2 * Copyright(c) 2015, 2016 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>
66 #include "user_sdma.h"
67 #include "verbs.h" /* for the headers */
68 #include "common.h" /* for struct hfi1_tid_info */
72 static uint hfi1_sdma_comp_ring_size = 128;
73 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
74 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
76 /* The maximum number of Data io vectors per message/request */
77 #define MAX_VECTORS_PER_REQ 8
79 * Maximum number of packet to send from each message/request
80 * before moving to the next one.
82 #define MAX_PKTS_PER_QUEUE 16
84 #define num_pages(x) (1 + ((((x) - 1) & PAGE_MASK) >> PAGE_SHIFT))
86 #define req_opcode(x) \
87 (((x) >> HFI1_SDMA_REQ_OPCODE_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
88 #define req_version(x) \
89 (((x) >> HFI1_SDMA_REQ_VERSION_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
90 #define req_iovcnt(x) \
91 (((x) >> HFI1_SDMA_REQ_IOVCNT_SHIFT) & HFI1_SDMA_REQ_IOVCNT_MASK)
93 /* Number of BTH.PSN bits used for sequence number in expected rcvs */
94 #define BTH_SEQ_MASK 0x7ffull
97 * Define fields in the KDETH header so we can update the header
100 #define KDETH_OFFSET_SHIFT 0
101 #define KDETH_OFFSET_MASK 0x7fff
102 #define KDETH_OM_SHIFT 15
103 #define KDETH_OM_MASK 0x1
104 #define KDETH_TID_SHIFT 16
105 #define KDETH_TID_MASK 0x3ff
106 #define KDETH_TIDCTRL_SHIFT 26
107 #define KDETH_TIDCTRL_MASK 0x3
108 #define KDETH_INTR_SHIFT 28
109 #define KDETH_INTR_MASK 0x1
110 #define KDETH_SH_SHIFT 29
111 #define KDETH_SH_MASK 0x1
112 #define KDETH_HCRC_UPPER_SHIFT 16
113 #define KDETH_HCRC_UPPER_MASK 0xff
114 #define KDETH_HCRC_LOWER_SHIFT 24
115 #define KDETH_HCRC_LOWER_MASK 0xff
117 #define PBC2LRH(x) ((((x) & 0xfff) << 2) - 4)
118 #define LRH2PBC(x) ((((x) >> 2) + 1) & 0xfff)
120 #define KDETH_GET(val, field) \
121 (((le32_to_cpu((val))) >> KDETH_##field##_SHIFT) & KDETH_##field##_MASK)
122 #define KDETH_SET(dw, field, val) do { \
123 u32 dwval = le32_to_cpu(dw); \
124 dwval &= ~(KDETH_##field##_MASK << KDETH_##field##_SHIFT); \
125 dwval |= (((val) & KDETH_##field##_MASK) << \
126 KDETH_##field##_SHIFT); \
127 dw = cpu_to_le32(dwval); \
130 #define AHG_HEADER_SET(arr, idx, dw, bit, width, value) \
132 if ((idx) < ARRAY_SIZE((arr))) \
133 (arr)[(idx++)] = sdma_build_ahg_descriptor( \
134 (__force u16)(value), (dw), (bit), \
140 /* KDETH OM multipliers and switch over point */
141 #define KDETH_OM_SMALL 4
142 #define KDETH_OM_LARGE 64
143 #define KDETH_OM_MAX_SIZE (1 << ((KDETH_OM_LARGE / KDETH_OM_SMALL) + 1))
145 /* Last packet in the request */
146 #define TXREQ_FLAGS_REQ_LAST_PKT BIT(0)
148 #define SDMA_REQ_IN_USE 0
149 #define SDMA_REQ_FOR_THREAD 1
150 #define SDMA_REQ_SEND_DONE 2
151 #define SDMA_REQ_HAVE_AHG 3
152 #define SDMA_REQ_HAS_ERROR 4
153 #define SDMA_REQ_DONE_ERROR 5
155 #define SDMA_PKT_Q_INACTIVE BIT(0)
156 #define SDMA_PKT_Q_ACTIVE BIT(1)
157 #define SDMA_PKT_Q_DEFERRED BIT(2)
160 * Maximum retry attempts to submit a TX request
161 * before putting the process to sleep.
163 #define MAX_DEFER_RETRY_COUNT 1
165 static unsigned initial_pkt_count = 8;
167 #define SDMA_IOWAIT_TIMEOUT 1000 /* in milliseconds */
169 struct sdma_mmu_node;
171 struct user_sdma_iovec {
172 struct list_head list;
174 /* number of pages in this vector */
176 /* array of pinned pages for this vector */
179 * offset into the virtual address space of the vector at
180 * which we last left off.
183 struct sdma_mmu_node *node;
186 #define SDMA_CACHE_NODE_EVICT 0
188 struct sdma_mmu_node {
189 struct mmu_rb_node rb;
190 struct list_head list;
191 struct hfi1_user_sdma_pkt_q *pq;
198 struct user_sdma_request {
199 struct sdma_req_info info;
200 struct hfi1_user_sdma_pkt_q *pq;
201 struct hfi1_user_sdma_comp_q *cq;
202 /* This is the original header from user space */
203 struct hfi1_pkt_header hdr;
205 * Pointer to the SDMA engine for this request.
206 * Since different request could be on different VLs,
207 * each request will need it's own engine pointer.
209 struct sdma_engine *sde;
213 * KDETH.Offset (Eager) field
214 * We need to remember the initial value so the headers
215 * can be updated properly.
