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(struct timer_list *t)
151 {
152 	struct iwl_txq *txq = from_timer(txq, t, stuck_timer);
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, 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 	timer_setup(&txq->stuck_timer, iwl_pcie_txq_stuck_timer, 0);
499 	txq->trans_pcie = trans_pcie;
500 
501 	txq->n_window = slots_num;
502 
503 	txq->entries = kcalloc(slots_num,
504 			       sizeof(struct iwl_pcie_txq_entry),
505 			       GFP_KERNEL);
506 
507 	if (!txq->entries)
508 		goto error;
509 
510 	if (cmd_queue)
511 		for (i = 0; i < slots_num; i++) {
512 			txq->entries[i].cmd =
513 				kmalloc(sizeof(struct iwl_device_cmd),
514 					GFP_KERNEL);
515 			if (!txq->entries[i].cmd)
516 				goto error;
517 		}
518 
519 	/* Circular buffer of transmit frame descriptors (TFDs),
520 	 * shared with device */
521 	txq->tfds = dma_alloc_coherent(trans->dev, tfd_sz,
522 				       &txq->dma_addr, GFP_KERNEL);
523 	if (!txq->tfds)
524 		goto error;
525 
526 	BUILD_BUG_ON(IWL_FIRST_TB_SIZE_ALIGN != sizeof(*txq->first_tb_bufs));
527 
528 	tb0_buf_sz = sizeof(*txq->first_tb_bufs) * slots_num;
529 
530 	txq->first_tb_bufs = dma_alloc_coherent(trans->dev, tb0_buf_sz,
531 					      &txq->first_tb_dma,
532 					      GFP_KERNEL);
533 	if (!txq->first_tb_bufs)
534 		goto err_free_tfds;
535 
536 	return 0;
537 err_free_tfds:
538 	dma_free_coherent(trans->dev, tfd_sz, txq->tfds, txq->dma_addr);
539 error:
540 	if (txq->entries && cmd_queue)
541 		for (i = 0; i < slots_num; i++)
542 			kfree(txq->entries[i].cmd);
543 	kfree(txq->entries);
544 	txq->entries = NULL;
545 
546 	return -ENOMEM;
547 
548 }
549 
550 int iwl_pcie_txq_init(struct iwl_trans *trans, struct iwl_txq *txq,
551 		      int slots_num, bool cmd_queue)
552 {
553 	int ret;
554 
555 	txq->need_update = false;
556 
557 	/* TFD_QUEUE_SIZE_MAX must be power-of-two size, otherwise
558 	 * iwl_queue_inc_wrap and iwl_queue_dec_wrap are broken. */
559 	BUILD_BUG_ON(TFD_QUEUE_SIZE_MAX & (TFD_QUEUE_SIZE_MAX - 1));
560 
561 	/* Initialize queue's high/low-water marks, and head/tail indexes */
562 	ret = iwl_queue_init(txq, slots_num);
563 	if (ret)
564 		return ret;
565 
566 	spin_lock_init(&txq->lock);
567 
568 	if (cmd_queue) {
569 		static struct lock_class_key iwl_pcie_cmd_queue_lock_class;
570 
571 		lockdep_set_class(&txq->lock, &iwl_pcie_cmd_queue_lock_class);
572 	}
573 
574 	__skb_queue_head_init(&txq->overflow_q);
575 
576 	return 0;
577 }
578 
579 void iwl_pcie_free_tso_page(struct iwl_trans_pcie *trans_pcie,
580 			    struct sk_buff *skb)
581 {
582 	struct page **page_ptr;
583 
584 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
585 
586 	if (*page_ptr) {
587 		__free_page(*page_ptr);
588 		*page_ptr = NULL;
589 	}
590 }
591 
592 static void iwl_pcie_clear_cmd_in_flight(struct iwl_trans *trans)
593 {
594 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
595 
596 	lockdep_assert_held(&trans_pcie->reg_lock);
597 
598 	if (trans_pcie->ref_cmd_in_flight) {
599 		trans_pcie->ref_cmd_in_flight = false;
600 		IWL_DEBUG_RPM(trans, "clear ref_cmd_in_flight - unref\n");
601 		iwl_trans_unref(trans);
602 	}
603 
604 	if (!trans->cfg->base_params->apmg_wake_up_wa)
605 		return;
606 	if (WARN_ON(!trans_pcie->cmd_hold_nic_awake))
607 		return;
608 
609 	trans_pcie->cmd_hold_nic_awake = false;
610 	__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
611 				   CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
612 }
613 
614 /*
615  * iwl_pcie_txq_unmap -  Unmap any remaining DMA mappings and free skb's
616  */
617 static void iwl_pcie_txq_unmap(struct iwl_trans *trans, int txq_id)
618 {
619 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
620 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
621 
622 	spin_lock_bh(&txq->lock);
623 	while (txq->write_ptr != txq->read_ptr) {
624 		IWL_DEBUG_TX_REPLY(trans, "Q %d Free %d\n",
625 				   txq_id, txq->read_ptr);
626 
627 		if (txq_id != trans_pcie->cmd_queue) {
628 			struct sk_buff *skb = txq->entries[txq->read_ptr].skb;
629 
630 			if (WARN_ON_ONCE(!skb))
631 				continue;
632 
633 			iwl_pcie_free_tso_page(trans_pcie, skb);
634 		}
635 		iwl_pcie_txq_free_tfd(trans, txq);
636 		txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr);
637 
638 		if (txq->read_ptr == txq->write_ptr) {
639 			unsigned long flags;
640 
641 			spin_lock_irqsave(&trans_pcie->reg_lock, flags);
642 			if (txq_id != trans_pcie->cmd_queue) {
643 				IWL_DEBUG_RPM(trans, "Q %d - last tx freed\n",
644 					      txq->id);
645 				iwl_trans_unref(trans);
646 			} else {
647 				iwl_pcie_clear_cmd_in_flight(trans);
648 			}
649 			spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
650 		}
651 	}
652 
653 	while (!skb_queue_empty(&txq->overflow_q)) {
654 		struct sk_buff *skb = __skb_dequeue(&txq->overflow_q);
655 
656 		iwl_op_mode_free_skb(trans->op_mode, skb);
657 	}
658 
659 	spin_unlock_bh(&txq->lock);
660 
661 	/* just in case - this queue may have been stopped */
662 	iwl_wake_queue(trans, txq);
663 }
664 
665 /*
666  * iwl_pcie_txq_free - Deallocate DMA queue.
667  * @txq: Transmit queue to deallocate.
668  *
669  * Empty queue by removing and destroying all BD's.
670  * Free all buffers.
671  * 0-fill, but do not free "txq" descriptor structure.
