
337 		}
338 		spin_unlock_bh(&txq->lock);
339 	}
340 }
341 
342 static inline dma_addr_t iwl_pcie_tfd_tb_get_addr(struct iwl_trans *trans,
343 						  void *_tfd, u8 idx)
344 {
345 
346 	if (trans->trans_cfg->use_tfh) {
347 		struct iwl_tfh_tfd *tfd = _tfd;
348 		struct iwl_tfh_tb *tb = &tfd->tbs[idx];
349 
350 		return (dma_addr_t)(le64_to_cpu(tb->addr));
351 	} else {
352 		struct iwl_tfd *tfd = _tfd;
353 		struct iwl_tfd_tb *tb = &tfd->tbs[idx];
354 		dma_addr_t addr = get_unaligned_le32(&tb->lo);
355 		dma_addr_t hi_len;
356 
357 		if (sizeof(dma_addr_t) <= sizeof(u32))
358 			return addr;
359 
360 		hi_len = le16_to_cpu(tb->hi_n_len) & 0xF;
361 
362 		/*
363 		 * shift by 16 twice to avoid warnings on 32-bit
364 		 * (where this code never runs anyway due to the
365 		 * if statement above)
366 		 */
367 		return addr | ((hi_len << 16) << 16);
368 	}
369 }
370 
371 static inline void iwl_pcie_tfd_set_tb(struct iwl_trans *trans, void *tfd,
372 				       u8 idx, dma_addr_t addr, u16 len)
373 {
374 	struct iwl_tfd *tfd_fh = (void *)tfd;
375 	struct iwl_tfd_tb *tb = &tfd_fh->tbs[idx];
376 
377 	u16 hi_n_len = len << 4;
378 
379 	put_unaligned_le32(addr, &tb->lo);
380 	hi_n_len |= iwl_get_dma_hi_addr(addr);
381 
382 	tb->hi_n_len = cpu_to_le16(hi_n_len);
383 
384 	tfd_fh->num_tbs = idx + 1;
385 }
386 
387 static inline u8 iwl_pcie_tfd_get_num_tbs(struct iwl_trans *trans, void *_tfd)
388 {
389 	if (trans->trans_cfg->use_tfh) {
390 		struct iwl_tfh_tfd *tfd = _tfd;
391 
392 		return le16_to_cpu(tfd->num_tbs) & 0x1f;
393 	} else {
394 		struct iwl_tfd *tfd = _tfd;
395 
396 		return tfd->num_tbs & 0x1f;
397 	}
398 }
399 
400 static void iwl_pcie_tfd_unmap(struct iwl_trans *trans,
401 			       struct iwl_cmd_meta *meta,
402 			       struct iwl_txq *txq, int index)
403 {
404 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
405 	int i, num_tbs;
406 	void *tfd = iwl_pcie_get_tfd(trans, txq, index);
407 
408 	/* Sanity check on number of chunks */
409 	num_tbs = iwl_pcie_tfd_get_num_tbs(trans, tfd);
410 
411 	if (num_tbs > trans_pcie->max_tbs) {
412 		IWL_ERR(trans, "Too many chunks: %i\n", num_tbs);
413 		/* @todo issue fatal error, it is quite serious situation */
414 		return;
415 	}
416 
417 	/* first TB is never freed - it's the bidirectional DMA data */
418 
419 	for (i = 1; i < num_tbs; i++) {
420 		if (meta->tbs & BIT(i))
421 			dma_unmap_page(trans->dev,
422 				       iwl_pcie_tfd_tb_get_addr(trans, tfd, i),
423 				       iwl_pcie_tfd_tb_get_len(trans, tfd, i),
424 				       DMA_TO_DEVICE);
425 		else
426 			dma_unmap_single(trans->dev,
427 					 iwl_pcie_tfd_tb_get_addr(trans, tfd,
428 								  i),
429 					 iwl_pcie_tfd_tb_get_len(trans, tfd,
430 								 i),
431 					 DMA_TO_DEVICE);
432 	}
433 
434 	meta->tbs = 0;
435 
436 	if (trans->trans_cfg->use_tfh) {
437 		struct iwl_tfh_tfd *tfd_fh = (void *)tfd;
438 
439 		tfd_fh->num_tbs = 0;
440 	} else {
441 		struct iwl_tfd *tfd_fh = (void *)tfd;
442 
443 		tfd_fh->num_tbs = 0;
444 	}
445 
446 }
447 
448 /*
449  * iwl_pcie_txq_free_tfd - Free all chunks referenced by TFD [txq->q.read_ptr]
450  * @trans - transport private data
451  * @txq - tx queue
452  * @dma_dir - the direction of the DMA mapping
453  *
454  * Does NOT advance any TFD circular buffer read/write indexes
455  * Does NOT free the TFD itself (which is within circular buffer)
456  */
457 void iwl_pcie_txq_free_tfd(struct iwl_trans *trans, struct iwl_txq *txq)
458 {
459 	/* rd_ptr is bounded by TFD_QUEUE_SIZE_MAX and
460 	 * idx is bounded by n_window
461 	 */
462 	int rd_ptr = txq->read_ptr;
463 	int idx = iwl_pcie_get_cmd_index(txq, rd_ptr);
464 
465 	lockdep_assert_held(&txq->lock);
466 
467 	/* We have only q->n_window txq->entries, but we use
468 	 * TFD_QUEUE_SIZE_MAX tfds
469 	 */
470 	iwl_pcie_tfd_unmap(trans, &txq->entries[idx].meta, txq, rd_ptr);
471 
472 	/* free SKB */
473 	if (txq->entries) {
474 		struct sk_buff *skb;
475 
476 		skb = txq->entries[idx].skb;
477 
478 		/* Can be called from irqs-disabled context
479 		 * If skb is not NULL, it means that the whole queue is being
480 		 * freed and that the queue is not empty - free the skb
481 		 */
482 		if (skb) {
483 			iwl_op_mode_free_skb(trans->op_mode, skb);
484 			txq->entries[idx].skb = NULL;
485 		}
486 	}
487 }
488 
489 static int iwl_pcie_txq_build_tfd(struct iwl_trans *trans, struct iwl_txq *txq,
490 				  dma_addr_t addr, u16 len, bool reset)
491 {
492 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
493 	void *tfd;
494 	u32 num_tbs;
495 
496 	tfd = txq->tfds + trans_pcie->tfd_size * txq->write_ptr;
497 
498 	if (reset)
499 		memset(tfd, 0, trans_pcie->tfd_size);
500 
501 	num_tbs = iwl_pcie_tfd_get_num_tbs(trans, tfd);
502 
503 	/* Each TFD can point to a maximum max_tbs Tx buffers */
504 	if (num_tbs >= trans_pcie->max_tbs) {
505 		IWL_ERR(trans, "Error can not send more than %d chunks\n",
506 			trans_pcie->max_tbs);
507 		return -EINVAL;
508 	}
509 
510 	if (WARN(addr & ~IWL_TX_DMA_MASK,
511 		 "Unaligned address = %llx\n", (unsigned long long)addr))
512 		return -EINVAL;
513 
514 	iwl_pcie_tfd_set_tb(trans, tfd, num_tbs, addr, len);
515 
516 	return num_tbs;
517 }
518 
519 int iwl_pcie_txq_alloc(struct iwl_trans *trans, struct iwl_txq *txq,
520 		       int slots_num, bool cmd_queue)
521 {
522 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
523 	size_t tfd_sz = trans_pcie->tfd_size *
524 		trans->trans_cfg->base_params->max_tfd_queue_size;
525 	size_t tb0_buf_sz;
526 	int i;
527 
528 	if (WARN_ON(txq->entries || txq->tfds))
529 		return -EINVAL;
530 
531 	if (trans->trans_cfg->use_tfh)
532 		tfd_sz = trans_pcie->tfd_size * slots_num;
533 
534 	timer_setup(&txq->stuck_timer, iwl_pcie_txq_stuck_timer, 0);
535 	txq->trans = trans;
536 
537 	txq->n_window = slots_num;
538 
539 	txq->entries = kcalloc(slots_num,
540 			       sizeof(struct iwl_pcie_txq_entry),
541 			       GFP_KERNEL);
542 
543 	if (!txq->entries)
544 		goto error;
545 
546 	if (cmd_queue)
547 		for (i = 0; i < slots_num; i++) {
548 			txq->entries[i].cmd =
549 				kmalloc(sizeof(struct iwl_device_cmd),
550 					GFP_KERNEL);
551 			if (!txq->entries[i].cmd)
552 				goto error;
553 		}
554 
555 	/* Circular buffer of transmit frame descriptors (TFDs),
556 	 * shared with device */
557 	txq->tfds = dma_alloc_coherent(trans->dev, tfd_sz,
558 				       &txq->dma_addr, GFP_KERNEL);
559 	if (!txq->tfds)
560 		goto error;
561 
562 	BUILD_BUG_ON(IWL_FIRST_TB_SIZE_ALIGN != sizeof(*txq->first_tb_bufs));
563 
564 	tb0_buf_sz = sizeof(*txq->first_tb_bufs) * slots_num;
565 
566 	txq->first_tb_bufs = dma_alloc_coherent(trans->dev, tb0_buf_sz,
567 					      &txq->first_tb_dma,
568 					      GFP_KERNEL);
569 	if (!txq->first_tb_bufs)
570 		goto err_free_tfds;
571 
572 	return 0;
573 err_free_tfds:
574 	dma_free_coherent(trans->dev, tfd_sz, txq->tfds, txq->dma_addr);
575 error:
576 	if (txq->entries && cmd_queue)
577 		for (i = 0; i < slots_num; i++)
578 			kfree(txq->entries[i].cmd);
579 	kfree(txq->entries);
580 	txq->entries = NULL;
581 
582 	return -ENOMEM;
583 
584 }
585 
586 int iwl_pcie_txq_init(struct iwl_trans *trans, struct iwl_txq *txq,
587 		      int slots_num, bool cmd_queue)
588 {
589 	int ret;
590 	u32 tfd_queue_max_size =
591 		trans->trans_cfg->base_params->max_tfd_queue_size;
592 
593 	txq->need_update = false;
594 
595 	/* max_tfd_queue_size must be power-of-two size, otherwise
596 	 * iwl_queue_inc_wrap and iwl_queue_dec_wrap are broken. */
597 	if (WARN_ONCE(tfd_queue_max_size & (tfd_queue_max_size - 1),
598 		      "Max tfd queue size must be a power of two, but is %d",
599 		      tfd_queue_max_size))
600 		return -EINVAL;
601 
602 	/* Initialize queue's high/low-water marks, and head/tail indexes */
603 	ret = iwl_queue_init(txq, slots_num);
604 	if (ret)
605 		return ret;
606 
607 	spin_lock_init(&txq->lock);
608 
609 	if (cmd_queue) {
610 		static struct lock_class_key iwl_pcie_cmd_queue_lock_class;
611 
612 		lockdep_set_class(&txq->lock, &iwl_pcie_cmd_queue_lock_class);
613 	}
614 
615 	__skb_queue_head_init(&txq->overflow_q);
616 
617 	return 0;
618 }
619 
620 void iwl_pcie_free_tso_page(struct iwl_trans_pcie *trans_pcie,
621 			    struct sk_buff *skb)
622 {
623 	struct page **page_ptr;
624 	struct page *next;
625 
626 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
627 	next = *page_ptr;
628 	*page_ptr = NULL;
629 
630 	while (next) {
631 		struct page *tmp = next;
632 
633 		next = *(void **)(page_address(next) + PAGE_SIZE -
634 				  sizeof(void *));
635 		__free_page(tmp);
636 	}
637 }
638 
639 static void iwl_pcie_clear_cmd_in_flight(struct iwl_trans *trans)
640 {
641 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
642 
643 	lockdep_assert_held(&trans_pcie->reg_lock);
644 
645 	if (!trans->trans_cfg->base_params->apmg_wake_up_wa)
646 		return;
647 	if (WARN_ON(!trans_pcie->cmd_hold_nic_awake))
648 		return;
649 
650 	trans_pcie->cmd_hold_nic_awake = false;
651 	__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
652 				   CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
653 }
654 
655 /*
656  * iwl_pcie_txq_unmap -  Unmap any remaining DMA mappings and free skb's
657  */
658 static void iwl_pcie_txq_unmap(struct iwl_trans *trans, int txq_id)
659 {
660 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
661 	struct iwl_txq *txq = trans->txqs.txq[txq_id];
662 
663 	spin_lock_bh(&txq->lock);
664 	while (txq->write_ptr != txq->read_ptr) {
665 		IWL_DEBUG_TX_REPLY(trans, "Q %d Free %d\n",
666 				   txq_id, txq->read_ptr);
667 
668 		if (txq_id != trans->txqs.cmd.q_id) {
669 			struct sk_buff *skb = txq->entries[txq->read_ptr].skb;
670 
671 			if (WARN_ON_ONCE(!skb))
672 				continue;
673 
674 			iwl_pcie_free_tso_page(trans_pcie, skb);
675 		}
676 		iwl_pcie_txq_free_tfd(trans, txq);
677 		txq->read_ptr = iwl_queue_inc_wrap(trans, txq->read_ptr);
678 
679 		if (txq->read_ptr == txq->write_ptr) {
680 			unsigned long flags;
681 
682 			spin_lock_irqsave(&trans_pcie->reg_lock, flags);
683 			if (txq_id == trans->txqs.cmd.q_id)
684 				iwl_pcie_clear_cmd_in_flight(trans);
685 			spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
686 		}
687 	}
688 
689 	while (!skb_queue_empty(&txq->overflow_q)) {
690 		struct sk_buff *skb = __skb_dequeue(&txq->overflow_q);
691 
692 		iwl_op_mode_free_skb(trans->op_mode, skb);
693 	}
694 
695 	spin_unlock_bh(&txq->lock);
696 
697 	/* just in case - this queue may have been stopped */
698 	iwl_wake_queue(trans, txq);
699 }
700 
701 /*
702  * iwl_pcie_txq_free - Deallocate DMA queue.
