1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (C) 2020-2024 Intel Corporation
10 #include "fw/api/commands.h"
11 #include "fw/api/tx.h"
12 #include "fw/api/datapath.h"
13 #include "fw/api/debug.h"
17 #include <linux/dmapool.h>
20 * iwl_txq_update_byte_tbl - Set up entry in Tx byte-count array
22 static void iwl_pcie_gen2_update_byte_tbl(struct iwl_trans *trans,
23 struct iwl_txq *txq, u16 byte_cnt,
26 int idx = iwl_txq_get_cmd_index(txq, txq->write_ptr);
27 u8 filled_tfd_size, num_fetch_chunks;
31 if (WARN(idx >= txq->n_window, "%d >= %d\n", idx, txq->n_window))
34 filled_tfd_size = offsetof(struct iwl_tfh_tfd, tbs) +
35 num_tbs * sizeof(struct iwl_tfh_tb);
37 * filled_tfd_size contains the number of filled bytes in the TFD.
38 * Dividing it by 64 will give the number of chunks to fetch
39 * to SRAM- 0 for one chunk, 1 for 2 and so on.
40 * If, for example, TFD contains only 3 TBs then 32 bytes
41 * of the TFD are used, and only one chunk of 64 bytes should
44 num_fetch_chunks = DIV_ROUND_UP(filled_tfd_size, 64) - 1;
46 if (trans->trans_cfg->device_family >= IWL_DEVICE_FAMILY_AX210) {
47 struct iwl_gen3_bc_tbl_entry *scd_bc_tbl_gen3 = txq->bc_tbl.addr;
49 /* Starting from AX210, the HW expects bytes */
50 WARN_ON(trans->txqs.bc_table_dword);
51 WARN_ON(len > 0x3FFF);
52 bc_ent = cpu_to_le16(len | (num_fetch_chunks << 14));
53 scd_bc_tbl_gen3[idx].tfd_offset = bc_ent;
55 struct iwlagn_scd_bc_tbl *scd_bc_tbl = txq->bc_tbl.addr;
57 /* Before AX210, the HW expects DW */
58 WARN_ON(!trans->txqs.bc_table_dword);
59 len = DIV_ROUND_UP(len, 4);
61 bc_ent = cpu_to_le16(len | (num_fetch_chunks << 12));
62 scd_bc_tbl->tfd_offset[idx] = bc_ent;
67 * iwl_txq_inc_wr_ptr - Send new write index to hardware
69 void iwl_txq_inc_wr_ptr(struct iwl_trans *trans, struct iwl_txq *txq)
71 lockdep_assert_held(&txq->lock);
73 IWL_DEBUG_TX(trans, "Q:%d WR: 0x%x\n", txq->id, txq->write_ptr);
76 * if not in power-save mode, uCode will never sleep when we're
77 * trying to tx (during RFKILL, we're not trying to tx).
79 iwl_write32(trans, HBUS_TARG_WRPTR, txq->write_ptr | (txq->id << 16));
82 static u8 iwl_txq_gen2_get_num_tbs(struct iwl_trans *trans,
83 struct iwl_tfh_tfd *tfd)
85 return le16_to_cpu(tfd->num_tbs) & 0x1f;
88 int iwl_txq_gen2_set_tb(struct iwl_trans *trans, struct iwl_tfh_tfd *tfd,
89 dma_addr_t addr, u16 len)
91 int idx = iwl_txq_gen2_get_num_tbs(trans, tfd);
92 struct iwl_tfh_tb *tb;
94 /* Only WARN here so we know about the issue, but we mess up our
95 * unmap path because not every place currently checks for errors
96 * returned from this function - it can only return an error if
97 * there's no more space, and so when we know there is enough we
98 * don't always check ...
100 WARN(iwl_txq_crosses_4g_boundary(addr, len),
101 "possible DMA problem with iova:0x%llx, len:%d\n",
102 (unsigned long long)addr, len);
104 if (WARN_ON(idx >= IWL_TFH_NUM_TBS))
108 /* Each TFD can point to a maximum max_tbs Tx buffers */
109 if (le16_to_cpu(tfd->num_tbs) >= trans->txqs.tfd.max_tbs) {
110 IWL_ERR(trans, "Error can not send more than %d chunks\n",
111 trans->txqs.tfd.max_tbs);
115 put_unaligned_le64(addr, &tb->addr);
116 tb->tb_len = cpu_to_le16(len);
118 tfd->num_tbs = cpu_to_le16(idx + 1);
123 static void iwl_txq_set_tfd_invalid_gen2(struct iwl_trans *trans,
124 struct iwl_tfh_tfd *tfd)
128 iwl_txq_gen2_set_tb(trans, tfd, trans->invalid_tx_cmd.dma,
129 trans->invalid_tx_cmd.size);
132 void iwl_txq_gen2_tfd_unmap(struct iwl_trans *trans, struct iwl_cmd_meta *meta,
133 struct iwl_tfh_tfd *tfd)
137 /* Sanity check on number of chunks */
138 num_tbs = iwl_txq_gen2_get_num_tbs(trans, tfd);
140 if (num_tbs > trans->txqs.tfd.max_tbs) {
141 IWL_ERR(trans, "Too many chunks: %i\n", num_tbs);
145 /* first TB is never freed - it's the bidirectional DMA data */
146 for (i = 1; i < num_tbs; i++) {
147 if (meta->tbs & BIT(i))
148 dma_unmap_page(trans->dev,
149 le64_to_cpu(tfd->tbs[i].addr),
150 le16_to_cpu(tfd->tbs[i].tb_len),
153 dma_unmap_single(trans->dev,
154 le64_to_cpu(tfd->tbs[i].addr),
155 le16_to_cpu(tfd->tbs[i].tb_len),
159 iwl_txq_set_tfd_invalid_gen2(trans, tfd);
162 void iwl_txq_gen2_free_tfd(struct iwl_trans *trans, struct iwl_txq *txq)
164 /* rd_ptr is bounded by TFD_QUEUE_SIZE_MAX and
165 * idx is bounded by n_window
167 int idx = iwl_txq_get_cmd_index(txq, txq->read_ptr);
170 lockdep_assert_held(&txq->lock);
175 iwl_txq_gen2_tfd_unmap(trans, &txq->entries[idx].meta,
176 iwl_txq_get_tfd(trans, txq, idx));
178 skb = txq->entries[idx].skb;
180 /* Can be called from irqs-disabled context
181 * If skb is not NULL, it means that the whole queue is being
182 * freed and that the queue is not empty - free the skb
185 iwl_op_mode_free_skb(trans->op_mode, skb);
186 txq->entries[idx].skb = NULL;
190 static struct page *get_workaround_page(struct iwl_trans *trans,
193 struct page **page_ptr;
196 page_ptr = (void *)((u8 *)skb->cb + trans->txqs.page_offs);
198 ret = alloc_page(GFP_ATOMIC);
202 /* set the chaining pointer to the previous page if there */
203 *(void **)((u8 *)page_address(ret) + PAGE_SIZE - sizeof(void *)) = *page_ptr;
210 * Add a TB and if needed apply the FH HW bug workaround;
211 * meta != NULL indicates that it's a page mapping and we
212 * need to dma_unmap_page() and set the meta->tbs bit in
215 static int iwl_txq_gen2_set_tb_with_wa(struct iwl_trans *trans,
217 struct iwl_tfh_tfd *tfd,
218 dma_addr_t phys, void *virt,
219 u16 len, struct iwl_cmd_meta *meta)
221 dma_addr_t oldphys = phys;
225 if (unlikely(dma_mapping_error(trans->dev, phys)))
228 if (likely(!iwl_txq_crosses_4g_boundary(phys, len))) {
229 ret = iwl_txq_gen2_set_tb(trans, tfd, phys, len);
235 meta->tbs |= BIT(ret);
242 * Work around a hardware bug. If (as expressed in the
243 * condition above) the TB ends on a 32-bit boundary,
244 * then the next TB may be accessed with the wrong
246 * To work around it, copy the data elsewhere and make
247 * a new mapping for it so the device will not fail.