672  */
673 static void iwl_pcie_txq_free(struct iwl_trans *trans, int txq_id)
674 {
675 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
676 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
677 	struct device *dev = trans->dev;
678 	int i;
679 
680 	if (WARN_ON(!txq))
681 		return;
682 
683 	iwl_pcie_txq_unmap(trans, txq_id);
684 
685 	/* De-alloc array of command/tx buffers */
686 	if (txq_id == trans_pcie->cmd_queue)
687 		for (i = 0; i < txq->n_window; i++) {
688 			kzfree(txq->entries[i].cmd);
689 			kzfree(txq->entries[i].free_buf);
690 		}
691 
692 	/* De-alloc circular buffer of TFDs */
693 	if (txq->tfds) {
694 		dma_free_coherent(dev,
695 				  trans_pcie->tfd_size * TFD_QUEUE_SIZE_MAX,
696 				  txq->tfds, txq->dma_addr);
697 		txq->dma_addr = 0;
698 		txq->tfds = NULL;
699 
700 		dma_free_coherent(dev,
701 				  sizeof(*txq->first_tb_bufs) * txq->n_window,
702 				  txq->first_tb_bufs, txq->first_tb_dma);
703 	}
704 
705 	kfree(txq->entries);
706 	txq->entries = NULL;
707 
708 	del_timer_sync(&txq->stuck_timer);
709 
710 	/* 0-fill queue descriptor structure */
711 	memset(txq, 0, sizeof(*txq));
712 }
713 
714 void iwl_pcie_tx_start(struct iwl_trans *trans, u32 scd_base_addr)
715 {
716 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
717 	int nq = trans->cfg->base_params->num_of_queues;
718 	int chan;
719 	u32 reg_val;
720 	int clear_dwords = (SCD_TRANS_TBL_OFFSET_QUEUE(nq) -
721 				SCD_CONTEXT_MEM_LOWER_BOUND) / sizeof(u32);
722 
723 	/* make sure all queue are not stopped/used */
724 	memset(trans_pcie->queue_stopped, 0, sizeof(trans_pcie->queue_stopped));
725 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
726 
727 	trans_pcie->scd_base_addr =
728 		iwl_read_prph(trans, SCD_SRAM_BASE_ADDR);
729 
730 	WARN_ON(scd_base_addr != 0 &&
731 		scd_base_addr != trans_pcie->scd_base_addr);
732 
733 	/* reset context data, TX status and translation data */
734 	iwl_trans_write_mem(trans, trans_pcie->scd_base_addr +
735 				   SCD_CONTEXT_MEM_LOWER_BOUND,
736 			    NULL, clear_dwords);
737 
738 	iwl_write_prph(trans, SCD_DRAM_BASE_ADDR,
739 		       trans_pcie->scd_bc_tbls.dma >> 10);
740 
741 	/* The chain extension of the SCD doesn't work well. This feature is
742 	 * enabled by default by the HW, so we need to disable it manually.
743 	 */
744 	if (trans->cfg->base_params->scd_chain_ext_wa)
745 		iwl_write_prph(trans, SCD_CHAINEXT_EN, 0);
746 
747 	iwl_trans_ac_txq_enable(trans, trans_pcie->cmd_queue,
748 				trans_pcie->cmd_fifo,
749 				trans_pcie->cmd_q_wdg_timeout);
750 
751 	/* Activate all Tx DMA/FIFO channels */
752 	iwl_scd_activate_fifos(trans);
753 
754 	/* Enable DMA channel */
755 	for (chan = 0; chan < FH_TCSR_CHNL_NUM; chan++)
756 		iwl_write_direct32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(chan),
757 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CHNL_ENABLE |
758 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CREDIT_ENABLE);
759 
760 	/* Update FH chicken bits */
761 	reg_val = iwl_read_direct32(trans, FH_TX_CHICKEN_BITS_REG);
762 	iwl_write_direct32(trans, FH_TX_CHICKEN_BITS_REG,
763 			   reg_val | FH_TX_CHICKEN_BITS_SCD_AUTO_RETRY_EN);
764 
765 	/* Enable L1-Active */
766 	if (trans->cfg->device_family < IWL_DEVICE_FAMILY_8000)
767 		iwl_clear_bits_prph(trans, APMG_PCIDEV_STT_REG,
768 				    APMG_PCIDEV_STT_VAL_L1_ACT_DIS);
769 }
770 
771 void iwl_trans_pcie_tx_reset(struct iwl_trans *trans)
772 {
773 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
774 	int txq_id;
775 
776 	/*
777 	 * we should never get here in gen2 trans mode return early to avoid
778 	 * having invalid accesses
779 	 */
780 	if (WARN_ON_ONCE(trans->cfg->gen2))
781 		return;
782 
783 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
784 	     txq_id++) {
785 		struct iwl_txq *txq = trans_pcie->txq[txq_id];
786 		if (trans->cfg->use_tfh)
787 			iwl_write_direct64(trans,
788 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
789 					   txq->dma_addr);
790 		else
791 			iwl_write_direct32(trans,
792 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
793 					   txq->dma_addr >> 8);
794 		iwl_pcie_txq_unmap(trans, txq_id);
795 		txq->read_ptr = 0;
796 		txq->write_ptr = 0;
797 	}
798 
799 	/* Tell NIC where to find the "keep warm" buffer */
800 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
801 			   trans_pcie->kw.dma >> 4);
802 
803 	/*
804 	 * Send 0 as the scd_base_addr since the device may have be reset
805 	 * while we were in WoWLAN in which case SCD_SRAM_BASE_ADDR will
806 	 * contain garbage.
807 	 */
808 	iwl_pcie_tx_start(trans, 0);
809 }
810 
811 static void iwl_pcie_tx_stop_fh(struct iwl_trans *trans)
812 {
813 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
814 	unsigned long flags;
815 	int ch, ret;
816 	u32 mask = 0;
817 
818 	spin_lock(&trans_pcie->irq_lock);
819 
820 	if (!iwl_trans_grab_nic_access(trans, &flags))
821 		goto out;
822 
823 	/* Stop each Tx DMA channel */
824 	for (ch = 0; ch < FH_TCSR_CHNL_NUM; ch++) {
825 		iwl_write32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(ch), 0x0);
826 		mask |= FH_TSSR_TX_STATUS_REG_MSK_CHNL_IDLE(ch);
827 	}
828 
829 	/* Wait for DMA channels to be idle */
830 	ret = iwl_poll_bit(trans, FH_TSSR_TX_STATUS_REG, mask, mask, 5000);
831 	if (ret < 0)
832 		IWL_ERR(trans,
833 			"Failing on timeout while stopping DMA channel %d [0x%08x]\n",
834 			ch, iwl_read32(trans, FH_TSSR_TX_STATUS_REG));
835 
836 	iwl_trans_release_nic_access(trans, &flags);
837 
838 out:
839 	spin_unlock(&trans_pcie->irq_lock);
840 }
841 
842 /*
843  * iwl_pcie_tx_stop - Stop all Tx DMA channels
844  */
845 int iwl_pcie_tx_stop(struct iwl_trans *trans)
846 {
847 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
848 	int txq_id;
849 
850 	/* Turn off all Tx DMA fifos */
851 	iwl_scd_deactivate_fifos(trans);
852 
853 	/* Turn off all Tx DMA channels */
854 	iwl_pcie_tx_stop_fh(trans);
855 
856 	/*
857 	 * This function can be called before the op_mode disabled the
858 	 * queues. This happens when we have an rfkill interrupt.
859 	 * Since we stop Tx altogether - mark the queues as stopped.
