2 * Copyright(c) 2015 - 2018 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/types.h>
49 #include <linux/device.h>
50 #include <linux/dmapool.h>
51 #include <linux/slab.h>
52 #include <linux/list.h>
53 #include <linux/highmem.h>
55 #include <linux/uio.h>
56 #include <linux/rbtree.h>
57 #include <linux/spinlock.h>
58 #include <linux/delay.h>
59 #include <linux/kthread.h>
60 #include <linux/mmu_context.h>
61 #include <linux/module.h>
62 #include <linux/vmalloc.h>
63 #include <linux/string.h>
68 #include "user_sdma.h"
69 #include "verbs.h" /* for the headers */
70 #include "common.h" /* for struct hfi1_tid_info */
73 static uint hfi1_sdma_comp_ring_size = 128;
74 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
75 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
77 static unsigned initial_pkt_count = 8;
79 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts);
80 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status);
81 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq);
82 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin);
83 static int pin_vector_pages(struct user_sdma_request *req,
84 struct user_sdma_iovec *iovec);
85 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
86 unsigned start, unsigned npages);
87 static int check_header_template(struct user_sdma_request *req,
88 struct hfi1_pkt_header *hdr, u32 lrhlen,
90 static int set_txreq_header(struct user_sdma_request *req,
91 struct user_sdma_txreq *tx, u32 datalen);
92 static int set_txreq_header_ahg(struct user_sdma_request *req,
93 struct user_sdma_txreq *tx, u32 len);
94 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
95 struct hfi1_user_sdma_comp_q *cq,
96 u16 idx, enum hfi1_sdma_comp_state state,
98 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags);
99 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
101 static int defer_packet_queue(
102 struct sdma_engine *sde,
103 struct iowait_work *wait,
104 struct sdma_txreq *txreq,
107 static void activate_packet_queue(struct iowait *wait, int reason);
108 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
110 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode);
111 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
112 void *arg2, bool *stop);
113 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode);
114 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
116 static struct mmu_rb_ops sdma_rb_ops = {
117 .filter = sdma_rb_filter,
118 .insert = sdma_rb_insert,
119 .evict = sdma_rb_evict,
120 .remove = sdma_rb_remove,
121 .invalidate = sdma_rb_invalidate
124 static int defer_packet_queue(
125 struct sdma_engine *sde,
126 struct iowait_work *wait,
127 struct sdma_txreq *txreq,
131 struct hfi1_user_sdma_pkt_q *pq =
132 container_of(wait->iow, struct hfi1_user_sdma_pkt_q, busy);
133 struct user_sdma_txreq *tx =
134 container_of(txreq, struct user_sdma_txreq, txreq);
136 if (sdma_progress(sde, seq, txreq)) {
137 if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
141 * We are assuming that if the list is enqueued somewhere, it
142 * is to the dmawait list since that is the only place where
143 * it is supposed to be enqueued.
145 xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
146 write_seqlock(&sde->waitlock);
147 if (list_empty(&pq->busy.list))
148 iowait_queue(pkts_sent, &pq->busy, &sde->dmawait);
149 write_sequnlock(&sde->waitlock);
155 static void activate_packet_queue(struct iowait *wait, int reason)
157 struct hfi1_user_sdma_pkt_q *pq =
158 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
159 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
160 wake_up(&wait->wait_dma);
163 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
164 struct hfi1_filedata *fd)
168 struct hfi1_devdata *dd;
169 struct hfi1_user_sdma_comp_q *cq;
170 struct hfi1_user_sdma_pkt_q *pq;
175 if (!hfi1_sdma_comp_ring_size)
180 pq = kzalloc(sizeof(*pq), GFP_KERNEL);
185 pq->ctxt = uctxt->ctxt;
186 pq->subctxt = fd->subctxt;
187 pq->n_max_reqs = hfi1_sdma_comp_ring_size;
188 atomic_set(&pq->n_reqs, 0);
189 init_waitqueue_head(&pq->wait);
190 atomic_set(&pq->n_locked, 0);
193 iowait_init(&pq->busy, 0, NULL, NULL, defer_packet_queue,
194 activate_packet_queue, NULL);
197 pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
203 pq->req_in_use = kcalloc(BITS_TO_LONGS(hfi1_sdma_comp_ring_size),
204 sizeof(*pq->req_in_use),
207 goto pq_reqs_no_in_use;
209 snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
211 pq->txreq_cache = kmem_cache_create(buf,
212 sizeof(struct user_sdma_txreq),
216 if (!pq->txreq_cache) {
217 dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