703  * @txq: Transmit queue to deallocate.
704  *
705  * Empty queue by removing and destroying all BD's.
706  * Free all buffers.
707  * 0-fill, but do not free "txq" descriptor structure.
708  */
709 static void iwl_pcie_txq_free(struct iwl_trans *trans, int txq_id)
710 {
711 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
712 	struct iwl_txq *txq = trans->txqs.txq[txq_id];
713 	struct device *dev = trans->dev;
714 	int i;
715 
716 	if (WARN_ON(!txq))
717 		return;
718 
719 	iwl_pcie_txq_unmap(trans, txq_id);
720 
721 	/* De-alloc array of command/tx buffers */
722 	if (txq_id == trans->txqs.cmd.q_id)
723 		for (i = 0; i < txq->n_window; i++) {
724 			kfree_sensitive(txq->entries[i].cmd);
725 			kfree_sensitive(txq->entries[i].free_buf);
726 		}
727 
728 	/* De-alloc circular buffer of TFDs */
729 	if (txq->tfds) {
730 		dma_free_coherent(dev,
731 				  trans_pcie->tfd_size *
732 				  trans->trans_cfg->base_params->max_tfd_queue_size,
733 				  txq->tfds, txq->dma_addr);
734 		txq->dma_addr = 0;
735 		txq->tfds = NULL;
736 
737 		dma_free_coherent(dev,
738 				  sizeof(*txq->first_tb_bufs) * txq->n_window,
739 				  txq->first_tb_bufs, txq->first_tb_dma);
740 	}
741 
742 	kfree(txq->entries);
743 	txq->entries = NULL;
744 
745 	del_timer_sync(&txq->stuck_timer);
746 
747 	/* 0-fill queue descriptor structure */
748 	memset(txq, 0, sizeof(*txq));
749 }
750 
751 void iwl_pcie_tx_start(struct iwl_trans *trans, u32 scd_base_addr)
752 {
753 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
754 	int nq = trans->trans_cfg->base_params->num_of_queues;
755 	int chan;
756 	u32 reg_val;
757 	int clear_dwords = (SCD_TRANS_TBL_OFFSET_QUEUE(nq) -
758 				SCD_CONTEXT_MEM_LOWER_BOUND) / sizeof(u32);
759 
760 	/* make sure all queue are not stopped/used */
761 	memset(trans->txqs.queue_stopped, 0,
762 	       sizeof(trans->txqs.queue_stopped));
763 	memset(trans->txqs.queue_used, 0, sizeof(trans->txqs.queue_used));
764 
765 	trans_pcie->scd_base_addr =
766 		iwl_read_prph(trans, SCD_SRAM_BASE_ADDR);
767 
768 	WARN_ON(scd_base_addr != 0 &&
769 		scd_base_addr != trans_pcie->scd_base_addr);
770 
771 	/* reset context data, TX status and translation data */
772 	iwl_trans_write_mem(trans, trans_pcie->scd_base_addr +
773 				   SCD_CONTEXT_MEM_LOWER_BOUND,
774 			    NULL, clear_dwords);
775 
776 	iwl_write_prph(trans, SCD_DRAM_BASE_ADDR,
777 		       trans_pcie->scd_bc_tbls.dma >> 10);
778 
779 	/* The chain extension of the SCD doesn't work well. This feature is
780 	 * enabled by default by the HW, so we need to disable it manually.
781 	 */
782 	if (trans->trans_cfg->base_params->scd_chain_ext_wa)
783 		iwl_write_prph(trans, SCD_CHAINEXT_EN, 0);
784 
785 	iwl_trans_ac_txq_enable(trans, trans->txqs.cmd.q_id,
786 				trans->txqs.cmd.fifo,
787 				trans->txqs.cmd.wdg_timeout);
788 
789 	/* Activate all Tx DMA/FIFO channels */
790 	iwl_scd_activate_fifos(trans);
791 
792 	/* Enable DMA channel */
793 	for (chan = 0; chan < FH_TCSR_CHNL_NUM; chan++)
794 		iwl_write_direct32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(chan),
795 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CHNL_ENABLE |
796 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CREDIT_ENABLE);
797 
798 	/* Update FH chicken bits */
799 	reg_val = iwl_read_direct32(trans, FH_TX_CHICKEN_BITS_REG);
800 	iwl_write_direct32(trans, FH_TX_CHICKEN_BITS_REG,
801 			   reg_val | FH_TX_CHICKEN_BITS_SCD_AUTO_RETRY_EN);
802 
803 	/* Enable L1-Active */
804 	if (trans->trans_cfg->device_family < IWL_DEVICE_FAMILY_8000)
805 		iwl_clear_bits_prph(trans, APMG_PCIDEV_STT_REG,
806 				    APMG_PCIDEV_STT_VAL_L1_ACT_DIS);
807 }
808 
809 void iwl_trans_pcie_tx_reset(struct iwl_trans *trans)
810 {
811 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
812 	int txq_id;
813 
814 	/*
815 	 * we should never get here in gen2 trans mode return early to avoid
816 	 * having invalid accesses
817 	 */
818 	if (WARN_ON_ONCE(trans->trans_cfg->gen2))
819 		return;
820 
821 	for (txq_id = 0; txq_id < trans->trans_cfg->base_params->num_of_queues;
822 	     txq_id++) {
823 		struct iwl_txq *txq = trans->txqs.txq[txq_id];
824 		if (trans->trans_cfg->use_tfh)
825 			iwl_write_direct64(trans,
826 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
827 					   txq->dma_addr);
828 		else
829 			iwl_write_direct32(trans,
830 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
831 					   txq->dma_addr >> 8);
832 		iwl_pcie_txq_unmap(trans, txq_id);
833 		txq->read_ptr = 0;
834 		txq->write_ptr = 0;
835 	}
836 
837 	/* Tell NIC where to find the "keep warm" buffer */
838 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
839 			   trans_pcie->kw.dma >> 4);
840 
841 	/*
842 	 * Send 0 as the scd_base_addr since the device may have be reset
843 	 * while we were in WoWLAN in which case SCD_SRAM_BASE_ADDR will
844 	 * contain garbage.
845 	 */
846 	iwl_pcie_tx_start(trans, 0);
847 }
848 
849 static void iwl_pcie_tx_stop_fh(struct iwl_trans *trans)
850 {
851 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
852 	unsigned long flags;
853 	int ch, ret;
854 	u32 mask = 0;
855 
856 	spin_lock(&trans_pcie->irq_lock);
857 
858 	if (!iwl_trans_grab_nic_access(trans, &flags))
859 		goto out;
860 
861 	/* Stop each Tx DMA channel */
862 	for (ch = 0; ch < FH_TCSR_CHNL_NUM; ch++) {
863 		iwl_write32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(ch), 0x0);
864 		mask |= FH_TSSR_TX_STATUS_REG_MSK_CHNL_IDLE(ch);
865 	}
866 
867 	/* Wait for DMA channels to be idle */
868 	ret = iwl_poll_bit(trans, FH_TSSR_TX_STATUS_REG, mask, mask, 5000);
869 	if (ret < 0)
870 		IWL_ERR(trans,
871 			"Failing on timeout while stopping DMA channel %d [0x%08x]\n",
872 			ch, iwl_read32(trans, FH_TSSR_TX_STATUS_REG));
873 
874 	iwl_trans_release_nic_access(trans, &flags);
875 
876 out:
877 	spin_unlock(&trans_pcie->irq_lock);
878 }
879 
880 /*
881  * iwl_pcie_tx_stop - Stop all Tx DMA channels
882  */
883 int iwl_pcie_tx_stop(struct iwl_trans *trans)
884 {
885 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
886 	int txq_id;
887 
888 	/* Turn off all Tx DMA fifos */
889 	iwl_scd_deactivate_fifos(trans);
890 
891 	/* Turn off all Tx DMA channels */
892 	iwl_pcie_tx_stop_fh(trans);
893 
894 	/*
895 	 * This function can be called before the op_mode disabled the
896 	 * queues. This happens when we have an rfkill interrupt.
897 	 * Since we stop Tx altogether - mark the queues as stopped.
898 	 */
899 	memset(trans->txqs.queue_stopped, 0,
900 	       sizeof(trans->txqs.queue_stopped));
901 	memset(trans->txqs.queue_used, 0, sizeof(trans->txqs.queue_used));
902 
903 	/* This can happen: start_hw, stop_device */
904 	if (!trans_pcie->txq_memory)
905 		return 0;
906 
907 	/* Unmap DMA from host system and free skb's */
908 	for (txq_id = 0; txq_id < trans->trans_cfg->base_params->num_of_queues;
909 	     txq_id++)
910 		iwl_pcie_txq_unmap(trans, txq_id);
911 
912 	return 0;
913 }
914 
915 /*
916  * iwl_trans_tx_free - Free TXQ Context
917  *
918  * Destroy all TX DMA queues and structures
919  */
920 void iwl_pcie_tx_free(struct iwl_trans *trans)
921 {
922 	int txq_id;
923 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
924 
925 	memset(trans->txqs.queue_used, 0, sizeof(trans->txqs.queue_used));
926 
927 	/* Tx queues */
928 	if (trans_pcie->txq_memory) {
929 		for (txq_id = 0;
930 		     txq_id < trans->trans_cfg->base_params->num_of_queues;
931 		     txq_id++) {
932 			iwl_pcie_txq_free(trans, txq_id);
933 			trans->txqs.txq[txq_id] = NULL;
934 		}
935 	}
936 
937 	kfree(trans_pcie->txq_memory);
938 	trans_pcie->txq_memory = NULL;
939 
940 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->kw);
941 
942 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->scd_bc_tbls);
943 }
944 
945 /*
946  * iwl_pcie_tx_alloc - allocate TX context
947  * Allocate all Tx DMA structures and initialize them
948  */
949 static int iwl_pcie_tx_alloc(struct iwl_trans *trans)
950 {
951 	int ret;
952 	int txq_id, slots_num;
953 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
954 	u16 bc_tbls_size = trans->trans_cfg->base_params->num_of_queues;
955 
956 	if (WARN_ON(trans->trans_cfg->device_family >= IWL_DEVICE_FAMILY_AX210))
957 		return -EINVAL;
958 
959 	bc_tbls_size *= sizeof(struct iwlagn_scd_bc_tbl);
960 
961 	/*It is not allowed to alloc twice, so warn when this happens.
