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_get_priority(&pq->busy);
149 iowait_queue(pkts_sent, &pq->busy, &sde->dmawait);
151 write_sequnlock(&sde->waitlock);
157 static void activate_packet_queue(struct iowait *wait, int reason)
159 struct hfi1_user_sdma_pkt_q *pq =
160 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
161 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
162 wake_up(&wait->wait_dma);
165 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
166 struct hfi1_filedata *fd)
170 struct hfi1_devdata *dd;
171 struct hfi1_user_sdma_comp_q *cq;
172 struct hfi1_user_sdma_pkt_q *pq;
177 if (!hfi1_sdma_comp_ring_size)
182 pq = kzalloc(sizeof(*pq), GFP_KERNEL);
187 pq->ctxt = uctxt->ctxt;
188 pq->subctxt = fd->subctxt;
189 pq->n_max_reqs = hfi1_sdma_comp_ring_size;
190 atomic_set(&pq->n_reqs, 0);
191 init_waitqueue_head(&pq->wait);
192 atomic_set(&pq->n_locked, 0);
195 iowait_init(&pq->busy, 0, NULL, NULL, defer_packet_queue,
196 activate_packet_queue, NULL, NULL);
199 pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
205 pq->req_in_use = kcalloc(BITS_TO_LONGS(hfi1_sdma_comp_ring_size),
206 sizeof(*pq->req_in_use),
209 goto pq_reqs_no_in_use;
211 snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
213 pq->txreq_cache = kmem_cache_create(buf,
214 sizeof(struct user_sdma_txreq),
218 if (!pq->txreq_cache) {
219 dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
224 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
228 cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
229 * hfi1_sdma_comp_ring_size));
233 cq->nentries = hfi1_sdma_comp_ring_size;
235 ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
238 dd_dev_err(dd, "Failed to register with MMU %d", ret);
252 kmem_cache_destroy(pq->txreq_cache);
254 kfree(pq->req_in_use);
263 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd,
264 struct hfi1_ctxtdata *uctxt)
266 struct hfi1_user_sdma_pkt_q *pq;
268 trace_hfi1_sdma_user_free_queues(uctxt->dd, uctxt->ctxt, fd->subctxt);
273 hfi1_mmu_rb_unregister(pq->handler);
274 iowait_sdma_drain(&pq->busy);
275 /* Wait until all requests have been freed. */
276 wait_event_interruptible(
278 !atomic_read(&pq->n_reqs));
280 kfree(pq->req_in_use);
281 kmem_cache_destroy(pq->txreq_cache);
286 vfree(fd->cq->comps);
293 static u8 dlid_to_selector(u16 dlid)
295 static u8 mapping[256];
296 static int initialized;
301 memset(mapping, 0xFF, 256);
305 hash = ((dlid >> 8) ^ dlid) & 0xFF;
306 if (mapping[hash] == 0xFF) {
307 mapping[hash] = next;
308 next = (next + 1) & 0x7F;
311 return mapping[hash];
315 * hfi1_user_sdma_process_request() - Process and start a user sdma request
316 * @fd: valid file descriptor
317 * @iovec: array of io vectors to process
318 * @dim: overall iovec array size
319 * @count: number of io vector array entries processed
321 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
322 struct iovec *iovec, unsigned long dim,
323 unsigned long *count)
326 struct hfi1_ctxtdata *uctxt = fd->uctxt;
327 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
328 struct hfi1_user_sdma_comp_q *cq = fd->cq;
329 struct hfi1_devdata *dd = pq->dd;
330 unsigned long idx = 0;
331 u8 pcount = initial_pkt_count;
332 struct sdma_req_info info;
333 struct user_sdma_request *req;
340 if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
343 "[%u:%u:%u] First vector not big enough for header %lu/%lu",
344 dd->unit, uctxt->ctxt, fd->subctxt,
345 iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
348 ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
350 hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
351 dd->unit, uctxt->ctxt, fd->subctxt, ret);
355 trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
357 if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
359 "[%u:%u:%u:%u] Invalid comp index",
360 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
365 * Sanity check the header io vector count. Need at least 1 vector
366 * (header) and cannot be larger than the actual io vector count.