222 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
226 cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
227 * hfi1_sdma_comp_ring_size));
231 cq->nentries = hfi1_sdma_comp_ring_size;
233 ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
236 dd_dev_err(dd, "Failed to register with MMU %d", ret);
250 kmem_cache_destroy(pq->txreq_cache);
252 kfree(pq->req_in_use);
261 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd,
262 struct hfi1_ctxtdata *uctxt)
264 struct hfi1_user_sdma_pkt_q *pq;
266 trace_hfi1_sdma_user_free_queues(uctxt->dd, uctxt->ctxt, fd->subctxt);
271 hfi1_mmu_rb_unregister(pq->handler);
272 iowait_sdma_drain(&pq->busy);
273 /* Wait until all requests have been freed. */
274 wait_event_interruptible(
276 !atomic_read(&pq->n_reqs));
278 kfree(pq->req_in_use);
279 kmem_cache_destroy(pq->txreq_cache);
284 vfree(fd->cq->comps);
291 static u8 dlid_to_selector(u16 dlid)
293 static u8 mapping[256];
294 static int initialized;
299 memset(mapping, 0xFF, 256);
303 hash = ((dlid >> 8) ^ dlid) & 0xFF;
304 if (mapping[hash] == 0xFF) {
305 mapping[hash] = next;
306 next = (next + 1) & 0x7F;
309 return mapping[hash];
313 * hfi1_user_sdma_process_request() - Process and start a user sdma request
314 * @fd: valid file descriptor
315 * @iovec: array of io vectors to process
316 * @dim: overall iovec array size
317 * @count: number of io vector array entries processed
319 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
320 struct iovec *iovec, unsigned long dim,
321 unsigned long *count)
324 struct hfi1_ctxtdata *uctxt = fd->uctxt;
325 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
326 struct hfi1_user_sdma_comp_q *cq = fd->cq;
327 struct hfi1_devdata *dd = pq->dd;
328 unsigned long idx = 0;
329 u8 pcount = initial_pkt_count;
330 struct sdma_req_info info;
331 struct user_sdma_request *req;
338 if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
341 "[%u:%u:%u] First vector not big enough for header %lu/%lu",
342 dd->unit, uctxt->ctxt, fd->subctxt,
343 iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
346 ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
348 hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
349 dd->unit, uctxt->ctxt, fd->subctxt, ret);
353 trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
355 if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
357 "[%u:%u:%u:%u] Invalid comp index",
358 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
363 * Sanity check the header io vector count. Need at least 1 vector
364 * (header) and cannot be larger than the actual io vector count.
366 if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
368 "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
369 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
370 req_iovcnt(info.ctrl), dim);
374 if (!info.fragsize) {
376 "[%u:%u:%u:%u] Request does not specify fragsize",
377 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
381 /* Try to claim the request. */
382 if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
383 hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
384 dd->unit, uctxt->ctxt, fd->subctxt,
389 * All safety checks have been done and this request has been claimed.
391 trace_hfi1_sdma_user_process_request(dd, uctxt->ctxt, fd->subctxt,
393 req = pq->reqs + info.comp_idx;
394 req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
403 req->seqsubmitted = 0;
406 INIT_LIST_HEAD(&req->txps);
408 memcpy(&req->info, &info, sizeof(info));
410 /* The request is initialized, count it */
411 atomic_inc(&pq->n_reqs);
413 if (req_opcode(info.ctrl) == EXPECTED) {
414 /* expected must have a TID info and at least one data vector */
415 if (req->data_iovs < 2) {
417 "Not enough vectors for expected request");
424 if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
425 SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
426 MAX_VECTORS_PER_REQ);
430 /* Copy the header from the user buffer */
431 ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
434 SDMA_DBG(req, "Failed to copy header template (%d)", ret);
439 /* If Static rate control is not enabled, sanitize the header. */
440 if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
443 /* Validate the opcode. Do not trust packets from user space blindly. */
444 opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
445 if ((opcode & USER_OPCODE_CHECK_MASK) !=
446 USER_OPCODE_CHECK_VAL) {
447 SDMA_DBG(req, "Invalid opcode (%d)", opcode);
452 * Validate the vl. Do not trust packets from user space blindly.
453 * VL comes from PBC, SC comes from LRH, and the VL needs to
454 * match the SC look up.
