xref: /linux/drivers/net/wireless/intel/iwlwifi/pcie/tx.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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 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 /* FIXME: need to abstract out TX command (once we know what it looks like) */
47 #include "dvm/commands.h"
48 
49 #define IWL_TX_CRC_SIZE 4
50 #define IWL_TX_DELIMITER_SIZE 4
51 
52 /*************** DMA-QUEUE-GENERAL-FUNCTIONS  *****
53  * DMA services
54  *
55  * Theory of operation
56  *
57  * A Tx or Rx queue resides in host DRAM, and is comprised of a circular buffer
58  * of buffer descriptors, each of which points to one or more data buffers for
59  * the device to read from or fill.  Driver and device exchange status of each
60  * queue via "read" and "write" pointers.  Driver keeps minimum of 2 empty
61  * entries in each circular buffer, to protect against confusing empty and full
62  * queue states.
63  *
64  * The device reads or writes the data in the queues via the device's several
65  * DMA/FIFO channels.  Each queue is mapped to a single DMA channel.
66  *
67  * For Tx queue, there are low mark and high mark limits. If, after queuing
68  * the packet for Tx, free space become < low mark, Tx queue stopped. When
69  * reclaiming packets (on 'tx done IRQ), if free space become > high mark,
70  * Tx queue resumed.
71  *
72  ***************************************************/
73 
74 static int iwl_queue_space(const struct iwl_txq *q)
75 {
76 	unsigned int max;
77 	unsigned int used;
78 
79 	/*
80 	 * To avoid ambiguity between empty and completely full queues, there
81 	 * should always be less than TFD_QUEUE_SIZE_MAX elements in the queue.
82 	 * If q->n_window is smaller than TFD_QUEUE_SIZE_MAX, there is no need
83 	 * to reserve any queue entries for this purpose.
84 	 */
85 	if (q->n_window < TFD_QUEUE_SIZE_MAX)
86 		max = q->n_window;
87 	else
88 		max = TFD_QUEUE_SIZE_MAX - 1;
89 
90 	/*
91 	 * TFD_QUEUE_SIZE_MAX is a power of 2, so the following is equivalent to
92 	 * modulo by TFD_QUEUE_SIZE_MAX and is well defined.
93 	 */
94 	used = (q->write_ptr - q->read_ptr) & (TFD_QUEUE_SIZE_MAX - 1);
95 
96 	if (WARN_ON(used > max))
97 		return 0;
98 
99 	return max - used;
100 }
101 
102 /*
103  * iwl_queue_init - Initialize queue's high/low-water and read/write indexes
104  */
105 static int iwl_queue_init(struct iwl_txq *q, int slots_num, u32 id)
106 {
107 	q->n_window = slots_num;
108 	q->id = id;
109 
110 	/* slots_num must be power-of-two size, otherwise
111 	 * get_cmd_index is broken. */
112 	if (WARN_ON(!is_power_of_2(slots_num)))
113 		return -EINVAL;
114 
115 	q->low_mark = q->n_window / 4;
116 	if (q->low_mark < 4)
117 		q->low_mark = 4;
118 
119 	q->high_mark = q->n_window / 8;
120 	if (q->high_mark < 2)
121 		q->high_mark = 2;
122 
123 	q->write_ptr = 0;
124 	q->read_ptr = 0;
125 
126 	return 0;
127 }
128 
129 static int iwl_pcie_alloc_dma_ptr(struct iwl_trans *trans,
130 				  struct iwl_dma_ptr *ptr, size_t size)
131 {
132 	if (WARN_ON(ptr->addr))
133 		return -EINVAL;
134 
135 	ptr->addr = dma_alloc_coherent(trans->dev, size,
136 				       &ptr->dma, GFP_KERNEL);
137 	if (!ptr->addr)
138 		return -ENOMEM;
139 	ptr->size = size;
140 	return 0;
141 }
142 
143 static void iwl_pcie_free_dma_ptr(struct iwl_trans *trans,
144 				  struct iwl_dma_ptr *ptr)
145 {
146 	if (unlikely(!ptr->addr))
147 		return;
148 
149 	dma_free_coherent(trans->dev, ptr->size, ptr->addr, ptr->dma);
150 	memset(ptr, 0, sizeof(*ptr));
151 }
152 
153 static void iwl_pcie_txq_stuck_timer(unsigned long data)
154 {
155 	struct iwl_txq *txq = (void *)data;
156 	struct iwl_trans_pcie *trans_pcie = txq->trans_pcie;
157 	struct iwl_trans *trans = iwl_trans_pcie_get_trans(trans_pcie);
158 
159 	spin_lock(&txq->lock);
160 	/* check if triggered erroneously */
161 	if (txq->read_ptr == txq->write_ptr) {
162 		spin_unlock(&txq->lock);
163 		return;
164 	}
165 	spin_unlock(&txq->lock);
166 
167 	IWL_ERR(trans, "Queue %d stuck for %u ms.\n", txq->id,
168 		jiffies_to_msecs(txq->wd_timeout));
169 
170 	iwl_trans_pcie_log_scd_error(trans, txq);
171 
172 	iwl_force_nmi(trans);
173 }
174 
175 /*
176  * iwl_pcie_txq_update_byte_cnt_tbl - Set up entry in Tx byte-count array
177  */
178 static void iwl_pcie_txq_update_byte_cnt_tbl(struct iwl_trans *trans,
179 					     struct iwl_txq *txq, u16 byte_cnt,
180 					     int num_tbs)
181 {
182 	struct iwlagn_scd_bc_tbl *scd_bc_tbl;
183 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
184 	int write_ptr = txq->write_ptr;
185 	int txq_id = txq->id;
186 	u8 sec_ctl = 0;
187 	u16 len = byte_cnt + IWL_TX_CRC_SIZE + IWL_TX_DELIMITER_SIZE;
188 	__le16 bc_ent;
189 	struct iwl_tx_cmd *tx_cmd =
190 		(void *)txq->entries[txq->write_ptr].cmd->payload;
191 
192 	scd_bc_tbl = trans_pcie->scd_bc_tbls.addr;
193 
194 	sec_ctl = tx_cmd->sec_ctl;
195 
196 	switch (sec_ctl & TX_CMD_SEC_MSK) {
197 	case TX_CMD_SEC_CCM:
198 		len += IEEE80211_CCMP_MIC_LEN;
199 		break;
200 	case TX_CMD_SEC_TKIP:
201 		len += IEEE80211_TKIP_ICV_LEN;
202 		break;
203 	case TX_CMD_SEC_WEP:
204 		len += IEEE80211_WEP_IV_LEN + IEEE80211_WEP_ICV_LEN;
205 		break;
206 	}
207 	if (trans_pcie->bc_table_dword)
208 		len = DIV_ROUND_UP(len, 4);
209 
210 	if (WARN_ON(len > 0xFFF || write_ptr >= TFD_QUEUE_SIZE_MAX))
211 		return;
212 
213 	if (trans->cfg->use_tfh) {
214 		u8 filled_tfd_size = offsetof(struct iwl_tfh_tfd, tbs) +
215 				     num_tbs * sizeof(struct iwl_tfh_tb);
216 		/*
217 		 * filled_tfd_size contains the number of filled bytes in the
218 		 * TFD.
219 		 * Dividing it by 64 will give the number of chunks to fetch
220 		 * to SRAM- 0 for one chunk, 1 for 2 and so on.
221 		 * If, for example, TFD contains only 3 TBs then 32 bytes
222 		 * of the TFD are used, and only one chunk of 64 bytes should
223 		 * be fetched
224 		 */
225 		u8 num_fetch_chunks = DIV_ROUND_UP(filled_tfd_size, 64) - 1;
226 
227 		bc_ent = cpu_to_le16(len | (num_fetch_chunks << 12));
228 	} else {
229 		u8 sta_id = tx_cmd->sta_id;
230 
231 		bc_ent = cpu_to_le16(len | (sta_id << 12));
232 	}
233 
234 	scd_bc_tbl[txq_id].tfd_offset[write_ptr] = bc_ent;
235 
236 	if (write_ptr < TFD_QUEUE_SIZE_BC_DUP)
237 		scd_bc_tbl[txq_id].
238 			tfd_offset[TFD_QUEUE_SIZE_MAX + write_ptr] = bc_ent;
239 }
240 
241 static void iwl_pcie_txq_inval_byte_cnt_tbl(struct iwl_trans *trans,
242 					    struct iwl_txq *txq)
243 {
244 	struct iwl_trans_pcie *trans_pcie =
245 		IWL_TRANS_GET_PCIE_TRANS(trans);
246 	struct iwlagn_scd_bc_tbl *scd_bc_tbl = trans_pcie->scd_bc_tbls.addr;
247 	int txq_id = txq->id;
248 	int read_ptr = txq->read_ptr;
249 	u8 sta_id = 0;
250 	__le16 bc_ent;
251 	struct iwl_tx_cmd *tx_cmd =
252 		(void *)txq->entries[read_ptr].cmd->payload;
253 
254 	WARN_ON(read_ptr >= TFD_QUEUE_SIZE_MAX);
255 
256 	if (txq_id != trans_pcie->cmd_queue)
257 		sta_id = tx_cmd->sta_id;
258 
259 	bc_ent = cpu_to_le16(1 | (sta_id << 12));
260 
261 	scd_bc_tbl[txq_id].tfd_offset[read_ptr] = bc_ent;
262 
263 	if (read_ptr < TFD_QUEUE_SIZE_BC_DUP)
264 		scd_bc_tbl[txq_id].
265 			tfd_offset[TFD_QUEUE_SIZE_MAX + read_ptr] = bc_ent;
266 }
267 
268 /*
269  * iwl_pcie_txq_inc_wr_ptr - Send new write index to hardware
270  */
271 static void iwl_pcie_txq_inc_wr_ptr(struct iwl_trans *trans,
272 				    struct iwl_txq *txq)
273 {
274 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
275 	u32 reg = 0;
276 	int txq_id = txq->id;
277 
278 	lockdep_assert_held(&txq->lock);
279 
280 	/*
281 	 * explicitly wake up the NIC if:
282 	 * 1. shadow registers aren't enabled
283 	 * 2. NIC is woken up for CMD regardless of shadow outside this function
284 	 * 3. there is a chance that the NIC is asleep
285 	 */
286 	if (!trans->cfg->base_params->shadow_reg_enable &&
287 	    txq_id != trans_pcie->cmd_queue &&
288 	    test_bit(STATUS_TPOWER_PMI, &trans->status)) {
289 		/*
290 		 * wake up nic if it's powered down ...
291 		 * uCode will wake up, and interrupt us again, so next
292 		 * time we'll skip this part.
293 		 */
294 		reg = iwl_read32(trans, CSR_UCODE_DRV_GP1);
295 
296 		if (reg & CSR_UCODE_DRV_GP1_BIT_MAC_SLEEP) {
297 			IWL_DEBUG_INFO(trans, "Tx queue %d requesting wakeup, GP1 = 0x%x\n",
298 				       txq_id, reg);
299 			iwl_set_bit(trans, CSR_GP_CNTRL,
300 				    CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
301 			txq->need_update = true;
302 			return;
303 		}
304 	}
305 
306 	/*
307 	 * if not in power-save mode, uCode will never sleep when we're
308 	 * trying to tx (during RFKILL, we're not trying to tx).