219 * KDETH.OFFSET (TID) field
220 * The offset can cover multiple packets, depending on the
221 * size of the TID entry.
226 * Remember this because the header template always sets it
231 * We copy the iovs for this request (based on
232 * info.iovcnt). These are only the data vectors
235 /* total length of the data in the request */
237 /* progress index moving along the iovs array */
239 struct user_sdma_iovec iovs[MAX_VECTORS_PER_REQ];
240 /* number of elements copied to the tids array */
242 /* TID array values copied from the tid_iov vector */
249 struct list_head txps;
251 /* status of the last txreq completed */
256 * A single txreq could span up to 3 physical pages when the MTU
257 * is sufficiently large (> 4K). Each of the IOV pointers also
258 * needs it's own set of flags so the vector has been handled
259 * independently of each other.
261 struct user_sdma_txreq {
262 /* Packet header for the txreq */
263 struct hfi1_pkt_header hdr;
264 struct sdma_txreq txreq;
265 struct list_head list;
266 struct user_sdma_request *req;
272 #define SDMA_DBG(req, fmt, ...) \
273 hfi1_cdbg(SDMA, "[%u:%u:%u:%u] " fmt, (req)->pq->dd->unit, \
274 (req)->pq->ctxt, (req)->pq->subctxt, (req)->info.comp_idx, \
276 #define SDMA_Q_DBG(pq, fmt, ...) \
277 hfi1_cdbg(SDMA, "[%u:%u:%u] " fmt, (pq)->dd->unit, (pq)->ctxt, \
278 (pq)->subctxt, ##__VA_ARGS__)
280 static int user_sdma_send_pkts(struct user_sdma_request *, unsigned);
281 static int num_user_pages(const struct iovec *);
282 static void user_sdma_txreq_cb(struct sdma_txreq *, int);
283 static inline void pq_update(struct hfi1_user_sdma_pkt_q *);
284 static void user_sdma_free_request(struct user_sdma_request *, bool);
285 static int pin_vector_pages(struct user_sdma_request *,
286 struct user_sdma_iovec *);
287 static void unpin_vector_pages(struct mm_struct *, struct page **, unsigned,
289 static int check_header_template(struct user_sdma_request *,
290 struct hfi1_pkt_header *, u32, u32);
291 static int set_txreq_header(struct user_sdma_request *,
292 struct user_sdma_txreq *, u32);
293 static int set_txreq_header_ahg(struct user_sdma_request *,
294 struct user_sdma_txreq *, u32);
295 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *,
296 struct hfi1_user_sdma_comp_q *,
297 u16, enum hfi1_sdma_comp_state, int);
298 static inline u32 set_pkt_bth_psn(__be32, u8, u32);
299 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
301 static int defer_packet_queue(
302 struct sdma_engine *,
306 static void activate_packet_queue(struct iowait *, int);
307 static bool sdma_rb_filter(struct mmu_rb_node *, unsigned long, unsigned long);
308 static int sdma_rb_insert(struct rb_root *, struct mmu_rb_node *);
309 static void sdma_rb_remove(struct rb_root *, struct mmu_rb_node *,
311 static int sdma_rb_invalidate(struct rb_root *, struct mmu_rb_node *);
313 static struct mmu_rb_ops sdma_rb_ops = {
314 .filter = sdma_rb_filter,
315 .insert = sdma_rb_insert,
316 .remove = sdma_rb_remove,
317 .invalidate = sdma_rb_invalidate
320 static int defer_packet_queue(
321 struct sdma_engine *sde,
323 struct sdma_txreq *txreq,
326 struct hfi1_user_sdma_pkt_q *pq =
327 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
328 struct hfi1_ibdev *dev = &pq->dd->verbs_dev;
329 struct user_sdma_txreq *tx =
330 container_of(txreq, struct user_sdma_txreq, txreq);
332 if (sdma_progress(sde, seq, txreq)) {
333 if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
337 * We are assuming that if the list is enqueued somewhere, it
338 * is to the dmawait list since that is the only place where
339 * it is supposed to be enqueued.