456 vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
457 sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
458 (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
459 if (vl >= dd->pport->vls_operational ||
460 vl != sc_to_vlt(dd, sc)) {
461 SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
466 /* Checking P_KEY for requests from user-space */
467 pkey = (u16)be32_to_cpu(req->hdr.bth[0]);
468 slid = be16_to_cpu(req->hdr.lrh[3]);
469 if (egress_pkey_check(dd->pport, slid, pkey, sc, PKEY_CHECK_INVALID)) {
475 * Also should check the BTH.lnh. If it says the next header is GRH then
476 * the RXE parsing will be off and will land in the middle of the KDETH
477 * or miss it entirely.
479 if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
480 SDMA_DBG(req, "User tried to pass in a GRH");
485 req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
487 * Calculate the initial TID offset based on the values of
488 * KDETH.OFFSET and KDETH.OM that are passed in.
490 req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
491 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
492 KDETH_OM_LARGE : KDETH_OM_SMALL);
493 trace_hfi1_sdma_user_initial_tidoffset(dd, uctxt->ctxt, fd->subctxt,
494 info.comp_idx, req->tidoffset);
497 /* Save all the IO vector structures */
498 for (i = 0; i < req->data_iovs; i++) {
499 req->iovs[i].offset = 0;
500 INIT_LIST_HEAD(&req->iovs[i].list);
501 memcpy(&req->iovs[i].iov,
503 sizeof(req->iovs[i].iov));
504 ret = pin_vector_pages(req, &req->iovs[i]);
509 req->data_len += req->iovs[i].iov.iov_len;
511 trace_hfi1_sdma_user_data_length(dd, uctxt->ctxt, fd->subctxt,
512 info.comp_idx, req->data_len);
513 if (pcount > req->info.npkts)
514 pcount = req->info.npkts;
517 * User space will provide the TID info only when the
518 * request type is EXPECTED. This is true even if there is
519 * only one packet in the request and the header is already
520 * setup. The reason for the singular TID case is that the
521 * driver needs to perform safety checks.
523 if (req_opcode(req->info.ctrl) == EXPECTED) {
524 u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
527 if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
533 * We have to copy all of the tids because they may vary
534 * in size and, therefore, the TID count might not be
535 * equal to the pkt count. However, there is no way to
536 * tell at this point.
538 tmp = memdup_user(iovec[idx].iov_base,
539 ntids * sizeof(*req->tids));
542 SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
552 dlid = be16_to_cpu(req->hdr.lrh[1]);
553 selector = dlid_to_selector(dlid);
554 selector += uctxt->ctxt + fd->subctxt;
555 req->sde = sdma_select_user_engine(dd, selector, vl);
557 if (!req->sde || !sdma_running(req->sde)) {
562 /* We don't need an AHG entry if the request contains only one packet */
563 if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
564 req->ahg_idx = sdma_ahg_alloc(req->sde);
566 set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
567 pq->state = SDMA_PKT_Q_ACTIVE;
568 /* Send the first N packets in the request to buy us some time */
569 ret = user_sdma_send_pkts(req, pcount);
570 if (unlikely(ret < 0 && ret != -EBUSY))
574 * This is a somewhat blocking send implementation.
575 * The driver will block the caller until all packets of the
576 * request have been submitted to the SDMA engine. However, it
577 * will not wait for send completions.
579 while (req->seqsubmitted != req->info.npkts) {
580 ret = user_sdma_send_pkts(req, pcount);
584 wait_event_interruptible_timeout(
586 (pq->state == SDMA_PKT_Q_ACTIVE),
588 SDMA_IOWAIT_TIMEOUT));
595 * If the submitted seqsubmitted == npkts, the completion routine
596 * controls the final state. If sequbmitted < npkts, wait for any
597 * outstanding packets to finish before cleaning up.
599 if (req->seqsubmitted < req->info.npkts) {
600 if (req->seqsubmitted)
601 wait_event(pq->busy.wait_dma,
602 (req->seqcomp == req->seqsubmitted - 1));
603 user_sdma_free_request(req, true);
605 set_comp_state(pq, cq, info.comp_idx, ERROR, ret);
610 static inline u32 compute_data_length(struct user_sdma_request *req,
611 struct user_sdma_txreq *tx)
614 * Determine the proper size of the packet data.
615 * The size of the data of the first packet is in the header
616 * template. However, it includes the header and ICRC, which need
618 * The minimum representable packet data length in a header is 4 bytes,
619 * therefore, when the data length request is less than 4 bytes, there's
620 * only one packet, and the packet data length is equal to that of the
621 * request data length.
622 * The size of the remaining packets is the minimum of the frag
623 * size (MTU) or remaining data in the request.