368 if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
370 "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
371 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
372 req_iovcnt(info.ctrl), dim);
376 if (!info.fragsize) {
378 "[%u:%u:%u:%u] Request does not specify fragsize",
379 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
383 /* Try to claim the request. */
384 if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
385 hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
386 dd->unit, uctxt->ctxt, fd->subctxt,
391 * All safety checks have been done and this request has been claimed.
393 trace_hfi1_sdma_user_process_request(dd, uctxt->ctxt, fd->subctxt,
395 req = pq->reqs + info.comp_idx;
396 req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
405 req->seqsubmitted = 0;
408 INIT_LIST_HEAD(&req->txps);
410 memcpy(&req->info, &info, sizeof(info));
412 /* The request is initialized, count it */
413 atomic_inc(&pq->n_reqs);
415 if (req_opcode(info.ctrl) == EXPECTED) {
416 /* expected must have a TID info and at least one data vector */
417 if (req->data_iovs < 2) {
419 "Not enough vectors for expected request");
426 if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
427 SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
428 MAX_VECTORS_PER_REQ);
432 /* Copy the header from the user buffer */
433 ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
436 SDMA_DBG(req, "Failed to copy header template (%d)", ret);
441 /* If Static rate control is not enabled, sanitize the header. */
442 if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
445 /* Validate the opcode. Do not trust packets from user space blindly. */
446 opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
447 if ((opcode & USER_OPCODE_CHECK_MASK) !=
448 USER_OPCODE_CHECK_VAL) {
449 SDMA_DBG(req, "Invalid opcode (%d)", opcode);
454 * Validate the vl. Do not trust packets from user space blindly.
455 * VL comes from PBC, SC comes from LRH, and the VL needs to
456 * match the SC look up.
458 vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
459 sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
460 (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
461 if (vl >= dd->pport->vls_operational ||
462 vl != sc_to_vlt(dd, sc)) {
463 SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
468 /* Checking P_KEY for requests from user-space */
469 pkey = (u16)be32_to_cpu(req->hdr.bth[0]);
470 slid = be16_to_cpu(req->hdr.lrh[3]);
471 if (egress_pkey_check(dd->pport, slid, pkey, sc, PKEY_CHECK_INVALID)) {
477 * Also should check the BTH.lnh. If it says the next header is GRH then
478 * the RXE parsing will be off and will land in the middle of the KDETH
479 * or miss it entirely.
481 if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
482 SDMA_DBG(req, "User tried to pass in a GRH");
487 req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
489 * Calculate the initial TID offset based on the values of
490 * KDETH.OFFSET and KDETH.OM that are passed in.
492 req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
493 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
494 KDETH_OM_LARGE : KDETH_OM_SMALL);
495 trace_hfi1_sdma_user_initial_tidoffset(dd, uctxt->ctxt, fd->subctxt,
496 info.comp_idx, req->tidoffset);
499 /* Save all the IO vector structures */
500 for (i = 0; i < req->data_iovs; i++) {
501 req->iovs[i].offset = 0;
502 INIT_LIST_HEAD(&req->iovs[i].list);
503 memcpy(&req->iovs[i].iov,
505 sizeof(req->iovs[i].iov));
506 ret = pin_vector_pages(req, &req->iovs[i]);
511 req->data_len += req->iovs[i].iov.iov_len;
513 trace_hfi1_sdma_user_data_length(dd, uctxt->ctxt, fd->subctxt,
514 info.comp_idx, req->data_len);
515 if (pcount > req->info.npkts)
516 pcount = req->info.npkts;
519 * User space will provide the TID info only when the
520 * request type is EXPECTED. This is true even if there is
521 * only one packet in the request and the header is already
522 * setup. The reason for the singular TID case is that the
523 * driver needs to perform safety checks.
525 if (req_opcode(req->info.ctrl) == EXPECTED) {
526 u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
529 if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
535 * We have to copy all of the tids because they may vary
536 * in size and, therefore, the TID count might not be
537 * equal to the pkt count. However, there is no way to
538 * tell at this point.