250 if (WARN_ON(len > PAGE_SIZE - sizeof(void *))) {
255 page = get_workaround_page(trans, skb);
261 memcpy(page_address(page), virt, len);
263 phys = dma_map_single(trans->dev, page_address(page), len,
265 if (unlikely(dma_mapping_error(trans->dev, phys)))
267 ret = iwl_txq_gen2_set_tb(trans, tfd, phys, len);
269 /* unmap the new allocation as single */
275 "TB bug workaround: copied %d bytes from 0x%llx to 0x%llx\n",
276 len, (unsigned long long)oldphys,
277 (unsigned long long)phys);
282 dma_unmap_page(trans->dev, oldphys, len, DMA_TO_DEVICE);
284 dma_unmap_single(trans->dev, oldphys, len, DMA_TO_DEVICE);
286 trace_iwlwifi_dev_tx_tb(trans->dev, skb, virt, phys, len);
292 struct iwl_tso_hdr_page *get_page_hdr(struct iwl_trans *trans, size_t len,
295 struct iwl_tso_hdr_page *p = this_cpu_ptr(trans->txqs.tso_hdr_page);
296 struct page **page_ptr;
298 page_ptr = (void *)((u8 *)skb->cb + trans->txqs.page_offs);
300 if (WARN_ON(*page_ptr))
307 * Check if there's enough room on this page
309 * Note that we put a page chaining pointer *last* in the
310 * page - we need it somewhere, and if it's there then we
311 * avoid DMA mapping the last bits of the page which may
312 * trigger the 32-bit boundary hardware bug.
314 * (see also get_workaround_page() in tx-gen2.c)
316 if (p->pos + len < (u8 *)page_address(p->page) + PAGE_SIZE -
320 /* We don't have enough room on this page, get a new one. */
321 __free_page(p->page);
324 p->page = alloc_page(GFP_ATOMIC);
327 p->pos = page_address(p->page);
328 /* set the chaining pointer to NULL */
329 *(void **)((u8 *)page_address(p->page) + PAGE_SIZE - sizeof(void *)) = NULL;
337 static int iwl_txq_gen2_build_amsdu(struct iwl_trans *trans,
339 struct iwl_tfh_tfd *tfd, int start_len,
341 struct iwl_device_tx_cmd *dev_cmd)
344 struct iwl_tx_cmd_gen2 *tx_cmd = (void *)dev_cmd->payload;
345 struct ieee80211_hdr *hdr = (void *)skb->data;
346 unsigned int snap_ip_tcp_hdrlen, ip_hdrlen, total_len, hdr_room;
347 unsigned int mss = skb_shinfo(skb)->gso_size;
348 u16 length, amsdu_pad;
350 struct iwl_tso_hdr_page *hdr_page;
353 trace_iwlwifi_dev_tx(trans->dev, skb, tfd, sizeof(*tfd),
354 &dev_cmd->hdr, start_len, 0);
356 ip_hdrlen = skb_network_header_len(skb);
357 snap_ip_tcp_hdrlen = 8 + ip_hdrlen + tcp_hdrlen(skb);
358 total_len = skb->len - snap_ip_tcp_hdrlen - hdr_len;
361 /* total amount of header we may need for this A-MSDU */
362 hdr_room = DIV_ROUND_UP(total_len, mss) *
363 (3 + snap_ip_tcp_hdrlen + sizeof(struct ethhdr));
365 /* Our device supports 9 segments at most, it will fit in 1 page */
366 hdr_page = get_page_hdr(trans, hdr_room, skb);
370 start_hdr = hdr_page->pos;
373 * Pull the ieee80211 header to be able to use TSO core,
374 * we will restore it for the tx_status flow.
376 skb_pull(skb, hdr_len);
379 * Remove the length of all the headers that we don't actually
380 * have in the MPDU by themselves, but that we duplicate into
381 * all the different MSDUs inside the A-MSDU.
383 le16_add_cpu(&tx_cmd->len, -snap_ip_tcp_hdrlen);
385 tso_start(skb, &tso);
388 /* this is the data left for this subframe */
389 unsigned int data_left = min_t(unsigned int, mss, total_len);
392 u8 *subf_hdrs_start = hdr_page->pos;
394 total_len -= data_left;
396 memset(hdr_page->pos, 0, amsdu_pad);
397 hdr_page->pos += amsdu_pad;
398 amsdu_pad = (4 - (sizeof(struct ethhdr) + snap_ip_tcp_hdrlen +
400 ether_addr_copy(hdr_page->pos, ieee80211_get_DA(hdr));
401 hdr_page->pos += ETH_ALEN;
402 ether_addr_copy(hdr_page->pos, ieee80211_get_SA(hdr));
403 hdr_page->pos += ETH_ALEN;
405 length = snap_ip_tcp_hdrlen + data_left;
406 *((__be16 *)hdr_page->pos) = cpu_to_be16(length);
407 hdr_page->pos += sizeof(length);
410 * This will copy the SNAP as well which will be considered
413 tso_build_hdr(skb, hdr_page->pos, &tso, data_left, !total_len);
415 hdr_page->pos += snap_ip_tcp_hdrlen;
417 tb_len = hdr_page->pos - start_hdr;
418 tb_phys = dma_map_single(trans->dev, start_hdr,
419 tb_len, DMA_TO_DEVICE);
420 if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
423 * No need for _with_wa, this is from the TSO page and
424 * we leave some space at the end of it so can't hit
425 * the buggy scenario.
427 iwl_txq_gen2_set_tb(trans, tfd, tb_phys, tb_len);
428 trace_iwlwifi_dev_tx_tb(trans->dev, skb, start_hdr,
430 /* add this subframe's headers' length to the tx_cmd */
431 le16_add_cpu(&tx_cmd->len, hdr_page->pos - subf_hdrs_start);
433 /* prepare the start_hdr for the next subframe */
434 start_hdr = hdr_page->pos;
436 /* put the payload */
440 tb_len = min_t(unsigned int, tso.size, data_left);
441 tb_phys = dma_map_single(trans->dev, tso.data,
442 tb_len, DMA_TO_DEVICE);
443 ret = iwl_txq_gen2_set_tb_with_wa(trans, skb, tfd,
450 tso_build_data(skb, &tso, tb_len);
454 /* re -add the WiFi header */
455 skb_push(skb, hdr_len);
465 iwl_tfh_tfd *iwl_txq_gen2_build_tx_amsdu(struct iwl_trans *trans,
467 struct iwl_device_tx_cmd *dev_cmd,
469 struct iwl_cmd_meta *out_meta,
473 int idx = iwl_txq_get_cmd_index(txq, txq->write_ptr);
474 struct iwl_tfh_tfd *tfd = iwl_txq_get_tfd(trans, txq, idx);
479 tb_phys = iwl_txq_get_first_tb_dma(txq, idx);
482 * No need for _with_wa, the first TB allocation is aligned up
483 * to a 64-byte boundary and thus can't be at the end or cross
484 * a page boundary (much less a 2^32 boundary).
486 iwl_txq_gen2_set_tb(trans, tfd, tb_phys, IWL_FIRST_TB_SIZE);
489 * The second TB (tb1) points to the remainder of the TX command
490 * and the 802.11 header - dword aligned size
491 * (This calculation modifies the TX command, so do it before the
492 * setup of the first TB)
494 len = tx_cmd_len + sizeof(struct iwl_cmd_header) + hdr_len -
497 /* do not align A-MSDU to dword as the subframe header aligns it */
499 /* map the data for TB1 */
500 tb1_addr = ((u8 *)&dev_cmd->hdr) + IWL_FIRST_TB_SIZE;
501 tb_phys = dma_map_single(trans->dev, tb1_addr, len, DMA_TO_DEVICE);
502 if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
505 * No need for _with_wa(), we ensure (via alignment) that the data
506 * here can never cross or end at a page boundary.