309 	 */
310 	IWL_DEBUG_TX(trans, "Q:%d WR: 0x%x\n", txq_id, txq->write_ptr);
311 	if (!txq->block)
312 		iwl_write32(trans, HBUS_TARG_WRPTR,
313 			    txq->write_ptr | (txq_id << 8));
314 }
315 
316 void iwl_pcie_txq_check_wrptrs(struct iwl_trans *trans)
317 {
318 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
319 	int i;
320 
321 	for (i = 0; i < trans->cfg->base_params->num_of_queues; i++) {
322 		struct iwl_txq *txq = &trans_pcie->txq[i];
323 
324 		spin_lock_bh(&txq->lock);
325 		if (trans_pcie->txq[i].need_update) {
326 			iwl_pcie_txq_inc_wr_ptr(trans, txq);
327 			trans_pcie->txq[i].need_update = false;
328 		}
329 		spin_unlock_bh(&txq->lock);
330 	}
331 }
332 
333 static inline void *iwl_pcie_get_tfd(struct iwl_trans_pcie *trans_pcie,
334 				     struct iwl_txq *txq, int idx)
335 {
336 	return txq->tfds + trans_pcie->tfd_size * idx;
337 }
338 
339 static inline dma_addr_t iwl_pcie_tfd_tb_get_addr(struct iwl_trans *trans,
340 						  void *_tfd, u8 idx)
341 {
342 
343 	if (trans->cfg->use_tfh) {
344 		struct iwl_tfh_tfd *tfd = _tfd;
345 		struct iwl_tfh_tb *tb = &tfd->tbs[idx];
346 
347 		return (dma_addr_t)(le64_to_cpu(tb->addr));
348 	} else {
349 		struct iwl_tfd *tfd = _tfd;
350 		struct iwl_tfd_tb *tb = &tfd->tbs[idx];
351 		dma_addr_t addr = get_unaligned_le32(&tb->lo);
352 		dma_addr_t hi_len;
353 
354 		if (sizeof(dma_addr_t) <= sizeof(u32))
355 			return addr;
356 
357 		hi_len = le16_to_cpu(tb->hi_n_len) & 0xF;
358 
359 		/*
360 		 * shift by 16 twice to avoid warnings on 32-bit
361 		 * (where this code never runs anyway due to the
362 		 * if statement above)
363 		 */
364 		return addr | ((hi_len << 16) << 16);
365 	}
366 }
367 
368 static inline void iwl_pcie_tfd_set_tb(struct iwl_trans *trans, void *tfd,
369 				       u8 idx, dma_addr_t addr, u16 len)
370 {
371 	if (trans->cfg->use_tfh) {
372 		struct iwl_tfh_tfd *tfd_fh = (void *)tfd;
373 		struct iwl_tfh_tb *tb = &tfd_fh->tbs[idx];
374 
375 		put_unaligned_le64(addr, &tb->addr);
376 		tb->tb_len = cpu_to_le16(len);
377 
378 		tfd_fh->num_tbs = cpu_to_le16(idx + 1);
379 	} else {
380 		struct iwl_tfd *tfd_fh = (void *)tfd;
381 		struct iwl_tfd_tb *tb = &tfd_fh->tbs[idx];
382 
383 		u16 hi_n_len = len << 4;
384 
385 		put_unaligned_le32(addr, &tb->lo);
386 		if (sizeof(dma_addr_t) > sizeof(u32))
387 			hi_n_len |= ((addr >> 16) >> 16) & 0xF;
388 
389 		tb->hi_n_len = cpu_to_le16(hi_n_len);
390 
391 		tfd_fh->num_tbs = idx + 1;
392 	}
393 }
394 
395 static inline u8 iwl_pcie_tfd_get_num_tbs(struct iwl_trans *trans, void *_tfd)
396 {
397 	if (trans->cfg->use_tfh) {
398 		struct iwl_tfh_tfd *tfd = _tfd;
399 
400 		return le16_to_cpu(tfd->num_tbs) & 0x1f;
401 	} else {
402 		struct iwl_tfd *tfd = _tfd;
403 
404 		return tfd->num_tbs & 0x1f;
405 	}
406 }
407 
408 static void iwl_pcie_tfd_unmap(struct iwl_trans *trans,
409 			       struct iwl_cmd_meta *meta,
410 			       struct iwl_txq *txq, int index)
411 {
412 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
413 	int i, num_tbs;
414 	void *tfd = iwl_pcie_get_tfd(trans_pcie, txq, index);
415 
416 	/* Sanity check on number of chunks */
417 	num_tbs = iwl_pcie_tfd_get_num_tbs(trans, tfd);
418 
419 	if (num_tbs >= trans_pcie->max_tbs) {
420 		IWL_ERR(trans, "Too many chunks: %i\n", num_tbs);
421 		/* @todo issue fatal error, it is quite serious situation */
422 		return;
423 	}
424 
425 	/* first TB is never freed - it's the bidirectional DMA data */
426 
427 	for (i = 1; i < num_tbs; i++) {
428 		if (meta->tbs & BIT(i))
429 			dma_unmap_page(trans->dev,
430 				       iwl_pcie_tfd_tb_get_addr(trans, tfd, i),
431 				       iwl_pcie_tfd_tb_get_len(trans, tfd, i),
432 				       DMA_TO_DEVICE);
433 		else
434 			dma_unmap_single(trans->dev,
435 					 iwl_pcie_tfd_tb_get_addr(trans, tfd,
436 								  i),
437 					 iwl_pcie_tfd_tb_get_len(trans, tfd,
438 								 i),
439 					 DMA_TO_DEVICE);
440 	}
441 
442 	if (trans->cfg->use_tfh) {
443 		struct iwl_tfh_tfd *tfd_fh = (void *)tfd;
444 
445 		tfd_fh->num_tbs = 0;
446 	} else {
447 		struct iwl_tfd *tfd_fh = (void *)tfd;
448 
449 		tfd_fh->num_tbs = 0;
450 	}
451 
452 }
453 
454 /*
455  * iwl_pcie_txq_free_tfd - Free all chunks referenced by TFD [txq->q.read_ptr]
456  * @trans - transport private data
457  * @txq - tx queue
458  * @dma_dir - the direction of the DMA mapping
459  *
460  * Does NOT advance any TFD circular buffer read/write indexes
461  * Does NOT free the TFD itself (which is within circular buffer)
462  */
463 static void iwl_pcie_txq_free_tfd(struct iwl_trans *trans, struct iwl_txq *txq)
464 {
465 	/* rd_ptr is bounded by TFD_QUEUE_SIZE_MAX and
466 	 * idx is bounded by n_window
467 	 */
468 	int rd_ptr = txq->read_ptr;
469 	int idx = get_cmd_index(txq, rd_ptr);
470 
471 	lockdep_assert_held(&txq->lock);
472 
473 	/* We have only q->n_window txq->entries, but we use
474 	 * TFD_QUEUE_SIZE_MAX tfds
475 	 */
476 	iwl_pcie_tfd_unmap(trans, &txq->entries[idx].meta, txq, rd_ptr);
477 
478 	/* free SKB */
479 	if (txq->entries) {
480 		struct sk_buff *skb;
481 
482 		skb = txq->entries[idx].skb;
483 
484 		/* Can be called from irqs-disabled context
485 		 * If skb is not NULL, it means that the whole queue is being
486 		 * freed and that the queue is not empty - free the skb
487 		 */
488 		if (skb) {
489 			iwl_op_mode_free_skb(trans->op_mode, skb);
490 			txq->entries[idx].skb = NULL;
491 		}
492 	}
493 }
494 
495 static int iwl_pcie_txq_build_tfd(struct iwl_trans *trans, struct iwl_txq *txq,
496 				  dma_addr_t addr, u16 len, bool reset)
497 {
498 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
499 	void *tfd;
500 	u32 num_tbs;
501 
502 	tfd = txq->tfds + trans_pcie->tfd_size * txq->write_ptr;
503 
504 	if (reset)
505 		memset(tfd, 0, trans_pcie->tfd_size);
506 
507 	num_tbs = iwl_pcie_tfd_get_num_tbs(trans, tfd);
508 
509 	/* Each TFD can point to a maximum max_tbs Tx buffers */
510 	if (num_tbs >= trans_pcie->max_tbs) {
511 		IWL_ERR(trans, "Error can not send more than %d chunks\n",
512 			trans_pcie->max_tbs);
513 		return -EINVAL;
514 	}
515 
516 	if (WARN(addr & ~IWL_TX_DMA_MASK,
517 		 "Unaligned address = %llx\n", (unsigned long long)addr))
518 		return -EINVAL;
519 
520 	iwl_pcie_tfd_set_tb(trans, tfd, num_tbs, addr, len);
521 
522 	return num_tbs;
523 }
524 
525 static int iwl_pcie_txq_alloc(struct iwl_trans *trans,
526 			       struct iwl_txq *txq, int slots_num,
527 			       u32 txq_id)
528 {
529 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
530 	size_t tfd_sz = trans_pcie->tfd_size * TFD_QUEUE_SIZE_MAX;
531 	size_t tb0_buf_sz;
532 	int i;
533 
534 	if (WARN_ON(txq->entries || txq->tfds))
535 		return -EINVAL;
536 
537 	setup_timer(&txq->stuck_timer, iwl_pcie_txq_stuck_timer,
538 		    (unsigned long)txq);
539 	txq->trans_pcie = trans_pcie;
540 
541 	txq->n_window = slots_num;
542 
543 	txq->entries = kcalloc(slots_num,
544 			       sizeof(struct iwl_pcie_txq_entry),
545 			       GFP_KERNEL);
546 
547 	if (!txq->entries)
548 		goto error;
549 
550 	if (txq_id == trans_pcie->cmd_queue)
551 		for (i = 0; i < slots_num; i++) {
552 			txq->entries[i].cmd =
553 				kmalloc(sizeof(struct iwl_device_cmd),
554 					GFP_KERNEL);
555 			if (!txq->entries[i].cmd)
556 				goto error;
557 		}
558 
559 	/* Circular buffer of transmit frame descriptors (TFDs),
560 	 * shared with device */
561 	txq->tfds = dma_alloc_coherent(trans->dev, tfd_sz,
562 				       &txq->dma_addr, GFP_KERNEL);
563 	if (!txq->tfds)
564 		goto error;
565 
566 	BUILD_BUG_ON(IWL_FIRST_TB_SIZE_ALIGN != sizeof(*txq->first_tb_bufs));
567 
568 	tb0_buf_sz = sizeof(*txq->first_tb_bufs) * slots_num;
569 
570 	txq->first_tb_bufs = dma_alloc_coherent(trans->dev, tb0_buf_sz,
571 					      &txq->first_tb_dma,
572 					      GFP_KERNEL);
573 	if (!txq->first_tb_bufs)
574 		goto err_free_tfds;
575 
576 	txq->id = txq_id;
577 
578 	return 0;
579 err_free_tfds:
580 	dma_free_coherent(trans->dev, tfd_sz, txq->tfds, txq->dma_addr);
581 error:
582 	if (txq->entries && txq_id == trans_pcie->cmd_queue)
583 		for (i = 0; i < slots_num; i++)
584 			kfree(txq->entries[i].cmd);
585 	kfree(txq->entries);
586 	txq->entries = NULL;
587 
588 	return -ENOMEM;
589 
590 }
591 
592 static int iwl_pcie_txq_init(struct iwl_trans *trans, struct iwl_txq *txq,
593 			      int slots_num, u32 txq_id)
594 {
595 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
596 	int ret;
597 
598 	txq->need_update = false;
599 
600 	/* TFD_QUEUE_SIZE_MAX must be power-of-two size, otherwise
601 	 * iwl_queue_inc_wrap and iwl_queue_dec_wrap are broken. */
602 	BUILD_BUG_ON(TFD_QUEUE_SIZE_MAX & (TFD_QUEUE_SIZE_MAX - 1));
603 
604 	/* Initialize queue's high/low-water marks, and head/tail indexes */
605 	ret = iwl_queue_init(txq, slots_num, txq_id);
606 	if (ret)
607 		return ret;
608 
609 	spin_lock_init(&txq->lock);
610 
611 	if (txq_id == trans_pcie->cmd_queue) {
612 		static struct lock_class_key iwl_pcie_cmd_queue_lock_class;
613 
614 		lockdep_set_class(&txq->lock, &iwl_pcie_cmd_queue_lock_class);
615 	}
616 
617 	__skb_queue_head_init(&txq->overflow_q);
618 
619 	/*
620 	 * Tell nic where to find circular buffer of Tx Frame Descriptors for
621 	 * given Tx queue, and enable the DMA channel used for that queue.
622 	 * Circular buffer (TFD queue in DRAM) physical base address */
623 	if (trans->cfg->use_tfh)
624 		iwl_write_direct64(trans,
625 				   FH_MEM_CBBC_QUEUE(trans, txq_id),
626 				   txq->dma_addr);
627 	else
628 		iwl_write_direct32(trans, FH_MEM_CBBC_QUEUE(trans, txq_id),
629 				   txq->dma_addr >> 8);
630 
631 	return 0;
632 }
633 
634 static void iwl_pcie_free_tso_page(struct iwl_trans_pcie *trans_pcie,
635 				   struct sk_buff *skb)
636 {
637 	struct page **page_ptr;
638 
639 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
640 
641 	if (*page_ptr) {
642 		__free_page(*page_ptr);
643 		*page_ptr = NULL;
644 	}
645 }
646 
647 static void iwl_pcie_clear_cmd_in_flight(struct iwl_trans *trans)
648 {
649 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
650 
651 	lockdep_assert_held(&trans_pcie->reg_lock);
652 
653 	if (trans_pcie->ref_cmd_in_flight) {
654 		trans_pcie->ref_cmd_in_flight = false;
655 		IWL_DEBUG_RPM(trans, "clear ref_cmd_in_flight - unref\n");
656 		iwl_trans_unref(trans);
657 	}
658 
659 	if (!trans->cfg->base_params->apmg_wake_up_wa)
660 		return;
661 	if (WARN_ON(!trans_pcie->cmd_hold_nic_awake))
662 		return;
663 
664 	trans_pcie->cmd_hold_nic_awake = false;
665 	__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
666 				   CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
667 }
668 
669 /*
670  * iwl_pcie_txq_unmap -  Unmap any remaining DMA mappings and free skb's
671  */
672 static void iwl_pcie_txq_unmap(struct iwl_trans *trans, int txq_id)
673 {
674 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
675 	struct iwl_txq *txq = &trans_pcie->txq[txq_id];
676 
677 	spin_lock_bh(&txq->lock);
678 	while (txq->write_ptr != txq->read_ptr) {
679 		IWL_DEBUG_TX_REPLY(trans, "Q %d Free %d\n",
680 				   txq_id, txq->read_ptr);
681 
682 		if (txq_id != trans_pcie->cmd_queue) {
683 			struct sk_buff *skb = txq->entries[txq->read_ptr].skb;
684 
685 			if (WARN_ON_ONCE(!skb))
686 				continue;
687 
688 			iwl_pcie_free_tso_page(trans_pcie, skb);
689 		}
690 		iwl_pcie_txq_free_tfd(trans, txq);
691 		txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr);
692 
693 		if (txq->read_ptr == txq->write_ptr) {
694 			unsigned long flags;
695 
696 			spin_lock_irqsave(&trans_pcie->reg_lock, flags);
697 			if (txq_id != trans_pcie->cmd_queue) {
698 				IWL_DEBUG_RPM(trans, "Q %d - last tx freed\n",
699 					      txq->id);
700 				iwl_trans_unref(trans);
701 			} else {
702 				iwl_pcie_clear_cmd_in_flight(trans);
703 			}
704 			spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
705 		}
706 	}
707 	txq->active = false;
708 
709 	while (!skb_queue_empty(&txq->overflow_q)) {
710 		struct sk_buff *skb = __skb_dequeue(&txq->overflow_q);
711 
712 		iwl_op_mode_free_skb(trans->op_mode, skb);
713 	}
714 
715 	spin_unlock_bh(&txq->lock);
716 
717 	/* just in case - this queue may have been stopped */
718 	iwl_wake_queue(trans, txq);
719 }
720 
721 /*
722  * iwl_pcie_txq_free - Deallocate DMA queue.