341 xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
342 write_seqlock(&dev->iowait_lock);
343 if (list_empty(&pq->busy.list))
344 list_add_tail(&pq->busy.list, &sde->dmawait);
345 write_sequnlock(&dev->iowait_lock);
351 static void activate_packet_queue(struct iowait *wait, int reason)
353 struct hfi1_user_sdma_pkt_q *pq =
354 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
355 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
356 wake_up(&wait->wait_dma);
359 static void sdma_kmem_cache_ctor(void *obj)
361 struct user_sdma_txreq *tx = obj;
363 memset(tx, 0, sizeof(*tx));
366 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt, struct file *fp)
368 struct hfi1_filedata *fd;
372 struct hfi1_devdata *dd;
373 struct hfi1_user_sdma_comp_q *cq;
374 struct hfi1_user_sdma_pkt_q *pq;
382 fd = fp->private_data;
384 if (!hfi1_sdma_comp_ring_size) {
391 pq = kzalloc(sizeof(*pq), GFP_KERNEL);
395 memsize = sizeof(*pq->reqs) * hfi1_sdma_comp_ring_size;
396 pq->reqs = kzalloc(memsize, GFP_KERNEL);
400 INIT_LIST_HEAD(&pq->list);
402 pq->ctxt = uctxt->ctxt;
403 pq->subctxt = fd->subctxt;
404 pq->n_max_reqs = hfi1_sdma_comp_ring_size;
405 pq->state = SDMA_PKT_Q_INACTIVE;
406 atomic_set(&pq->n_reqs, 0);
407 init_waitqueue_head(&pq->wait);
408 pq->sdma_rb_root = RB_ROOT;
409 INIT_LIST_HEAD(&pq->evict);
410 spin_lock_init(&pq->evict_lock);
412 iowait_init(&pq->busy, 0, NULL, defer_packet_queue,
413 activate_packet_queue, NULL);
415 snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
417 pq->txreq_cache = kmem_cache_create(buf,
418 sizeof(struct user_sdma_txreq),
421 sdma_kmem_cache_ctor);
422 if (!pq->txreq_cache) {
423 dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
428 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
432 memsize = PAGE_ALIGN(sizeof(*cq->comps) * hfi1_sdma_comp_ring_size);
433 cq->comps = vmalloc_user(memsize);
437 cq->nentries = hfi1_sdma_comp_ring_size;
440 ret = hfi1_mmu_rb_register(&pq->sdma_rb_root, &sdma_rb_ops);
442 dd_dev_err(dd, "Failed to register with MMU %d", ret);
446 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
447 list_add(&pq->list, &uctxt->sdma_queues);
448 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
454 kmem_cache_destroy(pq->txreq_cache);
466 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd)
468 struct hfi1_ctxtdata *uctxt = fd->uctxt;
469 struct hfi1_user_sdma_pkt_q *pq;
472 hfi1_cdbg(SDMA, "[%u:%u:%u] Freeing user SDMA queues", uctxt->dd->unit,
473 uctxt->ctxt, fd->subctxt);
475 hfi1_mmu_rb_unregister(&pq->sdma_rb_root);
477 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
478 if (!list_empty(&pq->list))
479 list_del_init(&pq->list);
480 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
481 iowait_sdma_drain(&pq->busy);
482 /* Wait until all requests have been freed. */
483 wait_event_interruptible(
485 (ACCESS_ONCE(pq->state) == SDMA_PKT_Q_INACTIVE));
487 kmem_cache_destroy(pq->txreq_cache);
492 vfree(fd->cq->comps);
499 int hfi1_user_sdma_process_request(struct file *fp, struct iovec *iovec,
500 unsigned long dim, unsigned long *count)
503 struct hfi1_filedata *fd = fp->private_data;
504 struct hfi1_ctxtdata *uctxt = fd->uctxt;
505 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
506 struct hfi1_user_sdma_comp_q *cq = fd->cq;
507 struct hfi1_devdata *dd = pq->dd;
508 unsigned long idx = 0;
509 u8 pcount = initial_pkt_count;
510 struct sdma_req_info info;
511 struct user_sdma_request *req;
515 if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
518 "[%u:%u:%u] First vector not big enough for header %lu/%lu",
519 dd->unit, uctxt->ctxt, fd->subctxt,
520 iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
523 ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
525 hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
526 dd->unit, uctxt->ctxt, fd->subctxt, ret);
530 trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
532 if (cq->comps[info.comp_idx].status == QUEUED ||
533 test_bit(SDMA_REQ_IN_USE, &pq->reqs[info.comp_idx].flags)) {
534 hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in QUEUED state",
535 dd->unit, uctxt->ctxt, fd->subctxt,
539 if (!info.fragsize) {
541 "[%u:%u:%u:%u] Request does not specify fragsize",
542 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
546 * We've done all the safety checks that we can up to this point,
547 * "allocate" the request entry.
549 hfi1_cdbg(SDMA, "[%u:%u:%u] Using req/comp entry %u\n", dd->unit,
550 uctxt->ctxt, fd->subctxt, info.comp_idx);
551 req = pq->reqs + info.comp_idx;
552 memset(req, 0, sizeof(*req));
553 /* Mark the request as IN_USE before we start filling it in. */
554 set_bit(SDMA_REQ_IN_USE, &req->flags);
555 req->data_iovs = req_iovcnt(info.ctrl) - 1;
559 INIT_LIST_HEAD(&req->txps);
561 memcpy(&req->info, &info, sizeof(info));
563 if (req_opcode(info.ctrl) == EXPECTED)
566 if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
567 SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
568 MAX_VECTORS_PER_REQ);
571 /* Copy the header from the user buffer */
572 ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
575 SDMA_DBG(req, "Failed to copy header template (%d)", ret);
580 /* If Static rate control is not enabled, sanitize the header. */
581 if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
584 /* Validate the opcode. Do not trust packets from user space blindly. */
585 opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
586 if ((opcode & USER_OPCODE_CHECK_MASK) !=
587 USER_OPCODE_CHECK_VAL) {
588 SDMA_DBG(req, "Invalid opcode (%d)", opcode);
593 * Validate the vl. Do not trust packets from user space blindly.
594 * VL comes from PBC, SC comes from LRH, and the VL needs to
595 * match the SC look up.
597 vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
598 sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
599 (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
600 if (vl >= dd->pport->vls_operational ||
601 vl != sc_to_vlt(dd, sc)) {
602 SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
607 /* Checking P_KEY for requests from user-space */
608 if (egress_pkey_check(dd->pport, req->hdr.lrh, req->hdr.bth, sc,
609 PKEY_CHECK_INVALID)) {
615 * Also should check the BTH.lnh. If it says the next header is GRH then
616 * the RXE parsing will be off and will land in the middle of the KDETH
617 * or miss it entirely.
619 if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
620 SDMA_DBG(req, "User tried to pass in a GRH");
625 req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
627 * Calculate the initial TID offset based on the values of
628 * KDETH.OFFSET and KDETH.OM that are passed in.