508 iwl_txq_gen2_set_tb(trans, tfd, tb_phys, len);
510 if (iwl_txq_gen2_build_amsdu(trans, skb, tfd, len + IWL_FIRST_TB_SIZE,
514 /* building the A-MSDU might have changed this data, memcpy it now */
515 memcpy(&txq->first_tb_bufs[idx], dev_cmd, IWL_FIRST_TB_SIZE);
519 iwl_txq_gen2_tfd_unmap(trans, out_meta, tfd);
523 static int iwl_txq_gen2_tx_add_frags(struct iwl_trans *trans,
525 struct iwl_tfh_tfd *tfd,
526 struct iwl_cmd_meta *out_meta)
530 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
531 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
533 unsigned int fragsz = skb_frag_size(frag);
539 tb_phys = skb_frag_dma_map(trans->dev, frag, 0,
540 fragsz, DMA_TO_DEVICE);
541 ret = iwl_txq_gen2_set_tb_with_wa(trans, skb, tfd, tb_phys,
542 skb_frag_address(frag),
552 iwl_tfh_tfd *iwl_txq_gen2_build_tx(struct iwl_trans *trans,
554 struct iwl_device_tx_cmd *dev_cmd,
556 struct iwl_cmd_meta *out_meta,
561 int idx = iwl_txq_get_cmd_index(txq, txq->write_ptr);
562 struct iwl_tfh_tfd *tfd = iwl_txq_get_tfd(trans, txq, idx);
564 int len, tb1_len, tb2_len;
566 struct sk_buff *frag;
568 tb_phys = iwl_txq_get_first_tb_dma(txq, idx);
570 /* The first TB points to bi-directional DMA data */
571 memcpy(&txq->first_tb_bufs[idx], dev_cmd, IWL_FIRST_TB_SIZE);
574 * No need for _with_wa, the first TB allocation is aligned up
575 * to a 64-byte boundary and thus can't be at the end or cross
576 * a page boundary (much less a 2^32 boundary).
578 iwl_txq_gen2_set_tb(trans, tfd, tb_phys, IWL_FIRST_TB_SIZE);
581 * The second TB (tb1) points to the remainder of the TX command
582 * and the 802.11 header - dword aligned size
583 * (This calculation modifies the TX command, so do it before the
584 * setup of the first TB)
586 len = tx_cmd_len + sizeof(struct iwl_cmd_header) + hdr_len -
590 tb1_len = ALIGN(len, 4);
594 /* map the data for TB1 */
595 tb1_addr = ((u8 *)&dev_cmd->hdr) + IWL_FIRST_TB_SIZE;
596 tb_phys = dma_map_single(trans->dev, tb1_addr, tb1_len, DMA_TO_DEVICE);
597 if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
600 * No need for _with_wa(), we ensure (via alignment) that the data
601 * here can never cross or end at a page boundary.
603 iwl_txq_gen2_set_tb(trans, tfd, tb_phys, tb1_len);
604 trace_iwlwifi_dev_tx(trans->dev, skb, tfd, sizeof(*tfd), &dev_cmd->hdr,
605 IWL_FIRST_TB_SIZE + tb1_len, hdr_len);
607 /* set up TFD's third entry to point to remainder of skb's head */
608 tb2_len = skb_headlen(skb) - hdr_len;
613 tb_phys = dma_map_single(trans->dev, skb->data + hdr_len,
614 tb2_len, DMA_TO_DEVICE);
615 ret = iwl_txq_gen2_set_tb_with_wa(trans, skb, tfd, tb_phys,
616 skb->data + hdr_len, tb2_len,
622 if (iwl_txq_gen2_tx_add_frags(trans, skb, tfd, out_meta))
625 skb_walk_frags(skb, frag) {
628 tb_phys = dma_map_single(trans->dev, frag->data,
629 skb_headlen(frag), DMA_TO_DEVICE);
630 ret = iwl_txq_gen2_set_tb_with_wa(trans, skb, tfd, tb_phys,
632 skb_headlen(frag), NULL);
635 if (iwl_txq_gen2_tx_add_frags(trans, frag, tfd, out_meta))
642 iwl_txq_gen2_tfd_unmap(trans, out_meta, tfd);
647 struct iwl_tfh_tfd *iwl_txq_gen2_build_tfd(struct iwl_trans *trans,
649 struct iwl_device_tx_cmd *dev_cmd,
651 struct iwl_cmd_meta *out_meta)
653 struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)skb->data;
654 int idx = iwl_txq_get_cmd_index(txq, txq->write_ptr);
655 struct iwl_tfh_tfd *tfd = iwl_txq_get_tfd(trans, txq, idx);
659 /* There must be data left over for TB1 or this code must be changed */
660 BUILD_BUG_ON(sizeof(struct iwl_tx_cmd_gen2) < IWL_FIRST_TB_SIZE);
661 BUILD_BUG_ON(sizeof(struct iwl_cmd_header) +
662 offsetofend(struct iwl_tx_cmd_gen2, dram_info) >
664 BUILD_BUG_ON(sizeof(struct iwl_tx_cmd_gen3) < IWL_FIRST_TB_SIZE);
665 BUILD_BUG_ON(sizeof(struct iwl_cmd_header) +
666 offsetofend(struct iwl_tx_cmd_gen3, dram_info) >
669 memset(tfd, 0, sizeof(*tfd));
671 if (trans->trans_cfg->device_family < IWL_DEVICE_FAMILY_AX210)
672 len = sizeof(struct iwl_tx_cmd_gen2);
674 len = sizeof(struct iwl_tx_cmd_gen3);
676 amsdu = ieee80211_is_data_qos(hdr->frame_control) &&
677 (*ieee80211_get_qos_ctl(hdr) &
678 IEEE80211_QOS_CTL_A_MSDU_PRESENT);
680 hdr_len = ieee80211_hdrlen(hdr->frame_control);
683 * Only build A-MSDUs here if doing so by GSO, otherwise it may be
684 * an A-MSDU for other reasons, e.g. NAN or an A-MSDU having been
685 * built in the higher layers already.
687 if (amsdu && skb_shinfo(skb)->gso_size)
688 return iwl_txq_gen2_build_tx_amsdu(trans, txq, dev_cmd, skb,
689 out_meta, hdr_len, len);
690 return iwl_txq_gen2_build_tx(trans, txq, dev_cmd, skb, out_meta,
691 hdr_len, len, !amsdu);
694 int iwl_txq_space(struct iwl_trans *trans, const struct iwl_txq *q)
700 * To avoid ambiguity between empty and completely full queues, there
701 * should always be less than max_tfd_queue_size elements in the queue.
702 * If q->n_window is smaller than max_tfd_queue_size, there is no need
703 * to reserve any queue entries for this purpose.
705 if (q->n_window < trans->trans_cfg->base_params->max_tfd_queue_size)
708 max = trans->trans_cfg->base_params->max_tfd_queue_size - 1;
711 * max_tfd_queue_size is a power of 2, so the following is equivalent to
712 * modulo by max_tfd_queue_size and is well defined.