723  * @txq: Transmit queue to deallocate.
724  *
725  * Empty queue by removing and destroying all BD's.
726  * Free all buffers.
727  * 0-fill, but do not free "txq" descriptor structure.
728  */
729 static void iwl_pcie_txq_free(struct iwl_trans *trans, int txq_id)
730 {
731 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
732 	struct iwl_txq *txq = &trans_pcie->txq[txq_id];
733 	struct device *dev = trans->dev;
734 	int i;
735 
736 	if (WARN_ON(!txq))
737 		return;
738 
739 	iwl_pcie_txq_unmap(trans, txq_id);
740 
741 	/* De-alloc array of command/tx buffers */
742 	if (txq_id == trans_pcie->cmd_queue)
743 		for (i = 0; i < txq->n_window; i++) {
744 			kzfree(txq->entries[i].cmd);
745 			kzfree(txq->entries[i].free_buf);
746 		}
747 
748 	/* De-alloc circular buffer of TFDs */
749 	if (txq->tfds) {
750 		dma_free_coherent(dev,
751 				  trans_pcie->tfd_size * TFD_QUEUE_SIZE_MAX,
752 				  txq->tfds, txq->dma_addr);
753 		txq->dma_addr = 0;
754 		txq->tfds = NULL;
755 
756 		dma_free_coherent(dev,
757 				  sizeof(*txq->first_tb_bufs) * txq->n_window,
758 				  txq->first_tb_bufs, txq->first_tb_dma);
759 	}
760 
761 	kfree(txq->entries);
762 	txq->entries = NULL;
763 
764 	del_timer_sync(&txq->stuck_timer);
765 
766 	/* 0-fill queue descriptor structure */
767 	memset(txq, 0, sizeof(*txq));
768 }
769 
770 void iwl_pcie_tx_start(struct iwl_trans *trans, u32 scd_base_addr)
771 {
772 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
773 	int nq = trans->cfg->base_params->num_of_queues;
774 	int chan;
775 	u32 reg_val;
776 	int clear_dwords = (SCD_TRANS_TBL_OFFSET_QUEUE(nq) -
777 				SCD_CONTEXT_MEM_LOWER_BOUND) / sizeof(u32);
778 
779 	/* make sure all queue are not stopped/used */
780 	memset(trans_pcie->queue_stopped, 0, sizeof(trans_pcie->queue_stopped));
781 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
782 
783 	if (trans->cfg->use_tfh)
784 		return;
785 
786 	trans_pcie->scd_base_addr =
787 		iwl_read_prph(trans, SCD_SRAM_BASE_ADDR);
788 
789 	WARN_ON(scd_base_addr != 0 &&
790 		scd_base_addr != trans_pcie->scd_base_addr);
791 
792 	/* reset context data, TX status and translation data */
793 	iwl_trans_write_mem(trans, trans_pcie->scd_base_addr +
794 				   SCD_CONTEXT_MEM_LOWER_BOUND,
795 			    NULL, clear_dwords);
796 
797 	iwl_write_prph(trans, SCD_DRAM_BASE_ADDR,
798 		       trans_pcie->scd_bc_tbls.dma >> 10);
799 
800 	/* The chain extension of the SCD doesn't work well. This feature is
801 	 * enabled by default by the HW, so we need to disable it manually.
802 	 */
803 	if (trans->cfg->base_params->scd_chain_ext_wa)
804 		iwl_write_prph(trans, SCD_CHAINEXT_EN, 0);
805 
806 	iwl_trans_ac_txq_enable(trans, trans_pcie->cmd_queue,
807 				trans_pcie->cmd_fifo,
808 				trans_pcie->cmd_q_wdg_timeout);
809 
810 	/* Activate all Tx DMA/FIFO channels */
811 	iwl_scd_activate_fifos(trans);
812 
813 	/* Enable DMA channel */
814 	for (chan = 0; chan < FH_TCSR_CHNL_NUM; chan++)
815 		iwl_write_direct32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(chan),
816 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CHNL_ENABLE |
817 				   FH_TCSR_TX_CONFIG_REG_VAL_DMA_CREDIT_ENABLE);
818 
819 	/* Update FH chicken bits */
820 	reg_val = iwl_read_direct32(trans, FH_TX_CHICKEN_BITS_REG);
821 	iwl_write_direct32(trans, FH_TX_CHICKEN_BITS_REG,
822 			   reg_val | FH_TX_CHICKEN_BITS_SCD_AUTO_RETRY_EN);
823 
824 	/* Enable L1-Active */
825 	if (trans->cfg->device_family != IWL_DEVICE_FAMILY_8000)
826 		iwl_clear_bits_prph(trans, APMG_PCIDEV_STT_REG,
827 				    APMG_PCIDEV_STT_VAL_L1_ACT_DIS);
828 }
829 
830 void iwl_trans_pcie_tx_reset(struct iwl_trans *trans)
831 {
832 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
833 	int txq_id;
834 
835 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
836 	     txq_id++) {
837 		struct iwl_txq *txq = &trans_pcie->txq[txq_id];
838 		if (trans->cfg->use_tfh)
839 			iwl_write_direct64(trans,
840 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
841 					   txq->dma_addr);
842 		else
843 			iwl_write_direct32(trans,
844 					   FH_MEM_CBBC_QUEUE(trans, txq_id),
845 					   txq->dma_addr >> 8);
846 		iwl_pcie_txq_unmap(trans, txq_id);
847 		txq->read_ptr = 0;
848 		txq->write_ptr = 0;
849 	}
850 
851 	/* Tell NIC where to find the "keep warm" buffer */
852 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
853 			   trans_pcie->kw.dma >> 4);
854 
855 	/*
856 	 * Send 0 as the scd_base_addr since the device may have be reset
857 	 * while we were in WoWLAN in which case SCD_SRAM_BASE_ADDR will
858 	 * contain garbage.
859 	 */
860 	iwl_pcie_tx_start(trans, 0);
861 }
862 
863 static void iwl_pcie_tx_stop_fh(struct iwl_trans *trans)
864 {
865 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
866 	unsigned long flags;
867 	int ch, ret;
868 	u32 mask = 0;
869 
870 	spin_lock(&trans_pcie->irq_lock);
871 
872 	if (!iwl_trans_grab_nic_access(trans, &flags))
873 		goto out;
874 
875 	/* Stop each Tx DMA channel */
876 	for (ch = 0; ch < FH_TCSR_CHNL_NUM; ch++) {
877 		iwl_write32(trans, FH_TCSR_CHNL_TX_CONFIG_REG(ch), 0x0);
878 		mask |= FH_TSSR_TX_STATUS_REG_MSK_CHNL_IDLE(ch);
879 	}
880 
881 	/* Wait for DMA channels to be idle */
882 	ret = iwl_poll_bit(trans, FH_TSSR_TX_STATUS_REG, mask, mask, 5000);
883 	if (ret < 0)
884 		IWL_ERR(trans,
885 			"Failing on timeout while stopping DMA channel %d [0x%08x]\n",
886 			ch, iwl_read32(trans, FH_TSSR_TX_STATUS_REG));
887 
888 	iwl_trans_release_nic_access(trans, &flags);
889 
890 out:
891 	spin_unlock(&trans_pcie->irq_lock);
892 }
893 
894 /*
895  * iwl_pcie_tx_stop - Stop all Tx DMA channels
896  */
897 int iwl_pcie_tx_stop(struct iwl_trans *trans)
898 {
899 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
900 	int txq_id;
901 
902 	/* Turn off all Tx DMA fifos */
903 	iwl_scd_deactivate_fifos(trans);
904 
905 	/* Turn off all Tx DMA channels */
906 	iwl_pcie_tx_stop_fh(trans);
907 
908 	/*
909 	 * This function can be called before the op_mode disabled the
910 	 * queues. This happens when we have an rfkill interrupt.
911 	 * Since we stop Tx altogether - mark the queues as stopped.
912 	 */
913 	memset(trans_pcie->queue_stopped, 0, sizeof(trans_pcie->queue_stopped));
914 	memset(trans_pcie->queue_used, 0, sizeof(trans_pcie->queue_used));
915 
916 	/* This can happen: start_hw, stop_device */
917 	if (!trans_pcie->txq)
918 		return 0;
919 
920 	/* Unmap DMA from host system and free skb's */
921 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
922 	     txq_id++)
923 		iwl_pcie_txq_unmap(trans, txq_id);
924 
925 	return 0;
926 }
927 
928 /*
929  * iwl_trans_tx_free - Free TXQ Context
930  *
931  * Destroy all TX DMA queues and structures
932  */
933 void iwl_pcie_tx_free(struct iwl_trans *trans)
934 {
935 	int txq_id;
936 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
937 
938 	/* Tx queues */
939 	if (trans_pcie->txq) {
940 		for (txq_id = 0;
941 		     txq_id < trans->cfg->base_params->num_of_queues; txq_id++)
942 			iwl_pcie_txq_free(trans, txq_id);
943 	}
944 
945 	kfree(trans_pcie->txq);
946 	trans_pcie->txq = NULL;
947 
948 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->kw);
949 
950 	iwl_pcie_free_dma_ptr(trans, &trans_pcie->scd_bc_tbls);
951 }
952 
953 /*
954  * iwl_pcie_tx_alloc - allocate TX context
955  * Allocate all Tx DMA structures and initialize them
956  */
957 static int iwl_pcie_tx_alloc(struct iwl_trans *trans)
958 {
959 	int ret;
960 	int txq_id, slots_num;
961 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
962 
963 	u16 scd_bc_tbls_size = trans->cfg->base_params->num_of_queues *
964 			sizeof(struct iwlagn_scd_bc_tbl);
965 
966 	/*It is not allowed to alloc twice, so warn when this happens.
967 	 * We cannot rely on the previous allocation, so free and fail */
968 	if (WARN_ON(trans_pcie->txq)) {
969 		ret = -EINVAL;
970 		goto error;
971 	}
972 
973 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->scd_bc_tbls,
974 				   scd_bc_tbls_size);
975 	if (ret) {
976 		IWL_ERR(trans, "Scheduler BC Table allocation failed\n");
977 		goto error;
978 	}
979 
980 	/* Alloc keep-warm buffer */
981 	ret = iwl_pcie_alloc_dma_ptr(trans, &trans_pcie->kw, IWL_KW_SIZE);
982 	if (ret) {
983 		IWL_ERR(trans, "Keep Warm allocation failed\n");
984 		goto error;
985 	}
986 
987 	trans_pcie->txq = kcalloc(trans->cfg->base_params->num_of_queues,
988 				  sizeof(struct iwl_txq), GFP_KERNEL);
989 	if (!trans_pcie->txq) {
990 		IWL_ERR(trans, "Not enough memory for txq\n");
991 		ret = -ENOMEM;
992 		goto error;
993 	}
994 
995 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
996 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
997 	     txq_id++) {
998 		slots_num = (txq_id == trans_pcie->cmd_queue) ?
999 					TFD_CMD_SLOTS : TFD_TX_CMD_SLOTS;
1000 		ret = iwl_pcie_txq_alloc(trans, &trans_pcie->txq[txq_id],
1001 					  slots_num, txq_id);
1002 		if (ret) {
1003 			IWL_ERR(trans, "Tx %d queue alloc failed\n", txq_id);
1004 			goto error;
1005 		}
1006 	}
1007 
1008 	return 0;
1009 
1010 error:
1011 	iwl_pcie_tx_free(trans);
1012 
1013 	return ret;
1014 }
1015 int iwl_pcie_tx_init(struct iwl_trans *trans)
1016 {
1017 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1018 	int ret;
1019 	int txq_id, slots_num;
1020 	bool alloc = false;
1021 
1022 	if (!trans_pcie->txq) {
1023 		ret = iwl_pcie_tx_alloc(trans);
1024 		if (ret)
1025 			goto error;
1026 		alloc = true;
1027 	}
1028 
1029 	spin_lock(&trans_pcie->irq_lock);
1030 
1031 	/* Turn off all Tx DMA fifos */
1032 	iwl_scd_deactivate_fifos(trans);
1033 
1034 	/* Tell NIC where to find the "keep warm" buffer */
1035 	iwl_write_direct32(trans, FH_KW_MEM_ADDR_REG,
1036 			   trans_pcie->kw.dma >> 4);
1037 
1038 	spin_unlock(&trans_pcie->irq_lock);
1039 
1040 	/* Alloc and init all Tx queues, including the command queue (#4/#9) */
1041 	for (txq_id = 0; txq_id < trans->cfg->base_params->num_of_queues;
1042 	     txq_id++) {
1043 		slots_num = (txq_id == trans_pcie->cmd_queue) ?