630 req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
631 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
632 KDETH_OM_LARGE : KDETH_OM_SMALL);
633 SDMA_DBG(req, "Initial TID offset %u", req->tidoffset);
636 /* Save all the IO vector structures */
637 while (i < req->data_iovs) {
638 INIT_LIST_HEAD(&req->iovs[i].list);
639 memcpy(&req->iovs[i].iov, iovec + idx++, sizeof(struct iovec));
640 ret = pin_vector_pages(req, &req->iovs[i]);
645 req->data_len += req->iovs[i++].iov.iov_len;
647 SDMA_DBG(req, "total data length %u", req->data_len);
649 if (pcount > req->info.npkts)
650 pcount = req->info.npkts;
653 * User space will provide the TID info only when the
654 * request type is EXPECTED. This is true even if there is
655 * only one packet in the request and the header is already
656 * setup. The reason for the singular TID case is that the
657 * driver needs to perform safety checks.
659 if (req_opcode(req->info.ctrl) == EXPECTED) {
660 u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
662 if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
666 req->tids = kcalloc(ntids, sizeof(*req->tids), GFP_KERNEL);
672 * We have to copy all of the tids because they may vary
673 * in size and, therefore, the TID count might not be
674 * equal to the pkt count. However, there is no way to
675 * tell at this point.
677 ret = copy_from_user(req->tids, iovec[idx].iov_base,
678 ntids * sizeof(*req->tids));
680 SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
689 /* Have to select the engine */
690 req->sde = sdma_select_engine_vl(dd,
691 (u32)(uctxt->ctxt + fd->subctxt),
693 if (!req->sde || !sdma_running(req->sde)) {
698 /* We don't need an AHG entry if the request contains only one packet */
699 if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG)) {
700 int ahg = sdma_ahg_alloc(req->sde);
702 if (likely(ahg >= 0)) {
703 req->ahg_idx = (u8)ahg;
704 set_bit(SDMA_REQ_HAVE_AHG, &req->flags);
708 set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
709 atomic_inc(&pq->n_reqs);
711 /* Send the first N packets in the request to buy us some time */
712 ret = user_sdma_send_pkts(req, pcount);
713 if (unlikely(ret < 0 && ret != -EBUSY)) {
719 * It is possible that the SDMA engine would have processed all the
720 * submitted packets by the time we get here. Therefore, only set
721 * packet queue state to ACTIVE if there are still uncompleted
724 if (atomic_read(&pq->n_reqs))
725 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
728 * This is a somewhat blocking send implementation.
729 * The driver will block the caller until all packets of the
730 * request have been submitted to the SDMA engine. However, it
731 * will not wait for send completions.
733 while (!test_bit(SDMA_REQ_SEND_DONE, &req->flags)) {
734 ret = user_sdma_send_pkts(req, pcount);
738 set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
739 if (ACCESS_ONCE(req->seqcomp) ==
740 req->seqsubmitted - 1)
744 wait_event_interruptible_timeout(
746 (pq->state == SDMA_PKT_Q_ACTIVE),
748 SDMA_IOWAIT_TIMEOUT));
754 user_sdma_free_request(req, true);
757 set_comp_state(pq, cq, info.comp_idx, ERROR, req->status);
761 static inline u32 compute_data_length(struct user_sdma_request *req,
762 struct user_sdma_txreq *tx)
765 * Determine the proper size of the packet data.
766 * The size of the data of the first packet is in the header
767 * template. However, it includes the header and ICRC, which need
769 * The size of the remaining packets is the minimum of the frag
770 * size (MTU) or remaining data in the request.
775 len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
776 (sizeof(tx->hdr) - 4));
777 } else if (req_opcode(req->info.ctrl) == EXPECTED) {
778 u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
781 * Get the data length based on the remaining space in the
784 len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
785 /* If we've filled up the TID pair, move to the next one. */
786 if (unlikely(!len) && ++req->tididx < req->n_tids &&
787 req->tids[req->tididx]) {
788 tidlen = EXP_TID_GET(req->tids[req->tididx],
791 len = min_t(u32, tidlen, req->info.fragsize);
794 * Since the TID pairs map entire pages, make sure that we
795 * are not going to try to send more data that we have
798 len = min(len, req->data_len - req->sent);
800 len = min(req->data_len - req->sent, (u32)req->info.fragsize);
802 SDMA_DBG(req, "Data Length = %u", len);
806 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
808 /* (Size of complete header - size of PBC) + 4B ICRC + data length */
809 return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
812 static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
816 struct user_sdma_txreq *tx = NULL;
817 struct hfi1_user_sdma_pkt_q *pq = NULL;
818 struct user_sdma_iovec *iovec = NULL;
825 /* If tx completion has reported an error, we are done. */
826 if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) {
827 set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
832 * Check if we might have sent the entire request already
834 if (unlikely(req->seqnum == req->info.npkts)) {
835 if (!list_empty(&req->txps))
840 if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
841 maxpkts = req->info.npkts - req->seqnum;
843 while (npkts < maxpkts) {
844 u32 datalen = 0, queued = 0, data_sent = 0;
848 * Check whether any of the completions have come back
849 * with errors. If so, we are not going to process any
850 * more packets from this request.
852 if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) {
853 set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
857 tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
864 INIT_LIST_HEAD(&tx->list);
866 if (req->seqnum == req->info.npkts - 1)
867 tx->flags |= TXREQ_FLAGS_REQ_LAST_PKT;
870 * Calculate the payload size - this is min of the fragment
871 * (MTU) size or the remaining bytes in the request but only
872 * if we have payload data.