714 used = (q->write_ptr - q->read_ptr) &
715 (trans->trans_cfg->base_params->max_tfd_queue_size - 1);
717 if (WARN_ON(used > max))
723 int iwl_txq_gen2_tx(struct iwl_trans *trans, struct sk_buff *skb,
724 struct iwl_device_tx_cmd *dev_cmd, int txq_id)
726 struct iwl_cmd_meta *out_meta;
727 struct iwl_txq *txq = trans->txqs.txq[txq_id];
732 if (WARN_ONCE(txq_id >= IWL_MAX_TVQM_QUEUES,
733 "queue %d out of range", txq_id))
736 if (WARN_ONCE(!test_bit(txq_id, trans->txqs.queue_used),
737 "TX on unused queue %d\n", txq_id))
740 if (skb_is_nonlinear(skb) &&
741 skb_shinfo(skb)->nr_frags > IWL_TRANS_MAX_FRAGS(trans) &&
742 __skb_linearize(skb))
745 spin_lock(&txq->lock);
747 if (iwl_txq_space(trans, txq) < txq->high_mark) {
748 iwl_txq_stop(trans, txq);
750 /* don't put the packet on the ring, if there is no room */
751 if (unlikely(iwl_txq_space(trans, txq) < 3)) {
752 struct iwl_device_tx_cmd **dev_cmd_ptr;
754 dev_cmd_ptr = (void *)((u8 *)skb->cb +
755 trans->txqs.dev_cmd_offs);
757 *dev_cmd_ptr = dev_cmd;
758 __skb_queue_tail(&txq->overflow_q, skb);
759 spin_unlock(&txq->lock);
764 idx = iwl_txq_get_cmd_index(txq, txq->write_ptr);
766 /* Set up driver data for this TFD */
767 txq->entries[idx].skb = skb;
768 txq->entries[idx].cmd = dev_cmd;
770 dev_cmd->hdr.sequence =
771 cpu_to_le16((u16)(QUEUE_TO_SEQ(txq_id) |
774 /* Set up first empty entry in queue's array of Tx/cmd buffers */
775 out_meta = &txq->entries[idx].meta;
778 tfd = iwl_txq_gen2_build_tfd(trans, txq, dev_cmd, skb, out_meta);
780 spin_unlock(&txq->lock);
784 if (trans->trans_cfg->device_family >= IWL_DEVICE_FAMILY_AX210) {
785 struct iwl_tx_cmd_gen3 *tx_cmd_gen3 =
786 (void *)dev_cmd->payload;
788 cmd_len = le16_to_cpu(tx_cmd_gen3->len);
790 struct iwl_tx_cmd_gen2 *tx_cmd_gen2 =
791 (void *)dev_cmd->payload;
793 cmd_len = le16_to_cpu(tx_cmd_gen2->len);
796 /* Set up entry for this TFD in Tx byte-count array */
797 iwl_pcie_gen2_update_byte_tbl(trans, txq, cmd_len,
798 iwl_txq_gen2_get_num_tbs(trans, tfd));
800 /* start timer if queue currently empty */
801 if (txq->read_ptr == txq->write_ptr && txq->wd_timeout)
802 mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
804 /* Tell device the write index *just past* this latest filled TFD */
805 txq->write_ptr = iwl_txq_inc_wrap(trans, txq->write_ptr);
806 iwl_txq_inc_wr_ptr(trans, txq);
808 * At this point the frame is "transmitted" successfully
809 * and we will get a TX status notification eventually.
811 spin_unlock(&txq->lock);
815 /*************** HOST COMMAND QUEUE FUNCTIONS *****/
818 * iwl_txq_gen2_unmap - Unmap any remaining DMA mappings and free skb's
820 void iwl_txq_gen2_unmap(struct iwl_trans *trans, int txq_id)
822 struct iwl_txq *txq = trans->txqs.txq[txq_id];
824 spin_lock_bh(&txq->lock);
825 while (txq->write_ptr != txq->read_ptr) {
826 IWL_DEBUG_TX_REPLY(trans, "Q %d Free %d\n",
827 txq_id, txq->read_ptr);
829 if (txq_id != trans->txqs.cmd.q_id) {
830 int idx = iwl_txq_get_cmd_index(txq, txq->read_ptr);
831 struct sk_buff *skb = txq->entries[idx].skb;
833 if (!WARN_ON_ONCE(!skb))
834 iwl_txq_free_tso_page(trans, skb);
836 iwl_txq_gen2_free_tfd(trans, txq);
837 txq->read_ptr = iwl_txq_inc_wrap(trans, txq->read_ptr);
840 while (!skb_queue_empty(&txq->overflow_q)) {
841 struct sk_buff *skb = __skb_dequeue(&txq->overflow_q);
843 iwl_op_mode_free_skb(trans->op_mode, skb);
846 spin_unlock_bh(&txq->lock);
848 /* just in case - this queue may have been stopped */
849 iwl_wake_queue(trans, txq);
852 static void iwl_txq_gen2_free_memory(struct iwl_trans *trans,
855 struct device *dev = trans->dev;
857 /* De-alloc circular buffer of TFDs */
859 dma_free_coherent(dev,
860 trans->txqs.tfd.size * txq->n_window,
861 txq->tfds, txq->dma_addr);
862 dma_free_coherent(dev,
863 sizeof(*txq->first_tb_bufs) * txq->n_window,
864 txq->first_tb_bufs, txq->first_tb_dma);
868 if (txq->bc_tbl.addr)
869 dma_pool_free(trans->txqs.bc_pool,
870 txq->bc_tbl.addr, txq->bc_tbl.dma);
875 * iwl_pcie_txq_free - Deallocate DMA queue.
876 * @txq: Transmit queue to deallocate.
878 * Empty queue by removing and destroying all BD's.
880 * 0-fill, but do not free "txq" descriptor structure.