1044 					TFD_CMD_SLOTS : TFD_TX_CMD_SLOTS;
1045 		ret = iwl_pcie_txq_init(trans, &trans_pcie->txq[txq_id],
1046 					 slots_num, txq_id);
1047 		if (ret) {
1048 			IWL_ERR(trans, "Tx %d queue init failed\n", txq_id);
1049 			goto error;
1050 		}
1051 	}
1052 
1053 	if (trans->cfg->use_tfh) {
1054 		iwl_write_direct32(trans, TFH_TRANSFER_MODE,
1055 				   TFH_TRANSFER_MAX_PENDING_REQ |
1056 				   TFH_CHUNK_SIZE_128 |
1057 				   TFH_CHUNK_SPLIT_MODE);
1058 		return 0;
1059 	}
1060 
1061 	iwl_set_bits_prph(trans, SCD_GP_CTRL, SCD_GP_CTRL_AUTO_ACTIVE_MODE);
1062 	if (trans->cfg->base_params->num_of_queues > 20)
1063 		iwl_set_bits_prph(trans, SCD_GP_CTRL,
1064 				  SCD_GP_CTRL_ENABLE_31_QUEUES);
1065 
1066 	return 0;
1067 error:
1068 	/*Upon error, free only if we allocated something */
1069 	if (alloc)
1070 		iwl_pcie_tx_free(trans);
1071 	return ret;
1072 }
1073 
1074 static inline void iwl_pcie_txq_progress(struct iwl_txq *txq)
1075 {
1076 	lockdep_assert_held(&txq->lock);
1077 
1078 	if (!txq->wd_timeout)
1079 		return;
1080 
1081 	/*
1082 	 * station is asleep and we send data - that must
1083 	 * be uAPSD or PS-Poll. Don't rearm the timer.
1084 	 */
1085 	if (txq->frozen)
1086 		return;
1087 
1088 	/*
1089 	 * if empty delete timer, otherwise move timer forward
1090 	 * since we're making progress on this queue
1091 	 */
1092 	if (txq->read_ptr == txq->write_ptr)
1093 		del_timer(&txq->stuck_timer);
1094 	else
1095 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1096 }
1097 
1098 /* Frees buffers until index _not_ inclusive */
1099 void iwl_trans_pcie_reclaim(struct iwl_trans *trans, int txq_id, int ssn,
1100 			    struct sk_buff_head *skbs)
1101 {
1102 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1103 	struct iwl_txq *txq = &trans_pcie->txq[txq_id];
1104 	int tfd_num = ssn & (TFD_QUEUE_SIZE_MAX - 1);
1105 	int last_to_free;
1106 
1107 	/* This function is not meant to release cmd queue*/
1108 	if (WARN_ON(txq_id == trans_pcie->cmd_queue))
1109 		return;
1110 
1111 	spin_lock_bh(&txq->lock);
1112 
1113 	if (!txq->active) {
1114 		IWL_DEBUG_TX_QUEUES(trans, "Q %d inactive - ignoring idx %d\n",
1115 				    txq_id, ssn);
1116 		goto out;
1117 	}
1118 
1119 	if (txq->read_ptr == tfd_num)
1120 		goto out;
1121 
1122 	IWL_DEBUG_TX_REPLY(trans, "[Q %d] %d -> %d (%d)\n",
1123 			   txq_id, txq->read_ptr, tfd_num, ssn);
1124 
1125 	/*Since we free until index _not_ inclusive, the one before index is
1126 	 * the last we will free. This one must be used */
1127 	last_to_free = iwl_queue_dec_wrap(tfd_num);
1128 
1129 	if (!iwl_queue_used(txq, last_to_free)) {
1130 		IWL_ERR(trans,
1131 			"%s: Read index for DMA queue txq id (%d), last_to_free %d is out of range [0-%d] %d %d.\n",
1132 			__func__, txq_id, last_to_free, TFD_QUEUE_SIZE_MAX,
1133 			txq->write_ptr, txq->read_ptr);
1134 		goto out;
1135 	}
1136 
1137 	if (WARN_ON(!skb_queue_empty(skbs)))
1138 		goto out;
1139 
1140 	for (;
1141 	     txq->read_ptr != tfd_num;
1142 	     txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr)) {
1143 		struct sk_buff *skb = txq->entries[txq->read_ptr].skb;
1144 
1145 		if (WARN_ON_ONCE(!skb))
1146 			continue;
1147 
1148 		iwl_pcie_free_tso_page(trans_pcie, skb);
1149 
1150 		__skb_queue_tail(skbs, skb);
1151 
1152 		txq->entries[txq->read_ptr].skb = NULL;
1153 
1154 		if (!trans->cfg->use_tfh)
1155 			iwl_pcie_txq_inval_byte_cnt_tbl(trans, txq);
1156 
1157 		iwl_pcie_txq_free_tfd(trans, txq);
1158 	}
1159 
1160 	iwl_pcie_txq_progress(txq);
1161 
1162 	if (iwl_queue_space(txq) > txq->low_mark &&
1163 	    test_bit(txq_id, trans_pcie->queue_stopped)) {
1164 		struct sk_buff_head overflow_skbs;
1165 
1166 		__skb_queue_head_init(&overflow_skbs);
1167 		skb_queue_splice_init(&txq->overflow_q, &overflow_skbs);
1168 
1169 		/*
1170 		 * This is tricky: we are in reclaim path which is non
1171 		 * re-entrant, so noone will try to take the access the
1172 		 * txq data from that path. We stopped tx, so we can't
1173 		 * have tx as well. Bottom line, we can unlock and re-lock
1174 		 * later.
1175 		 */
1176 		spin_unlock_bh(&txq->lock);
1177 
1178 		while (!skb_queue_empty(&overflow_skbs)) {
1179 			struct sk_buff *skb = __skb_dequeue(&overflow_skbs);
1180 			struct iwl_device_cmd *dev_cmd_ptr;
1181 
1182 			dev_cmd_ptr = *(void **)((u8 *)skb->cb +
1183 						 trans_pcie->dev_cmd_offs);
1184 
1185 			/*
1186 			 * Note that we can very well be overflowing again.
1187 			 * In that case, iwl_queue_space will be small again
1188 			 * and we won't wake mac80211's queue.
1189 			 */
1190 			iwl_trans_pcie_tx(trans, skb, dev_cmd_ptr, txq_id);
1191 		}
1192 		spin_lock_bh(&txq->lock);
1193 
1194 		if (iwl_queue_space(txq) > txq->low_mark)
1195 			iwl_wake_queue(trans, txq);
1196 	}
1197 
1198 	if (txq->read_ptr == txq->write_ptr) {
1199 		IWL_DEBUG_RPM(trans, "Q %d - last tx reclaimed\n", txq->id);
1200 		iwl_trans_unref(trans);
1201 	}
1202 
1203 out:
1204 	spin_unlock_bh(&txq->lock);
1205 }
1206 
1207 static int iwl_pcie_set_cmd_in_flight(struct iwl_trans *trans,
1208 				      const struct iwl_host_cmd *cmd)
1209 {
1210 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1211 	int ret;
1212 
1213 	lockdep_assert_held(&trans_pcie->reg_lock);
1214 
1215 	if (!(cmd->flags & CMD_SEND_IN_IDLE) &&
1216 	    !trans_pcie->ref_cmd_in_flight) {
1217 		trans_pcie->ref_cmd_in_flight = true;
1218 		IWL_DEBUG_RPM(trans, "set ref_cmd_in_flight - ref\n");
1219 		iwl_trans_ref(trans);
1220 	}
1221 
1222 	/*
1223 	 * wake up the NIC to make sure that the firmware will see the host
1224 	 * command - we will let the NIC sleep once all the host commands
1225 	 * returned. This needs to be done only on NICs that have
1226 	 * apmg_wake_up_wa set.
1227 	 */
1228 	if (trans->cfg->base_params->apmg_wake_up_wa &&
1229 	    !trans_pcie->cmd_hold_nic_awake) {
1230 		__iwl_trans_pcie_set_bit(trans, CSR_GP_CNTRL,
1231 					 CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1232 
1233 		ret = iwl_poll_bit(trans, CSR_GP_CNTRL,
1234 				   CSR_GP_CNTRL_REG_VAL_MAC_ACCESS_EN,
1235 				   (CSR_GP_CNTRL_REG_FLAG_MAC_CLOCK_READY |
1236 				    CSR_GP_CNTRL_REG_FLAG_GOING_TO_SLEEP),
1237 				   15000);
1238 		if (ret < 0) {
1239 			__iwl_trans_pcie_clear_bit(trans, CSR_GP_CNTRL,
1240 					CSR_GP_CNTRL_REG_FLAG_MAC_ACCESS_REQ);
1241 			IWL_ERR(trans, "Failed to wake NIC for hcmd\n");
1242 			return -EIO;
1243 		}
1244 		trans_pcie->cmd_hold_nic_awake = true;
1245 	}
1246 
1247 	return 0;
1248 }
1249 
1250 /*
1251  * iwl_pcie_cmdq_reclaim - Reclaim TX command queue entries already Tx'd
1252  *
1253  * When FW advances 'R' index, all entries between old and new 'R' index
1254  * need to be reclaimed. As result, some free space forms.  If there is
1255  * enough free space (> low mark), wake the stack that feeds us.
1256  */
1257 static void iwl_pcie_cmdq_reclaim(struct iwl_trans *trans, int txq_id, int idx)
1258 {
1259 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1260 	struct iwl_txq *txq = &trans_pcie->txq[txq_id];
1261 	unsigned long flags;
1262 	int nfreed = 0;
1263 
1264 	lockdep_assert_held(&txq->lock);
1265 
1266 	if ((idx >= TFD_QUEUE_SIZE_MAX) || (!iwl_queue_used(txq, idx))) {
1267 		IWL_ERR(trans,
1268 			"%s: Read index for DMA queue txq id (%d), index %d is out of range [0-%d] %d %d.\n",
1269 			__func__, txq_id, idx, TFD_QUEUE_SIZE_MAX,
1270 			txq->write_ptr, txq->read_ptr);
1271 		return;
1272 	}
1273 
1274 	for (idx = iwl_queue_inc_wrap(idx); txq->read_ptr != idx;
1275 	     txq->read_ptr = iwl_queue_inc_wrap(txq->read_ptr)) {
1276 
1277 		if (nfreed++ > 0) {
1278 			IWL_ERR(trans, "HCMD skipped: index (%d) %d %d\n",
1279 				idx, txq->write_ptr, txq->read_ptr);
1280 			iwl_force_nmi(trans);
1281 		}
1282 	}
1283 
1284 	if (txq->read_ptr == txq->write_ptr) {
1285 		spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1286 		iwl_pcie_clear_cmd_in_flight(trans);
1287 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1288 	}
1289 
1290 	iwl_pcie_txq_progress(txq);
1291 }
1292 
1293 static int iwl_pcie_txq_set_ratid_map(struct iwl_trans *trans, u16 ra_tid,
1294 				 u16 txq_id)
1295 {
1296 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1297 	u32 tbl_dw_addr;
1298 	u32 tbl_dw;
1299 	u16 scd_q2ratid;
1300 
1301 	scd_q2ratid = ra_tid & SCD_QUEUE_RA_TID_MAP_RATID_MSK;
1302 
1303 	tbl_dw_addr = trans_pcie->scd_base_addr +
1304 			SCD_TRANS_TBL_OFFSET_QUEUE(txq_id);
1305 
1306 	tbl_dw = iwl_trans_read_mem32(trans, tbl_dw_addr);
1307 
1308 	if (txq_id & 0x1)
1309 		tbl_dw = (scd_q2ratid << 16) | (tbl_dw & 0x0000FFFF);
1310 	else
1311 		tbl_dw = scd_q2ratid | (tbl_dw & 0xFFFF0000);
1312 
1313 	iwl_trans_write_mem32(trans, tbl_dw_addr, tbl_dw);
1314 
1315 	return 0;
1316 }
1317 
1318 /* Receiver address (actually, Rx station's index into station table),
1319  * combined with Traffic ID (QOS priority), in format used by Tx Scheduler */
1320 #define BUILD_RAxTID(sta_id, tid)	(((sta_id) << 4) + (tid))
1321 
1322 void iwl_trans_pcie_txq_enable(struct iwl_trans *trans, int txq_id, u16 ssn,
1323 			       const struct iwl_trans_txq_scd_cfg *cfg,
1324 			       unsigned int wdg_timeout)
1325 {
1326 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1327 	struct iwl_txq *txq = &trans_pcie->txq[txq_id];
1328 	int fifo = -1;
1329 
1330 	if (test_and_set_bit(txq_id, trans_pcie->queue_used))
1331 		WARN_ONCE(1, "queue %d already used - expect issues", txq_id);
1332 
1333 	if (cfg && trans->cfg->use_tfh)
1334 		WARN_ONCE(1, "Expected no calls to SCD configuration");
1335 
1336 	txq->wd_timeout = msecs_to_jiffies(wdg_timeout);
1337 
1338 	if (cfg) {
1339 		fifo = cfg->fifo;
1340 
1341 		/* Disable the scheduler prior configuring the cmd queue */
1342 		if (txq_id == trans_pcie->cmd_queue &&
1343 		    trans_pcie->scd_set_active)
1344 			iwl_scd_enable_set_active(trans, 0);
1345 
1346 		/* Stop this Tx queue before configuring it */
1347 		iwl_scd_txq_set_inactive(trans, txq_id);
1348 
1349 		/* Set this queue as a chain-building queue unless it is CMD */
1350 		if (txq_id != trans_pcie->cmd_queue)
1351 			iwl_scd_txq_set_chain(trans, txq_id);
1352 
1353 		if (cfg->aggregate) {
1354 			u16 ra_tid = BUILD_RAxTID(cfg->sta_id, cfg->tid);
1355 
1356 			/* Map receiver-address / traffic-ID to this queue */
1357 			iwl_pcie_txq_set_ratid_map(trans, ra_tid, txq_id);
1358 
1359 			/* enable aggregations for the queue */
1360 			iwl_scd_txq_enable_agg(trans, txq_id);
1361 			txq->ampdu = true;
1362 		} else {
1363 			/*
1364 			 * disable aggregations for the queue, this will also
1365 			 * make the ra_tid mapping configuration irrelevant
1366 			 * since it is now a non-AGG queue.