860 	 */
861 	memset(trans_pcie->queue_stopped, 0, sizeof(trans_pcie->queue_stopped));
862 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
863 
864 	/* This can happen: start_hw, stop_device */
865 	if (!trans_pcie->txq_memory)
866 		return 0;
867 
868 	/* Unmap DMA from host system and free skb's */
869 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
870 	     txq_id++)
871 		iwl_pcie_txq_unmap(trans, txq_id);
872 
873 	return 0;
874 }
875 
876 /*
877  * iwl_trans_tx_free - Free TXQ Context
878  *
879  * Destroy all TX DMA queues and structures
880  */
881 void iwl_pcie_tx_free(struct iwl_trans *trans)
882 {
883 	int txq_id;
884 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
885 
886 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
887 
888 	/* Tx queues */
889 	if (trans_pcie->txq_memory) {
890 		for (txq_id = 0;
891 		     txq_id < trans->cfg->base_params->num_of_queues;
892 		     txq_id++) {
893 			iwl_pcie_txq_free(trans, txq_id);
894 			trans_pcie->txq[txq_id] = NULL;
895 		}
896 	}
897 
898 	kfree(trans_pcie->txq_memory);
899 	trans_pcie->txq_memory = NULL;
900 
901 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->kw);
902 
903 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->scd_bc_tbls);
904 }
905 
906 /*
907  * iwl_pcie_tx_alloc - allocate TX context
908  * Allocate all Tx DMA structures and initialize them
909  */
910 static int iwl_pcie_tx_alloc(struct iwl_trans *trans)
911 {
912 	int ret;
913 	int txq_id, slots_num;
914 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
915 
916 	u16 scd_bc_tbls_size = trans->cfg->base_params->num_of_queues *
917 			sizeof(struct iwlagn_scd_bc_tbl);
918 
919 	/*It is not allowed to alloc twice, so warn when this happens.
920 	 * We cannot rely on the previous allocation, so free and fail */
921 	if (WARN_ON(trans_pcie->txq_memory)) {
922 		ret = -EINVAL;
923 		goto error;
924 	}
925 
926 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->scd_bc_tbls,
927 				   scd_bc_tbls_size);
928 	if (ret) {
929 		IWL_ERR(trans, "Scheduler BC Table allocation failed\n");
930 		goto error;
931 	}
932 
933 	/* Alloc keep-warm buffer */
934 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->kw, IWL_KW_SIZE);
935 	if (ret) {
936 		IWL_ERR(trans, "Keep Warm allocation failed\n");
937 		goto error;
938 	}
939 
940 	trans_pcie->txq_memory = kcalloc(trans->cfg->base_params->num_of_queues,
941 					 sizeof(struct iwl_txq), GFP_KERNEL);
942 	if (!trans_pcie->txq_memory) {
943 		IWL_ERR(trans, "Not enough memory for txq\n");
944 		ret = -ENOMEM;
945 		goto error;
946 	}
947 
948 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
949 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
950 	     txq_id++) {
951 		bool cmd_queue = (txq_id == trans_pcie->cmd_queue);
952 
953 		slots_num = cmd_queue ? trans_pcie->tx_cmd_queue_size :
954 			TFD_TX_CMD_SLOTS;
955 		trans_pcie->txq[txq_id] = &trans_pcie->txq_memory[txq_id];
956 		ret = iwl_pcie_txq_alloc(trans, trans_pcie->txq[txq_id],
957 					 slots_num, cmd_queue);
958 		if (ret) {
959 			IWL_ERR(trans, "Tx %d queue alloc failed\n", txq_id);
960 			goto error;
961 		}
962 		trans_pcie->txq[txq_id]->id = txq_id;
963 	}
964 
965 	return 0;
966 
967 error:
968 	iwl_pcie_tx_free(trans);
969 
970 	return ret;
971 }
972 
973 void iwl_pcie_set_tx_cmd_queue_size(struct iwl_trans *trans)
974 {
975 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
976 	int queue_size = TFD_CMD_SLOTS;
977 
978 	if (trans->cfg->tx_cmd_queue_size)
979 		queue_size = trans->cfg->tx_cmd_queue_size;
980 
981 	if (WARN_ON(!(is_power_of_2(queue_size) &&
982 		      TFD_QUEUE_CB_SIZE(queue_size) > 0)))
983 		trans_pcie->tx_cmd_queue_size = TFD_CMD_SLOTS;
984 	else
985 		trans_pcie->tx_cmd_queue_size = queue_size;
986 }
987 
988 int iwl_pcie_tx_init(struct iwl_trans *trans)
989 {
990 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
991 	int ret;
992 	int txq_id, slots_num;
993 	bool alloc = false;
994 
995 	iwl_pcie_set_tx_cmd_queue_size(trans);
996 
997 	if (!trans_pcie->txq_memory) {
998 		ret = iwl_pcie_tx_alloc(trans);
999 		if (ret)
1000 			goto error;
1001 		alloc = true;
1002 	}
1003 
1004 	spin_lock(&trans_pcie->irq_lock);
1005 
1006 	/* Turn off all Tx DMA fifos */
1007 	iwl_scd_deactivate_fifos(trans);
1008 
1009 	/* Tell NIC where to find the "keep warm" buffer */
1010 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
1011 			   trans_pcie->kw.dma >> 4);
1012 
1013 	spin_unlock(&trans_pcie->irq_lock);
1014 
1015 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
1016 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
1017 	     txq_id++) {
1018 		bool cmd_queue = (txq_id == trans_pcie->cmd_queue);
1019 
1020 		slots_num = cmd_queue ? trans_pcie->tx_cmd_queue_size :
1021 			TFD_TX_CMD_SLOTS;
1022 		ret = iwl_pcie_txq_init(trans, trans_pcie->txq[txq_id],
1023 					slots_num, cmd_queue);
1024 		if (ret) {
1025 			IWL_ERR(trans, "Tx %d queue init failed\n", txq_id);
1026 			goto error;
1027 		}
1028 
1029 		/*
1030 		 * Tell nic where to find circular buffer of TFDs for a
1031 		 * given Tx queue, and enable the DMA channel used for that
1032 		 * queue.
1033 		 * Circular buffer (TFD queue in DRAM) physical base address
1034 		 */
1035 		iwl_write_direct32(trans, FH_MEM_CBBC_QUEUE(trans, txq_id),
1036 				   trans_pcie->txq[txq_id]->dma_addr >> 8);
1037 	}
1038 
1039 	iwl_set_bits_prph(trans, SCD_GP_CTRL, SCD_GP_CTRL_AUTO_ACTIVE_MODE);
1040 	if (trans->cfg->base_params->num_of_queues > 20)
1041 		iwl_set_bits_prph(trans, SCD_GP_CTRL,
1042 				  SCD_GP_CTRL_ENABLE_31_QUEUES);
1043 
1044 	return 0;
1045 error:
1046 	/*Upon error, free only if we allocated something */
1047 	if (alloc)
1048 		iwl_pcie_tx_free(trans);
1049 	return ret;
1050 }
1051 
1052 static inline void iwl_pcie_txq_progress(struct iwl_txq *txq)
1053 {
1054 	lockdep_assert_held(&txq->lock);
1055 
1056 	if (!txq->wd_timeout)
1057 		return;
1058 
1059 	/*
1060 	 * station is asleep and we send data - that must
1061 	 * be uAPSD or PS-Poll. Don't rearm the timer.