540 tmp = memdup_user(iovec[idx].iov_base,
541 ntids * sizeof(*req->tids));
544 SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
554 dlid = be16_to_cpu(req->hdr.lrh[1]);
555 selector = dlid_to_selector(dlid);
556 selector += uctxt->ctxt + fd->subctxt;
557 req->sde = sdma_select_user_engine(dd, selector, vl);
559 if (!req->sde || !sdma_running(req->sde)) {
564 /* We don't need an AHG entry if the request contains only one packet */
565 if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
566 req->ahg_idx = sdma_ahg_alloc(req->sde);
568 set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
569 pq->state = SDMA_PKT_Q_ACTIVE;
570 /* Send the first N packets in the request to buy us some time */
571 ret = user_sdma_send_pkts(req, pcount);
572 if (unlikely(ret < 0 && ret != -EBUSY))
576 * This is a somewhat blocking send implementation.
577 * The driver will block the caller until all packets of the
578 * request have been submitted to the SDMA engine. However, it
579 * will not wait for send completions.
581 while (req->seqsubmitted != req->info.npkts) {
582 ret = user_sdma_send_pkts(req, pcount);
586 wait_event_interruptible_timeout(
588 (pq->state == SDMA_PKT_Q_ACTIVE),
590 SDMA_IOWAIT_TIMEOUT));
597 * If the submitted seqsubmitted == npkts, the completion routine
598 * controls the final state. If sequbmitted < npkts, wait for any
599 * outstanding packets to finish before cleaning up.
601 if (req->seqsubmitted < req->info.npkts) {
602 if (req->seqsubmitted)
603 wait_event(pq->busy.wait_dma,
604 (req->seqcomp == req->seqsubmitted - 1));
605 user_sdma_free_request(req, true);
607 set_comp_state(pq, cq, info.comp_idx, ERROR, ret);
612 static inline u32 compute_data_length(struct user_sdma_request *req,
613 struct user_sdma_txreq *tx)
616 * Determine the proper size of the packet data.
617 * The size of the data of the first packet is in the header
618 * template. However, it includes the header and ICRC, which need
620 * The minimum representable packet data length in a header is 4 bytes,
621 * therefore, when the data length request is less than 4 bytes, there's
622 * only one packet, and the packet data length is equal to that of the
623 * request data length.
624 * The size of the remaining packets is the minimum of the frag
625 * size (MTU) or remaining data in the request.
630 if (req->data_len < sizeof(u32))
633 len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
634 (sizeof(tx->hdr) - 4));
635 } else if (req_opcode(req->info.ctrl) == EXPECTED) {
636 u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
639 * Get the data length based on the remaining space in the
642 len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
643 /* If we've filled up the TID pair, move to the next one. */
644 if (unlikely(!len) && ++req->tididx < req->n_tids &&
645 req->tids[req->tididx]) {
646 tidlen = EXP_TID_GET(req->tids[req->tididx],
649 len = min_t(u32, tidlen, req->info.fragsize);
652 * Since the TID pairs map entire pages, make sure that we
653 * are not going to try to send more data that we have
656 len = min(len, req->data_len - req->sent);
658 len = min(req->data_len - req->sent, (u32)req->info.fragsize);
660 trace_hfi1_sdma_user_compute_length(req->pq->dd,
668 static inline u32 pad_len(u32 len)
670 if (len & (sizeof(u32) - 1))
671 len += sizeof(u32) - (len & (sizeof(u32) - 1));
675 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
677 /* (Size of complete header - size of PBC) + 4B ICRC + data length */
678 return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
681 static int user_sdma_txadd_ahg(struct user_sdma_request *req,
682 struct user_sdma_txreq *tx,
686 u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
687 u32 lrhlen = get_lrh_len(req->hdr, pad_len(datalen));
688 struct hfi1_user_sdma_pkt_q *pq = req->pq;
691 * Copy the request header into the tx header
692 * because the HW needs a cacheline-aligned
694 * This copy can be optimized out if the hdr
695 * member of user_sdma_request were also
698 memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
699 if (PBC2LRH(pbclen) != lrhlen) {
700 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
701 tx->hdr.pbc[0] = cpu_to_le16(pbclen);
703 ret = check_header_template(req, &tx->hdr, lrhlen, datalen);
706 ret = sdma_txinit_ahg(&tx->txreq, SDMA_TXREQ_F_AHG_COPY,
707 sizeof(tx->hdr) + datalen, req->ahg_idx,
708 0, NULL, 0, user_sdma_txreq_cb);
711 ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq, &tx->hdr, sizeof(tx->hdr));
713 sdma_txclean(pq->dd, &tx->txreq);
717 static int user_sdma_txadd(struct user_sdma_request *req,
718 struct user_sdma_txreq *tx,
719 struct user_sdma_iovec *iovec, u32 datalen,
720 u32 *queued_ptr, u32 *data_sent_ptr,
724 unsigned int pageidx, len;
725 unsigned long base, offset;
726 u64 iov_offset = *iov_offset_ptr;
727 u32 queued = *queued_ptr, data_sent = *data_sent_ptr;
728 struct hfi1_user_sdma_pkt_q *pq = req->pq;
730 base = (unsigned long)iovec->iov.iov_base;
731 offset = offset_in_page(base + iovec->offset + iov_offset);
732 pageidx = (((iovec->offset + iov_offset + base) - (base & PAGE_MASK)) >>
734 len = offset + req->info.fragsize > PAGE_SIZE ?