1367 			 */
1368 			iwl_scd_txq_disable_agg(trans, txq_id);
1369 
1370 			ssn = txq->read_ptr;
1371 		}
1372 	}
1373 
1374 	/* Place first TFD at index corresponding to start sequence number.
1375 	 * Assumes that ssn_idx is valid (!= 0xFFF) */
1376 	txq->read_ptr = (ssn & 0xff);
1377 	txq->write_ptr = (ssn & 0xff);
1378 	iwl_write_direct32(trans, HBUS_TARG_WRPTR,
1379 			   (ssn & 0xff) | (txq_id << 8));
1380 
1381 	if (cfg) {
1382 		u8 frame_limit = cfg->frame_limit;
1383 
1384 		iwl_write_prph(trans, SCD_QUEUE_RDPTR(txq_id), ssn);
1385 
1386 		/* Set up Tx window size and frame limit for this queue */
1387 		iwl_trans_write_mem32(trans, trans_pcie->scd_base_addr +
1388 				SCD_CONTEXT_QUEUE_OFFSET(txq_id), 0);
1389 		iwl_trans_write_mem32(trans,
1390 			trans_pcie->scd_base_addr +
1391 			SCD_CONTEXT_QUEUE_OFFSET(txq_id) + sizeof(u32),
1392 			((frame_limit << SCD_QUEUE_CTX_REG2_WIN_SIZE_POS) &
1393 					SCD_QUEUE_CTX_REG2_WIN_SIZE_MSK) |
1394 			((frame_limit << SCD_QUEUE_CTX_REG2_FRAME_LIMIT_POS) &
1395 					SCD_QUEUE_CTX_REG2_FRAME_LIMIT_MSK));
1396 
1397 		/* Set up status area in SRAM, map to Tx DMA/FIFO, activate */
1398 		iwl_write_prph(trans, SCD_QUEUE_STATUS_BITS(txq_id),
1399 			       (1 << SCD_QUEUE_STTS_REG_POS_ACTIVE) |
1400 			       (cfg->fifo << SCD_QUEUE_STTS_REG_POS_TXF) |
1401 			       (1 << SCD_QUEUE_STTS_REG_POS_WSL) |
1402 			       SCD_QUEUE_STTS_REG_MSK);
1403 
1404 		/* enable the scheduler for this queue (only) */
1405 		if (txq_id == trans_pcie->cmd_queue &&
1406 		    trans_pcie->scd_set_active)
1407 			iwl_scd_enable_set_active(trans, BIT(txq_id));
1408 
1409 		IWL_DEBUG_TX_QUEUES(trans,
1410 				    "Activate queue %d on FIFO %d WrPtr: %d\n",
1411 				    txq_id, fifo, ssn & 0xff);
1412 	} else {
1413 		IWL_DEBUG_TX_QUEUES(trans,
1414 				    "Activate queue %d WrPtr: %d\n",
1415 				    txq_id, ssn & 0xff);
1416 	}
1417 
1418 	txq->active = true;
1419 }
1420 
1421 void iwl_trans_pcie_txq_set_shared_mode(struct iwl_trans *trans, u32 txq_id,
1422 					bool shared_mode)
1423 {
1424 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1425 	struct iwl_txq *txq = &trans_pcie->txq[txq_id];
1426 
1427 	txq->ampdu = !shared_mode;
1428 }
1429 
1430 dma_addr_t iwl_trans_pcie_get_txq_byte_table(struct iwl_trans *trans, int txq)
1431 {
1432 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1433 
1434 	return trans_pcie->scd_bc_tbls.dma +
1435 	       txq * sizeof(struct iwlagn_scd_bc_tbl);
1436 }
1437 
1438 void iwl_trans_pcie_txq_disable(struct iwl_trans *trans, int txq_id,
1439 				bool configure_scd)
1440 {
1441 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1442 	u32 stts_addr = trans_pcie->scd_base_addr +
1443 			SCD_TX_STTS_QUEUE_OFFSET(txq_id);
1444 	static const u32 zero_val[4] = {};
1445 
1446 	trans_pcie->txq[txq_id].frozen_expiry_remainder = 0;
1447 	trans_pcie->txq[txq_id].frozen = false;
1448 
1449 	/*
1450 	 * Upon HW Rfkill - we stop the device, and then stop the queues
1451 	 * in the op_mode. Just for the sake of the simplicity of the op_mode,
1452 	 * allow the op_mode to call txq_disable after it already called
1453 	 * stop_device.
1454 	 */
1455 	if (!test_and_clear_bit(txq_id, trans_pcie->queue_used)) {
1456 		WARN_ONCE(test_bit(STATUS_DEVICE_ENABLED, &trans->status),
1457 			  "queue %d not used", txq_id);
1458 		return;
1459 	}
1460 
1461 	if (configure_scd && trans->cfg->use_tfh)
1462 		WARN_ONCE(1, "Expected no calls to SCD configuration");
1463 
1464 	if (configure_scd) {
1465 		iwl_scd_txq_set_inactive(trans, txq_id);
1466 
1467 		iwl_trans_write_mem(trans, stts_addr, (void *)zero_val,
1468 				    ARRAY_SIZE(zero_val));
1469 	}
1470 
1471 	iwl_pcie_txq_unmap(trans, txq_id);
1472 	trans_pcie->txq[txq_id].ampdu = false;
1473 
1474 	IWL_DEBUG_TX_QUEUES(trans, "Deactivate queue %d\n", txq_id);
1475 }
1476 
1477 /*************** HOST COMMAND QUEUE FUNCTIONS   *****/
1478 
1479 /*
1480  * iwl_pcie_enqueue_hcmd - enqueue a uCode command
1481  * @priv: device private data point
1482  * @cmd: a pointer to the ucode command structure
1483  *
1484  * The function returns < 0 values to indicate the operation
1485  * failed. On success, it returns the index (>= 0) of command in the
1486  * command queue.
1487  */
1488 static int iwl_pcie_enqueue_hcmd(struct iwl_trans *trans,
1489 				 struct iwl_host_cmd *cmd)
1490 {
1491 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1492 	struct iwl_txq *txq = &trans_pcie->txq[trans_pcie->cmd_queue];
1493 	struct iwl_device_cmd *out_cmd;
1494 	struct iwl_cmd_meta *out_meta;
1495 	unsigned long flags;
1496 	void *dup_buf = NULL;
1497 	dma_addr_t phys_addr;
1498 	int idx;
1499 	u16 copy_size, cmd_size, tb0_size;
1500 	bool had_nocopy = false;
1501 	u8 group_id = iwl_cmd_groupid(cmd->id);
1502 	int i, ret;
1503 	u32 cmd_pos;
1504 	const u8 *cmddata[IWL_MAX_CMD_TBS_PER_TFD];
1505 	u16 cmdlen[IWL_MAX_CMD_TBS_PER_TFD];
1506 
1507 	if (WARN(!trans->wide_cmd_header &&
1508 		 group_id > IWL_ALWAYS_LONG_GROUP,
1509 		 "unsupported wide command %#x\n", cmd->id))
1510 		return -EINVAL;
1511 
1512 	if (group_id != 0) {
1513 		copy_size = sizeof(struct iwl_cmd_header_wide);
1514 		cmd_size = sizeof(struct iwl_cmd_header_wide);
1515 	} else {
1516 		copy_size = sizeof(struct iwl_cmd_header);
1517 		cmd_size = sizeof(struct iwl_cmd_header);
1518 	}
1519 
1520 	/* need one for the header if the first is NOCOPY */
1521 	BUILD_BUG_ON(IWL_MAX_CMD_TBS_PER_TFD > IWL_NUM_OF_TBS - 1);
1522 
1523 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1524 		cmddata[i] = cmd->data[i];
1525 		cmdlen[i] = cmd->len[i];
1526 
1527 		if (!cmd->len[i])
1528 			continue;
1529 
1530 		/* need at least IWL_FIRST_TB_SIZE copied */
1531 		if (copy_size < IWL_FIRST_TB_SIZE) {
1532 			int copy = IWL_FIRST_TB_SIZE - copy_size;
1533 
1534 			if (copy > cmdlen[i])
1535 				copy = cmdlen[i];
1536 			cmdlen[i] -= copy;
1537 			cmddata[i] += copy;
1538 			copy_size += copy;
1539 		}
1540 
1541 		if (cmd->dataflags[i] & IWL_HCMD_DFL_NOCOPY) {
1542 			had_nocopy = true;
1543 			if (WARN_ON(cmd->dataflags[i] & IWL_HCMD_DFL_DUP)) {
1544 				idx = -EINVAL;
1545 				goto free_dup_buf;
1546 			}
1547 		} else if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP) {
1548 			/*
1549 			 * This is also a chunk that isn't copied
1550 			 * to the static buffer so set had_nocopy.
1551 			 */
1552 			had_nocopy = true;
1553 
1554 			/* only allowed once */
1555 			if (WARN_ON(dup_buf)) {
1556 				idx = -EINVAL;
1557 				goto free_dup_buf;
1558 			}
1559 
1560 			dup_buf = kmemdup(cmddata[i], cmdlen[i],
1561 					  GFP_ATOMIC);
1562 			if (!dup_buf)
1563 				return -ENOMEM;
1564 		} else {
1565 			/* NOCOPY must not be followed by normal! */
1566 			if (WARN_ON(had_nocopy)) {
1567 				idx = -EINVAL;
1568 				goto free_dup_buf;
1569 			}
1570 			copy_size += cmdlen[i];
1571 		}
1572 		cmd_size += cmd->len[i];
1573 	}
1574 
1575 	/*
1576 	 * If any of the command structures end up being larger than
1577 	 * the TFD_MAX_PAYLOAD_SIZE and they aren't dynamically
1578 	 * allocated into separate TFDs, then we will need to
1579 	 * increase the size of the buffers.
1580 	 */
1581 	if (WARN(copy_size > TFD_MAX_PAYLOAD_SIZE,
1582 		 "Command %s (%#x) is too large (%d bytes)\n",
1583 		 iwl_get_cmd_string(trans, cmd->id),
1584 		 cmd->id, copy_size)) {
1585 		idx = -EINVAL;
1586 		goto free_dup_buf;
1587 	}
1588 
1589 	spin_lock_bh(&txq->lock);
1590 
1591 	if (iwl_queue_space(txq) < ((cmd->flags & CMD_ASYNC) ? 2 : 1)) {
1592 		spin_unlock_bh(&txq->lock);
1593 
1594 		IWL_ERR(trans, "No space in command queue\n");
1595 		iwl_op_mode_cmd_queue_full(trans->op_mode);
1596 		idx = -ENOSPC;
1597 		goto free_dup_buf;
1598 	}
1599 
1600 	idx = get_cmd_index(txq, txq->write_ptr);
1601 	out_cmd = txq->entries[idx].cmd;
1602 	out_meta = &txq->entries[idx].meta;
1603 
1604 	memset(out_meta, 0, sizeof(*out_meta));	/* re-initialize to NULL */
1605 	if (cmd->flags & CMD_WANT_SKB)
1606 		out_meta->source = cmd;
1607 
1608 	/* set up the header */
1609 	if (group_id != 0) {
1610 		out_cmd->hdr_wide.cmd = iwl_cmd_opcode(cmd->id);
1611 		out_cmd->hdr_wide.group_id = group_id;
1612 		out_cmd->hdr_wide.version = iwl_cmd_version(cmd->id);
1613 		out_cmd->hdr_wide.length =
1614 			cpu_to_le16(cmd_size -
1615 				    sizeof(struct iwl_cmd_header_wide));
1616 		out_cmd->hdr_wide.reserved = 0;
1617 		out_cmd->hdr_wide.sequence =
1618 			cpu_to_le16(QUEUE_TO_SEQ(trans_pcie->cmd_queue) |
1619 						 INDEX_TO_SEQ(txq->write_ptr));
1620 
1621 		cmd_pos = sizeof(struct iwl_cmd_header_wide);
1622 		copy_size = sizeof(struct iwl_cmd_header_wide);
1623 	} else {
1624 		out_cmd->hdr.cmd = iwl_cmd_opcode(cmd->id);
1625 		out_cmd->hdr.sequence =
1626 			cpu_to_le16(QUEUE_TO_SEQ(trans_pcie->cmd_queue) |
1627 						 INDEX_TO_SEQ(txq->write_ptr));
1628 		out_cmd->hdr.group_id = 0;
1629 
1630 		cmd_pos = sizeof(struct iwl_cmd_header);
1631 		copy_size = sizeof(struct iwl_cmd_header);
1632 	}
1633 
1634 	/* and copy the data that needs to be copied */
1635 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1636 		int copy;
1637 
1638 		if (!cmd->len[i])
1639 			continue;
1640 
1641 		/* copy everything if not nocopy/dup */
1642 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1643 					   IWL_HCMD_DFL_DUP))) {
1644 			copy = cmd->len[i];
1645 
1646 			memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1647 			cmd_pos += copy;
1648 			copy_size += copy;
1649 			continue;
1650 		}
1651 
1652 		/*
1653 		 * Otherwise we need at least IWL_FIRST_TB_SIZE copied
1654 		 * in total (for bi-directional DMA), but copy up to what
1655 		 * we can fit into the payload for debug dump purposes.