882 static void iwl_txq_gen2_free(struct iwl_trans *trans, int txq_id)
887 if (WARN_ONCE(txq_id >= IWL_MAX_TVQM_QUEUES,
888 "queue %d out of range", txq_id))
891 txq = trans->txqs.txq[txq_id];
896 iwl_txq_gen2_unmap(trans, txq_id);
898 /* De-alloc array of command/tx buffers */
899 if (txq_id == trans->txqs.cmd.q_id)
900 for (i = 0; i < txq->n_window; i++) {
901 kfree_sensitive(txq->entries[i].cmd);
902 kfree_sensitive(txq->entries[i].free_buf);
904 del_timer_sync(&txq->stuck_timer);
906 iwl_txq_gen2_free_memory(trans, txq);
908 trans->txqs.txq[txq_id] = NULL;
910 clear_bit(txq_id, trans->txqs.queue_used);
914 * iwl_queue_init - Initialize queue's high/low-water and read/write indexes
916 static int iwl_queue_init(struct iwl_txq *q, int slots_num)
918 q->n_window = slots_num;
920 /* slots_num must be power-of-two size, otherwise
921 * iwl_txq_get_cmd_index is broken. */
922 if (WARN_ON(!is_power_of_2(slots_num)))
925 q->low_mark = q->n_window / 4;
929 q->high_mark = q->n_window / 8;
930 if (q->high_mark < 2)
939 int iwl_txq_init(struct iwl_trans *trans, struct iwl_txq *txq, int slots_num,
943 u32 tfd_queue_max_size =
944 trans->trans_cfg->base_params->max_tfd_queue_size;
946 txq->need_update = false;
948 /* max_tfd_queue_size must be power-of-two size, otherwise
949 * iwl_txq_inc_wrap and iwl_txq_dec_wrap are broken. */
950 if (WARN_ONCE(tfd_queue_max_size & (tfd_queue_max_size - 1),
951 "Max tfd queue size must be a power of two, but is %d",
955 /* Initialize queue's high/low-water marks, and head/tail indexes */
956 ret = iwl_queue_init(txq, slots_num);
960 spin_lock_init(&txq->lock);
963 static struct lock_class_key iwl_txq_cmd_queue_lock_class;
965 lockdep_set_class(&txq->lock, &iwl_txq_cmd_queue_lock_class);
968 __skb_queue_head_init(&txq->overflow_q);
973 void iwl_txq_free_tso_page(struct iwl_trans *trans, struct sk_buff *skb)
975 struct page **page_ptr;
978 page_ptr = (void *)((u8 *)skb->cb + trans->txqs.page_offs);
983 struct page *tmp = next;
985 next = *(void **)((u8 *)page_address(next) + PAGE_SIZE -
991 void iwl_txq_log_scd_error(struct iwl_trans *trans, struct iwl_txq *txq)
993 u32 txq_id = txq->id;
998 if (trans->trans_cfg->gen2) {
999 IWL_ERR(trans, "Queue %d is stuck %d %d\n", txq_id,
1000 txq->read_ptr, txq->write_ptr);
1001 /* TODO: access new SCD registers and dump them */
1005 status = iwl_read_prph(trans, SCD_QUEUE_STATUS_BITS(txq_id));
1006 fifo = (status >> SCD_QUEUE_STTS_REG_POS_TXF) & 0x7;
1007 active = !!(status & BIT(SCD_QUEUE_STTS_REG_POS_ACTIVE));
1010 "Queue %d is %sactive on fifo %d and stuck for %u ms. SW [%d, %d] HW [%d, %d] FH TRB=0x0%x\n",
1011 txq_id, active ? "" : "in", fifo,
1012 jiffies_to_msecs(txq->wd_timeout),
1013 txq->read_ptr, txq->write_ptr,
1014 iwl_read_prph(trans, SCD_QUEUE_RDPTR(txq_id)) &
1015 (trans->trans_cfg->base_params->max_tfd_queue_size - 1),
1016 iwl_read_prph(trans, SCD_QUEUE_WRPTR(txq_id)) &
1017 (trans->trans_cfg->base_params->max_tfd_queue_size - 1),
1018 iwl_read_direct32(trans, FH_TX_TRB_REG(fifo)));
1021 static void iwl_txq_stuck_timer(struct timer_list *t)
1023 struct iwl_txq *txq = from_timer(txq, t, stuck_timer);
1024 struct iwl_trans *trans = txq->trans;
1026 spin_lock(&txq->lock);
1027 /* check if triggered erroneously */
1028 if (txq->read_ptr == txq->write_ptr) {
1029 spin_unlock(&txq->lock);
1032 spin_unlock(&txq->lock);
1034 iwl_txq_log_scd_error(trans, txq);
1036 iwl_force_nmi(trans);
1039 static void iwl_txq_set_tfd_invalid_gen1(struct iwl_trans *trans,
1040 struct iwl_tfd *tfd)
1044 iwl_pcie_gen1_tfd_set_tb(trans, tfd, 0, trans->invalid_tx_cmd.dma,
1045 trans->invalid_tx_cmd.size);
1048 int iwl_txq_alloc(struct iwl_trans *trans, struct iwl_txq *txq, int slots_num,
1051 size_t num_entries = trans->trans_cfg->gen2 ?
1052 slots_num : trans->trans_cfg->base_params->max_tfd_queue_size;
1057 if (WARN_ONCE(slots_num <= 0, "Invalid slots num:%d\n", slots_num))
1060 if (WARN_ON(txq->entries || txq->tfds))
1063 tfd_sz = trans->txqs.tfd.size * num_entries;
1065 timer_setup(&txq->stuck_timer, iwl_txq_stuck_timer, 0);
1068 txq->n_window = slots_num;
1070 txq->entries = kcalloc(slots_num,
1071 sizeof(struct iwl_pcie_txq_entry),
1078 for (i = 0; i < slots_num; i++) {
1079 txq->entries[i].cmd =
1080 kmalloc(sizeof(struct iwl_device_cmd),
1082 if (!txq->entries[i].cmd)
1086 /* Circular buffer of transmit frame descriptors (TFDs),
1087 * shared with device */
1088 txq->tfds = dma_alloc_coherent(trans->dev, tfd_sz,
1089 &txq->dma_addr, GFP_KERNEL);
1093 BUILD_BUG_ON(sizeof(*txq->first_tb_bufs) != IWL_FIRST_TB_SIZE_ALIGN);
1095 tb0_buf_sz = sizeof(*txq->first_tb_bufs) * slots_num;
1097 txq->first_tb_bufs = dma_alloc_coherent(trans->dev, tb0_buf_sz,
1100 if (!txq->first_tb_bufs)
1103 for (i = 0; i < num_entries; i++) {
1104 void *tfd = iwl_txq_get_tfd(trans, txq, i);
1106 if (trans->trans_cfg->gen2)
1107 iwl_txq_set_tfd_invalid_gen2(trans, tfd);
1109 iwl_txq_set_tfd_invalid_gen1(trans, tfd);
1114 dma_free_coherent(trans->dev, tfd_sz, txq->tfds, txq->dma_addr);
1117 if (txq->entries && cmd_queue)
1118 for (i = 0; i < slots_num; i++)
1119 kfree(txq->entries[i].cmd);
1120 kfree(txq->entries);
1121 txq->entries = NULL;
1126 static struct iwl_txq *
1127 iwl_txq_dyn_alloc_dma(struct iwl_trans *trans, int size, unsigned int timeout)
1129 size_t bc_tbl_size, bc_tbl_entries;
1130 struct iwl_txq *txq;
1133 WARN_ON(!trans->txqs.bc_tbl_size);
1135 bc_tbl_size = trans->txqs.bc_tbl_size;
1136 bc_tbl_entries = bc_tbl_size / sizeof(u16);
1138 if (WARN_ON(size > bc_tbl_entries))
1139 return ERR_PTR(-EINVAL);
1141 txq = kzalloc(sizeof(*txq), GFP_KERNEL);
1143 return ERR_PTR(-ENOMEM);
1145 txq->bc_tbl.addr = dma_pool_alloc(trans->txqs.bc_pool, GFP_KERNEL,
1147 if (!txq->bc_tbl.addr) {
1148 IWL_ERR(trans, "Scheduler BC Table allocation failed\n");
1150 return ERR_PTR(-ENOMEM);
1153 ret = iwl_txq_alloc(trans, txq, size, false);
1155 IWL_ERR(trans, "Tx queue alloc failed\n");
1158 ret = iwl_txq_init(trans, txq, size, false);