628 if (req->data_len < sizeof(u32))
631 len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
632 (sizeof(tx->hdr) - 4));
633 } else if (req_opcode(req->info.ctrl) == EXPECTED) {
634 u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
637 * Get the data length based on the remaining space in the
640 len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
641 /* If we've filled up the TID pair, move to the next one. */
642 if (unlikely(!len) && ++req->tididx < req->n_tids &&
643 req->tids[req->tididx]) {
644 tidlen = EXP_TID_GET(req->tids[req->tididx],
647 len = min_t(u32, tidlen, req->info.fragsize);
650 * Since the TID pairs map entire pages, make sure that we
651 * are not going to try to send more data that we have
654 len = min(len, req->data_len - req->sent);
656 len = min(req->data_len - req->sent, (u32)req->info.fragsize);
658 trace_hfi1_sdma_user_compute_length(req->pq->dd,
666 static inline u32 pad_len(u32 len)
668 if (len & (sizeof(u32) - 1))
669 len += sizeof(u32) - (len & (sizeof(u32) - 1));
673 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
675 /* (Size of complete header - size of PBC) + 4B ICRC + data length */
676 return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
679 static int user_sdma_txadd_ahg(struct user_sdma_request *req,
680 struct user_sdma_txreq *tx,
684 u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
685 u32 lrhlen = get_lrh_len(req->hdr, pad_len(datalen));
686 struct hfi1_user_sdma_pkt_q *pq = req->pq;
689 * Copy the request header into the tx header
690 * because the HW needs a cacheline-aligned
692 * This copy can be optimized out if the hdr
693 * member of user_sdma_request were also
696 memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
697 if (PBC2LRH(pbclen) != lrhlen) {
698 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
699 tx->hdr.pbc[0] = cpu_to_le16(pbclen);
701 ret = check_header_template(req, &tx->hdr, lrhlen, datalen);
704 ret = sdma_txinit_ahg(&tx->txreq, SDMA_TXREQ_F_AHG_COPY,
705 sizeof(tx->hdr) + datalen, req->ahg_idx,
706 0, NULL, 0, user_sdma_txreq_cb);
709 ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq, &tx->hdr, sizeof(tx->hdr));
711 sdma_txclean(pq->dd, &tx->txreq);
715 static int user_sdma_txadd(struct user_sdma_request *req,
716 struct user_sdma_txreq *tx,
717 struct user_sdma_iovec *iovec, u32 datalen,
718 u32 *queued_ptr, u32 *data_sent_ptr,
722 unsigned int pageidx, len;
723 unsigned long base, offset;
724 u64 iov_offset = *iov_offset_ptr;
725 u32 queued = *queued_ptr, data_sent = *data_sent_ptr;
726 struct hfi1_user_sdma_pkt_q *pq = req->pq;
728 base = (unsigned long)iovec->iov.iov_base;
729 offset = offset_in_page(base + iovec->offset + iov_offset);
730 pageidx = (((iovec->offset + iov_offset + base) - (base & PAGE_MASK)) >>
732 len = offset + req->info.fragsize > PAGE_SIZE ?
733 PAGE_SIZE - offset : req->info.fragsize;
734 len = min((datalen - queued), len);
735 ret = sdma_txadd_page(pq->dd, &tx->txreq, iovec->pages[pageidx],
738 SDMA_DBG(req, "SDMA txreq add page failed %d\n", ret);
744 if (unlikely(queued < datalen && pageidx == iovec->npages &&
745 req->iov_idx < req->data_iovs - 1)) {
746 iovec->offset += iov_offset;
747 iovec = &req->iovs[++req->iov_idx];
751 *queued_ptr = queued;
752 *data_sent_ptr = data_sent;
753 *iov_offset_ptr = iov_offset;
757 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts)
762 struct user_sdma_txreq *tx = NULL;
763 struct hfi1_user_sdma_pkt_q *pq = NULL;
764 struct user_sdma_iovec *iovec = NULL;
771 /* If tx completion has reported an error, we are done. */
772 if (READ_ONCE(req->has_error))
776 * Check if we might have sent the entire request already
778 if (unlikely(req->seqnum == req->info.npkts)) {
779 if (!list_empty(&req->txps))
784 if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
785 maxpkts = req->info.npkts - req->seqnum;
787 while (npkts < maxpkts) {
788 u32 datalen = 0, queued = 0, data_sent = 0;
792 * Check whether any of the completions have come back
793 * with errors. If so, we are not going to process any
794 * more packets from this request.
796 if (READ_ONCE(req->has_error))
799 tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
806 INIT_LIST_HEAD(&tx->list);
809 * For the last packet set the ACK request
810 * and disable header suppression.