1656 		 */
1657 		copy = min_t(int, TFD_MAX_PAYLOAD_SIZE - cmd_pos, cmd->len[i]);
1658 
1659 		memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
1660 		cmd_pos += copy;
1661 
1662 		/* However, treat copy_size the proper way, we need it below */
1663 		if (copy_size < IWL_FIRST_TB_SIZE) {
1664 			copy = IWL_FIRST_TB_SIZE - copy_size;
1665 
1666 			if (copy > cmd->len[i])
1667 				copy = cmd->len[i];
1668 			copy_size += copy;
1669 		}
1670 	}
1671 
1672 	IWL_DEBUG_HC(trans,
1673 		     "Sending command %s (%.2x.%.2x), seq: 0x%04X, %d bytes at %d[%d]:%d\n",
1674 		     iwl_get_cmd_string(trans, cmd->id),
1675 		     group_id, out_cmd->hdr.cmd,
1676 		     le16_to_cpu(out_cmd->hdr.sequence),
1677 		     cmd_size, txq->write_ptr, idx, trans_pcie->cmd_queue);
1678 
1679 	/* start the TFD with the minimum copy bytes */
1680 	tb0_size = min_t(int, copy_size, IWL_FIRST_TB_SIZE);
1681 	memcpy(&txq->first_tb_bufs[idx], &out_cmd->hdr, tb0_size);
1682 	iwl_pcie_txq_build_tfd(trans, txq,
1683 			       iwl_pcie_get_first_tb_dma(txq, idx),
1684 			       tb0_size, true);
1685 
1686 	/* map first command fragment, if any remains */
1687 	if (copy_size > tb0_size) {
1688 		phys_addr = dma_map_single(trans->dev,
1689 					   ((u8 *)&out_cmd->hdr) + tb0_size,
1690 					   copy_size - tb0_size,
1691 					   DMA_TO_DEVICE);
1692 		if (dma_mapping_error(trans->dev, phys_addr)) {
1693 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1694 					   txq->write_ptr);
1695 			idx = -ENOMEM;
1696 			goto out;
1697 		}
1698 
1699 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr,
1700 				       copy_size - tb0_size, false);
1701 	}
1702 
1703 	/* map the remaining (adjusted) nocopy/dup fragments */
1704 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
1705 		const void *data = cmddata[i];
1706 
1707 		if (!cmdlen[i])
1708 			continue;
1709 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
1710 					   IWL_HCMD_DFL_DUP)))
1711 			continue;
1712 		if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP)
1713 			data = dup_buf;
1714 		phys_addr = dma_map_single(trans->dev, (void *)data,
1715 					   cmdlen[i], DMA_TO_DEVICE);
1716 		if (dma_mapping_error(trans->dev, phys_addr)) {
1717 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
1718 					   txq->write_ptr);
1719 			idx = -ENOMEM;
1720 			goto out;
1721 		}
1722 
1723 		iwl_pcie_txq_build_tfd(trans, txq, phys_addr, cmdlen[i], false);
1724 	}
1725 
1726 	BUILD_BUG_ON(IWL_TFH_NUM_TBS > sizeof(out_meta->tbs) * BITS_PER_BYTE);
1727 	out_meta->flags = cmd->flags;
1728 	if (WARN_ON_ONCE(txq->entries[idx].free_buf))
1729 		kzfree(txq->entries[idx].free_buf);
1730 	txq->entries[idx].free_buf = dup_buf;
1731 
1732 	trace_iwlwifi_dev_hcmd(trans->dev, cmd, cmd_size, &out_cmd->hdr_wide);
1733 
1734 	/* start timer if queue currently empty */
1735 	if (txq->read_ptr == txq->write_ptr && txq->wd_timeout)
1736 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1737 
1738 	spin_lock_irqsave(&trans_pcie->reg_lock, flags);
1739 	ret = iwl_pcie_set_cmd_in_flight(trans, cmd);
1740 	if (ret < 0) {
1741 		idx = ret;
1742 		spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1743 		goto out;
1744 	}
1745 
1746 	/* Increment and update queue's write index */
1747 	txq->write_ptr = iwl_queue_inc_wrap(txq->write_ptr);
1748 	iwl_pcie_txq_inc_wr_ptr(trans, txq);
1749 
1750 	spin_unlock_irqrestore(&trans_pcie->reg_lock, flags);
1751 
1752  out:
1753 	spin_unlock_bh(&txq->lock);
1754  free_dup_buf:
1755 	if (idx < 0)
1756 		kfree(dup_buf);
1757 	return idx;
1758 }
1759 
1760 /*
1761  * iwl_pcie_hcmd_complete - Pull unused buffers off the queue and reclaim them
1762  * @rxb: Rx buffer to reclaim
1763  */
1764 void iwl_pcie_hcmd_complete(struct iwl_trans *trans,
1765 			    struct iwl_rx_cmd_buffer *rxb)
1766 {
1767 	struct iwl_rx_packet *pkt = rxb_addr(rxb);
1768 	u16 sequence = le16_to_cpu(pkt->hdr.sequence);
1769 	u8 group_id = iwl_cmd_groupid(pkt->hdr.group_id);
1770 	u32 cmd_id;
1771 	int txq_id = SEQ_TO_QUEUE(sequence);
1772 	int index = SEQ_TO_INDEX(sequence);
1773 	int cmd_index;
1774 	struct iwl_device_cmd *cmd;
1775 	struct iwl_cmd_meta *meta;
1776 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1777 	struct iwl_txq *txq = &trans_pcie->txq[trans_pcie->cmd_queue];
1778 
1779 	/* If a Tx command is being handled and it isn't in the actual
1780 	 * command queue then there a command routing bug has been introduced
1781 	 * in the queue management code. */
1782 	if (WARN(txq_id != trans_pcie->cmd_queue,
1783 		 "wrong command queue %d (should be %d), sequence 0x%X readp=%d writep=%d\n",
1784 		 txq_id, trans_pcie->cmd_queue, sequence,
1785 		 trans_pcie->txq[trans_pcie->cmd_queue].read_ptr,
1786 		 trans_pcie->txq[trans_pcie->cmd_queue].write_ptr)) {
1787 		iwl_print_hex_error(trans, pkt, 32);
1788 		return;
1789 	}
1790 
1791 	spin_lock_bh(&txq->lock);
1792 
1793 	cmd_index = get_cmd_index(txq, index);
1794 	cmd = txq->entries[cmd_index].cmd;
1795 	meta = &txq->entries[cmd_index].meta;
1796 	cmd_id = iwl_cmd_id(cmd->hdr.cmd, group_id, 0);
1797 
1798 	iwl_pcie_tfd_unmap(trans, meta, txq, index);
1799 
1800 	/* Input error checking is done when commands are added to queue. */
1801 	if (meta->flags & CMD_WANT_SKB) {
1802 		struct page *p = rxb_steal_page(rxb);
1803 
1804 		meta->source->resp_pkt = pkt;
1805 		meta->source->_rx_page_addr = (unsigned long)page_address(p);
1806 		meta->source->_rx_page_order = trans_pcie->rx_page_order;
1807 	}
1808 
1809 	if (meta->flags & CMD_WANT_ASYNC_CALLBACK)
1810 		iwl_op_mode_async_cb(trans->op_mode, cmd);
1811 
1812 	iwl_pcie_cmdq_reclaim(trans, txq_id, index);
1813 
1814 	if (!(meta->flags & CMD_ASYNC)) {
1815 		if (!test_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status)) {
1816 			IWL_WARN(trans,
1817 				 "HCMD_ACTIVE already clear for command %s\n",
1818 				 iwl_get_cmd_string(trans, cmd_id));
1819 		}
1820 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1821 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1822 			       iwl_get_cmd_string(trans, cmd_id));
1823 		wake_up(&trans_pcie->wait_command_queue);
1824 	}
1825 
1826 	if (meta->flags & CMD_MAKE_TRANS_IDLE) {
1827 		IWL_DEBUG_INFO(trans, "complete %s - mark trans as idle\n",
1828 			       iwl_get_cmd_string(trans, cmd->hdr.cmd));
1829 		set_bit(STATUS_TRANS_IDLE, &trans->status);
1830 		wake_up(&trans_pcie->d0i3_waitq);
1831 	}
1832 
1833 	if (meta->flags & CMD_WAKE_UP_TRANS) {
1834 		IWL_DEBUG_INFO(trans, "complete %s - clear trans idle flag\n",
1835 			       iwl_get_cmd_string(trans, cmd->hdr.cmd));
1836 		clear_bit(STATUS_TRANS_IDLE, &trans->status);
1837 		wake_up(&trans_pcie->d0i3_waitq);
1838 	}
1839 
1840 	meta->flags = 0;
1841 
1842 	spin_unlock_bh(&txq->lock);
1843 }
1844 
1845 #define HOST_COMPLETE_TIMEOUT	(2 * HZ)
1846 
1847 static int iwl_pcie_send_hcmd_async(struct iwl_trans *trans,
1848 				    struct iwl_host_cmd *cmd)
1849 {
1850 	int ret;
1851 
1852 	/* An asynchronous command can not expect an SKB to be set. */
1853 	if (WARN_ON(cmd->flags & CMD_WANT_SKB))
1854 		return -EINVAL;
1855 
1856 	ret = iwl_pcie_enqueue_hcmd(trans, cmd);
1857 	if (ret < 0) {
1858 		IWL_ERR(trans,
1859 			"Error sending %s: enqueue_hcmd failed: %d\n",
1860 			iwl_get_cmd_string(trans, cmd->id), ret);
1861 		return ret;
1862 	}
1863 	return 0;
1864 }
1865 
1866 static int iwl_pcie_send_hcmd_sync(struct iwl_trans *trans,
1867 				   struct iwl_host_cmd *cmd)
1868 {
1869 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1870 	int cmd_idx;
1871 	int ret;
1872 
1873 	IWL_DEBUG_INFO(trans, "Attempting to send sync command %s\n",
1874 		       iwl_get_cmd_string(trans, cmd->id));
1875 
1876 	if (WARN(test_and_set_bit(STATUS_SYNC_HCMD_ACTIVE,
1877 				  &trans->status),
1878 		 "Command %s: a command is already active!\n",
1879 		 iwl_get_cmd_string(trans, cmd->id)))
1880 		return -EIO;
1881 
1882 	IWL_DEBUG_INFO(trans, "Setting HCMD_ACTIVE for command %s\n",
1883 		       iwl_get_cmd_string(trans, cmd->id));
1884 
1885 	if (pm_runtime_suspended(&trans_pcie->pci_dev->dev)) {
1886 		ret = wait_event_timeout(trans_pcie->d0i3_waitq,
1887 				 pm_runtime_active(&trans_pcie->pci_dev->dev),
1888 				 msecs_to_jiffies(IWL_TRANS_IDLE_TIMEOUT));
1889 		if (!ret) {
1890 			IWL_ERR(trans, "Timeout exiting D0i3 before hcmd\n");
1891 			return -ETIMEDOUT;
1892 		}
1893 	}
1894 
1895 	cmd_idx = iwl_pcie_enqueue_hcmd(trans, cmd);
1896 	if (cmd_idx < 0) {
1897 		ret = cmd_idx;
1898 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1899 		IWL_ERR(trans,
1900 			"Error sending %s: enqueue_hcmd failed: %d\n",
1901 			iwl_get_cmd_string(trans, cmd->id), ret);
1902 		return ret;
1903 	}
1904 
1905 	ret = wait_event_timeout(trans_pcie->wait_command_queue,
1906 				 !test_bit(STATUS_SYNC_HCMD_ACTIVE,
1907 					   &trans->status),
1908 				 HOST_COMPLETE_TIMEOUT);
1909 	if (!ret) {
1910 		struct iwl_txq *txq = &trans_pcie->txq[trans_pcie->cmd_queue];
1911 
1912 		IWL_ERR(trans, "Error sending %s: time out after %dms.\n",
1913 			iwl_get_cmd_string(trans, cmd->id),
1914 			jiffies_to_msecs(HOST_COMPLETE_TIMEOUT));
1915 
1916 		IWL_ERR(trans, "Current CMD queue read_ptr %d write_ptr %d\n",
1917 			txq->read_ptr, txq->write_ptr);
1918 
1919 		clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1920 		IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1921 			       iwl_get_cmd_string(trans, cmd->id));
1922 		ret = -ETIMEDOUT;
1923 
1924 		iwl_force_nmi(trans);
1925 		iwl_trans_fw_error(trans);
1926 
1927 		goto cancel;
1928 	}
1929 
1930 	if (test_bit(STATUS_FW_ERROR, &trans->status)) {
1931 		IWL_ERR(trans, "FW error in SYNC CMD %s\n",
1932 			iwl_get_cmd_string(trans, cmd->id));
1933 		dump_stack();
1934 		ret = -EIO;
1935 		goto cancel;
1936 	}
1937 
1938 	if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1939 	    test_bit(STATUS_RFKILL, &trans->status)) {
1940 		IWL_DEBUG_RF_KILL(trans, "RFKILL in SYNC CMD... no rsp\n");
1941 		ret = -ERFKILL;
1942 		goto cancel;
1943 	}
1944 
1945 	if ((cmd->flags & CMD_WANT_SKB) && !cmd->resp_pkt) {
1946 		IWL_ERR(trans, "Error: Response NULL in '%s'\n",
1947 			iwl_get_cmd_string(trans, cmd->id));
1948 		ret = -EIO;
1949 		goto cancel;
1950 	}
1951 
1952 	return 0;
1953 
1954 cancel:
1955 	if (cmd->flags & CMD_WANT_SKB) {
1956 		/*
1957 		 * Cancel the CMD_WANT_SKB flag for the cmd in the
1958 		 * TX cmd queue. Otherwise in case the cmd comes
1959 		 * in later, it will possibly set an invalid
1960 		 * address (cmd->meta.source).