875 iovec = &req->iovs[req->iov_idx];
876 if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
877 if (++req->iov_idx == req->data_iovs) {
881 iovec = &req->iovs[req->iov_idx];
882 WARN_ON(iovec->offset);
885 datalen = compute_data_length(req, tx);
888 "Request has data but pkt len is 0");
894 if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags)) {
896 u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
897 u32 lrhlen = get_lrh_len(req->hdr, datalen);
899 * Copy the request header into the tx header
900 * because the HW needs a cacheline-aligned
902 * This copy can be optimized out if the hdr
903 * member of user_sdma_request were also
906 memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
907 if (PBC2LRH(pbclen) != lrhlen) {
908 pbclen = (pbclen & 0xf000) |
910 tx->hdr.pbc[0] = cpu_to_le16(pbclen);
912 ret = sdma_txinit_ahg(&tx->txreq,
913 SDMA_TXREQ_F_AHG_COPY,
914 sizeof(tx->hdr) + datalen,
915 req->ahg_idx, 0, NULL, 0,
919 ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq,
927 changes = set_txreq_header_ahg(req, tx,
931 sdma_txinit_ahg(&tx->txreq,
932 SDMA_TXREQ_F_USE_AHG,
933 datalen, req->ahg_idx, changes,
934 req->ahg, sizeof(req->hdr),
938 ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
939 datalen, user_sdma_txreq_cb);
943 * Modify the header for this packet. This only needs
944 * to be done if we are not going to use AHG. Otherwise,
945 * the HW will do it based on the changes we gave it
946 * during sdma_txinit_ahg().
948 ret = set_txreq_header(req, tx, datalen);
954 * If the request contains any data vectors, add up to
955 * fragsize bytes to the descriptor.
957 while (queued < datalen &&
958 (req->sent + data_sent) < req->data_len) {
959 unsigned long base, offset;
960 unsigned pageidx, len;
962 base = (unsigned long)iovec->iov.iov_base;
963 offset = offset_in_page(base + iovec->offset +
965 pageidx = (((iovec->offset + iov_offset +
966 base) - (base & PAGE_MASK)) >> PAGE_SHIFT);
967 len = offset + req->info.fragsize > PAGE_SIZE ?
968 PAGE_SIZE - offset : req->info.fragsize;
969 len = min((datalen - queued), len);
970 ret = sdma_txadd_page(pq->dd, &tx->txreq,
971 iovec->pages[pageidx],
974 SDMA_DBG(req, "SDMA txreq add page failed %d\n",
981 if (unlikely(queued < datalen &&
982 pageidx == iovec->npages &&
983 req->iov_idx < req->data_iovs - 1)) {
984 iovec->offset += iov_offset;
985 iovec = &req->iovs[++req->iov_idx];
990 * The txreq was submitted successfully so we can update
993 req->koffset += datalen;
994 if (req_opcode(req->info.ctrl) == EXPECTED)
995 req->tidoffset += datalen;
996 req->sent += data_sent;
998 iovec->offset += iov_offset;
999 list_add_tail(&tx->txreq.list, &req->txps);
1001 * It is important to increment this here as it is used to
1002 * generate the BTH.PSN and, therefore, can't be bulk-updated
1003 * outside of the loop.
1005 tx->seqnum = req->seqnum++;
1009 ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps);
1010 if (list_empty(&req->txps)) {
1011 req->seqsubmitted = req->seqnum;
1012 if (req->seqnum == req->info.npkts) {
1013 set_bit(SDMA_REQ_SEND_DONE, &req->flags);
1015 * The txreq has already been submitted to the HW queue
1016 * so we can free the AHG entry now. Corruption will not
1017 * happen due to the sequential manner in which
1018 * descriptors are processed.
1020 if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags))
1021 sdma_ahg_free(req->sde, req->ahg_idx);
1023 } else if (ret > 0) {
1024 req->seqsubmitted += ret;
1030 sdma_txclean(pq->dd, &tx->txreq);
1032 kmem_cache_free(pq->txreq_cache, tx);
1037 * How many pages in this iovec element?