1160 IWL_ERR(trans, "Tx queue init failed\n");
1164 txq->wd_timeout = msecs_to_jiffies(timeout);
1169 iwl_txq_gen2_free_memory(trans, txq);
1170 return ERR_PTR(ret);
1173 static int iwl_txq_alloc_response(struct iwl_trans *trans, struct iwl_txq *txq,
1174 struct iwl_host_cmd *hcmd)
1176 struct iwl_tx_queue_cfg_rsp *rsp;
1180 if (WARN_ON(iwl_rx_packet_payload_len(hcmd->resp_pkt) !=
1183 goto error_free_resp;
1186 rsp = (void *)hcmd->resp_pkt->data;
1187 qid = le16_to_cpu(rsp->queue_number);
1188 wr_ptr = le16_to_cpu(rsp->write_pointer);
1190 if (qid >= ARRAY_SIZE(trans->txqs.txq)) {
1191 WARN_ONCE(1, "queue index %d unsupported", qid);
1193 goto error_free_resp;
1196 if (test_and_set_bit(qid, trans->txqs.queue_used)) {
1197 WARN_ONCE(1, "queue %d already used", qid);
1199 goto error_free_resp;
1202 if (WARN_ONCE(trans->txqs.txq[qid],
1203 "queue %d already allocated\n", qid)) {
1205 goto error_free_resp;
1209 trans->txqs.txq[qid] = txq;
1210 wr_ptr &= (trans->trans_cfg->base_params->max_tfd_queue_size - 1);
1212 /* Place first TFD at index corresponding to start sequence number */
1213 txq->read_ptr = wr_ptr;
1214 txq->write_ptr = wr_ptr;
1216 IWL_DEBUG_TX_QUEUES(trans, "Activate queue %d\n", qid);
1218 iwl_free_resp(hcmd);
1222 iwl_free_resp(hcmd);
1223 iwl_txq_gen2_free_memory(trans, txq);
1227 int iwl_txq_dyn_alloc(struct iwl_trans *trans, u32 flags, u32 sta_mask,
1228 u8 tid, int size, unsigned int timeout)
1230 struct iwl_txq *txq;
1232 struct iwl_tx_queue_cfg_cmd old;
1233 struct iwl_scd_queue_cfg_cmd new;
1235 struct iwl_host_cmd hcmd = {
1236 .flags = CMD_WANT_SKB,
1240 if (trans->trans_cfg->device_family == IWL_DEVICE_FAMILY_BZ &&
1241 trans->hw_rev_step == SILICON_A_STEP)
1244 txq = iwl_txq_dyn_alloc_dma(trans, size, timeout);
1246 return PTR_ERR(txq);
1248 if (trans->txqs.queue_alloc_cmd_ver == 0) {
1249 memset(&cmd.old, 0, sizeof(cmd.old));
1250 cmd.old.tfdq_addr = cpu_to_le64(txq->dma_addr);
1251 cmd.old.byte_cnt_addr = cpu_to_le64(txq->bc_tbl.dma);
1252 cmd.old.cb_size = cpu_to_le32(TFD_QUEUE_CB_SIZE(size));
1253 cmd.old.flags = cpu_to_le16(flags | TX_QUEUE_CFG_ENABLE_QUEUE);
1256 if (hweight32(sta_mask) != 1) {
1260 cmd.old.sta_id = ffs(sta_mask) - 1;
1262 hcmd.id = SCD_QUEUE_CFG;
1263 hcmd.len[0] = sizeof(cmd.old);
1264 hcmd.data[0] = &cmd.old;
1265 } else if (trans->txqs.queue_alloc_cmd_ver == 3) {
1266 memset(&cmd.new, 0, sizeof(cmd.new));
1267 cmd.new.operation = cpu_to_le32(IWL_SCD_QUEUE_ADD);
1268 cmd.new.u.add.tfdq_dram_addr = cpu_to_le64(txq->dma_addr);
1269 cmd.new.u.add.bc_dram_addr = cpu_to_le64(txq->bc_tbl.dma);
1270 cmd.new.u.add.cb_size = cpu_to_le32(TFD_QUEUE_CB_SIZE(size));
1271 cmd.new.u.add.flags = cpu_to_le32(flags);
1272 cmd.new.u.add.sta_mask = cpu_to_le32(sta_mask);
1273 cmd.new.u.add.tid = tid;
1275 hcmd.id = WIDE_ID(DATA_PATH_GROUP, SCD_QUEUE_CONFIG_CMD);
1276 hcmd.len[0] = sizeof(cmd.new);
1277 hcmd.data[0] = &cmd.new;
1283 ret = iwl_trans_send_cmd(trans, &hcmd);
1287 return iwl_txq_alloc_response(trans, txq, &hcmd);
1290 iwl_txq_gen2_free_memory(trans, txq);
1294 void iwl_txq_dyn_free(struct iwl_trans *trans, int queue)
1296 if (WARN(queue >= IWL_MAX_TVQM_QUEUES,
1297 "queue %d out of range", queue))
1301 * Upon HW Rfkill - we stop the device, and then stop the queues
1302 * in the op_mode. Just for the sake of the simplicity of the op_mode,
1303 * allow the op_mode to call txq_disable after it already called
1306 if (!test_and_clear_bit(queue, trans->txqs.queue_used)) {
1307 WARN_ONCE(test_bit(STATUS_DEVICE_ENABLED, &trans->status),
1308 "queue %d not used", queue);
1312 iwl_txq_gen2_free(trans, queue);
1314 IWL_DEBUG_TX_QUEUES(trans, "Deactivate queue %d\n", queue);
1317 void iwl_txq_gen2_tx_free(struct iwl_trans *trans)
1321 memset(trans->txqs.queue_used, 0, sizeof(trans->txqs.queue_used));
1323 /* Free all TX queues */
1324 for (i = 0; i < ARRAY_SIZE(trans->txqs.txq); i++) {
1325 if (!trans->txqs.txq[i])
1328 iwl_txq_gen2_free(trans, i);
1332 int iwl_txq_gen2_init(struct iwl_trans *trans, int txq_id, int queue_size)
1334 struct iwl_txq *queue;
1337 /* alloc and init the tx queue */
1338 if (!trans->txqs.txq[txq_id]) {
1339 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1341 IWL_ERR(trans, "Not enough memory for tx queue\n");
1344 trans->txqs.txq[txq_id] = queue;
1345 ret = iwl_txq_alloc(trans, queue, queue_size, true);
1347 IWL_ERR(trans, "Tx %d queue init failed\n", txq_id);
1351 queue = trans->txqs.txq[txq_id];
1354 ret = iwl_txq_init(trans, queue, queue_size,
1355 (txq_id == trans->txqs.cmd.q_id));
1357 IWL_ERR(trans, "Tx %d queue alloc failed\n", txq_id);
1360 trans->txqs.txq[txq_id]->id = txq_id;
1361 set_bit(txq_id, trans->txqs.queue_used);
1366 iwl_txq_gen2_tx_free(trans);
1370 static inline dma_addr_t iwl_txq_gen1_tfd_tb_get_addr(struct iwl_trans *trans,
1371 struct iwl_tfd *tfd, u8 idx)
1373 struct iwl_tfd_tb *tb = &tfd->tbs[idx];
1377 addr = get_unaligned_le32(&tb->lo);
1379 if (sizeof(dma_addr_t) <= sizeof(u32))
1382 hi_len = le16_to_cpu(tb->hi_n_len) & 0xF;
1385 * shift by 16 twice to avoid warnings on 32-bit
1386 * (where this code never runs anyway due to the
1387 * if statement above)
1389 return addr | ((hi_len << 16) << 16);
1392 void iwl_txq_gen1_tfd_unmap(struct iwl_trans *trans,
1393 struct iwl_cmd_meta *meta,
1394 struct iwl_txq *txq, int index)
1397 struct iwl_tfd *tfd = iwl_txq_get_tfd(trans, txq, index);
1399 /* Sanity check on number of chunks */
1400 num_tbs = iwl_txq_gen1_tfd_get_num_tbs(trans, tfd);
1402 if (num_tbs > trans->txqs.tfd.max_tbs) {
1403 IWL_ERR(trans, "Too many chunks: %i\n", num_tbs);
1404 /* @todo issue fatal error, it is quite serious situation */
1408 /* first TB is never freed - it's the bidirectional DMA data */
1410 for (i = 1; i < num_tbs; i++) {
1411 if (meta->tbs & BIT(i))
1412 dma_unmap_page(trans->dev,
1413 iwl_txq_gen1_tfd_tb_get_addr(trans,
1415 iwl_txq_gen1_tfd_tb_get_len(trans,
1419 dma_unmap_single(trans->dev,
1420 iwl_txq_gen1_tfd_tb_get_addr(trans,