962 	 * We cannot rely on the previous allocation, so free and fail */
963 	if (WARN_ON(trans_pcie->txq_memory)) {
964 		ret = -EINVAL;
965 		goto error;
966 	}
967 
968 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->scd_bc_tbls,
969 				     bc_tbls_size);
970 	if (ret) {
971 		IWL_ERR(trans, "Scheduler BC Table allocation failed\n");
972 		goto error;
973 	}
974 
975 	/* Alloc keep-warm buffer */
976 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->kw, IWL_KW_SIZE);
977 	if (ret) {
978 		IWL_ERR(trans, "Keep Warm allocation failed\n");
979 		goto error;
980 	}
981 
982 	trans_pcie->txq_memory =
983 		kcalloc(trans->trans_cfg->base_params->num_of_queues,
984 			sizeof(struct iwl_txq), GFP_KERNEL);
985 	if (!trans_pcie->txq_memory) {
986 		IWL_ERR(trans, "Not enough memory for txq\n");
987 		ret = -ENOMEM;
988 		goto error;
989 	}
990 
991 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
992 	for (txq_id = 0; txq_id < trans->trans_cfg->base_params->num_of_queues;
993 	     txq_id++) {
994 		bool cmd_queue = (txq_id == trans->txqs.cmd.q_id);
995 
996 		if (cmd_queue)
997 			slots_num = max_t(u32, IWL_CMD_QUEUE_SIZE,
998 					  trans->cfg->min_txq_size);
999 		else
1000 			slots_num = max_t(u32, IWL_DEFAULT_QUEUE_SIZE,
1001 					  trans->cfg->min_256_ba_txq_size);
1002 		trans->txqs.txq[txq_id] = &trans_pcie->txq_memory[txq_id];
1003 		ret = iwl_pcie_txq_alloc(trans, trans->txqs.txq[txq_id],
1004 					 slots_num, cmd_queue);
1005 		if (ret) {
1006 			IWL_ERR(trans, "Tx %d queue alloc failed\n", txq_id);
1007 			goto error;
1008 		}
1009 		trans->txqs.txq[txq_id]->id = txq_id;
1010 	}
1011 
1012 	return 0;
1013 
1014 error:
1015 	iwl_pcie_tx_free(trans);
1016 
1017 	return ret;
1018 }
1019 
1020 int iwl_pcie_tx_init(struct iwl_trans *trans)
1021 {
1022 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1023 	int ret;
1024 	int txq_id, slots_num;
1025 	bool alloc = false;
1026 
1027 	if (!trans_pcie->txq_memory) {
1028 		ret = iwl_pcie_tx_alloc(trans);
1029 		if (ret)
1030 			goto error;
1031 		alloc = true;
1032 	}
1033 
1034 	spin_lock(&trans_pcie->irq_lock);
1035 
1036 	/* Turn off all Tx DMA fifos */
1037 	iwl_scd_deactivate_fifos(trans);
1038 
1039 	/* Tell NIC where to find the "keep warm" buffer */
1040 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
1041 			   trans_pcie->kw.dma >> 4);
1042 
1043 	spin_unlock(&trans_pcie->irq_lock);
1044 
1045 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
1046 	for (txq_id = 0; txq_id < trans->trans_cfg->base_params->num_of_queues;
1047 	     txq_id++) {
1048 		bool cmd_queue = (txq_id == trans->txqs.cmd.q_id);
1049 
1050 		if (cmd_queue)
1051 			slots_num = max_t(u32, IWL_CMD_QUEUE_SIZE,
1052 					  trans->cfg->min_txq_size);
1053 		else
1054 			slots_num = max_t(u32, IWL_DEFAULT_QUEUE_SIZE,
1055 					  trans->cfg->min_256_ba_txq_size);
1056 		ret = iwl_pcie_txq_init(trans, trans->txqs.txq[txq_id],
1057 					slots_num, cmd_queue);
1058 		if (ret) {
1059 			IWL_ERR(trans, "Tx %d queue init failed\n", txq_id);
1060 			goto error;
1061 		}
1062 
1063 		/*
1064 		 * Tell nic where to find circular buffer of TFDs for a
1065 		 * given Tx queue, and enable the DMA channel used for that
1066 		 * queue.
1067 		 * Circular buffer (TFD queue in DRAM) physical base address
1068 		 */
1069 		iwl_write_direct32(trans, FH_MEM_CBBC_QUEUE(trans, txq_id),
1070 				   trans->txqs.txq[txq_id]->dma_addr >> 8);
1071 	}
1072 
1073 	iwl_set_bits_prph(trans, SCD_GP_CTRL, SCD_GP_CTRL_AUTO_ACTIVE_MODE);
1074 	if (trans->trans_cfg->base_params->num_of_queues > 20)
1075 		iwl_set_bits_prph(trans, SCD_GP_CTRL,
1076 				  SCD_GP_CTRL_ENABLE_31_QUEUES);
1077 
1078 	return 0;
1079 error:
1080 	/*Upon error, free only if we allocated something */
1081 	if (alloc)
1082 		iwl_pcie_tx_free(trans);
1083 	return ret;
1084 }
1085 
1086 static inline void iwl_pcie_txq_progress(struct iwl_txq *txq)
1087 {
1088 	lockdep_assert_held(&txq->lock);
1089 
1090 	if (!txq->wd_timeout)
1091 		return;
1092 
1093 	/*
1094 	 * station is asleep and we send data - that must
1095 	 * be uAPSD or PS-Poll. Don't rearm the timer.
1096 	 */
1097 	if (txq->frozen)
1098 		return;
1099 
1100 	/*
1101 	 * if empty delete timer, otherwise move timer forward
1102 	 * since we're making progress on this queue
1103 	 */
1104 	if (txq->read_ptr == txq->write_ptr)
1105 		del_timer(&txq->stuck_timer);
1106 	else
1107 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1108 }
1109 
1110 /* Frees buffers until index _not_ inclusive */
1111 void iwl_trans_pcie_reclaim(struct iwl_trans *trans, int txq_id, int ssn,
1112 			    struct sk_buff_head *skbs)
1113 {
1114 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1115 	struct iwl_txq *txq = trans->txqs.txq[txq_id];
1116 	int tfd_num = iwl_pcie_get_cmd_index(txq, ssn);
1117 	int read_ptr = iwl_pcie_get_cmd_index(txq, txq->read_ptr);
1118 	int last_to_free;
1119 
1120 	/* This function is not meant to release cmd queue*/
1121 	if (WARN_ON(txq_id == trans->txqs.cmd.q_id))
1122 		return;
1123 
1124 	spin_lock_bh(&txq->lock);
1125 
1126 	if (!test_bit(txq_id, trans->txqs.queue_used)) {
1127 		IWL_DEBUG_TX_QUEUES(trans, "Q %d inactive - ignoring idx %d\n",
1128 				    txq_id, ssn);
1129 		goto out;
1130 	}
1131 
1132 	if (read_ptr == tfd_num)
1133 		goto out;
1134 
1135 	IWL_DEBUG_TX_REPLY(trans, "[Q %d] %d -> %d (%d)\n",
1136 			   txq_id, txq->read_ptr, tfd_num, ssn);
1137 
1138 	/*Since we free until index _not_ inclusive, the one before index is
1139 	 * the last we will free. This one must be used */
1140 	last_to_free = iwl_queue_dec_wrap(trans, tfd_num);
1141 
1142 	if (!iwl_queue_used(txq, last_to_free)) {
1143 		IWL_ERR(trans,
1144 			"%s: Read index for txq id (%d), last_to_free %d is out of range [0-%d] %d %d.\n",
1145 			__func__, txq_id, last_to_free,
1146 			trans->trans_cfg->base_params->max_tfd_queue_size,
1147 			txq->write_ptr, txq->read_ptr);
1148 		goto out;
1149 	}
1150 
1151 	if (WARN_ON(!skb_queue_empty(skbs)))
1152 		goto out;
1153 
1154 	for (;
1155 	     read_ptr != tfd_num;
1156 	     txq->read_ptr = iwl_queue_inc_wrap(trans, txq->read_ptr),
1157 	     read_ptr = iwl_pcie_get_cmd_index(txq, txq->read_ptr)) {
1158 		struct sk_buff *skb = txq->entries[read_ptr].skb;
1159 
1160 		if (WARN_ON_ONCE(!skb))
1161 			continue;
1162 
1163 		iwl_pcie_free_tso_page(trans_pcie, skb);
1164 
1165 		__skb_queue_tail(skbs, skb);
1166 
1167 		txq->entries[read_ptr].skb = NULL;
1168 
1169 		if (!trans->trans_cfg->use_tfh)
1170 			iwl_pcie_txq_inval_byte_cnt_tbl(trans, txq);
1171 
1172 		iwl_pcie_txq_free_tfd(trans, txq);
1173 	}
1174 
1175 	iwl_pcie_txq_progress(txq);
1176 
1177 	if (iwl_queue_space(trans, txq) > txq->low_mark &&
1178 	    test_bit(txq_id, trans->txqs.queue_stopped)) {
1179 		struct sk_buff_head overflow_skbs;
1180 
1181 		__skb_queue_head_init(&overflow_skbs);
1182 		skb_queue_splice_init(&txq->overflow_q, &overflow_skbs);
1183 
1184 		/*
1185 		 * We are going to transmit from the overflow queue.
1186 		 * Remember this state so that wait_for_txq_empty will know we
1187 		 * are adding more packets to the TFD queue. It cannot rely on
1188 		 * the state of &txq->overflow_q, as we just emptied it, but
1189 		 * haven't TXed the content yet.
1190 		 */
1191 		txq->overflow_tx = true;
1192 
1193 		/*
1194 		 * This is tricky: we are in reclaim path which is non
1195 		 * re-entrant, so noone will try to take the access the
1196 		 * txq data from that path. We stopped tx, so we can't
1197 		 * have tx as well. Bottom line, we can unlock and re-lock
1198 		 * later.
1199 		 */
1200 		spin_unlock_bh(&txq->lock);
1201 
1202 		while (!skb_queue_empty(&overflow_skbs)) {
1203 			struct sk_buff *skb = __skb_dequeue(&overflow_skbs);
1204 			struct iwl_device_tx_cmd *dev_cmd_ptr;
1205 
1206 			dev_cmd_ptr = *(void **)((u8 *)skb->cb +
1207 						 trans_pcie->dev_cmd_offs);
1208 
1209 			/*
1210 			 * Note that we can very well be overflowing again.
1211 			 * In that case, iwl_queue_space will be small again
1212 			 * and we won't wake mac80211's queue.