812 if (req->seqnum == req->info.npkts - 1)
813 tx->flags |= (TXREQ_FLAGS_REQ_ACK |
814 TXREQ_FLAGS_REQ_DISABLE_SH);
817 * Calculate the payload size - this is min of the fragment
818 * (MTU) size or the remaining bytes in the request but only
819 * if we have payload data.
822 iovec = &req->iovs[req->iov_idx];
823 if (READ_ONCE(iovec->offset) == iovec->iov.iov_len) {
824 if (++req->iov_idx == req->data_iovs) {
828 iovec = &req->iovs[req->iov_idx];
829 WARN_ON(iovec->offset);
832 datalen = compute_data_length(req, tx);
835 * Disable header suppression for the payload <= 8DWS.
836 * If there is an uncorrectable error in the receive
837 * data FIFO when the received payload size is less than
838 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
839 * not reported.There is set RHF.EccErr if the header
844 "Request has data but pkt len is 0");
847 } else if (datalen <= 32) {
848 tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
852 if (req->ahg_idx >= 0) {
854 ret = user_sdma_txadd_ahg(req, tx, datalen);
860 changes = set_txreq_header_ahg(req, tx,
868 ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
869 datalen, user_sdma_txreq_cb);
873 * Modify the header for this packet. This only needs
874 * to be done if we are not going to use AHG. Otherwise,
875 * the HW will do it based on the changes we gave it
876 * during sdma_txinit_ahg().
878 ret = set_txreq_header(req, tx, datalen);
884 * If the request contains any data vectors, add up to
885 * fragsize bytes to the descriptor.
887 while (queued < datalen &&
888 (req->sent + data_sent) < req->data_len) {
889 ret = user_sdma_txadd(req, tx, iovec, datalen,
890 &queued, &data_sent, &iov_offset);
895 * The txreq was submitted successfully so we can update
898 req->koffset += datalen;
899 if (req_opcode(req->info.ctrl) == EXPECTED)
900 req->tidoffset += datalen;
901 req->sent += data_sent;
903 iovec->offset += iov_offset;
904 list_add_tail(&tx->txreq.list, &req->txps);
906 * It is important to increment this here as it is used to
907 * generate the BTH.PSN and, therefore, can't be bulk-updated
908 * outside of the loop.
910 tx->seqnum = req->seqnum++;
914 ret = sdma_send_txlist(req->sde,
915 iowait_get_ib_work(&pq->busy),
917 req->seqsubmitted += count;
918 if (req->seqsubmitted == req->info.npkts) {
920 * The txreq has already been submitted to the HW queue
921 * so we can free the AHG entry now. Corruption will not
922 * happen due to the sequential manner in which
923 * descriptors are processed.
925 if (req->ahg_idx >= 0)
926 sdma_ahg_free(req->sde, req->ahg_idx);
931 sdma_txclean(pq->dd, &tx->txreq);
933 kmem_cache_free(pq->txreq_cache, tx);
937 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
939 struct evict_data evict_data;
941 evict_data.cleared = 0;
942 evict_data.target = npages;
943 hfi1_mmu_rb_evict(pq->handler, &evict_data);
944 return evict_data.cleared;
947 static int pin_sdma_pages(struct user_sdma_request *req,
948 struct user_sdma_iovec *iovec,
949 struct sdma_mmu_node *node,
954 struct hfi1_user_sdma_pkt_q *pq = req->pq;
956 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
959 memcpy(pages, node->pages, node->npages * sizeof(*pages));
961 npages -= node->npages;
963 if (!hfi1_can_pin_pages(pq->dd, pq->mm,
964 atomic_read(&pq->n_locked), npages)) {
965 cleared = sdma_cache_evict(pq, npages);
966 if (cleared >= npages)
969 pinned = hfi1_acquire_user_pages(pq->mm,
970 ((unsigned long)iovec->iov.iov_base +
971 (node->npages * PAGE_SIZE)), npages, 0,
972 pages + node->npages);
977 if (pinned != npages) {
978 unpin_vector_pages(pq->mm, pages, node->npages, pinned);
982 node->rb.len = iovec->iov.iov_len;
984 atomic_add(pinned, &pq->n_locked);
988 static void unpin_sdma_pages(struct sdma_mmu_node *node)
991 unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
992 atomic_sub(node->npages, &node->pq->n_locked);
996 static int pin_vector_pages(struct user_sdma_request *req,
997 struct user_sdma_iovec *iovec)
999 int ret = 0, pinned, npages;
1000 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1001 struct sdma_mmu_node *node = NULL;
1002 struct mmu_rb_node *rb_node;
1007 hfi1_mmu_rb_remove_unless_exact(pq->handler,
1009 iovec->iov.iov_base,
1010 iovec->iov.iov_len, &rb_node);