735 PAGE_SIZE - offset : req->info.fragsize;
736 len = min((datalen - queued), len);
737 ret = sdma_txadd_page(pq->dd, &tx->txreq, iovec->pages[pageidx],
740 SDMA_DBG(req, "SDMA txreq add page failed %d\n", ret);
746 if (unlikely(queued < datalen && pageidx == iovec->npages &&
747 req->iov_idx < req->data_iovs - 1)) {
748 iovec->offset += iov_offset;
749 iovec = &req->iovs[++req->iov_idx];
753 *queued_ptr = queued;
754 *data_sent_ptr = data_sent;
755 *iov_offset_ptr = iov_offset;
759 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts)
764 struct user_sdma_txreq *tx = NULL;
765 struct hfi1_user_sdma_pkt_q *pq = NULL;
766 struct user_sdma_iovec *iovec = NULL;
773 /* If tx completion has reported an error, we are done. */
774 if (READ_ONCE(req->has_error))
778 * Check if we might have sent the entire request already
780 if (unlikely(req->seqnum == req->info.npkts)) {
781 if (!list_empty(&req->txps))
786 if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
787 maxpkts = req->info.npkts - req->seqnum;
789 while (npkts < maxpkts) {
790 u32 datalen = 0, queued = 0, data_sent = 0;
794 * Check whether any of the completions have come back
795 * with errors. If so, we are not going to process any
796 * more packets from this request.
798 if (READ_ONCE(req->has_error))
801 tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
808 INIT_LIST_HEAD(&tx->list);
811 * For the last packet set the ACK request
812 * and disable header suppression.
814 if (req->seqnum == req->info.npkts - 1)
815 tx->flags |= (TXREQ_FLAGS_REQ_ACK |
816 TXREQ_FLAGS_REQ_DISABLE_SH);
819 * Calculate the payload size - this is min of the fragment
820 * (MTU) size or the remaining bytes in the request but only
821 * if we have payload data.
824 iovec = &req->iovs[req->iov_idx];
825 if (READ_ONCE(iovec->offset) == iovec->iov.iov_len) {
826 if (++req->iov_idx == req->data_iovs) {
830 iovec = &req->iovs[req->iov_idx];
831 WARN_ON(iovec->offset);
834 datalen = compute_data_length(req, tx);
837 * Disable header suppression for the payload <= 8DWS.
838 * If there is an uncorrectable error in the receive
839 * data FIFO when the received payload size is less than
840 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
841 * not reported.There is set RHF.EccErr if the header
846 "Request has data but pkt len is 0");
849 } else if (datalen <= 32) {
850 tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
854 if (req->ahg_idx >= 0) {
856 ret = user_sdma_txadd_ahg(req, tx, datalen);
862 changes = set_txreq_header_ahg(req, tx,
870 ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
871 datalen, user_sdma_txreq_cb);
875 * Modify the header for this packet. This only needs
876 * to be done if we are not going to use AHG. Otherwise,
877 * the HW will do it based on the changes we gave it
878 * during sdma_txinit_ahg().