1213 			 */
1214 			iwl_trans_tx(trans, skb, dev_cmd_ptr, txq_id);
1215 		}
1216 
1217 		if (iwl_queue_space(trans, txq) > txq->low_mark)
1218 			iwl_wake_queue(trans, txq);
1219 
1220 		spin_lock_bh(&txq->lock);
1221 		txq->overflow_tx = false;
1222 	}
1223 
1224 out:
1225 	spin_unlock_bh(&txq->lock);
1226 }
1227 
1228 /* Set wr_ptr of specific device and txq  */
1229 void iwl_trans_pcie_set_q_ptrs(struct iwl_trans *trans, int txq_id, int ptr)
1230 {
1231 	struct iwl_txq *txq = trans->txqs.txq[txq_id];
1232 
1233 	spin_lock_bh(&txq->lock);
1234 
1235 	txq->write_ptr = ptr;
1236 	txq->read_ptr = txq->write_ptr;
1237 
1238 	spin_unlock_bh(&txq->lock);
1239 }
1240 
1241 static int iwl_pcie_set_cmd_in_flight(struct iwl_trans *trans,
1242 				      const struct iwl_host_cmd *cmd)
1243 {
1244 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1245 	int ret;
1246 
1247 	lockdep_assert_held(&trans_pcie->reg_lock);
1248 
1249 	/* Make sure the NIC is still alive in the bus */
1250 	if (test_bit(STATUS_TRANS_DEAD, &trans->status))
1251 		return -ENODEV;
1252 
1253 	/*
1254 	 * wake up the NIC to make sure that the firmware will see the host
1255 	 * command - we will let the NIC sleep once all the host commands
1256 	 * returned. This needs to be done only on NICs that have
1257 	 * apmg_wake_up_wa set.
1258 	 */
1259 	if (trans->trans_cfg->base_params->apmg_wake_up_wa &&
1260 	    !trans_pcie->cmd_hold_nic_awake) {
1261 		__iwl_trans_pcie_set_bit(trans, CSR_GP_CNTRL,
1262 					 CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1263 
1264 		ret = iwl_poll_bit(trans, CSR_GP_CNTRL,
1265 				   CSR_GP_CNTRL_REG_VAL_MAC_ACCESS_EN,
1266 				   (CSR_GP_CNTRL_REG_FLAG_MAC_CLOCK_READY |
1267 				    CSR_GP_CNTRL_REG_FLAG_GOING_TO_SLEEP),
1268 				   15000);
1269 		if (ret < 0) {
1270 			__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
1271 					CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1272 			IWL_ERR(trans, "Failed to wake NIC for hcmd\n");
1273 			return -EIO;
1274 		}
1275 		trans_pcie->cmd_hold_nic_awake = true;
1276 	}
1277 
1278 	return 0;
1279 }
1280 
1281 /*
1282  * iwl_pcie_cmdq_reclaim - Reclaim TX command queue entries already Tx'd
1283  *
1284  * When FW advances 'R' index, all entries between old and new 'R' index
1285  * need to be reclaimed. As result, some free space forms.  If there is
1286  * enough free space (> low mark), wake the stack that feeds us.
1287  */
1288 static void iwl_pcie_cmdq_reclaim(struct iwl_trans *trans, int txq_id, int idx)
1289 {
1290 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1291 	struct iwl_txq *txq = trans->txqs.txq[txq_id];
1292 	unsigned long flags;
1293 	int nfreed = 0;
1294 	u16 r;
1295 
1296 	lockdep_assert_held(&txq->lock);
1297 
1298 	idx = iwl_pcie_get_cmd_index(txq, idx);
1299 	r = iwl_pcie_get_cmd_index(txq, txq->read_ptr);
1300 
1301 	if (idx >= trans->trans_cfg->base_params->max_tfd_queue_size ||
1302 	    (!iwl_queue_used(txq, idx))) {
1303 		WARN_ONCE(test_bit(txq_id, trans->txqs.queue_used),
1304 			  "%s: Read index for DMA queue txq id (%d), index %d is out of range [0-%d] %d %d.\n",
1305 			  __func__, txq_id, idx,
1306 			  trans->trans_cfg->base_params->max_tfd_queue_size,
1307 			  txq->write_ptr, txq->read_ptr);
1308 		return;
1309 	}
1310 
1311 	for (idx = iwl_queue_inc_wrap(trans, idx); r != idx;
1312 	     r = iwl_queue_inc_wrap(trans, r)) {
1313 		txq->read_ptr = iwl_queue_inc_wrap(trans, txq->read_ptr);
1314 
1315 		if (nfreed++ > 0) {
1316 			IWL_ERR(trans, "HCMD skipped: index (%d) %d %d\n",
1317 				idx, txq->write_ptr, r);
1318 			iwl_force_nmi(trans);
1319 		}
1320 	}
1321 
1322 	if (txq->read_ptr == txq->write_ptr) {
1323 		spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1324 		iwl_pcie_clear_cmd_in_flight(trans);
1325 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1326 	}
1327 
1328 	iwl_pcie_txq_progress(txq);
1329 }
1330 
1331 static int iwl_pcie_txq_set_ratid_map(struct iwl_trans *trans, u16 ra_tid,
1332 				 u16 txq_id)
1333 {
1334 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1335 	u32 tbl_dw_addr;
1336 	u32 tbl_dw;
1337 	u16 scd_q2ratid;
1338 
1339 	scd_q2ratid = ra_tid & SCD_QUEUE_RA_TID_MAP_RATID_MSK;
1340 
1341 	tbl_dw_addr = trans_pcie->scd_base_addr +
1342 			SCD_TRANS_TBL_OFFSET_QUEUE(txq_id);
1343 
1344 	tbl_dw = iwl_trans_read_mem32(trans, tbl_dw_addr);
1345 
1346 	if (txq_id & 0x1)
1347 		tbl_dw = (scd_q2ratid << 16) | (tbl_dw & 0x0000FFFF);
1348 	else
1349 		tbl_dw = scd_q2ratid | (tbl_dw & 0xFFFF0000);
1350 
1351 	iwl_trans_write_mem32(trans, tbl_dw_addr, tbl_dw);
1352 
1353 	return 0;
1354 }
1355 
1356 /* Receiver address (actually, Rx station's index into station table),
1357  * combined with Traffic ID (QOS priority), in format used by Tx Scheduler */
1358 #define BUILD_RAxTID(sta_id, tid)	(((sta_id) << 4) + (tid))
1359 
1360 bool iwl_trans_pcie_txq_enable(struct iwl_trans *trans, int txq_id, u16 ssn,
1361 			       const struct iwl_trans_txq_scd_cfg *cfg,
1362 			       unsigned int wdg_timeout)
1363 {
1364 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1365 	struct iwl_txq *txq = trans->txqs.txq[txq_id];
1366 	int fifo = -1;
1367 	bool scd_bug = false;
1368 
1369 	if (test_and_set_bit(txq_id, trans->txqs.queue_used))
1370 		WARN_ONCE(1, "queue %d already used - expect issues", txq_id);
1371 
1372 	txq->wd_timeout = msecs_to_jiffies(wdg_timeout);
1373 
1374 	if (cfg) {
1375 		fifo = cfg->fifo;
1376 
1377 		/* Disable the scheduler prior configuring the cmd queue */
1378 		if (txq_id == trans->txqs.cmd.q_id &&
1379 		    trans_pcie->scd_set_active)
1380 			iwl_scd_enable_set_active(trans, 0);
1381 
1382 		/* Stop this Tx queue before configuring it */
1383 		iwl_scd_txq_set_inactive(trans, txq_id);
1384 
1385 		/* Set this queue as a chain-building queue unless it is CMD */
1386 		if (txq_id != trans->txqs.cmd.q_id)
1387 			iwl_scd_txq_set_chain(trans, txq_id);
1388 
1389 		if (cfg->aggregate) {
1390 			u16 ra_tid = BUILD_RAxTID(cfg->sta_id, cfg->tid);
1391 
1392 			/* Map receiver-address / traffic-ID to this queue */
1393 			iwl_pcie_txq_set_ratid_map(trans, ra_tid, txq_id);
1394 
1395 			/* enable aggregations for the queue */
1396 			iwl_scd_txq_enable_agg(trans, txq_id);
1397 			txq->ampdu = true;
1398 		} else {
1399 			/*
1400 			 * disable aggregations for the queue, this will also
1401 			 * make the ra_tid mapping configuration irrelevant
1402 			 * since it is now a non-AGG queue.
1403 			 */
1404 			iwl_scd_txq_disable_agg(trans, txq_id);
1405 
1406 			ssn = txq->read_ptr;
1407 		}
1408 	} else {
1409 		/*
1410 		 * If we need to move the SCD write pointer by steps of
1411 		 * 0x40, 0x80 or 0xc0, it gets stuck. Avoids this and let
1412 		 * the op_mode know by returning true later.
1413 		 * Do this only in case cfg is NULL since this trick can
1414 		 * be done only if we have DQA enabled which is true for mvm
1415 		 * only. And mvm never sets a cfg pointer.
1416 		 * This is really ugly, but this is the easiest way out for
1417 		 * this sad hardware issue.
1418 		 * This bug has been fixed on devices 9000 and up.
1419 		 */
1420 		scd_bug = !trans->trans_cfg->mq_rx_supported &&
1421 			!((ssn - txq->write_ptr) & 0x3f) &&
1422 			(ssn != txq->write_ptr);
1423 		if (scd_bug)
1424 			ssn++;
1425 	}
1426 
1427 	/* Place first TFD at index corresponding to start sequence number.
1428 	 * Assumes that ssn_idx is valid (!= 0xFFF) */
1429 	txq->read_ptr = (ssn & 0xff);
1430 	txq->write_ptr = (ssn & 0xff);
1431 	iwl_write_direct32(trans, HBUS_TARG_WRPTR,
1432 			   (ssn & 0xff) | (txq_id << 8));
1433 
1434 	if (cfg) {
1435 		u8 frame_limit = cfg->frame_limit;
1436 
1437 		iwl_write_prph(trans, SCD_QUEUE_RDPTR(txq_id), ssn);
1438 
1439 		/* Set up Tx window size and frame limit for this queue */
1440 		iwl_trans_write_mem32(trans, trans_pcie->scd_base_addr +
1441 				SCD_CONTEXT_QUEUE_OFFSET(txq_id), 0);
1442 		iwl_trans_write_mem32(trans,
1443 			trans_pcie->scd_base_addr +
1444 			SCD_CONTEXT_QUEUE_OFFSET(txq_id) + sizeof(u32),
1445 			SCD_QUEUE_CTX_REG2_VAL(WIN_SIZE, frame_limit) |
1446 			SCD_QUEUE_CTX_REG2_VAL(FRAME_LIMIT, frame_limit));
1447 
1448 		/* Set up status area in SRAM, map to Tx DMA/FIFO, activate */
1449 		iwl_write_prph(trans, SCD_QUEUE_STATUS_BITS(txq_id),
1450 			       (1 << SCD_QUEUE_STTS_REG_POS_ACTIVE) |
1451 			       (cfg->fifo << SCD_QUEUE_STTS_REG_POS_TXF) |
1452 			       (1 << SCD_QUEUE_STTS_REG_POS_WSL) |
1453 			       SCD_QUEUE_STTS_REG_MSK);
1454 
1455 		/* enable the scheduler for this queue (only) */
1456 		if (txq_id == trans->txqs.cmd.q_id &&
1457 		    trans_pcie->scd_set_active)
1458 			iwl_scd_enable_set_active(trans, BIT(txq_id));
1459 
1460 		IWL_DEBUG_TX_QUEUES(trans,
1461 				    "Activate queue %d on FIFO %d WrPtr: %d\n",
1462 				    txq_id, fifo, ssn & 0xff);
1463 	} else {
1464 		IWL_DEBUG_TX_QUEUES(trans,
1465 				    "Activate queue %d WrPtr: %d\n",
1466 				    txq_id, ssn & 0xff);
1467 	}
1468 
1469 	return scd_bug;
1470 }
1471 
1472 void iwl_trans_pcie_txq_set_shared_mode(struct iwl_trans *trans, u32 txq_id,
1473 					bool shared_mode)
1474 {
1475 	struct iwl_txq *txq = trans->txqs.txq[txq_id];
1476 
1477 	txq->ampdu = !shared_mode;
1478 }
1479 
1480 void iwl_trans_pcie_txq_disable(struct iwl_trans *trans, int txq_id,
1481 				bool configure_scd)
1482 {
1483 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1484 	u32 stts_addr = trans_pcie->scd_base_addr +
1485 			SCD_TX_STTS_QUEUE_OFFSET(txq_id);
1486 	static const u32 zero_val[4] = {};
1487 
1488 	trans->txqs.txq[txq_id]->frozen_expiry_remainder = 0;
1489 	trans->txqs.txq[txq_id]->frozen = false;
1490 
1491 	/*
1492 	 * Upon HW Rfkill - we stop the device, and then stop the queues
1493 	 * in the op_mode. Just for the sake of the simplicity of the op_mode,
1494 	 * allow the op_mode to call txq_disable after it already called
1495 	 * stop_device.