1012 node = container_of(rb_node, struct sdma_mmu_node, rb);
1014 atomic_inc(&node->refcount);
1015 iovec->pages = node->pages;
1016 iovec->npages = node->npages;
1023 node = kzalloc(sizeof(*node), GFP_KERNEL);
1027 node->rb.addr = (unsigned long)iovec->iov.iov_base;
1029 atomic_set(&node->refcount, 0);
1033 npages = num_user_pages((unsigned long)iov->iov_base, iov->iov_len);
1034 if (node->npages < npages) {
1035 pinned = pin_sdma_pages(req, iovec, node, npages);
1040 node->npages += pinned;
1041 npages = node->npages;
1043 iovec->pages = node->pages;
1044 iovec->npages = npages;
1047 ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1054 unpin_sdma_pages(node);
1059 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1060 unsigned start, unsigned npages)
1062 hfi1_release_user_pages(mm, pages + start, npages, false);
1066 static int check_header_template(struct user_sdma_request *req,
1067 struct hfi1_pkt_header *hdr, u32 lrhlen,
1071 * Perform safety checks for any type of packet:
1072 * - transfer size is multiple of 64bytes
1073 * - packet length is multiple of 4 bytes
1074 * - packet length is not larger than MTU size
1076 * These checks are only done for the first packet of the
1077 * transfer since the header is "given" to us by user space.
1078 * For the remainder of the packets we compute the values.
1080 if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1081 lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1084 if (req_opcode(req->info.ctrl) == EXPECTED) {
1086 * The header is checked only on the first packet. Furthermore,
1087 * we ensure that at least one TID entry is copied when the
1088 * request is submitted. Therefore, we don't have to verify that
1089 * tididx points to something sane.
1091 u32 tidval = req->tids[req->tididx],
1092 tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1093 tididx = EXP_TID_GET(tidval, IDX),
1094 tidctrl = EXP_TID_GET(tidval, CTRL),
1096 __le32 kval = hdr->kdeth.ver_tid_offset;
1098 tidoff = KDETH_GET(kval, OFFSET) *
1099 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1100 KDETH_OM_LARGE : KDETH_OM_SMALL);
1102 * Expected receive packets have the following
1103 * additional checks:
1104 * - offset is not larger than the TID size
1105 * - TIDCtrl values match between header and TID array
1106 * - TID indexes match between header and TID array
1108 if ((tidoff + datalen > tidlen) ||
1109 KDETH_GET(kval, TIDCTRL) != tidctrl ||
1110 KDETH_GET(kval, TID) != tididx)
1117 * Correctly set the BTH.PSN field based on type of
1118 * transfer - eager packets can just increment the PSN but
1119 * expected packets encode generation and sequence in the
1120 * BTH.PSN field so just incrementing will result in errors.
1122 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1124 u32 val = be32_to_cpu(bthpsn),
1125 mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1129 psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1135 static int set_txreq_header(struct user_sdma_request *req,
1136 struct user_sdma_txreq *tx, u32 datalen)
1138 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1139 struct hfi1_pkt_header *hdr = &tx->hdr;
1140 u8 omfactor; /* KDETH.OM */
1143 u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1145 /* Copy the header template to the request before modification */
1146 memcpy(hdr, &req->hdr, sizeof(*hdr));
1149 * Check if the PBC and LRH length are mismatched. If so
1150 * adjust both in the header.
1152 pbclen = le16_to_cpu(hdr->pbc[0]);
1153 if (PBC2LRH(pbclen) != lrhlen) {
1154 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1155 hdr->pbc[0] = cpu_to_le16(pbclen);
1156 hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1159 * This is the first packet in the sequence that has
1160 * a "static" size that can be used for the rest of
1161 * the packets (besides the last one).
1163 if (unlikely(req->seqnum == 2)) {
1165 * From this point on the lengths in both the
1166 * PBC and LRH are the same until the last
1168 * Adjust the template so we don't have to update
1171 req->hdr.pbc[0] = hdr->pbc[0];
1172 req->hdr.lrh[2] = hdr->lrh[2];
1176 * We only have to modify the header if this is not the
1177 * first packet in the request. Otherwise, we use the
1178 * header given to us.