1062 	 */
1063 	if (txq->frozen)
1064 		return;
1065 
1066 	/*
1067 	 * if empty delete timer, otherwise move timer forward
1068 	 * since we're making progress on this queue
1069 	 */
1070 	if (txq->read_ptr == txq->write_ptr)
1071 		del_timer(&txq->stuck_timer);
1072 	else
1073 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1074 }
1075 
1076 /* Frees buffers until index _not_ inclusive */
1077 void iwl_trans_pcie_reclaim(struct iwl_trans *trans, int txq_id, int ssn,
1078 			    struct sk_buff_head *skbs)
1079 {
1080 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1081 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1082 	int tfd_num = ssn & (TFD_QUEUE_SIZE_MAX - 1);
1083 	int last_to_free;
1084 
1085 	/* This function is not meant to release cmd queue*/
1086 	if (WARN_ON(txq_id == trans_pcie->cmd_queue))
1087 		return;
1088 
1089 	spin_lock_bh(&txq->lock);
1090 
1091 	if (!test_bit(txq_id, trans_pcie->queue_used)) {
1092 		IWL_DEBUG_TX_QUEUES(trans, "Q %d inactive - ignoring idx %d\n",
1093 				    txq_id, ssn);
1094 		goto out;
1095 	}
1096 
1097 	if (txq->read_ptr == tfd_num)
1098 		goto out;
1099 
1100 	IWL_DEBUG_TX_REPLY(trans, "[Q %d] %d -> %d (%d)\n",
1101 			   txq_id, txq->read_ptr, tfd_num, ssn);
1102 
1103 	/*Since we free until index _not_ inclusive, the one before index is
1104 	 * the last we will free. This one must be used */
1105 	last_to_free = iwl_queue_dec_wrap(tfd_num);
1106 
1107 	if (!iwl_queue_used(txq, last_to_free)) {
1108 		IWL_ERR(trans,
1109 			"%s: Read index for DMA queue txq id (%d), last_to_free %d is out of range [0-%d] %d %d.\n",
1110 			__func__, txq_id, last_to_free, TFD_QUEUE_SIZE_MAX,
1111 			txq->write_ptr, txq->read_ptr);
1112 		goto out;
1113 	}
1114 
1115 	if (WARN_ON(!skb_queue_empty(skbs)))
1116 		goto out;
1117 
1118 	for (;
1119 	     txq->read_ptr != tfd_num;
1120 	     txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr)) {
1121 		int idx = iwl_pcie_get_cmd_index(txq, txq->read_ptr);
1122 		struct sk_buff *skb = txq->entries[idx].skb;
1123 
1124 		if (WARN_ON_ONCE(!skb))
1125 			continue;
1126 
1127 		iwl_pcie_free_tso_page(trans_pcie, skb);
1128 
1129 		__skb_queue_tail(skbs, skb);
1130 
1131 		txq->entries[idx].skb = NULL;
1132 
1133 		if (!trans->cfg->use_tfh)
1134 			iwl_pcie_txq_inval_byte_cnt_tbl(trans, txq);
1135 
1136 		iwl_pcie_txq_free_tfd(trans, txq);
1137 	}
1138 
1139 	iwl_pcie_txq_progress(txq);
1140 
1141 	if (iwl_queue_space(txq) > txq->low_mark &&
1142 	    test_bit(txq_id, trans_pcie->queue_stopped)) {
1143 		struct sk_buff_head overflow_skbs;
1144 
1145 		__skb_queue_head_init(&overflow_skbs);
1146 		skb_queue_splice_init(&txq->overflow_q, &overflow_skbs);
1147 
1148 		/*
1149 		 * This is tricky: we are in reclaim path which is non
1150 		 * re-entrant, so noone will try to take the access the
1151 		 * txq data from that path. We stopped tx, so we can't
1152 		 * have tx as well. Bottom line, we can unlock and re-lock
1153 		 * later.
1154 		 */
1155 		spin_unlock_bh(&txq->lock);
1156 
1157 		while (!skb_queue_empty(&overflow_skbs)) {
1158 			struct sk_buff *skb = __skb_dequeue(&overflow_skbs);
1159 			struct iwl_device_cmd *dev_cmd_ptr;
1160 
1161 			dev_cmd_ptr = *(void **)((u8 *)skb->cb +
1162 						 trans_pcie->dev_cmd_offs);
1163 
1164 			/*
1165 			 * Note that we can very well be overflowing again.
1166 			 * In that case, iwl_queue_space will be small again
1167 			 * and we won't wake mac80211's queue.
1168 			 */
1169 			iwl_trans_pcie_tx(trans, skb, dev_cmd_ptr, txq_id);
1170 		}
1171 		spin_lock_bh(&txq->lock);
1172 
1173 		if (iwl_queue_space(txq) > txq->low_mark)
1174 			iwl_wake_queue(trans, txq);
1175 	}
1176 
1177 	if (txq->read_ptr == txq->write_ptr) {
1178 		IWL_DEBUG_RPM(trans, "Q %d - last tx reclaimed\n", txq->id);
1179 		iwl_trans_unref(trans);
1180 	}
1181 
1182 out:
1183 	spin_unlock_bh(&txq->lock);
1184 }
1185 
1186 static int iwl_pcie_set_cmd_in_flight(struct iwl_trans *trans,
1187 				      const struct iwl_host_cmd *cmd)
1188 {
1189 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1190 	int ret;
1191 
1192 	lockdep_assert_held(&trans_pcie->reg_lock);
1193 
1194 	if (!(cmd->flags & CMD_SEND_IN_IDLE) &&
1195 	    !trans_pcie->ref_cmd_in_flight) {
1196 		trans_pcie->ref_cmd_in_flight = true;
1197 		IWL_DEBUG_RPM(trans, "set ref_cmd_in_flight - ref\n");
1198 		iwl_trans_ref(trans);
1199 	}
1200 
1201 	/*
1202 	 * wake up the NIC to make sure that the firmware will see the host
1203 	 * command - we will let the NIC sleep once all the host commands
1204 	 * returned. This needs to be done only on NICs that have
1205 	 * apmg_wake_up_wa set.
1206 	 */
1207 	if (trans->cfg->base_params->apmg_wake_up_wa &&
1208 	    !trans_pcie->cmd_hold_nic_awake) {
1209 		__iwl_trans_pcie_set_bit(trans, CSR_GP_CNTRL,
1210 					 CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1211 
1212 		ret = iwl_poll_bit(trans, CSR_GP_CNTRL,
1213 				   CSR_GP_CNTRL_REG_VAL_MAC_ACCESS_EN,
1214 				   (CSR_GP_CNTRL_REG_FLAG_MAC_CLOCK_READY |
1215 				    CSR_GP_CNTRL_REG_FLAG_GOING_TO_SLEEP),
1216 				   15000);
1217 		if (ret < 0) {
1218 			__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
1219 					CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1220 			IWL_ERR(trans, "Failed to wake NIC for hcmd\n");
1221 			return -EIO;
1222 		}
1223 		trans_pcie->cmd_hold_nic_awake = true;
1224 	}
1225 
1226 	return 0;
1227 }
1228 
1229 /*
1230  * iwl_pcie_cmdq_reclaim - Reclaim TX command queue entries already Tx'd
1231  *
1232  * When FW advances 'R' index, all entries between old and new 'R' index
1233  * need to be reclaimed. As result, some free space forms.  If there is
1234  * enough free space (> low mark), wake the stack that feeds us.