1496 	 */
1497 	if (!test_and_clear_bit(txq_id, trans->txqs.queue_used)) {
1498 		WARN_ONCE(test_bit(STATUS_DEVICE_ENABLED, &trans->status),
1499 			  "queue %d not used", txq_id);
1500 		return;
1501 	}
1502 
1503 	if (configure_scd) {
1504 		iwl_scd_txq_set_inactive(trans, txq_id);
1505 
1506 		iwl_trans_write_mem(trans, stts_addr, (void *)zero_val,
1507 				    ARRAY_SIZE(zero_val));
1508 	}
1509 
1510 	iwl_pcie_txq_unmap(trans, txq_id);
1511 	trans->txqs.txq[txq_id]->ampdu = false;
1512 
1513 	IWL_DEBUG_TX_QUEUES(trans, "Deactivate queue %d\n", txq_id);
1514 }
1515 
1516 /*************** HOST COMMAND QUEUE FUNCTIONS   *****/
1517 
1518 /*
1519  * iwl_pcie_enqueue_hcmd - enqueue a uCode command
1520  * @priv: device private data point
1521  * @cmd: a pointer to the ucode command structure
1522  *
1523  * The function returns < 0 values to indicate the operation
1524  * failed. On success, it returns the index (>= 0) of command in the
1525  * command queue.
1526  */
1527 static int iwl_pcie_enqueue_hcmd(struct iwl_trans *trans,
1528 				 struct iwl_host_cmd *cmd)
1529 {
1530 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1531 	struct iwl_txq *txq = trans->txqs.txq[trans->txqs.cmd.q_id];
1532 	struct iwl_device_cmd *out_cmd;
1533 	struct iwl_cmd_meta *out_meta;
1534 	unsigned long flags;
1535 	void *dup_buf = NULL;
1536 	dma_addr_t phys_addr;
1537 	int idx;
1538 	u16 copy_size, cmd_size, tb0_size;
1539 	bool had_nocopy = false;
1540 	u8 group_id = iwl_cmd_groupid(cmd->id);
1541 	int i, ret;
1542 	u32 cmd_pos;
1543 	const u8 *cmddata[IWL_MAX_CMD_TBS_PER_TFD];
1544 	u16 cmdlen[IWL_MAX_CMD_TBS_PER_TFD];
1545 
1546 	if (WARN(!trans->wide_cmd_header &&
1547 		 group_id > IWL_ALWAYS_LONG_GROUP,
1548 		 "unsupported wide command %#x\n", cmd->id))
1549 		return -EINVAL;
1550 
1551 	if (group_id != 0) {
1552 		copy_size = sizeof(struct iwl_cmd_header_wide);
1553 		cmd_size = sizeof(struct iwl_cmd_header_wide);
1554 	} else {
1555 		copy_size = sizeof(struct iwl_cmd_header);
1556 		cmd_size = sizeof(struct iwl_cmd_header);
1557 	}
1558 
1559 	/* need one for the header if the first is NOCOPY */
1560 	BUILD_BUG_ON(IWL_MAX_CMD_TBS_PER_TFD > IWL_NUM_OF_TBS - 1);
1561 
1562 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1563 		cmddata[i] = cmd->data[i];
1564 		cmdlen[i] = cmd->len[i];
1565 
1566 		if (!cmd->len[i])
1567 			continue;
1568 
1569 		/* need at least IWL_FIRST_TB_SIZE copied */
1570 		if (copy_size < IWL_FIRST_TB_SIZE) {
1571 			int copy = IWL_FIRST_TB_SIZE - copy_size;
1572 
1573 			if (copy > cmdlen[i])
1574 				copy = cmdlen[i];
1575 			cmdlen[i] -= copy;
1576 			cmddata[i] += copy;
1577 			copy_size += copy;
1578 		}
1579 
1580 		if (cmd->dataflags[i] & IWL_HCMD_DFL_NOCOPY) {
1581 			had_nocopy = true;
1582 			if (WARN_ON(cmd->dataflags[i] & IWL_HCMD_DFL_DUP)) {
1583 				idx = -EINVAL;
1584 				goto free_dup_buf;
1585 			}
1586 		} else if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP) {
1587 			/*
1588 			 * This is also a chunk that isn't copied
1589 			 * to the static buffer so set had_nocopy.
1590 			 */
1591 			had_nocopy = true;
1592 
1593 			/* only allowed once */
1594 			if (WARN_ON(dup_buf)) {
1595 				idx = -EINVAL;
1596 				goto free_dup_buf;
1597 			}
1598 
1599 			dup_buf = kmemdup(cmddata[i], cmdlen[i],
1600 					  GFP_ATOMIC);
1601 			if (!dup_buf)
1602 				return -ENOMEM;
1603 		} else {
1604 			/* NOCOPY must not be followed by normal! */
1605 			if (WARN_ON(had_nocopy)) {
1606 				idx = -EINVAL;
1607 				goto free_dup_buf;
1608 			}
1609 			copy_size += cmdlen[i];
1610 		}
1611 		cmd_size += cmd->len[i];
1612 	}
1613 
1614 	/*
1615 	 * If any of the command structures end up being larger than
1616 	 * the TFD_MAX_PAYLOAD_SIZE and they aren't dynamically
1617 	 * allocated into separate TFDs, then we will need to
1618 	 * increase the size of the buffers.
1619 	 */
1620 	if (WARN(copy_size > TFD_MAX_PAYLOAD_SIZE,
1621 		 "Command %s (%#x) is too large (%d bytes)\n",
1622 		 iwl_get_cmd_string(trans, cmd->id),
1623 		 cmd->id, copy_size)) {
1624 		idx = -EINVAL;
1625 		goto free_dup_buf;
1626 	}
1627 
1628 	spin_lock_bh(&txq->lock);
1629 
1630 	if (iwl_queue_space(trans, txq) < ((cmd->flags & CMD_ASYNC) ? 2 : 1)) {
1631 		spin_unlock_bh(&txq->lock);
1632 
1633 		IWL_ERR(trans, "No space in command queue\n");
1634 		iwl_op_mode_cmd_queue_full(trans->op_mode);
1635 		idx = -ENOSPC;
1636 		goto free_dup_buf;
1637 	}
1638 
1639 	idx = iwl_pcie_get_cmd_index(txq, txq->write_ptr);
1640 	out_cmd = txq->entries[idx].cmd;
1641 	out_meta = &txq->entries[idx].meta;
1642 
1643 	memset(out_meta, 0, sizeof(*out_meta));	/* re-initialize to NULL */
1644 	if (cmd->flags & CMD_WANT_SKB)
1645 		out_meta->source = cmd;
1646 
1647 	/* set up the header */
1648 	if (group_id != 0) {
1649 		out_cmd->hdr_wide.cmd = iwl_cmd_opcode(cmd->id);
1650 		out_cmd->hdr_wide.group_id = group_id;
1651 		out_cmd->hdr_wide.version = iwl_cmd_version(cmd->id);
1652 		out_cmd->hdr_wide.length =
1653 			cpu_to_le16(cmd_size -
1654 				    sizeof(struct iwl_cmd_header_wide));
1655 		out_cmd->hdr_wide.reserved = 0;
1656 		out_cmd->hdr_wide.sequence =
1657 			cpu_to_le16(QUEUE_TO_SEQ(trans->txqs.cmd.q_id) |
1658 						 INDEX_TO_SEQ(txq->write_ptr));
1659 
1660 		cmd_pos = sizeof(struct iwl_cmd_header_wide);
1661 		copy_size = sizeof(struct iwl_cmd_header_wide);
1662 	} else {
1663 		out_cmd->hdr.cmd = iwl_cmd_opcode(cmd->id);
1664 		out_cmd->hdr.sequence =
1665 			cpu_to_le16(QUEUE_TO_SEQ(trans->txqs.cmd.q_id) |
1666 						 INDEX_TO_SEQ(txq->write_ptr));
1667 		out_cmd->hdr.group_id = 0;
1668 
1669 		cmd_pos = sizeof(struct iwl_cmd_header);
1670 		copy_size = sizeof(struct iwl_cmd_header);
1671 	}
1672 
1673 	/* and copy the data that needs to be copied */
1674 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1675 		int copy;
1676 
1677 		if (!cmd->len[i])
1678 			continue;
1679 
1680 		/* copy everything if not nocopy/dup */
1681 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1682 					   IWL_HCMD_DFL_DUP))) {
1683 			copy = cmd->len[i];
1684 
1685 			memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1686 			cmd_pos += copy;
1687 			copy_size += copy;
1688 			continue;
1689 		}
1690 
1691 		/*
1692 		 * Otherwise we need at least IWL_FIRST_TB_SIZE copied
1693 		 * in total (for bi-directional DMA), but copy up to what
1694 		 * we can fit into the payload for debug dump purposes.