1961 		 */
1962 		trans_pcie->txq[trans_pcie->cmd_queue].
1963 			entries[cmd_idx].meta.flags &= ~CMD_WANT_SKB;
1964 	}
1965 
1966 	if (cmd->resp_pkt) {
1967 		iwl_free_resp(cmd);
1968 		cmd->resp_pkt = NULL;
1969 	}
1970 
1971 	return ret;
1972 }
1973 
1974 int iwl_trans_pcie_send_hcmd(struct iwl_trans *trans, struct iwl_host_cmd *cmd)
1975 {
1976 	if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1977 	    test_bit(STATUS_RFKILL, &trans->status)) {
1978 		IWL_DEBUG_RF_KILL(trans, "Dropping CMD 0x%x: RF KILL\n",
1979 				  cmd->id);
1980 		return -ERFKILL;
1981 	}
1982 
1983 	if (cmd->flags & CMD_ASYNC)
1984 		return iwl_pcie_send_hcmd_async(trans, cmd);
1985 
1986 	/* We still can fail on RFKILL that can be asserted while we wait */
1987 	return iwl_pcie_send_hcmd_sync(trans, cmd);
1988 }
1989 
1990 static int iwl_fill_data_tbs(struct iwl_trans *trans, struct sk_buff *skb,
1991 			     struct iwl_txq *txq, u8 hdr_len,
1992 			     struct iwl_cmd_meta *out_meta,
1993 			     struct iwl_device_cmd *dev_cmd, u16 tb1_len)
1994 {
1995 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
1996 	u16 tb2_len;
1997 	int i;
1998 
1999 	/*
2000 	 * Set up TFD's third entry to point directly to remainder
2001 	 * of skb's head, if any
2002 	 */
2003 	tb2_len = skb_headlen(skb) - hdr_len;
2004 
2005 	if (tb2_len > 0) {
2006 		dma_addr_t tb2_phys = dma_map_single(trans->dev,
2007 						     skb->data + hdr_len,
2008 						     tb2_len, DMA_TO_DEVICE);
2009 		if (unlikely(dma_mapping_error(trans->dev, tb2_phys))) {
2010 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
2011 					   txq->write_ptr);
2012 			return -EINVAL;
2013 		}
2014 		iwl_pcie_txq_build_tfd(trans, txq, tb2_phys, tb2_len, false);
2015 	}
2016 
2017 	/* set up the remaining entries to point to the data */
2018 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2019 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2020 		dma_addr_t tb_phys;
2021 		int tb_idx;
2022 
2023 		if (!skb_frag_size(frag))
2024 			continue;
2025 
2026 		tb_phys = skb_frag_dma_map(trans->dev, frag, 0,
2027 					   skb_frag_size(frag), DMA_TO_DEVICE);
2028 
2029 		if (unlikely(dma_mapping_error(trans->dev, tb_phys))) {
2030 			iwl_pcie_tfd_unmap(trans, out_meta, txq,
2031 					   txq->write_ptr);
2032 			return -EINVAL;
2033 		}
2034 		tb_idx = iwl_pcie_txq_build_tfd(trans, txq, tb_phys,
2035 						skb_frag_size(frag), false);
2036 
2037 		out_meta->tbs |= BIT(tb_idx);
2038 	}
2039 
2040 	trace_iwlwifi_dev_tx(trans->dev, skb,
2041 			     iwl_pcie_get_tfd(trans_pcie, txq, txq->write_ptr),
2042 			     trans_pcie->tfd_size,
2043 			     &dev_cmd->hdr, IWL_FIRST_TB_SIZE + tb1_len,
2044 			     skb->data + hdr_len, tb2_len);
2045 	trace_iwlwifi_dev_tx_data(trans->dev, skb,
2046 				  hdr_len, skb->len - hdr_len);
2047 	return 0;
2048 }
2049 
2050 #ifdef CONFIG_INET
2051 static struct iwl_tso_hdr_page *
2052 get_page_hdr(struct iwl_trans *trans, size_t len)
2053 {
2054 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
2055 	struct iwl_tso_hdr_page *p = this_cpu_ptr(trans_pcie->tso_hdr_page);
2056 
2057 	if (!p->page)
2058 		goto alloc;
2059 
2060 	/* enough room on this page */
2061 	if (p->pos + len < (u8 *)page_address(p->page) + PAGE_SIZE)
2062 		return p;
2063 
2064 	/* We don't have enough room on this page, get a new one. */
2065 	__free_page(p->page);
2066 
2067 alloc:
2068 	p->page = alloc_page(GFP_ATOMIC);
2069 	if (!p->page)
2070 		return NULL;
2071 	p->pos = page_address(p->page);
2072 	return p;
2073 }
2074 
2075 static void iwl_compute_pseudo_hdr_csum(void *iph, struct tcphdr *tcph,
2076 					bool ipv6, unsigned int len)
2077 {
2078 	if (ipv6) {
2079 		struct ipv6hdr *iphv6 = iph;
2080 
2081 		tcph->check = ~csum_ipv6_magic(&iphv6->saddr, &iphv6->daddr,
2082 					       len + tcph->doff * 4,
2083 					       IPPROTO_TCP, 0);
2084 	} else {
2085 		struct iphdr *iphv4 = iph;
2086 
2087 		ip_send_check(iphv4);
2088 		tcph->check = ~csum_tcpudp_magic(iphv4->saddr, iphv4->daddr,
2089 						 len + tcph->doff * 4,
2090 						 IPPROTO_TCP, 0);
2091 	}
2092 }
2093 
2094 static int iwl_fill_data_tbs_amsdu(struct iwl_trans *trans, struct sk_buff *skb,
2095 				   struct iwl_txq *txq, u8 hdr_len,
2096 				   struct iwl_cmd_meta *out_meta,
2097 				   struct iwl_device_cmd *dev_cmd, u16 tb1_len)
2098 {
2099 	struct iwl_tx_cmd *tx_cmd = (void *)dev_cmd->payload;
2100 	struct iwl_trans_pcie *trans_pcie = txq->trans_pcie;
2101 	struct ieee80211_hdr *hdr = (void *)skb->data;
2102 	unsigned int snap_ip_tcp_hdrlen, ip_hdrlen, total_len, hdr_room;
2103 	unsigned int mss = skb_shinfo(skb)->gso_size;
2104 	u16 length, iv_len, amsdu_pad;
2105 	u8 *start_hdr;
2106 	struct iwl_tso_hdr_page *hdr_page;
2107 	struct page **page_ptr;
2108 	int ret;
2109 	struct tso_t tso;
2110 
2111 	/* if the packet is protected, then it must be CCMP or GCMP */
2112 	BUILD_BUG_ON(IEEE80211_CCMP_HDR_LEN != IEEE80211_GCMP_HDR_LEN);
2113 	iv_len = ieee80211_has_protected(hdr->frame_control) ?
2114 		IEEE80211_CCMP_HDR_LEN : 0;
2115 
2116 	trace_iwlwifi_dev_tx(trans->dev, skb,
2117 			     iwl_pcie_get_tfd(trans_pcie, txq, txq->write_ptr),
2118 			     trans_pcie->tfd_size,
2119 			     &dev_cmd->hdr, IWL_FIRST_TB_SIZE + tb1_len,
2120 			     NULL, 0);
2121 
2122 	ip_hdrlen = skb_transport_header(skb) - skb_network_header(skb);
2123 	snap_ip_tcp_hdrlen = 8 + ip_hdrlen + tcp_hdrlen(skb);
2124 	total_len = skb->len - snap_ip_tcp_hdrlen - hdr_len - iv_len;
2125 	amsdu_pad = 0;
2126 
2127 	/* total amount of header we may need for this A-MSDU */
2128 	hdr_room = DIV_ROUND_UP(total_len, mss) *
2129 		(3 + snap_ip_tcp_hdrlen + sizeof(struct ethhdr)) + iv_len;
2130 
2131 	/* Our device supports 9 segments at most, it will fit in 1 page */
2132 	hdr_page = get_page_hdr(trans, hdr_room);
2133 	if (!hdr_page)
2134 		return -ENOMEM;
2135 
2136 	get_page(hdr_page->page);
2137 	start_hdr = hdr_page->pos;
2138 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
2139 	*page_ptr = hdr_page->page;
2140 	memcpy(hdr_page->pos, skb->data + hdr_len, iv_len);
2141 	hdr_page->pos += iv_len;
2142 
2143 	/*
2144 	 * Pull the ieee80211 header + IV to be able to use TSO core,
2145 	 * we will restore it for the tx_status flow.
2146 	 */
2147 	skb_pull(skb, hdr_len + iv_len);
2148 
2149 	/*
2150 	 * Remove the length of all the headers that we don't actually
2151 	 * have in the MPDU by themselves, but that we duplicate into
2152 	 * all the different MSDUs inside the A-MSDU.
2153 	 */
2154 	le16_add_cpu(&tx_cmd->len, -snap_ip_tcp_hdrlen);
2155 
2156 	tso_start(skb, &tso);
2157 
2158 	while (total_len) {
2159 		/* this is the data left for this subframe */
2160 		unsigned int data_left =
2161 			min_t(unsigned int, mss, total_len);
2162 		struct sk_buff *csum_skb = NULL;
2163 		unsigned int hdr_tb_len;
2164 		dma_addr_t hdr_tb_phys;
2165 		struct tcphdr *tcph;
2166 		u8 *iph, *subf_hdrs_start = hdr_page->pos;
2167 
2168 		total_len -= data_left;
2169 
2170 		memset(hdr_page->pos, 0, amsdu_pad);
2171 		hdr_page->pos += amsdu_pad;
2172 		amsdu_pad = (4 - (sizeof(struct ethhdr) + snap_ip_tcp_hdrlen +
2173 				  data_left)) & 0x3;
2174 		ether_addr_copy(hdr_page->pos, ieee80211_get_DA(hdr));
2175 		hdr_page->pos += ETH_ALEN;
2176 		ether_addr_copy(hdr_page->pos, ieee80211_get_SA(hdr));
2177 		hdr_page->pos += ETH_ALEN;
2178 
2179 		length = snap_ip_tcp_hdrlen + data_left;
2180 		*((__be16 *)hdr_page->pos) = cpu_to_be16(length);
2181 		hdr_page->pos += sizeof(length);
2182 
2183 		/*
2184 		 * This will copy the SNAP as well which will be considered
2185 		 * as MAC header.