1039 static inline int num_user_pages(const struct iovec *iov)
1041 const unsigned long addr = (unsigned long)iov->iov_base;
1042 const unsigned long len = iov->iov_len;
1043 const unsigned long spage = addr & PAGE_MASK;
1044 const unsigned long epage = (addr + len - 1) & PAGE_MASK;
1046 return 1 + ((epage - spage) >> PAGE_SHIFT);
1049 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
1052 struct sdma_mmu_node *node, *ptr;
1053 struct list_head to_evict = LIST_HEAD_INIT(to_evict);
1055 spin_lock(&pq->evict_lock);
1056 list_for_each_entry_safe_reverse(node, ptr, &pq->evict, list) {
1057 /* Make sure that no one is still using the node. */
1058 if (!atomic_read(&node->refcount)) {
1059 set_bit(SDMA_CACHE_NODE_EVICT, &node->flags);
1060 list_del_init(&node->list);
1061 list_add(&node->list, &to_evict);
1062 cleared += node->npages;
1063 if (cleared >= npages)
1067 spin_unlock(&pq->evict_lock);
1069 list_for_each_entry_safe(node, ptr, &to_evict, list)
1070 hfi1_mmu_rb_remove(&pq->sdma_rb_root, &node->rb);
1075 static int pin_vector_pages(struct user_sdma_request *req,
1076 struct user_sdma_iovec *iovec) {
1077 int ret = 0, pinned, npages, cleared;
1078 struct page **pages;
1079 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1080 struct sdma_mmu_node *node = NULL;
1081 struct mmu_rb_node *rb_node;
1083 rb_node = hfi1_mmu_rb_extract(&pq->sdma_rb_root,
1084 (unsigned long)iovec->iov.iov_base,
1085 iovec->iov.iov_len);
1086 if (rb_node && !IS_ERR(rb_node))
1087 node = container_of(rb_node, struct sdma_mmu_node, rb);
1092 node = kzalloc(sizeof(*node), GFP_KERNEL);
1096 node->rb.addr = (unsigned long)iovec->iov.iov_base;
1098 atomic_set(&node->refcount, 0);
1099 INIT_LIST_HEAD(&node->list);
1102 npages = num_user_pages(&iovec->iov);
1103 if (node->npages < npages) {
1104 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1106 SDMA_DBG(req, "Failed page array alloc");
1110 memcpy(pages, node->pages, node->npages * sizeof(*pages));
1112 npages -= node->npages;
1115 * If rb_node is NULL, it means that this is brand new node
1116 * and, therefore not on the eviction list.
1117 * If, however, the rb_node is non-NULL, it means that the
1118 * node is already in RB tree and, therefore on the eviction
1119 * list (nodes are unconditionally inserted in the eviction
1120 * list). In that case, we have to remove the node prior to
1121 * calling the eviction function in order to prevent it from
1122 * freeing this node.
1125 spin_lock(&pq->evict_lock);
1126 list_del_init(&node->list);
1127 spin_unlock(&pq->evict_lock);
1130 if (!hfi1_can_pin_pages(pq->dd, pq->n_locked, npages)) {
1131 cleared = sdma_cache_evict(pq, npages);
1132 if (cleared >= npages)
1135 pinned = hfi1_acquire_user_pages(
1136 ((unsigned long)iovec->iov.iov_base +
1137 (node->npages * PAGE_SIZE)), npages, 0,
1138 pages + node->npages);
1144 if (pinned != npages) {
1145 unpin_vector_pages(current->mm, pages, node->npages,
1151 node->rb.len = iovec->iov.iov_len;
1152 node->pages = pages;
1153 node->npages += pinned;
1154 npages = node->npages;
1155 spin_lock(&pq->evict_lock);
1156 list_add(&node->list, &pq->evict);
1157 pq->n_locked += pinned;
1158 spin_unlock(&pq->evict_lock);
1160 iovec->pages = node->pages;
1161 iovec->npages = npages;
1164 ret = hfi1_mmu_rb_insert(&req->pq->sdma_rb_root, &node->rb);
1166 spin_lock(&pq->evict_lock);
1167 if (!list_empty(&node->list))
1168 list_del(&node->list);
1169 pq->n_locked -= node->npages;
1170 spin_unlock(&pq->evict_lock);
1176 unpin_vector_pages(current->mm, node->pages, 0, node->npages);
1181 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1182 unsigned start, unsigned npages)
1184 hfi1_release_user_pages(mm, pages + start, npages, 0);
1188 static int check_header_template(struct user_sdma_request *req,
1189 struct hfi1_pkt_header *hdr, u32 lrhlen,
1193 * Perform safety checks for any type of packet:
1194 * - transfer size is multiple of 64bytes
1195 * - packet length is multiple of 4bytes
1196 * - entire request length is multiple of 4bytes
1197 * - packet length is not larger than MTU size
1199 * These checks are only done for the first packet of the
1200 * transfer since the header is "given" to us by user space.
1201 * For the remainder of the packets we compute the values.
1203 if (req->info.fragsize % PIO_BLOCK_SIZE ||
1204 lrhlen & 0x3 || req->data_len & 0x3 ||
1205 lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1208 if (req_opcode(req->info.ctrl) == EXPECTED) {
1210 * The header is checked only on the first packet. Furthermore,
1211 * we ensure that at least one TID entry is copied when the
1212 * request is submitted. Therefore, we don't have to verify that
1213 * tididx points to something sane.
1215 u32 tidval = req->tids[req->tididx],
1216 tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1217 tididx = EXP_TID_GET(tidval, IDX),
1218 tidctrl = EXP_TID_GET(tidval, CTRL),
1220 __le32 kval = hdr->kdeth.ver_tid_offset;
1222 tidoff = KDETH_GET(kval, OFFSET) *
1223 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1224 KDETH_OM_LARGE : KDETH_OM_SMALL);
1226 * Expected receive packets have the following
1227 * additional checks:
1228 * - offset is not larger than the TID size
1229 * - TIDCtrl values match between header and TID array
1230 * - TID indexes match between header and TID array
1232 if ((tidoff + datalen > tidlen) ||
1233 KDETH_GET(kval, TIDCTRL) != tidctrl ||
1234 KDETH_GET(kval, TID) != tididx)
1241 * Correctly set the BTH.PSN field based on type of
1242 * transfer - eager packets can just increment the PSN but
1243 * expected packets encode generation and sequence in the
1244 * BTH.PSN field so just incrementing will result in errors.