1422 iwl_txq_gen1_tfd_tb_get_len(trans,
1429 iwl_txq_set_tfd_invalid_gen1(trans, tfd);
1432 #define IWL_TX_CRC_SIZE 4
1433 #define IWL_TX_DELIMITER_SIZE 4
1436 * iwl_txq_gen1_update_byte_cnt_tbl - Set up entry in Tx byte-count array
1438 void iwl_txq_gen1_update_byte_cnt_tbl(struct iwl_trans *trans,
1439 struct iwl_txq *txq, u16 byte_cnt,
1442 struct iwlagn_scd_bc_tbl *scd_bc_tbl;
1443 int write_ptr = txq->write_ptr;
1444 int txq_id = txq->id;
1446 u16 len = byte_cnt + IWL_TX_CRC_SIZE + IWL_TX_DELIMITER_SIZE;
1448 struct iwl_device_tx_cmd *dev_cmd = txq->entries[txq->write_ptr].cmd;
1449 struct iwl_tx_cmd *tx_cmd = (void *)dev_cmd->payload;
1450 u8 sta_id = tx_cmd->sta_id;
1452 scd_bc_tbl = trans->txqs.scd_bc_tbls.addr;
1454 sec_ctl = tx_cmd->sec_ctl;
1456 switch (sec_ctl & TX_CMD_SEC_MSK) {
1457 case TX_CMD_SEC_CCM:
1458 len += IEEE80211_CCMP_MIC_LEN;
1460 case TX_CMD_SEC_TKIP:
1461 len += IEEE80211_TKIP_ICV_LEN;
1463 case TX_CMD_SEC_WEP:
1464 len += IEEE80211_WEP_IV_LEN + IEEE80211_WEP_ICV_LEN;
1467 if (trans->txqs.bc_table_dword)
1468 len = DIV_ROUND_UP(len, 4);
1470 if (WARN_ON(len > 0xFFF || write_ptr >= TFD_QUEUE_SIZE_MAX))
1473 bc_ent = cpu_to_le16(len | (sta_id << 12));
1475 scd_bc_tbl[txq_id].tfd_offset[write_ptr] = bc_ent;
1477 if (write_ptr < TFD_QUEUE_SIZE_BC_DUP)
1478 scd_bc_tbl[txq_id].tfd_offset[TFD_QUEUE_SIZE_MAX + write_ptr] =
1482 void iwl_txq_gen1_inval_byte_cnt_tbl(struct iwl_trans *trans,
1483 struct iwl_txq *txq)
1485 struct iwlagn_scd_bc_tbl *scd_bc_tbl = trans->txqs.scd_bc_tbls.addr;
1486 int txq_id = txq->id;
1487 int read_ptr = txq->read_ptr;
1490 struct iwl_device_tx_cmd *dev_cmd = txq->entries[read_ptr].cmd;
1491 struct iwl_tx_cmd *tx_cmd = (void *)dev_cmd->payload;
1493 WARN_ON(read_ptr >= TFD_QUEUE_SIZE_MAX);
1495 if (txq_id != trans->txqs.cmd.q_id)
1496 sta_id = tx_cmd->sta_id;
1498 bc_ent = cpu_to_le16(1 | (sta_id << 12));
1500 scd_bc_tbl[txq_id].tfd_offset[read_ptr] = bc_ent;
1502 if (read_ptr < TFD_QUEUE_SIZE_BC_DUP)
1503 scd_bc_tbl[txq_id].tfd_offset[TFD_QUEUE_SIZE_MAX + read_ptr] =
1508 * iwl_txq_free_tfd - Free all chunks referenced by TFD [txq->q.read_ptr]
1509 * @trans - transport private data
1511 * @dma_dir - the direction of the DMA mapping
1513 * Does NOT advance any TFD circular buffer read/write indexes
1514 * Does NOT free the TFD itself (which is within circular buffer)
1516 void iwl_txq_free_tfd(struct iwl_trans *trans, struct iwl_txq *txq)
1518 /* rd_ptr is bounded by TFD_QUEUE_SIZE_MAX and
1519 * idx is bounded by n_window
1521 int rd_ptr = txq->read_ptr;
1522 int idx = iwl_txq_get_cmd_index(txq, rd_ptr);
1523 struct sk_buff *skb;
1525 lockdep_assert_held(&txq->lock);
1530 /* We have only q->n_window txq->entries, but we use
1531 * TFD_QUEUE_SIZE_MAX tfds
1533 if (trans->trans_cfg->gen2)
1534 iwl_txq_gen2_tfd_unmap(trans, &txq->entries[idx].meta,
1535 iwl_txq_get_tfd(trans, txq, rd_ptr));
1537 iwl_txq_gen1_tfd_unmap(trans, &txq->entries[idx].meta,
1541 skb = txq->entries[idx].skb;
1543 /* Can be called from irqs-disabled context
1544 * If skb is not NULL, it means that the whole queue is being
1545 * freed and that the queue is not empty - free the skb
1548 iwl_op_mode_free_skb(trans->op_mode, skb);
1549 txq->entries[idx].skb = NULL;
1553 void iwl_txq_progress(struct iwl_txq *txq)
1555 lockdep_assert_held(&txq->lock);
1557 if (!txq->wd_timeout)
1561 * station is asleep and we send data - that must
1562 * be uAPSD or PS-Poll. Don't rearm the timer.
1568 * if empty delete timer, otherwise move timer forward
1569 * since we're making progress on this queue
1571 if (txq->read_ptr == txq->write_ptr)
1572 del_timer(&txq->stuck_timer);
1574 mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
1577 /* Frees buffers until index _not_ inclusive */
1578 void iwl_txq_reclaim(struct iwl_trans *trans, int txq_id, int ssn,
1579 struct sk_buff_head *skbs, bool is_flush)
1581 struct iwl_txq *txq = trans->txqs.txq[txq_id];
1582 int tfd_num, read_ptr, last_to_free;
1584 /* This function is not meant to release cmd queue*/
1585 if (WARN_ON(txq_id == trans->txqs.cmd.q_id))
1591 tfd_num = iwl_txq_get_cmd_index(txq, ssn);
1593 spin_lock_bh(&txq->lock);
1594 read_ptr = iwl_txq_get_cmd_index(txq, txq->read_ptr);
1596 if (!test_bit(txq_id, trans->txqs.queue_used)) {
1597 IWL_DEBUG_TX_QUEUES(trans, "Q %d inactive - ignoring idx %d\n",
1602 if (read_ptr == tfd_num)
1605 IWL_DEBUG_TX_REPLY(trans, "[Q %d] %d (%d) -> %d (%d)\n",
1606 txq_id, read_ptr, txq->read_ptr, tfd_num, ssn);
1608 /*Since we free until index _not_ inclusive, the one before index is
1609 * the last we will free. This one must be used */
1610 last_to_free = iwl_txq_dec_wrap(trans, tfd_num);
1612 if (!iwl_txq_used(txq, last_to_free)) {
1614 "%s: Read index for txq id (%d), last_to_free %d is out of range [0-%d] %d %d.\n",
1615 __func__, txq_id, last_to_free,
1616 trans->trans_cfg->base_params->max_tfd_queue_size,
1617 txq->write_ptr, txq->read_ptr);
1619 iwl_op_mode_time_point(trans->op_mode,
1620 IWL_FW_INI_TIME_POINT_FAKE_TX,
1625 if (WARN_ON(!skb_queue_empty(skbs)))
1629 read_ptr != tfd_num;
1630 txq->read_ptr = iwl_txq_inc_wrap(trans, txq->read_ptr),
1631 read_ptr = iwl_txq_get_cmd_index(txq, txq->read_ptr)) {
1632 struct sk_buff *skb = txq->entries[read_ptr].skb;
1634 if (WARN_ONCE(!skb, "no SKB at %d (%d) on queue %d\n",
1635 read_ptr, txq->read_ptr, txq_id))
1638 iwl_txq_free_tso_page(trans, skb);
1640 __skb_queue_tail(skbs, skb);
1642 txq->entries[read_ptr].skb = NULL;
1644 if (!trans->trans_cfg->gen2)
1645 iwl_txq_gen1_inval_byte_cnt_tbl(trans, txq);
1647 iwl_txq_free_tfd(trans, txq);
1650 iwl_txq_progress(txq);
1652 if (iwl_txq_space(trans, txq) > txq->low_mark &&
1653 test_bit(txq_id, trans->txqs.queue_stopped)) {
1654 struct sk_buff_head overflow_skbs;
1655 struct sk_buff *skb;
1657 __skb_queue_head_init(&overflow_skbs);
1658 skb_queue_splice_init(&txq->overflow_q,
1659 is_flush ? skbs : &overflow_skbs);
1662 * We are going to transmit from the overflow queue.