2186 		 */
2187 		tso_build_hdr(skb, hdr_page->pos, &tso, data_left, !total_len);
2188 		iph = hdr_page->pos + 8;
2189 		tcph = (void *)(iph + ip_hdrlen);
2190 
2191 		/* For testing on current hardware only */
2192 		if (trans_pcie->sw_csum_tx) {
2193 			csum_skb = alloc_skb(data_left + tcp_hdrlen(skb),
2194 					     GFP_ATOMIC);
2195 			if (!csum_skb) {
2196 				ret = -ENOMEM;
2197 				goto out_unmap;
2198 			}
2199 
2200 			iwl_compute_pseudo_hdr_csum(iph, tcph,
2201 						    skb->protocol ==
2202 							htons(ETH_P_IPV6),
2203 						    data_left);
2204 
2205 			memcpy(skb_put(csum_skb, tcp_hdrlen(skb)),
2206 			       tcph, tcp_hdrlen(skb));
2207 			skb_reset_transport_header(csum_skb);
2208 			csum_skb->csum_start =
2209 				(unsigned char *)tcp_hdr(csum_skb) -
2210 						 csum_skb->head;
2211 		}
2212 
2213 		hdr_page->pos += snap_ip_tcp_hdrlen;
2214 
2215 		hdr_tb_len = hdr_page->pos - start_hdr;
2216 		hdr_tb_phys = dma_map_single(trans->dev, start_hdr,
2217 					     hdr_tb_len, DMA_TO_DEVICE);
2218 		if (unlikely(dma_mapping_error(trans->dev, hdr_tb_phys))) {
2219 			dev_kfree_skb(csum_skb);
2220 			ret = -EINVAL;
2221 			goto out_unmap;
2222 		}
2223 		iwl_pcie_txq_build_tfd(trans, txq, hdr_tb_phys,
2224 				       hdr_tb_len, false);
2225 		trace_iwlwifi_dev_tx_tso_chunk(trans->dev, start_hdr,
2226 					       hdr_tb_len);
2227 		/* add this subframe's headers' length to the tx_cmd */
2228 		le16_add_cpu(&tx_cmd->len, hdr_page->pos - subf_hdrs_start);
2229 
2230 		/* prepare the start_hdr for the next subframe */
2231 		start_hdr = hdr_page->pos;
2232 
2233 		/* put the payload */
2234 		while (data_left) {
2235 			unsigned int size = min_t(unsigned int, tso.size,
2236 						  data_left);
2237 			dma_addr_t tb_phys;
2238 
2239 			if (trans_pcie->sw_csum_tx)
2240 				memcpy(skb_put(csum_skb, size), tso.data, size);
2241 
2242 			tb_phys = dma_map_single(trans->dev, tso.data,
2243 						 size, DMA_TO_DEVICE);
2244 			if (unlikely(dma_mapping_error(trans->dev, tb_phys))) {
2245 				dev_kfree_skb(csum_skb);
2246 				ret = -EINVAL;
2247 				goto out_unmap;
2248 			}
2249 
2250 			iwl_pcie_txq_build_tfd(trans, txq, tb_phys,
2251 					       size, false);
2252 			trace_iwlwifi_dev_tx_tso_chunk(trans->dev, tso.data,
2253 						       size);
2254 
2255 			data_left -= size;
2256 			tso_build_data(skb, &tso, size);
2257 		}
2258 
2259 		/* For testing on early hardware only */
2260 		if (trans_pcie->sw_csum_tx) {
2261 			__wsum csum;
2262 
2263 			csum = skb_checksum(csum_skb,
2264 					    skb_checksum_start_offset(csum_skb),
2265 					    csum_skb->len -
2266 					    skb_checksum_start_offset(csum_skb),
2267 					    0);
2268 			dev_kfree_skb(csum_skb);
2269 			dma_sync_single_for_cpu(trans->dev, hdr_tb_phys,
2270 						hdr_tb_len, DMA_TO_DEVICE);
2271 			tcph->check = csum_fold(csum);
2272 			dma_sync_single_for_device(trans->dev, hdr_tb_phys,
2273 						   hdr_tb_len, DMA_TO_DEVICE);
2274 		}
2275 	}
2276 
2277 	/* re -add the WiFi header and IV */
2278 	skb_push(skb, hdr_len + iv_len);
2279 
2280 	return 0;
2281 
2282 out_unmap:
2283 	iwl_pcie_tfd_unmap(trans, out_meta, txq, txq->write_ptr);
2284 	return ret;
2285 }
2286 #else /* CONFIG_INET */
2287 static int iwl_fill_data_tbs_amsdu(struct iwl_trans *trans, struct sk_buff *skb,
2288 				   struct iwl_txq *txq, u8 hdr_len,
2289 				   struct iwl_cmd_meta *out_meta,
2290 				   struct iwl_device_cmd *dev_cmd, u16 tb1_len)
2291 {
2292 	/* No A-MSDU without CONFIG_INET */
2293 	WARN_ON(1);
2294 
2295 	return -1;
2296 }
2297 #endif /* CONFIG_INET */
2298 
2299 int iwl_trans_pcie_tx(struct iwl_trans *trans, struct sk_buff *skb,
2300 		      struct iwl_device_cmd *dev_cmd, int txq_id)
2301 {
2302 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
2303 	struct ieee80211_hdr *hdr;
2304 	struct iwl_tx_cmd *tx_cmd = (struct iwl_tx_cmd *)dev_cmd->payload;
2305 	struct iwl_cmd_meta *out_meta;
2306 	struct iwl_txq *txq;
2307 	dma_addr_t tb0_phys, tb1_phys, scratch_phys;
2308 	void *tb1_addr;
2309 	void *tfd;
2310 	u16 len, tb1_len;
2311 	bool wait_write_ptr;
2312 	__le16 fc;
2313 	u8 hdr_len;
2314 	u16 wifi_seq;
2315 	bool amsdu;
2316 
2317 	txq = &trans_pcie->txq[txq_id];
2318 
2319 	if (WARN_ONCE(!test_bit(txq_id, trans_pcie->queue_used),
2320 		      "TX on unused queue %d\n", txq_id))
2321 		return -EINVAL;
2322 
2323 	if (unlikely(trans_pcie->sw_csum_tx &&
2324 		     skb->ip_summed == CHECKSUM_PARTIAL)) {
2325 		int offs = skb_checksum_start_offset(skb);
2326 		int csum_offs = offs + skb->csum_offset;
2327 		__wsum csum;
2328 
2329 		if (skb_ensure_writable(skb, csum_offs + sizeof(__sum16)))
2330 			return -1;
2331 
2332 		csum = skb_checksum(skb, offs, skb->len - offs, 0);
2333 		*(__sum16 *)(skb->data + csum_offs) = csum_fold(csum);
2334 
2335 		skb->ip_summed = CHECKSUM_UNNECESSARY;
2336 	}
2337 
2338 	if (skb_is_nonlinear(skb) &&
2339 	    skb_shinfo(skb)->nr_frags > IWL_PCIE_MAX_FRAGS(trans_pcie) &&
2340 	    __skb_linearize(skb))
2341 		return -ENOMEM;
2342 
2343 	/* mac80211 always puts the full header into the SKB's head,
2344 	 * so there's no need to check if it's readable there
2345 	 */
2346 	hdr = (struct ieee80211_hdr *)skb->data;
2347 	fc = hdr->frame_control;
2348 	hdr_len = ieee80211_hdrlen(fc);
2349 
2350 	spin_lock(&txq->lock);
2351 
2352 	if (iwl_queue_space(txq) < txq->high_mark) {
2353 		iwl_stop_queue(trans, txq);
2354 
2355 		/* don't put the packet on the ring, if there is no room */
2356 		if (unlikely(iwl_queue_space(txq) < 3)) {
2357 			struct iwl_device_cmd **dev_cmd_ptr;
2358 
2359 			dev_cmd_ptr = (void *)((u8 *)skb->cb +
2360 					       trans_pcie->dev_cmd_offs);
2361 
2362 			*dev_cmd_ptr = dev_cmd;
2363 			__skb_queue_tail(&txq->overflow_q, skb);
2364 
2365 			spin_unlock(&txq->lock);
2366 			return 0;
2367 		}
2368 	}
2369 
2370 	/* In AGG mode, the index in the ring must correspond to the WiFi
2371 	 * sequence number. This is a HW requirements to help the SCD to parse
2372 	 * the BA.
2373 	 * Check here that the packets are in the right place on the ring.
2374 	 */
2375 	wifi_seq = IEEE80211_SEQ_TO_SN(le16_to_cpu(hdr->seq_ctrl));
2376 	WARN_ONCE(txq->ampdu &&
2377 		  (wifi_seq & 0xff) != txq->write_ptr,
2378 		  "Q: %d WiFi Seq %d tfdNum %d",
2379 		  txq_id, wifi_seq, txq->write_ptr);
2380 
2381 	/* Set up driver data for this TFD */
2382 	txq->entries[txq->write_ptr].skb = skb;
2383 	txq->entries[txq->write_ptr].cmd = dev_cmd;
2384 
2385 	dev_cmd->hdr.sequence =
2386 		cpu_to_le16((u16)(QUEUE_TO_SEQ(txq_id) |
2387 			    INDEX_TO_SEQ(txq->write_ptr)));
2388 
2389 	tb0_phys = iwl_pcie_get_first_tb_dma(txq, txq->write_ptr);
2390 	scratch_phys = tb0_phys + sizeof(struct iwl_cmd_header) +
2391 		       offsetof(struct iwl_tx_cmd, scratch);
2392 
2393 	tx_cmd->dram_lsb_ptr = cpu_to_le32(scratch_phys);
2394 	tx_cmd->dram_msb_ptr = iwl_get_dma_hi_addr(scratch_phys);
2395 
2396 	/* Set up first empty entry in queue's array of Tx/cmd buffers */
2397 	out_meta = &txq->entries[txq->write_ptr].meta;
2398 	out_meta->flags = 0;
2399 
2400 	/*
2401 	 * The second TB (tb1) points to the remainder of the TX command
2402 	 * and the 802.11 header - dword aligned size
2403 	 * (This calculation modifies the TX command, so do it before the
2404 	 * setup of the first TB)
2405 	 */
2406 	len = sizeof(struct iwl_tx_cmd) + sizeof(struct iwl_cmd_header) +
2407 	      hdr_len - IWL_FIRST_TB_SIZE;
2408 	/* do not align A-MSDU to dword as the subframe header aligns it */
2409 	amsdu = ieee80211_is_data_qos(fc) &&
2410 		(*ieee80211_get_qos_ctl(hdr) &
2411 		 IEEE80211_QOS_CTL_A_MSDU_PRESENT);
2412 	if (trans_pcie->sw_csum_tx || !amsdu) {
2413 		tb1_len = ALIGN(len, 4);
2414 		/* Tell NIC about any 2-byte padding after MAC header */
2415 		if (tb1_len != len)
2416 			tx_cmd->tx_flags |= TX_CMD_FLG_MH_PAD_MSK;
2417 	} else {
2418 		tb1_len = len;
2419 	}
2420 
2421 	/*
2422 	 * The first TB points to bi-directional DMA data, we'll
2423 	 * memcpy the data into it later.
2424 	 */
2425 	iwl_pcie_txq_build_tfd(trans, txq, tb0_phys,
2426 			       IWL_FIRST_TB_SIZE, true);
2427 
2428 	/* there must be data left over for TB1 or this code must be changed */
2429 	BUILD_BUG_ON(sizeof(struct iwl_tx_cmd) < IWL_FIRST_TB_SIZE);
2430 
2431 	/* map the data for TB1 */
2432 	tb1_addr = ((u8 *)&dev_cmd->hdr) + IWL_FIRST_TB_SIZE;
2433 	tb1_phys = dma_map_single(trans->dev, tb1_addr, tb1_len, DMA_TO_DEVICE);
2434 	if (unlikely(dma_mapping_error(trans->dev, tb1_phys)))
2435 		goto out_err;
2436 	iwl_pcie_txq_build_tfd(trans, txq, tb1_phys, tb1_len, false);
2437 
2438 	if (amsdu) {
2439 		if (unlikely(iwl_fill_data_tbs_amsdu(trans, skb, txq, hdr_len,
2440 						     out_meta, dev_cmd,
2441 						     tb1_len)))
2442 			goto out_err;
2443 	} else if (unlikely(iwl_fill_data_tbs(trans, skb, txq, hdr_len,
2444 				       out_meta, dev_cmd, tb1_len))) {
2445 		goto out_err;
2446 	}
2447 
2448 	/* building the A-MSDU might have changed this data, so memcpy it now */
2449 	memcpy(&txq->first_tb_bufs[txq->write_ptr], &dev_cmd->hdr,
2450 	       IWL_FIRST_TB_SIZE);
2451 
2452 	tfd = iwl_pcie_get_tfd(trans_pcie, txq, txq->write_ptr);
2453 	/* Set up entry for this TFD in Tx byte-count array */
2454 	iwl_pcie_txq_update_byte_cnt_tbl(trans, txq, le16_to_cpu(tx_cmd->len),
2455 					 iwl_pcie_tfd_get_num_tbs(trans, tfd));
2456 
2457 	wait_write_ptr = ieee80211_has_morefrags(fc);
2458 
2459 	/* start timer if queue currently empty */
2460 	if (txq->read_ptr == txq->write_ptr) {
2461 		if (txq->wd_timeout) {
2462 			/*
2463 			 * If the TXQ is active, then set the timer, if not,
2464 			 * set the timer in remainder so that the timer will
2465 			 * be armed with the right value when the station will
2466 			 * wake up.
2467 			 */
2468 			if (!txq->frozen)
2469 				mod_timer(&txq->stuck_timer,
2470 					  jiffies + txq->wd_timeout);
2471 			else
2472 				txq->frozen_expiry_remainder = txq->wd_timeout;
2473 		}
2474 		IWL_DEBUG_RPM(trans, "Q: %d first tx - take ref\n", txq->id);
2475 		iwl_trans_ref(trans);
2476 	}
2477 
2478 	/* Tell device the write index *just past* this latest filled TFD */
2479 	txq->write_ptr = iwl_queue_inc_wrap(txq->write_ptr);
2480 	if (!wait_write_ptr)
2481 		iwl_pcie_txq_inc_wr_ptr(trans, txq);
2482 
2483 	/*
2484 	 * At this point the frame is "transmitted" successfully
2485 	 * and we will get a TX status notification eventually.
2486 	 */
2487 	spin_unlock(&txq->lock);
2488 	return 0;
2489 out_err:
2490 	spin_unlock(&txq->lock);
2491 	return -1;
2492 }
2493