1235  */
1236 static void iwl_pcie_cmdq_reclaim(struct iwl_trans *trans, int txq_id, int idx)
1237 {
1238 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1239 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1240 	unsigned long flags;
1241 	int nfreed = 0;
1242 
1243 	lockdep_assert_held(&txq->lock);
1244 
1245 	if ((idx >= TFD_QUEUE_SIZE_MAX) || (!iwl_queue_used(txq, idx))) {
1246 		IWL_ERR(trans,
1247 			"%s: Read index for DMA queue txq id (%d), index %d is out of range [0-%d] %d %d.\n",
1248 			__func__, txq_id, idx, TFD_QUEUE_SIZE_MAX,
1249 			txq->write_ptr, txq->read_ptr);
1250 		return;
1251 	}
1252 
1253 	for (idx = iwl_queue_inc_wrap(idx); txq->read_ptr != idx;
1254 	     txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr)) {
1255 
1256 		if (nfreed++ > 0) {
1257 			IWL_ERR(trans, "HCMD skipped: index (%d) %d %d\n",
1258 				idx, txq->write_ptr, txq->read_ptr);
1259 			iwl_force_nmi(trans);
1260 		}
1261 	}
1262 
1263 	if (txq->read_ptr == txq->write_ptr) {
1264 		spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1265 		iwl_pcie_clear_cmd_in_flight(trans);
1266 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1267 	}
1268 
1269 	iwl_pcie_txq_progress(txq);
1270 }
1271 
1272 static int iwl_pcie_txq_set_ratid_map(struct iwl_trans *trans, u16 ra_tid,
1273 				 u16 txq_id)
1274 {
1275 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1276 	u32 tbl_dw_addr;
1277 	u32 tbl_dw;
1278 	u16 scd_q2ratid;
1279 
1280 	scd_q2ratid = ra_tid & SCD_QUEUE_RA_TID_MAP_RATID_MSK;
1281 
1282 	tbl_dw_addr = trans_pcie->scd_base_addr +
1283 			SCD_TRANS_TBL_OFFSET_QUEUE(txq_id);
1284 
1285 	tbl_dw = iwl_trans_read_mem32(trans, tbl_dw_addr);
1286 
1287 	if (txq_id & 0x1)
1288 		tbl_dw = (scd_q2ratid << 16) | (tbl_dw & 0x0000FFFF);
1289 	else
1290 		tbl_dw = scd_q2ratid | (tbl_dw & 0xFFFF0000);
1291 
1292 	iwl_trans_write_mem32(trans, tbl_dw_addr, tbl_dw);
1293 
1294 	return 0;
1295 }
1296 
1297 /* Receiver address (actually, Rx station's index into station table),
1298  * combined with Traffic ID (QOS priority), in format used by Tx Scheduler */
1299 #define BUILD_RAxTID(sta_id, tid)	(((sta_id) << 4) + (tid))
1300 
1301 bool iwl_trans_pcie_txq_enable(struct iwl_trans *trans, int txq_id, u16 ssn,
1302 			       const struct iwl_trans_txq_scd_cfg *cfg,
1303 			       unsigned int wdg_timeout)
1304 {
1305 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1306 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1307 	int fifo = -1;
1308 	bool scd_bug = false;
1309 
1310 	if (test_and_set_bit(txq_id, trans_pcie->queue_used))
1311 		WARN_ONCE(1, "queue %d already used - expect issues", txq_id);
1312 
1313 	txq->wd_timeout = msecs_to_jiffies(wdg_timeout);
1314 
1315 	if (cfg) {
1316 		fifo = cfg->fifo;
1317 
1318 		/* Disable the scheduler prior configuring the cmd queue */
1319 		if (txq_id == trans_pcie->cmd_queue &&
1320 		    trans_pcie->scd_set_active)
1321 			iwl_scd_enable_set_active(trans, 0);
1322 
1323 		/* Stop this Tx queue before configuring it */
1324 		iwl_scd_txq_set_inactive(trans, txq_id);
1325 
1326 		/* Set this queue as a chain-building queue unless it is CMD */
1327 		if (txq_id != trans_pcie->cmd_queue)
1328 			iwl_scd_txq_set_chain(trans, txq_id);
1329 
1330 		if (cfg->aggregate) {
1331 			u16 ra_tid = BUILD_RAxTID(cfg->sta_id, cfg->tid);
1332 
1333 			/* Map receiver-address / traffic-ID to this queue */
1334 			iwl_pcie_txq_set_ratid_map(trans, ra_tid, txq_id);
1335 
1336 			/* enable aggregations for the queue */
1337 			iwl_scd_txq_enable_agg(trans, txq_id);
1338 			txq->ampdu = true;
1339 		} else {
1340 			/*
1341 			 * disable aggregations for the queue, this will also
1342 			 * make the ra_tid mapping configuration irrelevant
1343 			 * since it is now a non-AGG queue.
1344 			 */
1345 			iwl_scd_txq_disable_agg(trans, txq_id);
1346 
1347 			ssn = txq->read_ptr;
1348 		}
1349 	} else {
1350 		/*
1351 		 * If we need to move the SCD write pointer by steps of
1352 		 * 0x40, 0x80 or 0xc0, it gets stuck. Avoids this and let
1353 		 * the op_mode know by returning true later.
1354 		 * Do this only in case cfg is NULL since this trick can
1355 		 * be done only if we have DQA enabled which is true for mvm
1356 		 * only. And mvm never sets a cfg pointer.
1357 		 * This is really ugly, but this is the easiest way out for
1358 		 * this sad hardware issue.
1359 		 * This bug has been fixed on devices 9000 and up.
1360 		 */
1361 		scd_bug = !trans->cfg->mq_rx_supported &&
1362 			!((ssn - txq->write_ptr) & 0x3f) &&
1363 			(ssn != txq->write_ptr);
1364 		if (scd_bug)
1365 			ssn++;
1366 	}
1367 
1368 	/* Place first TFD at index corresponding to start sequence number.
1369 	 * Assumes that ssn_idx is valid (!= 0xFFF) */
1370 	txq->read_ptr = (ssn & 0xff);
1371 	txq->write_ptr = (ssn & 0xff);
1372 	iwl_write_direct32(trans, HBUS_TARG_WRPTR,
1373 			   (ssn & 0xff) | (txq_id << 8));
1374 
1375 	if (cfg) {
1376 		u8 frame_limit = cfg->frame_limit;
1377 
1378 		iwl_write_prph(trans, SCD_QUEUE_RDPTR(txq_id), ssn);
1379 
1380 		/* Set up Tx window size and frame limit for this queue */
1381 		iwl_trans_write_mem32(trans, trans_pcie->scd_base_addr +
1382 				SCD_CONTEXT_QUEUE_OFFSET(txq_id), 0);
1383 		iwl_trans_write_mem32(trans,
1384 			trans_pcie->scd_base_addr +
1385 			SCD_CONTEXT_QUEUE_OFFSET(txq_id) + sizeof(u32),
1386 			SCD_QUEUE_CTX_REG2_VAL(WIN_SIZE, frame_limit) |
1387 			SCD_QUEUE_CTX_REG2_VAL(FRAME_LIMIT, frame_limit));
1388 
1389 		/* Set up status area in SRAM, map to Tx DMA/FIFO, activate */
1390 		iwl_write_prph(trans, SCD_QUEUE_STATUS_BITS(txq_id),
1391 			       (1 << SCD_QUEUE_STTS_REG_POS_ACTIVE) |
1392 			       (cfg->fifo << SCD_QUEUE_STTS_REG_POS_TXF) |
1393 			       (1 << SCD_QUEUE_STTS_REG_POS_WSL) |
1394 			       SCD_QUEUE_STTS_REG_MSK);
1395 
1396 		/* enable the scheduler for this queue (only) */
1397 		if (txq_id == trans_pcie->cmd_queue &&
1398 		    trans_pcie->scd_set_active)
1399 			iwl_scd_enable_set_active(trans, BIT(txq_id));
1400 
1401 		IWL_DEBUG_TX_QUEUES(trans,
1402 				    "Activate queue %d on FIFO %d WrPtr: %d\n",
1403 				    txq_id, fifo, ssn & 0xff);
1404 	} else {
1405 		IWL_DEBUG_TX_QUEUES(trans,
1406 				    "Activate queue %d WrPtr: %d\n",
1407 				    txq_id, ssn & 0xff);
1408 	}
1409 
1410 	return scd_bug;
1411 }
1412 
1413 void iwl_trans_pcie_txq_set_shared_mode(struct iwl_trans *trans, u32 txq_id,
1414 					bool shared_mode)
1415 {
1416 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1417 	struct iwl_txq *txq = trans_pcie->txq[txq_id];
1418 
1419 	txq->ampdu = !shared_mode;
1420 }
1421 
1422 void iwl_trans_pcie_txq_disable(struct iwl_trans *trans, int txq_id,
1423 				bool configure_scd)
1424 {
1425 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1426 	u32 stts_addr = trans_pcie->scd_base_addr +
1427 			SCD_TX_STTS_QUEUE_OFFSET(txq_id);
1428 	static const u32 zero_val[4] = {};
1429 
1430 	trans_pcie->txq[txq_id]->frozen_expiry_remainder = 0;
1431 	trans_pcie->txq[txq_id]->frozen = false;
1432 
1433 	/*
1434 	 * Upon HW Rfkill - we stop the device, and then stop the queues
1435 	 * in the op_mode. Just for the sake of the simplicity of the op_mode,
1436 	 * allow the op_mode to call txq_disable after it already called
1437 	 * stop_device.