1180 if (unlikely(!req->seqnum)) {
1181 ret = check_header_template(req, hdr, lrhlen, datalen);
1187 hdr->bth[2] = cpu_to_be32(
1188 set_pkt_bth_psn(hdr->bth[2],
1189 (req_opcode(req->info.ctrl) == EXPECTED),
1192 /* Set ACK request on last packet */
1193 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1194 hdr->bth[2] |= cpu_to_be32(1UL << 31);
1196 /* Set the new offset */
1197 hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1198 /* Expected packets have to fill in the new TID information */
1199 if (req_opcode(req->info.ctrl) == EXPECTED) {
1200 tidval = req->tids[req->tididx];
1202 * If the offset puts us at the end of the current TID,
1203 * advance everything.
1205 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1209 * Since we don't copy all the TIDs, all at once,
1210 * we have to check again.
1212 if (++req->tididx > req->n_tids - 1 ||
1213 !req->tids[req->tididx]) {
1216 tidval = req->tids[req->tididx];
1218 omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1219 KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
1220 KDETH_OM_SMALL_SHIFT;
1221 /* Set KDETH.TIDCtrl based on value for this TID. */
1222 KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1223 EXP_TID_GET(tidval, CTRL));
1224 /* Set KDETH.TID based on value for this TID */
1225 KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1226 EXP_TID_GET(tidval, IDX));
1227 /* Clear KDETH.SH when DISABLE_SH flag is set */
1228 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1229 KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1231 * Set the KDETH.OFFSET and KDETH.OM based on size of
1234 trace_hfi1_sdma_user_tid_info(
1235 pq->dd, pq->ctxt, pq->subctxt, req->info.comp_idx,
1236 req->tidoffset, req->tidoffset >> omfactor,
1237 omfactor != KDETH_OM_SMALL_SHIFT);
1238 KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1239 req->tidoffset >> omfactor);
1240 KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1241 omfactor != KDETH_OM_SMALL_SHIFT);
1244 trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1245 req->info.comp_idx, hdr, tidval);
1246 return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1249 static int set_txreq_header_ahg(struct user_sdma_request *req,
1250 struct user_sdma_txreq *tx, u32 datalen)
1252 u32 ahg[AHG_KDETH_ARRAY_SIZE];
1254 u8 omfactor; /* KDETH.OM */
1255 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1256 struct hfi1_pkt_header *hdr = &req->hdr;
1257 u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1258 u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1259 size_t array_size = ARRAY_SIZE(ahg);
1261 if (PBC2LRH(pbclen) != lrhlen) {
1262 /* PBC.PbcLengthDWs */
1263 idx = ahg_header_set(ahg, idx, array_size, 0, 0, 12,
1264 (__force u16)cpu_to_le16(LRH2PBC(lrhlen)));
1267 /* LRH.PktLen (we need the full 16 bits due to byte swap) */
1268 idx = ahg_header_set(ahg, idx, array_size, 3, 0, 16,
1269 (__force u16)cpu_to_be16(lrhlen >> 2));
1275 * Do the common updates
1277 /* BTH.PSN and BTH.A */
1278 val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1279 (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1280 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1282 idx = ahg_header_set(ahg, idx, array_size, 6, 0, 16,
1283 (__force u16)cpu_to_be16(val32 >> 16));
1286 idx = ahg_header_set(ahg, idx, array_size, 6, 16, 16,
1287 (__force u16)cpu_to_be16(val32 & 0xffff));
1291 idx = ahg_header_set(ahg, idx, array_size, 15, 0, 16,
1292 (__force u16)cpu_to_le16(req->koffset & 0xffff));
1295 idx = ahg_header_set(ahg, idx, array_size, 15, 16, 16,
1296 (__force u16)cpu_to_le16(req->koffset >> 16));
1299 if (req_opcode(req->info.ctrl) == EXPECTED) {
1302 tidval = req->tids[req->tididx];
1305 * If the offset puts us at the end of the current TID,
1306 * advance everything.
1308 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1312 * Since we don't copy all the TIDs, all at once,
1313 * we have to check again.