880 ret = set_txreq_header(req, tx, datalen);
886 * If the request contains any data vectors, add up to
887 * fragsize bytes to the descriptor.
889 while (queued < datalen &&
890 (req->sent + data_sent) < req->data_len) {
891 ret = user_sdma_txadd(req, tx, iovec, datalen,
892 &queued, &data_sent, &iov_offset);
897 * The txreq was submitted successfully so we can update
900 req->koffset += datalen;
901 if (req_opcode(req->info.ctrl) == EXPECTED)
902 req->tidoffset += datalen;
903 req->sent += data_sent;
905 iovec->offset += iov_offset;
906 list_add_tail(&tx->txreq.list, &req->txps);
908 * It is important to increment this here as it is used to
909 * generate the BTH.PSN and, therefore, can't be bulk-updated
910 * outside of the loop.
912 tx->seqnum = req->seqnum++;
916 ret = sdma_send_txlist(req->sde,
917 iowait_get_ib_work(&pq->busy),
919 req->seqsubmitted += count;
920 if (req->seqsubmitted == req->info.npkts) {
922 * The txreq has already been submitted to the HW queue
923 * so we can free the AHG entry now. Corruption will not
924 * happen due to the sequential manner in which
925 * descriptors are processed.
927 if (req->ahg_idx >= 0)
928 sdma_ahg_free(req->sde, req->ahg_idx);
933 sdma_txclean(pq->dd, &tx->txreq);
935 kmem_cache_free(pq->txreq_cache, tx);
939 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
941 struct evict_data evict_data;
943 evict_data.cleared = 0;
944 evict_data.target = npages;
945 hfi1_mmu_rb_evict(pq->handler, &evict_data);
946 return evict_data.cleared;
949 static int pin_sdma_pages(struct user_sdma_request *req,
950 struct user_sdma_iovec *iovec,
951 struct sdma_mmu_node *node,
956 struct hfi1_user_sdma_pkt_q *pq = req->pq;
958 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
961 memcpy(pages, node->pages, node->npages * sizeof(*pages));
963 npages -= node->npages;
965 if (!hfi1_can_pin_pages(pq->dd, pq->mm,
966 atomic_read(&pq->n_locked), npages)) {
967 cleared = sdma_cache_evict(pq, npages);
968 if (cleared >= npages)
971 pinned = hfi1_acquire_user_pages(pq->mm,
972 ((unsigned long)iovec->iov.iov_base +
973 (node->npages * PAGE_SIZE)), npages, 0,
974 pages + node->npages);
979 if (pinned != npages) {
980 unpin_vector_pages(pq->mm, pages, node->npages, pinned);
984 node->rb.len = iovec->iov.iov_len;
986 atomic_add(pinned, &pq->n_locked);
990 static void unpin_sdma_pages(struct sdma_mmu_node *node)
993 unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
994 atomic_sub(node->npages, &node->pq->n_locked);
998 static int pin_vector_pages(struct user_sdma_request *req,
999 struct user_sdma_iovec *iovec)
1001 int ret = 0, pinned, npages;
1002 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1003 struct sdma_mmu_node *node = NULL;
1004 struct mmu_rb_node *rb_node;
1009 hfi1_mmu_rb_remove_unless_exact(pq->handler,
1011 iovec->iov.iov_base,
1012 iovec->iov.iov_len, &rb_node);
1014 node = container_of(rb_node, struct sdma_mmu_node, rb);
1016 atomic_inc(&node->refcount);
1017 iovec->pages = node->pages;
1018 iovec->npages = node->npages;
1025 node = kzalloc(sizeof(*node), GFP_KERNEL);
1029 node->rb.addr = (unsigned long)iovec->iov.iov_base;
1031 atomic_set(&node->refcount, 0);
1035 npages = num_user_pages((unsigned long)iov->iov_base, iov->iov_len);
1036 if (node->npages < npages) {
1037 pinned = pin_sdma_pages(req, iovec, node, npages);
1042 node->npages += pinned;
1043 npages = node->npages;
1045 iovec->pages = node->pages;
1046 iovec->npages = npages;
1049 ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1056 unpin_sdma_pages(node);
1061 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1062 unsigned start, unsigned npages)
1064 hfi1_release_user_pages(mm, pages + start, npages, false);
1068 static int check_header_template(struct user_sdma_request *req,
1069 struct hfi1_pkt_header *hdr, u32 lrhlen,
1073 * Perform safety checks for any type of packet:
1074 * - transfer size is multiple of 64bytes
1075 * - packet length is multiple of 4 bytes
1076 * - packet length is not larger than MTU size
1078 * These checks are only done for the first packet of the
1079 * transfer since the header is "given" to us by user space.