1246 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1248 u32 val = be32_to_cpu(bthpsn),
1249 mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1253 psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1259 static int set_txreq_header(struct user_sdma_request *req,
1260 struct user_sdma_txreq *tx, u32 datalen)
1262 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1263 struct hfi1_pkt_header *hdr = &tx->hdr;
1266 u32 tidval = 0, lrhlen = get_lrh_len(*hdr, datalen);
1268 /* Copy the header template to the request before modification */
1269 memcpy(hdr, &req->hdr, sizeof(*hdr));
1272 * Check if the PBC and LRH length are mismatched. If so
1273 * adjust both in the header.
1275 pbclen = le16_to_cpu(hdr->pbc[0]);
1276 if (PBC2LRH(pbclen) != lrhlen) {
1277 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1278 hdr->pbc[0] = cpu_to_le16(pbclen);
1279 hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1282 * This is the first packet in the sequence that has
1283 * a "static" size that can be used for the rest of
1284 * the packets (besides the last one).
1286 if (unlikely(req->seqnum == 2)) {
1288 * From this point on the lengths in both the
1289 * PBC and LRH are the same until the last
1291 * Adjust the template so we don't have to update
1294 req->hdr.pbc[0] = hdr->pbc[0];
1295 req->hdr.lrh[2] = hdr->lrh[2];
1299 * We only have to modify the header if this is not the
1300 * first packet in the request. Otherwise, we use the
1301 * header given to us.
1303 if (unlikely(!req->seqnum)) {
1304 ret = check_header_template(req, hdr, lrhlen, datalen);
1310 hdr->bth[2] = cpu_to_be32(
1311 set_pkt_bth_psn(hdr->bth[2],
1312 (req_opcode(req->info.ctrl) == EXPECTED),
1315 /* Set ACK request on last packet */
1316 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
1317 hdr->bth[2] |= cpu_to_be32(1UL << 31);
1319 /* Set the new offset */
1320 hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1321 /* Expected packets have to fill in the new TID information */
1322 if (req_opcode(req->info.ctrl) == EXPECTED) {
1323 tidval = req->tids[req->tididx];
1325 * If the offset puts us at the end of the current TID,
1326 * advance everything.
1328 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1332 * Since we don't copy all the TIDs, all at once,
1333 * we have to check again.
1335 if (++req->tididx > req->n_tids - 1 ||
1336 !req->tids[req->tididx]) {
1339 tidval = req->tids[req->tididx];
1341 req->omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1342 KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE : KDETH_OM_SMALL;
1343 /* Set KDETH.TIDCtrl based on value for this TID. */
1344 KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1345 EXP_TID_GET(tidval, CTRL));
1346 /* Set KDETH.TID based on value for this TID */
1347 KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1348 EXP_TID_GET(tidval, IDX));
1349 /* Clear KDETH.SH only on the last packet */
1350 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
1351 KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1353 * Set the KDETH.OFFSET and KDETH.OM based on size of
1356 SDMA_DBG(req, "TID offset %ubytes %uunits om%u",
1357 req->tidoffset, req->tidoffset / req->omfactor,
1358 req->omfactor != KDETH_OM_SMALL);
1359 KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1360 req->tidoffset / req->omfactor);
1361 KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1362 req->omfactor != KDETH_OM_SMALL);
1365 trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1366 req->info.comp_idx, hdr, tidval);
1367 return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1370 static int set_txreq_header_ahg(struct user_sdma_request *req,
1371 struct user_sdma_txreq *tx, u32 len)
1374 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1375 struct hfi1_pkt_header *hdr = &req->hdr;
1376 u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1377 u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, len);
1379 if (PBC2LRH(pbclen) != lrhlen) {
1380 /* PBC.PbcLengthDWs */
1381 AHG_HEADER_SET(req->ahg, diff, 0, 0, 12,
1382 cpu_to_le16(LRH2PBC(lrhlen)));
1383 /* LRH.PktLen (we need the full 16 bits due to byte swap) */
1384 AHG_HEADER_SET(req->ahg, diff, 3, 0, 16,
1385 cpu_to_be16(lrhlen >> 2));
1389 * Do the common updates
1391 /* BTH.PSN and BTH.A */
1392 val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1393 (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1394 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
1396 AHG_HEADER_SET(req->ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16));
1397 AHG_HEADER_SET(req->ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff));
1399 AHG_HEADER_SET(req->ahg, diff, 15, 0, 16,
1400 cpu_to_le16(req->koffset & 0xffff));
1401 AHG_HEADER_SET(req->ahg, diff, 15, 16, 16,
1402 cpu_to_le16(req->koffset >> 16));
1403 if (req_opcode(req->info.ctrl) == EXPECTED) {
1406 tidval = req->tids[req->tididx];
1409 * If the offset puts us at the end of the current TID,
1410 * advance everything.
1412 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1416 * Since we don't copy all the TIDs, all at once,
1417 * we have to check again.
1419 if (++req->tididx > req->n_tids - 1 ||
1420 !req->tids[req->tididx]) {
1423 tidval = req->tids[req->tididx];
1425 req->omfactor = ((EXP_TID_GET(tidval, LEN) *
1427 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE :
1429 /* KDETH.OM and KDETH.OFFSET (TID) */
1430 AHG_HEADER_SET(req->ahg, diff, 7, 0, 16,
1431 ((!!(req->omfactor - KDETH_OM_SMALL)) << 15 |
1432 ((req->tidoffset / req->omfactor) & 0x7fff)));
1433 /* KDETH.TIDCtrl, KDETH.TID */
1434 val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1435 (EXP_TID_GET(tidval, IDX) & 0x3ff));
1436 /* Clear KDETH.SH on last packet */
1437 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT)) {
1438 val |= cpu_to_le16(KDETH_GET(hdr->kdeth.ver_tid_offset,
1440 val &= cpu_to_le16(~(1U << 13));
1441 AHG_HEADER_SET(req->ahg, diff, 7, 16, 14, val);
1443 AHG_HEADER_SET(req->ahg, diff, 7, 16, 12, val);
1447 trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1448 req->info.comp_idx, req->sde->this_idx,
1449 req->ahg_idx, req->ahg, diff, tidval);
1454 * SDMA tx request completion callback. Called when the SDMA progress
1455 * state machine gets notification that the SDMA descriptors for this
1456 * tx request have been processed by the DMA engine. Called in
1457 * interrupt context.