1695 		 */
1696 		copy = min_t(int, TFD_MAX_PAYLOAD_SIZE - cmd_pos, cmd->len[i]);
1697 
1698 		memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1699 		cmd_pos += copy;
1700 
1701 		/* However, treat copy_size the proper way, we need it below */
1702 		if (copy_size < IWL_FIRST_TB_SIZE) {
1703 			copy = IWL_FIRST_TB_SIZE - copy_size;
1704 
1705 			if (copy > cmd->len[i])
1706 				copy = cmd->len[i];
1707 			copy_size += copy;
1708 		}
1709 	}
1710 
1711 	IWL_DEBUG_HC(trans,
1712 		     "Sending command %s (%.2x.%.2x), seq: 0x%04X, %d bytes at %d[%d]:%d\n",
1713 		     iwl_get_cmd_string(trans, cmd->id),
1714 		     group_id, out_cmd->hdr.cmd,
1715 		     le16_to_cpu(out_cmd->hdr.sequence),
1716 		     cmd_size, txq->write_ptr, idx, trans->txqs.cmd.q_id);
1717 
1718 	/* start the TFD with the minimum copy bytes */
1719 	tb0_size = min_t(int, copy_size, IWL_FIRST_TB_SIZE);
1720 	memcpy(&txq->first_tb_bufs[idx], &out_cmd->hdr, tb0_size);
1721 	iwl_pcie_txq_build_tfd(trans, txq,
1722 			       iwl_pcie_get_first_tb_dma(txq, idx),
1723 			       tb0_size, true);
1724 
1725 	/* map first command fragment, if any remains */
1726 	if (copy_size > tb0_size) {
1727 		phys_addr = dma_map_single(trans->dev,
1728 					   ((u8 *)&out_cmd->hdr) + tb0_size,
1729 					   copy_size - tb0_size,
1730 					   DMA_TO_DEVICE);
1731 		if (dma_mapping_error(trans->dev, phys_addr)) {
1732 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1733 					   txq->write_ptr);
1734 			idx = -ENOMEM;
1735 			goto out;
1736 		}
1737 
1738 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr,
1739 				       copy_size - tb0_size, false);
1740 	}
1741 
1742 	/* map the remaining (adjusted) nocopy/dup fragments */
1743 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1744 		const void *data = cmddata[i];
1745 
1746 		if (!cmdlen[i])
1747 			continue;
1748 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1749 					   IWL_HCMD_DFL_DUP)))
1750 			continue;
1751 		if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP)
1752 			data = dup_buf;
1753 		phys_addr = dma_map_single(trans->dev, (void *)data,
1754 					   cmdlen[i], DMA_TO_DEVICE);
1755 		if (dma_mapping_error(trans->dev, phys_addr)) {
1756 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1757 					   txq->write_ptr);
1758 			idx = -ENOMEM;
1759 			goto out;
1760 		}
1761 
1762 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr, cmdlen[i], false);
1763 	}
1764 
1765 	BUILD_BUG_ON(IWL_TFH_NUM_TBS > sizeof(out_meta->tbs) * BITS_PER_BYTE);
1766 	out_meta->flags = cmd->flags;
1767 	if (WARN_ON_ONCE(txq->entries[idx].free_buf))
1768 		kfree_sensitive(txq->entries[idx].free_buf);
1769 	txq->entries[idx].free_buf = dup_buf;
1770 
1771 	trace_iwlwifi_dev_hcmd(trans->dev, cmd, cmd_size, &out_cmd->hdr_wide);
1772 
1773 	/* start timer if queue currently empty */
1774 	if (txq->read_ptr == txq->write_ptr && txq->wd_timeout)
1775 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1776 
1777 	spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1778 	ret = iwl_pcie_set_cmd_in_flight(trans, cmd);
1779 	if (ret < 0) {
1780 		idx = ret;
1781 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1782 		goto out;
1783 	}
1784 
1785 	/* Increment and update queue's write index */
1786 	txq->write_ptr = iwl_queue_inc_wrap(trans, txq->write_ptr);
1787 	iwl_pcie_txq_inc_wr_ptr(trans, txq);
1788 
1789 	spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1790 
1791  out:
1792 	spin_unlock_bh(&txq->lock);
1793  free_dup_buf:
1794 	if (idx < 0)
1795 		kfree(dup_buf);
1796 	return idx;
1797 }
1798 
1799 /*
1800  * iwl_pcie_hcmd_complete - Pull unused buffers off the queue and reclaim them
1801  * @rxb: Rx buffer to reclaim
1802  */
1803 void iwl_pcie_hcmd_complete(struct iwl_trans *trans,
1804 			    struct iwl_rx_cmd_buffer *rxb)
1805 {
1806 	struct iwl_rx_packet *pkt = rxb_addr(rxb);
1807 	u16 sequence = le16_to_cpu(pkt->hdr.sequence);
1808 	u8 group_id;
1809 	u32 cmd_id;
1810 	int txq_id = SEQ_TO_QUEUE(sequence);
1811 	int index = SEQ_TO_INDEX(sequence);
1812 	int cmd_index;
1813 	struct iwl_device_cmd *cmd;
1814 	struct iwl_cmd_meta *meta;
1815 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1816 	struct iwl_txq *txq = trans->txqs.txq[trans->txqs.cmd.q_id];
1817 
1818 	/* If a Tx command is being handled and it isn't in the actual
1819 	 * command queue then there a command routing bug has been introduced
1820 	 * in the queue management code. */
1821 	if (WARN(txq_id != trans->txqs.cmd.q_id,
1822 		 "wrong command queue %d (should be %d), sequence 0x%X readp=%d writep=%d\n",
1823 		 txq_id, trans->txqs.cmd.q_id, sequence, txq->read_ptr,
1824 		 txq->write_ptr)) {
1825 		iwl_print_hex_error(trans, pkt, 32);
1826 		return;
1827 	}
1828 
1829 	spin_lock_bh(&txq->lock);
1830 
1831 	cmd_index = iwl_pcie_get_cmd_index(txq, index);
1832 	cmd = txq->entries[cmd_index].cmd;
1833 	meta = &txq->entries[cmd_index].meta;
1834 	group_id = cmd->hdr.group_id;
1835 	cmd_id = iwl_cmd_id(cmd->hdr.cmd, group_id, 0);
1836 
1837 	iwl_pcie_tfd_unmap(trans, meta, txq, index);
1838 
1839 	/* Input error checking is done when commands are added to queue. */
1840 	if (meta->flags & CMD_WANT_SKB) {
1841 		struct page *p = rxb_steal_page(rxb);
1842 
1843 		meta->source->resp_pkt = pkt;
1844 		meta->source->_rx_page_addr = (unsigned long)page_address(p);
1845 		meta->source->_rx_page_order = trans_pcie->rx_page_order;
1846 	}
1847 
1848 	if (meta->flags & CMD_WANT_ASYNC_CALLBACK)
1849 		iwl_op_mode_async_cb(trans->op_mode, cmd);
1850 
1851 	iwl_pcie_cmdq_reclaim(trans, txq_id, index);
1852 
1853 	if (!(meta->flags & CMD_ASYNC)) {
1854 		if (!test_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status)) {
1855 			IWL_WARN(trans,
1856 				 "HCMD_ACTIVE already clear for command %s\n",
1857 				 iwl_get_cmd_string(trans, cmd_id));
1858 		}
1859 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1860 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1861 			       iwl_get_cmd_string(trans, cmd_id));
1862 		wake_up(&trans_pcie->wait_command_queue);
1863 	}
1864 
1865 	meta->flags = 0;
1866 
1867 	spin_unlock_bh(&txq->lock);
1868 }
1869 
1870 #define HOST_COMPLETE_TIMEOUT	(2 * HZ)
1871 
1872 static int iwl_pcie_send_hcmd_async(struct iwl_trans *trans,
1873 				    struct iwl_host_cmd *cmd)
1874 {
1875 	int ret;
1876 
1877 	/* An asynchronous command can not expect an SKB to be set. */
1878 	if (WARN_ON(cmd->flags & CMD_WANT_SKB))
1879 		return -EINVAL;
1880 
1881 	ret = iwl_pcie_enqueue_hcmd(trans, cmd);
1882 	if (ret < 0) {
1883 		IWL_ERR(trans,
1884 			"Error sending %s: enqueue_hcmd failed: %d\n",
1885 			iwl_get_cmd_string(trans, cmd->id), ret);
1886 		return ret;
1887 	}
1888 	return 0;
1889 }
1890 
1891 static int iwl_pcie_send_hcmd_sync(struct iwl_trans *trans,
1892 				   struct iwl_host_cmd *cmd)
1893 {
1894 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1895 	struct iwl_txq *txq = trans->txqs.txq[trans->txqs.cmd.q_id];
1896 	int cmd_idx;
1897 	int ret;
1898 
1899 	IWL_DEBUG_INFO(trans, "Attempting to send sync command %s\n",
1900 		       iwl_get_cmd_string(trans, cmd->id));
1901 
1902 	if (WARN(test_and_set_bit(STATUS_SYNC_HCMD_ACTIVE,
1903 				  &trans->status),
1904 		 "Command %s: a command is already active!\n",
1905 		 iwl_get_cmd_string(trans, cmd->id)))
1906 		return -EIO;
1907 
1908 	IWL_DEBUG_INFO(trans, "Setting HCMD_ACTIVE for command %s\n",
1909 		       iwl_get_cmd_string(trans, cmd->id));
1910 
1911 	cmd_idx = iwl_pcie_enqueue_hcmd(trans, cmd);
1912 	if (cmd_idx < 0) {
1913 		ret = cmd_idx;
1914 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1915 		IWL_ERR(trans,
1916 			"Error sending %s: enqueue_hcmd failed: %d\n",
1917 			iwl_get_cmd_string(trans, cmd->id), ret);
1918 		return ret;
1919 	}
1920 
1921 	ret = wait_event_timeout(trans_pcie->wait_command_queue,
1922 				 !test_bit(STATUS_SYNC_HCMD_ACTIVE,
1923 					   &trans->status),
1924 				 HOST_COMPLETE_TIMEOUT);
1925 	if (!ret) {
1926 		IWL_ERR(trans, "Error sending %s: time out after %dms.\n",
1927 			iwl_get_cmd_string(trans, cmd->id),
1928 			jiffies_to_msecs(HOST_COMPLETE_TIMEOUT));
1929 
1930 		IWL_ERR(trans, "Current CMD queue read_ptr %d write_ptr %d\n",
1931 			txq->read_ptr, txq->write_ptr);
1932 
1933 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1934 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1935 			       iwl_get_cmd_string(trans, cmd->id));
1936 		ret = -ETIMEDOUT;
1937 
1938 		iwl_trans_pcie_sync_nmi(trans);
1939 		goto cancel;
1940 	}
1941 
1942 	if (test_bit(STATUS_FW_ERROR, &trans->status)) {
1943 		iwl_trans_pcie_dump_regs(trans);
1944 		IWL_ERR(trans, "FW error in SYNC CMD %s\n",
1945 			iwl_get_cmd_string(trans, cmd->id));
1946 		dump_stack();
1947 		ret = -EIO;
1948 		goto cancel;
1949 	}
1950 
1951 	if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1952 	    test_bit(STATUS_RFKILL_OPMODE, &trans->status)) {
1953 		IWL_DEBUG_RF_KILL(trans, "RFKILL in SYNC CMD... no rsp\n");
1954 		ret = -ERFKILL;
1955 		goto cancel;
1956 	}
1957 
1958 	if ((cmd->flags & CMD_WANT_SKB) && !cmd->resp_pkt) {
1959 		IWL_ERR(trans, "Error: Response NULL in '%s'\n",
1960 			iwl_get_cmd_string(trans, cmd->id));
1961 		ret = -EIO;
1962 		goto cancel;
1963 	}
1964 
1965 	return 0;
1966 
1967 cancel:
1968 	if (cmd->flags & CMD_WANT_SKB) {
1969 		/*
1970 		 * Cancel the CMD_WANT_SKB flag for the cmd in the
1971 		 * TX cmd queue. Otherwise in case the cmd comes
1972 		 * in later, it will possibly set an invalid
1973 		 * address (cmd->meta.source).
1974 		 */
1975 		txq->entries[cmd_idx].meta.flags &= ~CMD_WANT_SKB;
1976 	}
1977 
1978 	if (cmd->resp_pkt) {
1979 		iwl_free_resp(cmd);
1980 		cmd->resp_pkt = NULL;
1981 	}
1982 
1983 	return ret;
1984 }
1985 
1986 int iwl_trans_pcie_send_hcmd(struct iwl_trans *trans, struct iwl_host_cmd *cmd)
1987 {
1988 	/* Make sure the NIC is still alive in the bus */
1989 	if (test_bit(STATUS_TRANS_DEAD, &trans->status))
1990 		return -ENODEV;
1991 
1992 	if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1993 	    test_bit(STATUS_RFKILL_OPMODE, &trans->status)) {
1994 		IWL_DEBUG_RF_KILL(trans, "Dropping CMD 0x%x: RF KILL\n",
1995 				  cmd->id);
1996 		return -ERFKILL;
1997 	}
1998 
1999 	if (cmd->flags & CMD_ASYNC)
2000 		return iwl_pcie_send_hcmd_async(trans, cmd);
2001 
2002 	/* We still can fail on RFKILL that can be asserted while we wait */
2003 	return iwl_pcie_send_hcmd_sync(trans, cmd);
2004 }
2005 
2006 static int iwl_fill_data_tbs(struct iwl_trans *trans, struct sk_buff *skb,
2007 			     struct iwl_txq *txq, u8 hdr_len,
2008 			     struct iwl_cmd_meta *out_meta)
2009 {
2010 	u16 head_tb_len;
2011 	int i;
2012 
2013 	/*
2014 	 * Set up TFD's third entry to point directly to remainder
2015 	 * of skb's head, if any
2016 	 */
2017 	head_tb_len = skb_headlen(skb) - hdr_len;
2018 
2019 	if (head_tb_len > 0) {
2020 		dma_addr_t tb_phys = dma_map_single(trans->dev,
2021 						    skb->data + hdr_len,
2022 						    head_tb_len, DMA_TO_DEVICE);
2023 		if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
2024 			return -EINVAL;
2025 		trace_iwlwifi_dev_tx_tb(trans->dev, skb, skb->data + hdr_len,
2026 					tb_phys, head_tb_len);
2027 		iwl_pcie_txq_build_tfd(trans, txq, tb_phys, head_tb_len, false);
2028 	}
2029 
2030 	/* set up the remaining entries to point to the data */
2031 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2032 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2033 		dma_addr_t tb_phys;
2034 		int tb_idx;
2035 
2036 		if (!skb_frag_size(frag))
2037 			continue;
2038 
2039 		tb_phys = skb_frag_dma_map(trans->dev, frag, 0,
2040 					   skb_frag_size(frag), DMA_TO_DEVICE);
2041 
2042 		if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
2043 			return -EINVAL;
2044 		trace_iwlwifi_dev_tx_tb(trans->dev, skb, skb_frag_address(frag),
2045 					tb_phys, skb_frag_size(frag));
2046 		tb_idx = iwl_pcie_txq_build_tfd(trans, txq, tb_phys,
2047 						skb_frag_size(frag), false);
2048 		if (tb_idx < 0)
2049 			return tb_idx;
2050 
2051 		out_meta->tbs |= BIT(tb_idx);
2052 	}
2053 
2054 	return 0;
2055 }
2056 
2057 #ifdef CONFIG_INET
2058 struct iwl_tso_hdr_page *get_page_hdr(struct iwl_trans *trans, size_t len,
2059 				      struct sk_buff *skb)
2060 {
2061 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
2062 	struct iwl_tso_hdr_page *p = this_cpu_ptr(trans_pcie->tso_hdr_page);
2063 	struct page **page_ptr;
2064 
2065 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
2066 
2067 	if (WARN_ON(*page_ptr))
2068 		return NULL;
2069 
2070 	if (!p->page)
2071 		goto alloc;
2072 
2073 	/*
2074 	 * Check if there's enough room on this page
2075 	 *
2076 	 * Note that we put a page chaining pointer *last* in the
2077 	 * page - we need it somewhere, and if it's there then we
2078 	 * avoid DMA mapping the last bits of the page which may
2079 	 * trigger the 32-bit boundary hardware bug.