1080 * For the remainder of the packets we compute the values.
1082 if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1083 lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1086 if (req_opcode(req->info.ctrl) == EXPECTED) {
1088 * The header is checked only on the first packet. Furthermore,
1089 * we ensure that at least one TID entry is copied when the
1090 * request is submitted. Therefore, we don't have to verify that
1091 * tididx points to something sane.
1093 u32 tidval = req->tids[req->tididx],
1094 tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1095 tididx = EXP_TID_GET(tidval, IDX),
1096 tidctrl = EXP_TID_GET(tidval, CTRL),
1098 __le32 kval = hdr->kdeth.ver_tid_offset;
1100 tidoff = KDETH_GET(kval, OFFSET) *
1101 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1102 KDETH_OM_LARGE : KDETH_OM_SMALL);
1104 * Expected receive packets have the following
1105 * additional checks:
1106 * - offset is not larger than the TID size
1107 * - TIDCtrl values match between header and TID array
1108 * - TID indexes match between header and TID array
1110 if ((tidoff + datalen > tidlen) ||
1111 KDETH_GET(kval, TIDCTRL) != tidctrl ||
1112 KDETH_GET(kval, TID) != tididx)
1119 * Correctly set the BTH.PSN field based on type of
1120 * transfer - eager packets can just increment the PSN but
1121 * expected packets encode generation and sequence in the
1122 * BTH.PSN field so just incrementing will result in errors.
1124 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1126 u32 val = be32_to_cpu(bthpsn),
1127 mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1131 psn = (psn & ~HFI1_KDETH_BTH_SEQ_MASK) |
1132 ((psn + frags) & HFI1_KDETH_BTH_SEQ_MASK);
1138 static int set_txreq_header(struct user_sdma_request *req,
1139 struct user_sdma_txreq *tx, u32 datalen)
1141 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1142 struct hfi1_pkt_header *hdr = &tx->hdr;
1143 u8 omfactor; /* KDETH.OM */
1146 u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1148 /* Copy the header template to the request before modification */
1149 memcpy(hdr, &req->hdr, sizeof(*hdr));
1152 * Check if the PBC and LRH length are mismatched. If so
1153 * adjust both in the header.
1155 pbclen = le16_to_cpu(hdr->pbc[0]);
1156 if (PBC2LRH(pbclen) != lrhlen) {
1157 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1158 hdr->pbc[0] = cpu_to_le16(pbclen);
1159 hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1162 * This is the first packet in the sequence that has
1163 * a "static" size that can be used for the rest of
1164 * the packets (besides the last one).
1166 if (unlikely(req->seqnum == 2)) {
1168 * From this point on the lengths in both the
1169 * PBC and LRH are the same until the last
1171 * Adjust the template so we don't have to update
1174 req->hdr.pbc[0] = hdr->pbc[0];
1175 req->hdr.lrh[2] = hdr->lrh[2];
1179 * We only have to modify the header if this is not the
1180 * first packet in the request. Otherwise, we use the
1181 * header given to us.
1183 if (unlikely(!req->seqnum)) {
1184 ret = check_header_template(req, hdr, lrhlen, datalen);
1190 hdr->bth[2] = cpu_to_be32(
1191 set_pkt_bth_psn(hdr->bth[2],
1192 (req_opcode(req->info.ctrl) == EXPECTED),
1195 /* Set ACK request on last packet */
1196 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1197 hdr->bth[2] |= cpu_to_be32(1UL << 31);
1199 /* Set the new offset */
1200 hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1201 /* Expected packets have to fill in the new TID information */
1202 if (req_opcode(req->info.ctrl) == EXPECTED) {
1203 tidval = req->tids[req->tididx];
1205 * If the offset puts us at the end of the current TID,
1206 * advance everything.