1438 	 */
1439 	if (!test_and_clear_bit(txq_id, trans_pcie->queue_used)) {
1440 		WARN_ONCE(test_bit(STATUS_DEVICE_ENABLED, &trans->status),
1441 			  "queue %d not used", txq_id);
1442 		return;
1443 	}
1444 
1445 	if (configure_scd) {
1446 		iwl_scd_txq_set_inactive(trans, txq_id);
1447 
1448 		iwl_trans_write_mem(trans, stts_addr, (void *)zero_val,
1449 				    ARRAY_SIZE(zero_val));
1450 	}
1451 
1452 	iwl_pcie_txq_unmap(trans, txq_id);
1453 	trans_pcie->txq[txq_id]->ampdu = false;
1454 
1455 	IWL_DEBUG_TX_QUEUES(trans, "Deactivate queue %d\n", txq_id);
1456 }
1457 
1458 /*************** HOST COMMAND QUEUE FUNCTIONS   *****/
1459 
1460 /*
1461  * iwl_pcie_enqueue_hcmd - enqueue a uCode command
1462  * @priv: device private data point
1463  * @cmd: a pointer to the ucode command structure
1464  *
1465  * The function returns < 0 values to indicate the operation
1466  * failed. On success, it returns the index (>= 0) of command in the
1467  * command queue.
1468  */
1469 static int iwl_pcie_enqueue_hcmd(struct iwl_trans *trans,
1470 				 struct iwl_host_cmd *cmd)
1471 {
1472 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1473 	struct iwl_txq *txq = trans_pcie->txq[trans_pcie->cmd_queue];
1474 	struct iwl_device_cmd *out_cmd;
1475 	struct iwl_cmd_meta *out_meta;
1476 	unsigned long flags;
1477 	void *dup_buf = NULL;
1478 	dma_addr_t phys_addr;
1479 	int idx;
1480 	u16 copy_size, cmd_size, tb0_size;
1481 	bool had_nocopy = false;
1482 	u8 group_id = iwl_cmd_groupid(cmd->id);
1483 	int i, ret;
1484 	u32 cmd_pos;
1485 	const u8 *cmddata[IWL_MAX_CMD_TBS_PER_TFD];
1486 	u16 cmdlen[IWL_MAX_CMD_TBS_PER_TFD];
1487 
1488 	if (WARN(!trans->wide_cmd_header &&
1489 		 group_id > IWL_ALWAYS_LONG_GROUP,
1490 		 "unsupported wide command %#x\n", cmd->id))
1491 		return -EINVAL;
1492 
1493 	if (group_id != 0) {
1494 		copy_size = sizeof(struct iwl_cmd_header_wide);
1495 		cmd_size = sizeof(struct iwl_cmd_header_wide);
1496 	} else {
1497 		copy_size = sizeof(struct iwl_cmd_header);
1498 		cmd_size = sizeof(struct iwl_cmd_header);
1499 	}
1500 
1501 	/* need one for the header if the first is NOCOPY */
1502 	BUILD_BUG_ON(IWL_MAX_CMD_TBS_PER_TFD > IWL_NUM_OF_TBS - 1);
1503 
1504 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1505 		cmddata[i] = cmd->data[i];
1506 		cmdlen[i] = cmd->len[i];
1507 
1508 		if (!cmd->len[i])
1509 			continue;
1510 
1511 		/* need at least IWL_FIRST_TB_SIZE copied */
1512 		if (copy_size < IWL_FIRST_TB_SIZE) {
1513 			int copy = IWL_FIRST_TB_SIZE - copy_size;
1514 
1515 			if (copy > cmdlen[i])
1516 				copy = cmdlen[i];
1517 			cmdlen[i] -= copy;
1518 			cmddata[i] += copy;
1519 			copy_size += copy;
1520 		}
1521 
1522 		if (cmd->dataflags[i] & IWL_HCMD_DFL_NOCOPY) {
1523 			had_nocopy = true;
1524 			if (WARN_ON(cmd->dataflags[i] & IWL_HCMD_DFL_DUP)) {
1525 				idx = -EINVAL;
1526 				goto free_dup_buf;
1527 			}
1528 		} else if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP) {
1529 			/*
1530 			 * This is also a chunk that isn't copied
1531 			 * to the static buffer so set had_nocopy.
1532 			 */
1533 			had_nocopy = true;
1534 
1535 			/* only allowed once */
1536 			if (WARN_ON(dup_buf)) {
1537 				idx = -EINVAL;
1538 				goto free_dup_buf;
1539 			}
1540 
1541 			dup_buf = kmemdup(cmddata[i], cmdlen[i],
1542 					  GFP_ATOMIC);
1543 			if (!dup_buf)
1544 				return -ENOMEM;
1545 		} else {
1546 			/* NOCOPY must not be followed by normal! */
1547 			if (WARN_ON(had_nocopy)) {
1548 				idx = -EINVAL;
1549 				goto free_dup_buf;
1550 			}
1551 			copy_size += cmdlen[i];
1552 		}
1553 		cmd_size += cmd->len[i];
1554 	}
1555 
1556 	/*
1557 	 * If any of the command structures end up being larger than
1558 	 * the TFD_MAX_PAYLOAD_SIZE and they aren't dynamically
1559 	 * allocated into separate TFDs, then we will need to
1560 	 * increase the size of the buffers.