2080 	 *
2081 	 * (see also get_workaround_page() in tx-gen2.c)
2082 	 */
2083 	if (p->pos + len < (u8 *)page_address(p->page) + PAGE_SIZE -
2084 			   sizeof(void *))
2085 		goto out;
2086 
2087 	/* We don't have enough room on this page, get a new one. */
2088 	__free_page(p->page);
2089 
2090 alloc:
2091 	p->page = alloc_page(GFP_ATOMIC);
2092 	if (!p->page)
2093 		return NULL;
2094 	p->pos = page_address(p->page);
2095 	/* set the chaining pointer to NULL */
2096 	*(void **)(page_address(p->page) + PAGE_SIZE - sizeof(void *)) = NULL;
2097 out:
2098 	*page_ptr = p->page;
2099 	get_page(p->page);
2100 	return p;
2101 }
2102 
2103 static void iwl_compute_pseudo_hdr_csum(void *iph, struct tcphdr *tcph,
2104 					bool ipv6, unsigned int len)
2105 {
2106 	if (ipv6) {
2107 		struct ipv6hdr *iphv6 = iph;
2108 
2109 		tcph->check = ~csum_ipv6_magic(&iphv6->saddr, &iphv6->daddr,
2110 					       len + tcph->doff * 4,
2111 					       IPPROTO_TCP, 0);
2112 	} else {
2113 		struct iphdr *iphv4 = iph;
2114 
2115 		ip_send_check(iphv4);
2116 		tcph->check = ~csum_tcpudp_magic(iphv4->saddr, iphv4->daddr,
2117 						 len + tcph->doff * 4,
2118 						 IPPROTO_TCP, 0);
2119 	}
2120 }
2121 
2122 static int iwl_fill_data_tbs_amsdu(struct iwl_trans *trans, struct sk_buff *skb,
2123 				   struct iwl_txq *txq, u8 hdr_len,
2124 				   struct iwl_cmd_meta *out_meta,
2125 				   struct iwl_device_tx_cmd *dev_cmd,
2126 				   u16 tb1_len)
2127 {
2128 	struct iwl_tx_cmd *tx_cmd = (void *)dev_cmd->payload;
2129 	struct iwl_trans_pcie *trans_pcie =
2130 		IWL_TRANS_GET_PCIE_TRANS(txq->trans);
2131 	struct ieee80211_hdr *hdr = (void *)skb->data;
2132 	unsigned int snap_ip_tcp_hdrlen, ip_hdrlen, total_len, hdr_room;
2133 	unsigned int mss = skb_shinfo(skb)->gso_size;
2134 	u16 length, iv_len, amsdu_pad;
2135 	u8 *start_hdr;
2136 	struct iwl_tso_hdr_page *hdr_page;
2137 	struct tso_t tso;
2138 
2139 	/* if the packet is protected, then it must be CCMP or GCMP */
2140 	BUILD_BUG_ON(IEEE80211_CCMP_HDR_LEN != IEEE80211_GCMP_HDR_LEN);
2141 	iv_len = ieee80211_has_protected(hdr->frame_control) ?
2142 		IEEE80211_CCMP_HDR_LEN : 0;
2143 
2144 	trace_iwlwifi_dev_tx(trans->dev, skb,
2145 			     iwl_pcie_get_tfd(trans, txq, txq->write_ptr),
2146 			     trans_pcie->tfd_size,
2147 			     &dev_cmd->hdr, IWL_FIRST_TB_SIZE + tb1_len, 0);
2148 
2149 	ip_hdrlen = skb_transport_header(skb) - skb_network_header(skb);
2150 	snap_ip_tcp_hdrlen = 8 + ip_hdrlen + tcp_hdrlen(skb);
2151 	total_len = skb->len - snap_ip_tcp_hdrlen - hdr_len - iv_len;
2152 	amsdu_pad = 0;
2153 
2154 	/* total amount of header we may need for this A-MSDU */
2155 	hdr_room = DIV_ROUND_UP(total_len, mss) *
2156 		(3 + snap_ip_tcp_hdrlen + sizeof(struct ethhdr)) + iv_len;
2157 
2158 	/* Our device supports 9 segments at most, it will fit in 1 page */
2159 	hdr_page = get_page_hdr(trans, hdr_room, skb);
2160 	if (!hdr_page)
2161 		return -ENOMEM;
2162 
2163 	start_hdr = hdr_page->pos;
2164 	memcpy(hdr_page->pos, skb->data + hdr_len, iv_len);
2165 	hdr_page->pos += iv_len;
2166 
2167 	/*
2168 	 * Pull the ieee80211 header + IV to be able to use TSO core,
2169 	 * we will restore it for the tx_status flow.
2170 	 */
2171 	skb_pull(skb, hdr_len + iv_len);
2172 
2173 	/*
2174 	 * Remove the length of all the headers that we don't actually
2175 	 * have in the MPDU by themselves, but that we duplicate into
2176 	 * all the different MSDUs inside the A-MSDU.
2177 	 */
2178 	le16_add_cpu(&tx_cmd->len, -snap_ip_tcp_hdrlen);
2179 
2180 	tso_start(skb, &tso);
2181 
2182 	while (total_len) {
2183 		/* this is the data left for this subframe */
2184 		unsigned int data_left =
2185 			min_t(unsigned int, mss, total_len);
2186 		struct sk_buff *csum_skb = NULL;
2187 		unsigned int hdr_tb_len;
2188 		dma_addr_t hdr_tb_phys;
2189 		struct tcphdr *tcph;
2190 		u8 *iph, *subf_hdrs_start = hdr_page->pos;
2191 
2192 		total_len -= data_left;
2193 
2194 		memset(hdr_page->pos, 0, amsdu_pad);
2195 		hdr_page->pos += amsdu_pad;
2196 		amsdu_pad = (4 - (sizeof(struct ethhdr) + snap_ip_tcp_hdrlen +
2197 				  data_left)) & 0x3;
2198 		ether_addr_copy(hdr_page->pos, ieee80211_get_DA(hdr));
2199 		hdr_page->pos += ETH_ALEN;
2200 		ether_addr_copy(hdr_page->pos, ieee80211_get_SA(hdr));
2201 		hdr_page->pos += ETH_ALEN;
2202 
2203 		length = snap_ip_tcp_hdrlen + data_left;
2204 		*((__be16 *)hdr_page->pos) = cpu_to_be16(length);
2205 		hdr_page->pos += sizeof(length);
2206 
2207 		/*
2208 		 * This will copy the SNAP as well which will be considered
2209 		 * as MAC header.