1208 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1212 * Since we don't copy all the TIDs, all at once,
1213 * we have to check again.
1215 if (++req->tididx > req->n_tids - 1 ||
1216 !req->tids[req->tididx]) {
1219 tidval = req->tids[req->tididx];
1221 omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1222 KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
1223 KDETH_OM_SMALL_SHIFT;
1224 /* Set KDETH.TIDCtrl based on value for this TID. */
1225 KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1226 EXP_TID_GET(tidval, CTRL));
1227 /* Set KDETH.TID based on value for this TID */
1228 KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1229 EXP_TID_GET(tidval, IDX));
1230 /* Clear KDETH.SH when DISABLE_SH flag is set */
1231 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1232 KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1234 * Set the KDETH.OFFSET and KDETH.OM based on size of
1237 trace_hfi1_sdma_user_tid_info(
1238 pq->dd, pq->ctxt, pq->subctxt, req->info.comp_idx,
1239 req->tidoffset, req->tidoffset >> omfactor,
1240 omfactor != KDETH_OM_SMALL_SHIFT);
1241 KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1242 req->tidoffset >> omfactor);
1243 KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1244 omfactor != KDETH_OM_SMALL_SHIFT);
1247 trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1248 req->info.comp_idx, hdr, tidval);
1249 return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1252 static int set_txreq_header_ahg(struct user_sdma_request *req,
1253 struct user_sdma_txreq *tx, u32 datalen)
1255 u32 ahg[AHG_KDETH_ARRAY_SIZE];
1257 u8 omfactor; /* KDETH.OM */
1258 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1259 struct hfi1_pkt_header *hdr = &req->hdr;
1260 u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1261 u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1262 size_t array_size = ARRAY_SIZE(ahg);
1264 if (PBC2LRH(pbclen) != lrhlen) {
1265 /* PBC.PbcLengthDWs */
1266 idx = ahg_header_set(ahg, idx, array_size, 0, 0, 12,
1267 (__force u16)cpu_to_le16(LRH2PBC(lrhlen)));
1270 /* LRH.PktLen (we need the full 16 bits due to byte swap) */
1271 idx = ahg_header_set(ahg, idx, array_size, 3, 0, 16,
1272 (__force u16)cpu_to_be16(lrhlen >> 2));
1278 * Do the common updates
1280 /* BTH.PSN and BTH.A */
1281 val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1282 (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1283 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1285 idx = ahg_header_set(ahg, idx, array_size, 6, 0, 16,
1286 (__force u16)cpu_to_be16(val32 >> 16));
1289 idx = ahg_header_set(ahg, idx, array_size, 6, 16, 16,
1290 (__force u16)cpu_to_be16(val32 & 0xffff));
1294 idx = ahg_header_set(ahg, idx, array_size, 15, 0, 16,
1295 (__force u16)cpu_to_le16(req->koffset & 0xffff));
1298 idx = ahg_header_set(ahg, idx, array_size, 15, 16, 16,
1299 (__force u16)cpu_to_le16(req->koffset >> 16));
1302 if (req_opcode(req->info.ctrl) == EXPECTED) {
1305 tidval = req->tids[req->tididx];
1308 * If the offset puts us at the end of the current TID,
1309 * advance everything.
1311 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1315 * Since we don't copy all the TIDs, all at once,
1316 * we have to check again.
1318 if (++req->tididx > req->n_tids - 1 ||
1319 !req->tids[req->tididx])
1321 tidval = req->tids[req->tididx];
1323 omfactor = ((EXP_TID_GET(tidval, LEN) *
1325 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1326 KDETH_OM_SMALL_SHIFT;
1327 /* KDETH.OM and KDETH.OFFSET (TID) */
1328 idx = ahg_header_set(
1329 ahg, idx, array_size, 7, 0, 16,
1330 ((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1331 ((req->tidoffset >> omfactor)
1335 /* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1336 val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1337 (EXP_TID_GET(tidval, IDX) & 0x3ff));
1339 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1340 val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1342 AHG_KDETH_INTR_SHIFT));
1344 val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1345 cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1346 cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1348 AHG_KDETH_INTR_SHIFT));
1351 idx = ahg_header_set(ahg, idx, array_size,
1352 7, 16, 14, (__force u16)val);
1357 trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1358 req->info.comp_idx, req->sde->this_idx,
1359 req->ahg_idx, ahg, idx, tidval);
1360 sdma_txinit_ahg(&tx->txreq,
1361 SDMA_TXREQ_F_USE_AHG,
1362 datalen, req->ahg_idx, idx,
1363 ahg, sizeof(req->hdr),
1364 user_sdma_txreq_cb);
1370 * user_sdma_txreq_cb() - SDMA tx request completion callback.