1561 	 */
1562 	if (WARN(copy_size > TFD_MAX_PAYLOAD_SIZE,
1563 		 "Command %s (%#x) is too large (%d bytes)\n",
1564 		 iwl_get_cmd_string(trans, cmd->id),
1565 		 cmd->id, copy_size)) {
1566 		idx = -EINVAL;
1567 		goto free_dup_buf;
1568 	}
1569 
1570 	spin_lock_bh(&txq->lock);
1571 
1572 	if (iwl_queue_space(txq) < ((cmd->flags & CMD_ASYNC) ? 2 : 1)) {
1573 		spin_unlock_bh(&txq->lock);
1574 
1575 		IWL_ERR(trans, "No space in command queue\n");
1576 		iwl_op_mode_cmd_queue_full(trans->op_mode);
1577 		idx = -ENOSPC;
1578 		goto free_dup_buf;
1579 	}
1580 
1581 	idx = iwl_pcie_get_cmd_index(txq, txq->write_ptr);
1582 	out_cmd = txq->entries[idx].cmd;
1583 	out_meta = &txq->entries[idx].meta;
1584 
1585 	memset(out_meta, 0, sizeof(*out_meta));	/* re-initialize to NULL */
1586 	if (cmd->flags & CMD_WANT_SKB)
1587 		out_meta->source = cmd;
1588 
1589 	/* set up the header */
1590 	if (group_id != 0) {
1591 		out_cmd->hdr_wide.cmd = iwl_cmd_opcode(cmd->id);
1592 		out_cmd->hdr_wide.group_id = group_id;
1593 		out_cmd->hdr_wide.version = iwl_cmd_version(cmd->id);
1594 		out_cmd->hdr_wide.length =
1595 			cpu_to_le16(cmd_size -
1596 				    sizeof(struct iwl_cmd_header_wide));
1597 		out_cmd->hdr_wide.reserved = 0;
1598 		out_cmd->hdr_wide.sequence =
1599 			cpu_to_le16(QUEUE_TO_SEQ(trans_pcie->cmd_queue) |
1600 						 INDEX_TO_SEQ(txq->write_ptr));
1601 
1602 		cmd_pos = sizeof(struct iwl_cmd_header_wide);
1603 		copy_size = sizeof(struct iwl_cmd_header_wide);
1604 	} else {
1605 		out_cmd->hdr.cmd = iwl_cmd_opcode(cmd->id);
1606 		out_cmd->hdr.sequence =
1607 			cpu_to_le16(QUEUE_TO_SEQ(trans_pcie->cmd_queue) |
1608 						 INDEX_TO_SEQ(txq->write_ptr));
1609 		out_cmd->hdr.group_id = 0;
1610 
1611 		cmd_pos = sizeof(struct iwl_cmd_header);
1612 		copy_size = sizeof(struct iwl_cmd_header);
1613 	}
1614 
1615 	/* and copy the data that needs to be copied */
1616 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1617 		int copy;
1618 
1619 		if (!cmd->len[i])
1620 			continue;
1621 
1622 		/* copy everything if not nocopy/dup */
1623 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1624 					   IWL_HCMD_DFL_DUP))) {
1625 			copy = cmd->len[i];
1626 
1627 			memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1628 			cmd_pos += copy;
1629 			copy_size += copy;
1630 			continue;
1631 		}
1632 
1633 		/*
1634 		 * Otherwise we need at least IWL_FIRST_TB_SIZE copied
1635 		 * in total (for bi-directional DMA), but copy up to what
1636 		 * we can fit into the payload for debug dump purposes.
1637 		 */
1638 		copy = min_t(int, TFD_MAX_PAYLOAD_SIZE - cmd_pos, cmd->len[i]);
1639 
1640 		memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1641 		cmd_pos += copy;
1642 
1643 		/* However, treat copy_size the proper way, we need it below */
1644 		if (copy_size < IWL_FIRST_TB_SIZE) {
1645 			copy = IWL_FIRST_TB_SIZE - copy_size;
1646 
1647 			if (copy > cmd->len[i])
1648 				copy = cmd->len[i];
1649 			copy_size += copy;
1650 		}
1651 	}
1652 
1653 	IWL_DEBUG_HC(trans,
1654 		     "Sending command %s (%.2x.%.2x), seq: 0x%04X, %d bytes at %d[%d]:%d\n",
1655 		     iwl_get_cmd_string(trans, cmd->id),
1656 		     group_id, out_cmd->hdr.cmd,
1657 		     le16_to_cpu(out_cmd->hdr.sequence),
1658 		     cmd_size, txq->write_ptr, idx, trans_pcie->cmd_queue);
1659 
1660 	/* start the TFD with the minimum copy bytes */
1661 	tb0_size = min_t(int, copy_size, IWL_FIRST_TB_SIZE);
1662 	memcpy(&txq->first_tb_bufs[idx], &out_cmd->hdr, tb0_size);
1663 	iwl_pcie_txq_build_tfd(trans, txq,
1664 			       iwl_pcie_get_first_tb_dma(txq, idx),
1665 			       tb0_size, true);
1666 
1667 	/* map first command fragment, if any remains */
1668 	if (copy_size > tb0_size) {
1669 		phys_addr = dma_map_single(trans->dev,
1670 					   ((u8 *)&out_cmd->hdr) + tb0_size,
1671 					   copy_size - tb0_size,
1672 					   DMA_TO_DEVICE);
1673 		if (dma_mapping_error(trans->dev, phys_addr)) {
1674 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1675 					   txq->write_ptr);
1676 			idx = -ENOMEM;
1677 			goto out;
1678 		}
1679 
1680 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr,
1681 				       copy_size - tb0_size, false);
1682 	}
1683 
1684 	/* map the remaining (adjusted) nocopy/dup fragments */
1685 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1686 		const void *data = cmddata[i];
1687 
1688 		if (!cmdlen[i])
1689 			continue;
1690 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1691 					   IWL_HCMD_DFL_DUP)))
1692 			continue;
1693 		if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP)
1694 			data = dup_buf;
1695 		phys_addr = dma_map_single(trans->dev, (void *)data,
1696 					   cmdlen[i], DMA_TO_DEVICE);
1697 		if (dma_mapping_error(trans->dev, phys_addr)) {
1698 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1699 					   txq->write_ptr);
1700 			idx = -ENOMEM;
1701 			goto out;
1702 		}
1703 
1704 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr, cmdlen[i], false);
1705 	}
1706 
1707 	BUILD_BUG_ON(IWL_TFH_NUM_TBS > sizeof(out_meta->tbs) * BITS_PER_BYTE);
1708 	out_meta->flags = cmd->flags;
1709 	if (WARN_ON_ONCE(txq->entries[idx].free_buf))
1710 		kzfree(txq->entries[idx].free_buf);
1711 	txq->entries[idx].free_buf = dup_buf;
1712 
1713 	trace_iwlwifi_dev_hcmd(trans->dev, cmd, cmd_size, &out_cmd->hdr_wide);
1714 
1715 	/* start timer if queue currently empty */
1716 	if (txq->read_ptr == txq->write_ptr && txq->wd_timeout)
1717 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1718 
1719 	spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1720 	ret = iwl_pcie_set_cmd_in_flight(trans, cmd);
1721 	if (ret < 0) {
1722 		idx = ret;
1723 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1724 		goto out;
1725 	}
1726 
1727 	/* Increment and update queue's write index */
1728 	txq->write_ptr = iwl_queue_inc_wrap(txq->write_ptr);
1729 	iwl_pcie_txq_inc_wr_ptr(trans, txq);
1730 
1731 	spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1732 
1733  out:
1734 	spin_unlock_bh(&txq->lock);
1735  free_dup_buf:
1736 	if (idx < 0)
1737 		kfree(dup_buf);
1738 	return idx;
1739 }
1740 
1741 /*