1315 if (++req->tididx > req->n_tids - 1 ||
1316 !req->tids[req->tididx])
1318 tidval = req->tids[req->tididx];
1320 omfactor = ((EXP_TID_GET(tidval, LEN) *
1322 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1323 KDETH_OM_SMALL_SHIFT;
1324 /* KDETH.OM and KDETH.OFFSET (TID) */
1325 idx = ahg_header_set(
1326 ahg, idx, array_size, 7, 0, 16,
1327 ((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1328 ((req->tidoffset >> omfactor)
1332 /* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1333 val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1334 (EXP_TID_GET(tidval, IDX) & 0x3ff));
1336 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1337 val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1339 AHG_KDETH_INTR_SHIFT));
1341 val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1342 cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1343 cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1345 AHG_KDETH_INTR_SHIFT));
1348 idx = ahg_header_set(ahg, idx, array_size,
1349 7, 16, 14, (__force u16)val);
1354 trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1355 req->info.comp_idx, req->sde->this_idx,
1356 req->ahg_idx, ahg, idx, tidval);
1357 sdma_txinit_ahg(&tx->txreq,
1358 SDMA_TXREQ_F_USE_AHG,
1359 datalen, req->ahg_idx, idx,
1360 ahg, sizeof(req->hdr),
1361 user_sdma_txreq_cb);
1367 * user_sdma_txreq_cb() - SDMA tx request completion callback.
1368 * @txreq: valid sdma tx request
1369 * @status: success/failure of request
1371 * Called when the SDMA progress state machine gets notification that
1372 * the SDMA descriptors for this tx request have been processed by the
1373 * DMA engine. Called in interrupt context.
1374 * Only do work on completed sequences.
1376 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1378 struct user_sdma_txreq *tx =
1379 container_of(txreq, struct user_sdma_txreq, txreq);
1380 struct user_sdma_request *req;
1381 struct hfi1_user_sdma_pkt_q *pq;
1382 struct hfi1_user_sdma_comp_q *cq;
1383 enum hfi1_sdma_comp_state state = COMPLETE;
1392 if (status != SDMA_TXREQ_S_OK) {
1393 SDMA_DBG(req, "SDMA completion with error %d",
1395 WRITE_ONCE(req->has_error, 1);
1399 req->seqcomp = tx->seqnum;
1400 kmem_cache_free(pq->txreq_cache, tx);
1402 /* sequence isn't complete? We are done */
1403 if (req->seqcomp != req->info.npkts - 1)
1406 user_sdma_free_request(req, false);
1407 set_comp_state(pq, cq, req->info.comp_idx, state, status);
1411 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1413 if (atomic_dec_and_test(&pq->n_reqs))
1417 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1421 if (!list_empty(&req->txps)) {
1422 struct sdma_txreq *t, *p;
1424 list_for_each_entry_safe(t, p, &req->txps, list) {
1425 struct user_sdma_txreq *tx =
1426 container_of(t, struct user_sdma_txreq, txreq);
1427 list_del_init(&t->list);
1428 sdma_txclean(req->pq->dd, t);
1429 kmem_cache_free(req->pq->txreq_cache, tx);
1433 for (i = 0; i < req->data_iovs; i++) {
1434 struct sdma_mmu_node *node = req->iovs[i].node;
1439 req->iovs[i].node = NULL;
1442 hfi1_mmu_rb_remove(req->pq->handler,
1445 atomic_dec(&node->refcount);
1449 clear_bit(req->info.comp_idx, req->pq->req_in_use);
1452 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1453 struct hfi1_user_sdma_comp_q *cq,
1454 u16 idx, enum hfi1_sdma_comp_state state,
1458 cq->comps[idx].errcode = -ret;
1459 smp_wmb(); /* make sure errcode is visible first */
1460 cq->comps[idx].status = state;
1461 trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1465 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1468 return (bool)(node->addr == addr);
1471 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1473 struct sdma_mmu_node *node =
1474 container_of(mnode, struct sdma_mmu_node, rb);
1476 atomic_inc(&node->refcount);
1481 * Return 1 to remove the node from the rb tree and call the remove op.
1483 * Called with the rb tree lock held.
1485 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1486 void *evict_arg, bool *stop)
1488 struct sdma_mmu_node *node =
1489 container_of(mnode, struct sdma_mmu_node, rb);
1490 struct evict_data *evict_data = evict_arg;
1492 /* is this node still being used? */
1493 if (atomic_read(&node->refcount))
1494 return 0; /* keep this node */
1496 /* this node will be evicted, add its pages to our count */
1497 evict_data->cleared += node->npages;
1499 /* have enough pages been cleared? */
1500 if (evict_data->cleared >= evict_data->target)
1503 return 1; /* remove this node */
1506 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1508 struct sdma_mmu_node *node =
1509 container_of(mnode, struct sdma_mmu_node, rb);
1511 unpin_sdma_pages(node);
1515 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1517 struct sdma_mmu_node *node =
1518 container_of(mnode, struct sdma_mmu_node, rb);
1520 if (!atomic_read(&node->refcount))