1371 * @txreq: valid sdma tx request
1372 * @status: success/failure of request
1374 * Called when the SDMA progress state machine gets notification that
1375 * the SDMA descriptors for this tx request have been processed by the
1376 * DMA engine. Called in interrupt context.
1377 * Only do work on completed sequences.
1379 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1381 struct user_sdma_txreq *tx =
1382 container_of(txreq, struct user_sdma_txreq, txreq);
1383 struct user_sdma_request *req;
1384 struct hfi1_user_sdma_pkt_q *pq;
1385 struct hfi1_user_sdma_comp_q *cq;
1386 enum hfi1_sdma_comp_state state = COMPLETE;
1395 if (status != SDMA_TXREQ_S_OK) {
1396 SDMA_DBG(req, "SDMA completion with error %d",
1398 WRITE_ONCE(req->has_error, 1);
1402 req->seqcomp = tx->seqnum;
1403 kmem_cache_free(pq->txreq_cache, tx);
1405 /* sequence isn't complete? We are done */
1406 if (req->seqcomp != req->info.npkts - 1)
1409 user_sdma_free_request(req, false);
1410 set_comp_state(pq, cq, req->info.comp_idx, state, status);
1414 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1416 if (atomic_dec_and_test(&pq->n_reqs))
1420 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1424 if (!list_empty(&req->txps)) {
1425 struct sdma_txreq *t, *p;
1427 list_for_each_entry_safe(t, p, &req->txps, list) {
1428 struct user_sdma_txreq *tx =
1429 container_of(t, struct user_sdma_txreq, txreq);
1430 list_del_init(&t->list);
1431 sdma_txclean(req->pq->dd, t);
1432 kmem_cache_free(req->pq->txreq_cache, tx);
1436 for (i = 0; i < req->data_iovs; i++) {
1437 struct sdma_mmu_node *node = req->iovs[i].node;
1442 req->iovs[i].node = NULL;
1445 hfi1_mmu_rb_remove(req->pq->handler,
1448 atomic_dec(&node->refcount);
1452 clear_bit(req->info.comp_idx, req->pq->req_in_use);
1455 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1456 struct hfi1_user_sdma_comp_q *cq,
1457 u16 idx, enum hfi1_sdma_comp_state state,
1461 cq->comps[idx].errcode = -ret;
1462 smp_wmb(); /* make sure errcode is visible first */
1463 cq->comps[idx].status = state;
1464 trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1468 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1471 return (bool)(node->addr == addr);
1474 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1476 struct sdma_mmu_node *node =
1477 container_of(mnode, struct sdma_mmu_node, rb);
1479 atomic_inc(&node->refcount);
1484 * Return 1 to remove the node from the rb tree and call the remove op.
1486 * Called with the rb tree lock held.
1488 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1489 void *evict_arg, bool *stop)
1491 struct sdma_mmu_node *node =
1492 container_of(mnode, struct sdma_mmu_node, rb);
1493 struct evict_data *evict_data = evict_arg;
1495 /* is this node still being used? */
1496 if (atomic_read(&node->refcount))
1497 return 0; /* keep this node */
1499 /* this node will be evicted, add its pages to our count */
1500 evict_data->cleared += node->npages;
1502 /* have enough pages been cleared? */
1503 if (evict_data->cleared >= evict_data->target)
1506 return 1; /* remove this node */
1509 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1511 struct sdma_mmu_node *node =
1512 container_of(mnode, struct sdma_mmu_node, rb);
1514 unpin_sdma_pages(node);
1518 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1520 struct sdma_mmu_node *node =
1521 container_of(mnode, struct sdma_mmu_node, rb);
1523 if (!atomic_read(&node->refcount))