1742  * iwl_pcie_hcmd_complete - Pull unused buffers off the queue and reclaim them
1743  * @rxb: Rx buffer to reclaim
1744  */
1745 void iwl_pcie_hcmd_complete(struct iwl_trans *trans,
1746 			    struct iwl_rx_cmd_buffer *rxb)
1747 {
1748 	struct iwl_rx_packet *pkt = rxb_addr(rxb);
1749 	u16 sequence = le16_to_cpu(pkt->hdr.sequence);
1750 	u8 group_id;
1751 	u32 cmd_id;
1752 	int txq_id = SEQ_TO_QUEUE(sequence);
1753 	int index = SEQ_TO_INDEX(sequence);
1754 	int cmd_index;
1755 	struct iwl_device_cmd *cmd;
1756 	struct iwl_cmd_meta *meta;
1757 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1758 	struct iwl_txq *txq = trans_pcie->txq[trans_pcie->cmd_queue];
1759 
1760 	/* If a Tx command is being handled and it isn't in the actual
1761 	 * command queue then there a command routing bug has been introduced
1762 	 * in the queue management code. */
1763 	if (WARN(txq_id != trans_pcie->cmd_queue,
1764 		 "wrong command queue %d (should be %d), sequence 0x%X readp=%d writep=%d\n",
1765 		 txq_id, trans_pcie->cmd_queue, sequence, txq->read_ptr,
1766 		 txq->write_ptr)) {
1767 		iwl_print_hex_error(trans, pkt, 32);
1768 		return;
1769 	}
1770 
1771 	spin_lock_bh(&txq->lock);
1772 
1773 	cmd_index = iwl_pcie_get_cmd_index(txq, index);
1774 	cmd = txq->entries[cmd_index].cmd;
1775 	meta = &txq->entries[cmd_index].meta;
1776 	group_id = cmd->hdr.group_id;
1777 	cmd_id = iwl_cmd_id(cmd->hdr.cmd, group_id, 0);
1778 
1779 	iwl_pcie_tfd_unmap(trans, meta, txq, index);
1780 
1781 	/* Input error checking is done when commands are added to queue. */
1782 	if (meta->flags & CMD_WANT_SKB) {
1783 		struct page *p = rxb_steal_page(rxb);
1784 
1785 		meta->source->resp_pkt = pkt;
1786 		meta->source->_rx_page_addr = (unsigned long)page_address(p);
1787 		meta->source->_rx_page_order = trans_pcie->rx_page_order;
1788 	}
1789 
1790 	if (meta->flags & CMD_WANT_ASYNC_CALLBACK)
1791 		iwl_op_mode_async_cb(trans->op_mode, cmd);
1792 
1793 	iwl_pcie_cmdq_reclaim(trans, txq_id, index);
1794 
1795 	if (!(meta->flags & CMD_ASYNC)) {
1796 		if (!test_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status)) {
1797 			IWL_WARN(trans,
1798 				 "HCMD_ACTIVE already clear for command %s\n",
1799 				 iwl_get_cmd_string(trans, cmd_id));
1800 		}
1801 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1802 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1803 			       iwl_get_cmd_string(trans, cmd_id));
1804 		wake_up(&trans_pcie->wait_command_queue);
1805 	}
1806 
1807 	if (meta->flags & CMD_MAKE_TRANS_IDLE) {
1808 		IWL_DEBUG_INFO(trans, "complete %s - mark trans as idle\n",
1809 			       iwl_get_cmd_string(trans, cmd->hdr.cmd));
1810 		set_bit(STATUS_TRANS_IDLE, &trans->status);
1811 		wake_up(&trans_pcie->d0i3_waitq);
1812 	}
1813 
1814 	if (meta->flags & CMD_WAKE_UP_TRANS) {
1815 		IWL_DEBUG_INFO(trans, "complete %s - clear trans idle flag\n",
1816 			       iwl_get_cmd_string(trans, cmd->hdr.cmd));
1817 		clear_bit(STATUS_TRANS_IDLE, &trans->status);
1818 		wake_up(&trans_pcie->d0i3_waitq);
1819 	}
1820 
1821 	meta->flags = 0;
1822 
1823 	spin_unlock_bh(&txq->lock);
1824 }
1825 
1826 #define HOST_COMPLETE_TIMEOUT	(2 * HZ)
1827 
1828 static int iwl_pcie_send_hcmd_async(struct iwl_trans *trans,
1829 				    struct iwl_host_cmd *cmd)
1830 {
1831 	int ret;
1832 
1833 	/* An asynchronous command can not expect an SKB to be set. */
1834 	if (WARN_ON(cmd->flags & CMD_WANT_SKB))
1835 		return -EINVAL;
1836 
1837 	ret = iwl_pcie_enqueue_hcmd(trans, cmd);
1838 	if (ret < 0) {
1839 		IWL_ERR(trans,
1840 			"Error sending %s: enqueue_hcmd failed: %d\n",
1841 			iwl_get_cmd_string(trans, cmd->id), ret);
1842 		return ret;
1843 	}
1844 	return 0;
1845 }
1846 
1847 static int iwl_pcie_send_hcmd_sync(struct iwl_trans *trans,
1848 				   struct iwl_host_cmd *cmd)
1849 {
1850 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1851 	struct iwl_txq *txq = trans_pcie->txq[trans_pcie->cmd_queue];
1852 	int cmd_idx;
1853 	int ret;
1854 
1855 	IWL_DEBUG_INFO(trans, "Attempting to send sync command %s\n",
1856 		       iwl_get_cmd_string(trans, cmd->id));
1857 
1858 	if (WARN(test_and_set_bit(STATUS_SYNC_HCMD_ACTIVE,
1859 				  &trans->status),
1860 		 "Command %s: a command is already active!\n",
1861 		 iwl_get_cmd_string(trans, cmd->id)))
1862 		return -EIO;
1863 
1864 	IWL_DEBUG_INFO(trans, "Setting HCMD_ACTIVE for command %s\n",
1865 		       iwl_get_cmd_string(trans, cmd->id));
1866 
1867 	if (pm_runtime_suspended(&trans_pcie->pci_dev->dev)) {
1868 		ret = wait_event_timeout(trans_pcie->d0i3_waitq,
1869 				 pm_runtime_active(&trans_pcie->pci_dev->dev),
1870 				 msecs_to_jiffies(IWL_TRANS_IDLE_TIMEOUT));
1871 		if (!ret) {
1872 			IWL_ERR(trans, "Timeout exiting D0i3 before hcmd\n");
1873 			return -ETIMEDOUT;
1874 		}
1875 	}
1876 
1877 	cmd_idx = iwl_pcie_enqueue_hcmd(trans, cmd);
1878 	if (cmd_idx < 0) {
1879 		ret = cmd_idx;
1880 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1881 		IWL_ERR(trans,
1882 			"Error sending %s: enqueue_hcmd failed: %d\n",
1883 			iwl_get_cmd_string(trans, cmd->id), ret);
1884 		return ret;
1885 	}
1886 
1887 	ret = wait_event_timeout(trans_pcie->wait_command_queue,
1888 				 !test_bit(STATUS_SYNC_HCMD_ACTIVE,
1889 					   &trans->status),
1890 				 HOST_COMPLETE_TIMEOUT);
1891 	if (!ret) {
1892 		IWL_ERR(trans, "Error sending %s: time out after %dms.\n",
1893 			iwl_get_cmd_string(trans, cmd->id),
1894 			jiffies_to_msecs(HOST_COMPLETE_TIMEOUT));
1895 
1896 		IWL_ERR(trans, "Current CMD queue read_ptr %d write_ptr %d\n",
1897 			txq->read_ptr, txq->write_ptr);
1898 
1899 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1900 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1901 			       iwl_get_cmd_string(trans, cmd->id));
1902 		ret = -ETIMEDOUT;
1903 
1904 		iwl_force_nmi(trans);
1905 		iwl_trans_fw_error(trans);
1906 
1907 		goto cancel;
1908 	}
1909 
1910 	if (test_bit(STATUS_FW_ERROR, &trans->status)) {
1911 		iwl_trans_dump_regs(trans);
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, 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, 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, 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