1459 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1461 struct user_sdma_txreq *tx =
1462 container_of(txreq, struct user_sdma_txreq, txreq);
1463 struct user_sdma_request *req;
1464 struct hfi1_user_sdma_pkt_q *pq;
1465 struct hfi1_user_sdma_comp_q *cq;
1475 if (status != SDMA_TXREQ_S_OK) {
1476 SDMA_DBG(req, "SDMA completion with error %d",
1478 set_bit(SDMA_REQ_HAS_ERROR, &req->flags);
1481 req->seqcomp = tx->seqnum;
1482 kmem_cache_free(pq->txreq_cache, tx);
1485 idx = req->info.comp_idx;
1486 if (req->status == -1 && status == SDMA_TXREQ_S_OK) {
1487 if (req->seqcomp == req->info.npkts - 1) {
1489 user_sdma_free_request(req, false);
1491 set_comp_state(pq, cq, idx, COMPLETE, 0);
1494 if (status != SDMA_TXREQ_S_OK)
1495 req->status = status;
1496 if (req->seqcomp == (ACCESS_ONCE(req->seqsubmitted) - 1) &&
1497 (test_bit(SDMA_REQ_SEND_DONE, &req->flags) ||
1498 test_bit(SDMA_REQ_DONE_ERROR, &req->flags))) {
1499 user_sdma_free_request(req, false);
1501 set_comp_state(pq, cq, idx, ERROR, req->status);
1506 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1508 if (atomic_dec_and_test(&pq->n_reqs)) {
1509 xchg(&pq->state, SDMA_PKT_Q_INACTIVE);
1514 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1516 if (!list_empty(&req->txps)) {
1517 struct sdma_txreq *t, *p;
1519 list_for_each_entry_safe(t, p, &req->txps, list) {
1520 struct user_sdma_txreq *tx =
1521 container_of(t, struct user_sdma_txreq, txreq);
1522 list_del_init(&t->list);
1523 sdma_txclean(req->pq->dd, t);
1524 kmem_cache_free(req->pq->txreq_cache, tx);
1527 if (req->data_iovs) {
1528 struct sdma_mmu_node *node;
1531 for (i = 0; i < req->data_iovs; i++) {
1532 node = req->iovs[i].node;
1537 hfi1_mmu_rb_remove(&req->pq->sdma_rb_root,
1540 atomic_dec(&node->refcount);
1544 clear_bit(SDMA_REQ_IN_USE, &req->flags);
1547 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1548 struct hfi1_user_sdma_comp_q *cq,
1549 u16 idx, enum hfi1_sdma_comp_state state,
1552 hfi1_cdbg(SDMA, "[%u:%u:%u:%u] Setting completion status %u %d",
1553 pq->dd->unit, pq->ctxt, pq->subctxt, idx, state, ret);
1554 cq->comps[idx].status = state;
1556 cq->comps[idx].errcode = -ret;
1557 trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1561 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1564 return (bool)(node->addr == addr);
1567 static int sdma_rb_insert(struct rb_root *root, struct mmu_rb_node *mnode)
1569 struct sdma_mmu_node *node =
1570 container_of(mnode, struct sdma_mmu_node, rb);
1572 atomic_inc(&node->refcount);
1576 static void sdma_rb_remove(struct rb_root *root, struct mmu_rb_node *mnode,
1577 struct mm_struct *mm)
1579 struct sdma_mmu_node *node =
1580 container_of(mnode, struct sdma_mmu_node, rb);
1582 spin_lock(&node->pq->evict_lock);
1584 * We've been called by the MMU notifier but this node has been
1585 * scheduled for eviction. The eviction function will take care
1586 * of freeing this node.
1587 * We have to take the above lock first because we are racing
1588 * against the setting of the bit in the eviction function.
1590 if (mm && test_bit(SDMA_CACHE_NODE_EVICT, &node->flags)) {
1591 spin_unlock(&node->pq->evict_lock);
1595 if (!list_empty(&node->list))
1596 list_del(&node->list);
1597 node->pq->n_locked -= node->npages;
1598 spin_unlock(&node->pq->evict_lock);
1601 * If mm is set, we are being called by the MMU notifier and we
1602 * should not pass a mm_struct to unpin_vector_page(). This is to
1603 * prevent a deadlock when hfi1_release_user_pages() attempts to
1604 * take the mmap_sem, which the MMU notifier has already taken.
1606 unpin_vector_pages(mm ? NULL : current->mm, node->pages, 0,
1609 * If called by the MMU notifier, we have to adjust the pinned
1610 * page count ourselves.
1613 mm->pinned_vm -= node->npages;
1617 static int sdma_rb_invalidate(struct rb_root *root, struct mmu_rb_node *mnode)
1619 struct sdma_mmu_node *node =
1620 container_of(mnode, struct sdma_mmu_node, rb);
1622 if (!atomic_read(&node->refcount))