2210 		 */
2211 		tso_build_hdr(skb, hdr_page->pos, &tso, data_left, !total_len);
2212 		iph = hdr_page->pos + 8;
2213 		tcph = (void *)(iph + ip_hdrlen);
2214 
2215 		/* For testing on current hardware only */
2216 		if (trans_pcie->sw_csum_tx) {
2217 			csum_skb = alloc_skb(data_left + tcp_hdrlen(skb),
2218 					     GFP_ATOMIC);
2219 			if (!csum_skb)
2220 				return -ENOMEM;
2221 
2222 			iwl_compute_pseudo_hdr_csum(iph, tcph,
2223 						    skb->protocol ==
2224 							htons(ETH_P_IPV6),
2225 						    data_left);
2226 
2227 			skb_put_data(csum_skb, tcph, tcp_hdrlen(skb));
2228 			skb_reset_transport_header(csum_skb);
2229 			csum_skb->csum_start =
2230 				(unsigned char *)tcp_hdr(csum_skb) -
2231 						 csum_skb->head;
2232 		}
2233 
2234 		hdr_page->pos += snap_ip_tcp_hdrlen;
2235 
2236 		hdr_tb_len = hdr_page->pos - start_hdr;
2237 		hdr_tb_phys = dma_map_single(trans->dev, start_hdr,
2238 					     hdr_tb_len, DMA_TO_DEVICE);
2239 		if (unlikely(dma_mapping_error(trans->dev, hdr_tb_phys))) {
2240 			dev_kfree_skb(csum_skb);
2241 			return -EINVAL;
2242 		}
2243 		iwl_pcie_txq_build_tfd(trans, txq, hdr_tb_phys,
2244 				       hdr_tb_len, false);
2245 		trace_iwlwifi_dev_tx_tb(trans->dev, skb, start_hdr,
2246 					hdr_tb_phys, hdr_tb_len);
2247 		/* add this subframe's headers' length to the tx_cmd */
2248 		le16_add_cpu(&tx_cmd->len, hdr_page->pos - subf_hdrs_start);
2249 
2250 		/* prepare the start_hdr for the next subframe */
2251 		start_hdr = hdr_page->pos;
2252 
2253 		/* put the payload */
2254 		while (data_left) {
2255 			unsigned int size = min_t(unsigned int, tso.size,
2256 						  data_left);
2257 			dma_addr_t tb_phys;
2258 
2259 			if (trans_pcie->sw_csum_tx)
2260 				skb_put_data(csum_skb, tso.data, size);
2261 
2262 			tb_phys = dma_map_single(trans->dev, tso.data,
2263 						 size, DMA_TO_DEVICE);
2264 			if (unlikely(dma_mapping_error(trans->dev, tb_phys))) {
2265 				dev_kfree_skb(csum_skb);
2266 				return -EINVAL;
2267 			}
2268 
2269 			iwl_pcie_txq_build_tfd(trans, txq, tb_phys,
2270 					       size, false);
2271 			trace_iwlwifi_dev_tx_tb(trans->dev, skb, tso.data,
2272 						tb_phys, size);
2273 
2274 			data_left -= size;
2275 			tso_build_data(skb, &tso, size);
2276 		}
2277 
2278 		/* For testing on early hardware only */
2279 		if (trans_pcie->sw_csum_tx) {
2280 			__wsum csum;
2281 
2282 			csum = skb_checksum(csum_skb,
2283 					    skb_checksum_start_offset(csum_skb),
2284 					    csum_skb->len -
2285 					    skb_checksum_start_offset(csum_skb),
2286 					    0);
2287 			dev_kfree_skb(csum_skb);
2288 			dma_sync_single_for_cpu(trans->dev, hdr_tb_phys,
2289 						hdr_tb_len, DMA_TO_DEVICE);
2290 			tcph->check = csum_fold(csum);
2291 			dma_sync_single_for_device(trans->dev, hdr_tb_phys,
2292 						   hdr_tb_len, DMA_TO_DEVICE);
2293 		}
2294 	}
2295 
2296 	/* re -add the WiFi header and IV */
2297 	skb_push(skb, hdr_len + iv_len);
2298 
2299 	return 0;
2300 }
2301 #else /* CONFIG_INET */
2302 static int iwl_fill_data_tbs_amsdu(struct iwl_trans *trans, struct sk_buff *skb,
2303 				   struct iwl_txq *txq, u8 hdr_len,
2304 				   struct iwl_cmd_meta *out_meta,
2305 				   struct iwl_device_tx_cmd *dev_cmd,
2306 				   u16 tb1_len)
2307 {
2308 	/* No A-MSDU without CONFIG_INET */
2309 	WARN_ON(1);
2310 
2311 	return -1;
2312 }
2313 #endif /* CONFIG_INET */
2314 
2315 int iwl_trans_pcie_tx(struct iwl_trans *trans, struct sk_buff *skb,
2316 		      struct iwl_device_tx_cmd *dev_cmd, int txq_id)
2317 {
2318 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
2319 	struct ieee80211_hdr *hdr;
2320 	struct iwl_tx_cmd *tx_cmd = (struct iwl_tx_cmd *)dev_cmd->payload;
2321 	struct iwl_cmd_meta *out_meta;
2322 	struct iwl_txq *txq;
2323 	dma_addr_t tb0_phys, tb1_phys, scratch_phys;
2324 	void *tb1_addr;
2325 	void *tfd;
2326 	u16 len, tb1_len;
2327 	bool wait_write_ptr;
2328 	__le16 fc;
2329 	u8 hdr_len;
2330 	u16 wifi_seq;
2331 	bool amsdu;
2332 
2333 	txq = trans->txqs.txq[txq_id];
2334 
2335 	if (WARN_ONCE(!test_bit(txq_id, trans->txqs.queue_used),
2336 		      "TX on unused queue %d\n", txq_id))
2337 		return -EINVAL;
2338 
2339 	if (unlikely(trans_pcie->sw_csum_tx &&
2340 		     skb->ip_summed == CHECKSUM_PARTIAL)) {
2341 		int offs = skb_checksum_start_offset(skb);
2342 		int csum_offs = offs + skb->csum_offset;
2343 		__wsum csum;
2344 
2345 		if (skb_ensure_writable(skb, csum_offs + sizeof(__sum16)))
2346 			return -1;
2347 
2348 		csum = skb_checksum(skb, offs, skb->len - offs, 0);
2349 		*(__sum16 *)(skb->data + csum_offs) = csum_fold(csum);
2350 
2351 		skb->ip_summed = CHECKSUM_UNNECESSARY;
2352 	}
2353 
2354 	if (skb_is_nonlinear(skb) &&
2355 	    skb_shinfo(skb)->nr_frags > IWL_PCIE_MAX_FRAGS(trans_pcie) &&
2356 	    __skb_linearize(skb))
2357 		return -ENOMEM;
2358 
2359 	/* mac80211 always puts the full header into the SKB's head,
2360 	 * so there's no need to check if it's readable there
2361 	 */
2362 	hdr = (struct ieee80211_hdr *)skb->data;
2363 	fc = hdr->frame_control;
2364 	hdr_len = ieee80211_hdrlen(fc);
2365 
2366 	spin_lock(&txq->lock);
2367 
2368 	if (iwl_queue_space(trans, txq) < txq->high_mark) {
2369 		iwl_stop_queue(trans, txq);
2370 
2371 		/* don't put the packet on the ring, if there is no room */
2372 		if (unlikely(iwl_queue_space(trans, txq) < 3)) {
2373 			struct iwl_device_tx_cmd **dev_cmd_ptr;
2374 
2375 			dev_cmd_ptr = (void *)((u8 *)skb->cb +
2376 					       trans_pcie->dev_cmd_offs);
2377 
2378 			*dev_cmd_ptr = dev_cmd;
2379 			__skb_queue_tail(&txq->overflow_q, skb);
2380 
2381 			spin_unlock(&txq->lock);
2382 			return 0;
2383 		}
2384 	}
2385 
2386 	/* In AGG mode, the index in the ring must correspond to the WiFi
2387 	 * sequence number. This is a HW requirements to help the SCD to parse
2388 	 * the BA.
2389 	 * Check here that the packets are in the right place on the ring.
2390 	 */
2391 	wifi_seq = IEEE80211_SEQ_TO_SN(le16_to_cpu(hdr->seq_ctrl));
2392 	WARN_ONCE(txq->ampdu &&
2393 		  (wifi_seq & 0xff) != txq->write_ptr,
2394 		  "Q: %d WiFi Seq %d tfdNum %d",
2395 		  txq_id, wifi_seq, txq->write_ptr);
2396 
2397 	/* Set up driver data for this TFD */
2398 	txq->entries[txq->write_ptr].skb = skb;
2399 	txq->entries[txq->write_ptr].cmd = dev_cmd;
2400 
2401 	dev_cmd->hdr.sequence =
2402 		cpu_to_le16((u16)(QUEUE_TO_SEQ(txq_id) |
2403 			    INDEX_TO_SEQ(txq->write_ptr)));
2404 
2405 	tb0_phys = iwl_pcie_get_first_tb_dma(txq, txq->write_ptr);
2406 	scratch_phys = tb0_phys + sizeof(struct iwl_cmd_header) +
2407 		       offsetof(struct iwl_tx_cmd, scratch);
2408 
2409 	tx_cmd->dram_lsb_ptr = cpu_to_le32(scratch_phys);
2410 	tx_cmd->dram_msb_ptr = iwl_get_dma_hi_addr(scratch_phys);
2411 
2412 	/* Set up first empty entry in queue's array of Tx/cmd buffers */
2413 	out_meta = &txq->entries[txq->write_ptr].meta;
2414 	out_meta->flags = 0;
2415 
2416 	/*
2417 	 * The second TB (tb1) points to the remainder of the TX command
2418 	 * and the 802.11 header - dword aligned size
2419 	 * (This calculation modifies the TX command, so do it before the
2420 	 * setup of the first TB)
2421 	 */
2422 	len = sizeof(struct iwl_tx_cmd) + sizeof(struct iwl_cmd_header) +
2423 	      hdr_len - IWL_FIRST_TB_SIZE;
2424 	/* do not align A-MSDU to dword as the subframe header aligns it */
2425 	amsdu = ieee80211_is_data_qos(fc) &&
2426 		(*ieee80211_get_qos_ctl(hdr) &
2427 		 IEEE80211_QOS_CTL_A_MSDU_PRESENT);
2428 	if (trans_pcie->sw_csum_tx || !amsdu) {
2429 		tb1_len = ALIGN(len, 4);
2430 		/* Tell NIC about any 2-byte padding after MAC header */
2431 		if (tb1_len != len)
2432 			tx_cmd->tx_flags |= cpu_to_le32(TX_CMD_FLG_MH_PAD);
2433 	} else {
2434 		tb1_len = len;
2435 	}
2436 
2437 	/*
2438 	 * The first TB points to bi-directional DMA data, we'll
2439 	 * memcpy the data into it later.
2440 	 */
2441 	iwl_pcie_txq_build_tfd(trans, txq, tb0_phys,
2442 			       IWL_FIRST_TB_SIZE, true);
2443 
2444 	/* there must be data left over for TB1 or this code must be changed */
2445 	BUILD_BUG_ON(sizeof(struct iwl_tx_cmd) < IWL_FIRST_TB_SIZE);
2446 
2447 	/* map the data for TB1 */
2448 	tb1_addr = ((u8 *)&dev_cmd->hdr) + IWL_FIRST_TB_SIZE;
2449 	tb1_phys = dma_map_single(trans->dev, tb1_addr, tb1_len, DMA_TO_DEVICE);
2450 	if (unlikely(dma_mapping_error(trans->dev, tb1_phys)))
2451 		goto out_err;
2452 	iwl_pcie_txq_build_tfd(trans, txq, tb1_phys, tb1_len, false);
2453 
2454 	trace_iwlwifi_dev_tx(trans->dev, skb,
2455 			     iwl_pcie_get_tfd(trans, txq,
2456 					      txq->write_ptr),
2457 			     trans_pcie->tfd_size,
2458 			     &dev_cmd->hdr, IWL_FIRST_TB_SIZE + tb1_len,
2459 			     hdr_len);
2460 
2461 	/*
2462 	 * If gso_size wasn't set, don't give the frame "amsdu treatment"
2463 	 * (adding subframes, etc.).
2464 	 * This can happen in some testing flows when the amsdu was already
2465 	 * pre-built, and we just need to send the resulting skb.
2466 	 */
2467 	if (amsdu && skb_shinfo(skb)->gso_size) {
2468 		if (unlikely(iwl_fill_data_tbs_amsdu(trans, skb, txq, hdr_len,
2469 						     out_meta, dev_cmd,
2470 						     tb1_len)))
2471 			goto out_err;
2472 	} else {
2473 		struct sk_buff *frag;
2474 
2475 		if (unlikely(iwl_fill_data_tbs(trans, skb, txq, hdr_len,
2476 					       out_meta)))
2477 			goto out_err;
2478 
2479 		skb_walk_frags(skb, frag) {
2480 			if (unlikely(iwl_fill_data_tbs(trans, frag, txq, 0,
2481 						       out_meta)))
2482 				goto out_err;
2483 		}
2484 	}
2485 
2486 	/* building the A-MSDU might have changed this data, so memcpy it now */
2487 	memcpy(&txq->first_tb_bufs[txq->write_ptr], dev_cmd, IWL_FIRST_TB_SIZE);
2488 
2489 	tfd = iwl_pcie_get_tfd(trans, txq, txq->write_ptr);
2490 	/* Set up entry for this TFD in Tx byte-count array */
2491 	iwl_pcie_txq_update_byte_cnt_tbl(trans, txq, le16_to_cpu(tx_cmd->len),
2492 					 iwl_pcie_tfd_get_num_tbs(trans, tfd));
2493 
2494 	wait_write_ptr = ieee80211_has_morefrags(fc);
2495 
2496 	/* start timer if queue currently empty */
2497 	if (txq->read_ptr == txq->write_ptr && txq->wd_timeout) {
2498 		/*
2499 		 * If the TXQ is active, then set the timer, if not,
2500 		 * set the timer in remainder so that the timer will
2501 		 * be armed with the right value when the station will
2502 		 * wake up.
2503 		 */
2504 		if (!txq->frozen)
2505 			mod_timer(&txq->stuck_timer,
2506 				  jiffies + txq->wd_timeout);
2507 		else
2508 			txq->frozen_expiry_remainder = txq->wd_timeout;
2509 	}
2510 
2511 	/* Tell device the write index *just past* this latest filled TFD */
2512 	txq->write_ptr = iwl_queue_inc_wrap(trans, txq->write_ptr);
2513 	if (!wait_write_ptr)
2514 		iwl_pcie_txq_inc_wr_ptr(trans, txq);
2515 
2516 	/*
2517 	 * At this point the frame is "transmitted" successfully
2518 	 * and we will get a TX status notification eventually.
2519 	 */
2520 	spin_unlock(&txq->lock);
2521 	return 0;
2522 out_err:
2523 	iwl_pcie_tfd_unmap(trans, out_meta, txq, txq->write_ptr);
2524 	spin_unlock(&txq->lock);
2525 	return -1;
2526 }
2527