1663 * Remember this state so that wait_for_txq_empty will know we
1664 * are adding more packets to the TFD queue. It cannot rely on
1665 * the state of &txq->overflow_q, as we just emptied it, but
1666 * haven't TXed the content yet.
1668 txq->overflow_tx = true;
1671 * This is tricky: we are in reclaim path which is non
1672 * re-entrant, so noone will try to take the access the
1673 * txq data from that path. We stopped tx, so we can't
1674 * have tx as well. Bottom line, we can unlock and re-lock
1677 spin_unlock_bh(&txq->lock);
1679 while ((skb = __skb_dequeue(&overflow_skbs))) {
1680 struct iwl_device_tx_cmd *dev_cmd_ptr;
1682 dev_cmd_ptr = *(void **)((u8 *)skb->cb +
1683 trans->txqs.dev_cmd_offs);
1686 * Note that we can very well be overflowing again.
1687 * In that case, iwl_txq_space will be small again
1688 * and we won't wake mac80211's queue.
1690 iwl_trans_tx(trans, skb, dev_cmd_ptr, txq_id);
1693 if (iwl_txq_space(trans, txq) > txq->low_mark)
1694 iwl_wake_queue(trans, txq);
1696 spin_lock_bh(&txq->lock);
1697 txq->overflow_tx = false;
1701 spin_unlock_bh(&txq->lock);
1704 /* Set wr_ptr of specific device and txq */
1705 void iwl_txq_set_q_ptrs(struct iwl_trans *trans, int txq_id, int ptr)
1707 struct iwl_txq *txq = trans->txqs.txq[txq_id];
1709 spin_lock_bh(&txq->lock);
1711 txq->write_ptr = ptr;
1712 txq->read_ptr = txq->write_ptr;
1714 spin_unlock_bh(&txq->lock);
1717 void iwl_trans_txq_freeze_timer(struct iwl_trans *trans, unsigned long txqs,
1722 for_each_set_bit(queue, &txqs, BITS_PER_LONG) {
1723 struct iwl_txq *txq = trans->txqs.txq[queue];
1726 spin_lock_bh(&txq->lock);
1730 if (txq->frozen == freeze)
1733 IWL_DEBUG_TX_QUEUES(trans, "%s TXQ %d\n",
1734 freeze ? "Freezing" : "Waking", queue);
1736 txq->frozen = freeze;
1738 if (txq->read_ptr == txq->write_ptr)
1742 if (unlikely(time_after(now,
1743 txq->stuck_timer.expires))) {
1745 * The timer should have fired, maybe it is
1746 * spinning right now on the lock.
1750 /* remember how long until the timer fires */
1751 txq->frozen_expiry_remainder =
1752 txq->stuck_timer.expires - now;
1753 del_timer(&txq->stuck_timer);
1758 * Wake a non-empty queue -> arm timer with the
1759 * remainder before it froze
1761 mod_timer(&txq->stuck_timer,
1762 now + txq->frozen_expiry_remainder);
1765 spin_unlock_bh(&txq->lock);
1769 #define HOST_COMPLETE_TIMEOUT (2 * HZ)
1771 static int iwl_trans_txq_send_hcmd_sync(struct iwl_trans *trans,
1772 struct iwl_host_cmd *cmd)
1774 const char *cmd_str = iwl_get_cmd_string(trans, cmd->id);
1775 struct iwl_txq *txq = trans->txqs.txq[trans->txqs.cmd.q_id];
1779 IWL_DEBUG_INFO(trans, "Attempting to send sync command %s\n", cmd_str);
1781 if (WARN(test_and_set_bit(STATUS_SYNC_HCMD_ACTIVE,
1783 "Command %s: a command is already active!\n", cmd_str))
1786 IWL_DEBUG_INFO(trans, "Setting HCMD_ACTIVE for command %s\n", cmd_str);
1788 cmd_idx = trans->ops->send_cmd(trans, cmd);
1791 clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1792 IWL_ERR(trans, "Error sending %s: enqueue_hcmd failed: %d\n",
1797 ret = wait_event_timeout(trans->wait_command_queue,
1798 !test_bit(STATUS_SYNC_HCMD_ACTIVE,
1800 HOST_COMPLETE_TIMEOUT);
1802 IWL_ERR(trans, "Error sending %s: time out after %dms.\n",
1803 cmd_str, jiffies_to_msecs(HOST_COMPLETE_TIMEOUT));
1805 IWL_ERR(trans, "Current CMD queue read_ptr %d write_ptr %d\n",
1806 txq->read_ptr, txq->write_ptr);
1808 clear_bit(STATUS_SYNC_HCMD_ACTIVE, &trans->status);
1809 IWL_DEBUG_INFO(trans, "Clearing HCMD_ACTIVE for command %s\n",
1813 iwl_trans_sync_nmi(trans);
1817 if (test_bit(STATUS_FW_ERROR, &trans->status)) {
1818 if (!test_and_clear_bit(STATUS_SUPPRESS_CMD_ERROR_ONCE,
1820 IWL_ERR(trans, "FW error in SYNC CMD %s\n", cmd_str);
1827 if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1828 test_bit(STATUS_RFKILL_OPMODE, &trans->status)) {
1829 IWL_DEBUG_RF_KILL(trans, "RFKILL in SYNC CMD... no rsp\n");
1834 if ((cmd->flags & CMD_WANT_SKB) && !cmd->resp_pkt) {
1835 IWL_ERR(trans, "Error: Response NULL in '%s'\n", cmd_str);
1843 if (cmd->flags & CMD_WANT_SKB) {
1845 * Cancel the CMD_WANT_SKB flag for the cmd in the
1846 * TX cmd queue. Otherwise in case the cmd comes
1847 * in later, it will possibly set an invalid
1848 * address (cmd->meta.source).
1850 txq->entries[cmd_idx].meta.flags &= ~CMD_WANT_SKB;
1853 if (cmd->resp_pkt) {
1855 cmd->resp_pkt = NULL;
1861 int iwl_trans_txq_send_hcmd(struct iwl_trans *trans,
1862 struct iwl_host_cmd *cmd)
1864 /* Make sure the NIC is still alive in the bus */
1865 if (test_bit(STATUS_TRANS_DEAD, &trans->status))
1868 if (!(cmd->flags & CMD_SEND_IN_RFKILL) &&
1869 test_bit(STATUS_RFKILL_OPMODE, &trans->status)) {
1870 IWL_DEBUG_RF_KILL(trans, "Dropping CMD 0x%x: RF KILL\n",
1875 if (unlikely(trans->system_pm_mode == IWL_PLAT_PM_MODE_D3 &&
1876 !(cmd->flags & CMD_SEND_IN_D3))) {
1877 IWL_DEBUG_WOWLAN(trans, "Dropping CMD 0x%x: D3\n", cmd->id);
1881 if (cmd->flags & CMD_ASYNC) {
1884 /* An asynchronous command can not expect an SKB to be set. */
1885 if (WARN_ON(cmd->flags & CMD_WANT_SKB))
1888 ret = trans->ops->send_cmd(trans, cmd);
1891 "Error sending %s: enqueue_hcmd failed: %d\n",
1892 iwl_get_cmd_string(trans, cmd->id), ret);
1898 return iwl_trans_txq_send_hcmd_sync(trans, cmd);