3 * This file is provided under a dual BSD/GPLv2 license. When using or
4 * redistributing this file, you may do so under either license.
8 * Copyright(c) 2015 Intel Corporation.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of version 2 of the GNU General Public License as
12 * published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful, but
15 * WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * General Public License for more details.
21 * Copyright(c) 2015 Intel Corporation.
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
27 * - Redistributions of source code must retain the above copyright
28 * notice, this list of conditions and the following disclaimer.
29 * - Redistributions in binary form must reproduce the above copyright
30 * notice, this list of conditions and the following disclaimer in
31 * the documentation and/or other materials provided with the
33 * - Neither the name of Intel Corporation nor the names of its
34 * contributors may be used to endorse or promote products derived
35 * from this software without specific prior written permission.
37 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
38 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
39 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
40 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
41 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
42 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
43 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
44 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
45 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
46 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
47 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
51 #include <linux/types.h>
52 #include <linux/device.h>
53 #include <linux/dmapool.h>
54 #include <linux/slab.h>
55 #include <linux/list.h>
56 #include <linux/highmem.h>
58 #include <linux/uio.h>
59 #include <linux/rbtree.h>
60 #include <linux/spinlock.h>
61 #include <linux/delay.h>
62 #include <linux/kthread.h>
63 #include <linux/mmu_context.h>
64 #include <linux/module.h>
65 #include <linux/vmalloc.h>
69 #include "user_sdma.h"
71 #include "verbs.h" /* for the headers */
72 #include "common.h" /* for struct hfi1_tid_info */
75 static uint hfi1_sdma_comp_ring_size = 128;
76 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
77 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
79 /* The maximum number of Data io vectors per message/request */
80 #define MAX_VECTORS_PER_REQ 8
82 * Maximum number of packet to send from each message/request
83 * before moving to the next one.
85 #define MAX_PKTS_PER_QUEUE 16
87 #define num_pages(x) (1 + ((((x) - 1) & PAGE_MASK) >> PAGE_SHIFT))
89 #define req_opcode(x) \
90 (((x) >> HFI1_SDMA_REQ_OPCODE_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
91 #define req_version(x) \
92 (((x) >> HFI1_SDMA_REQ_VERSION_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
93 #define req_iovcnt(x) \
94 (((x) >> HFI1_SDMA_REQ_IOVCNT_SHIFT) & HFI1_SDMA_REQ_IOVCNT_MASK)
96 /* Number of BTH.PSN bits used for sequence number in expected rcvs */
97 #define BTH_SEQ_MASK 0x7ffull
100 * Define fields in the KDETH header so we can update the header
103 #define KDETH_OFFSET_SHIFT 0
104 #define KDETH_OFFSET_MASK 0x7fff
105 #define KDETH_OM_SHIFT 15
106 #define KDETH_OM_MASK 0x1
107 #define KDETH_TID_SHIFT 16
108 #define KDETH_TID_MASK 0x3ff
109 #define KDETH_TIDCTRL_SHIFT 26
110 #define KDETH_TIDCTRL_MASK 0x3
111 #define KDETH_INTR_SHIFT 28
112 #define KDETH_INTR_MASK 0x1
113 #define KDETH_SH_SHIFT 29
114 #define KDETH_SH_MASK 0x1
115 #define KDETH_HCRC_UPPER_SHIFT 16
116 #define KDETH_HCRC_UPPER_MASK 0xff
117 #define KDETH_HCRC_LOWER_SHIFT 24
118 #define KDETH_HCRC_LOWER_MASK 0xff
120 #define PBC2LRH(x) ((((x) & 0xfff) << 2) - 4)
121 #define LRH2PBC(x) ((((x) >> 2) + 1) & 0xfff)
123 #define KDETH_GET(val, field) \
124 (((le32_to_cpu((val))) >> KDETH_##field##_SHIFT) & KDETH_##field##_MASK)
125 #define KDETH_SET(dw, field, val) do { \
126 u32 dwval = le32_to_cpu(dw); \
127 dwval &= ~(KDETH_##field##_MASK << KDETH_##field##_SHIFT); \
128 dwval |= (((val) & KDETH_##field##_MASK) << \
129 KDETH_##field##_SHIFT); \
130 dw = cpu_to_le32(dwval); \
133 #define AHG_HEADER_SET(arr, idx, dw, bit, width, value) \
135 if ((idx) < ARRAY_SIZE((arr))) \
136 (arr)[(idx++)] = sdma_build_ahg_descriptor( \
137 (__force u16)(value), (dw), (bit), \
143 /* KDETH OM multipliers and switch over point */
144 #define KDETH_OM_SMALL 4
145 #define KDETH_OM_LARGE 64
146 #define KDETH_OM_MAX_SIZE (1 << ((KDETH_OM_LARGE / KDETH_OM_SMALL) + 1))
148 /* Last packet in the request */
149 #define USER_SDMA_TXREQ_FLAGS_LAST_PKT (1 << 0)
151 #define SDMA_REQ_IN_USE 0
152 #define SDMA_REQ_FOR_THREAD 1
153 #define SDMA_REQ_SEND_DONE 2
154 #define SDMA_REQ_HAVE_AHG 3
155 #define SDMA_REQ_HAS_ERROR 4
156 #define SDMA_REQ_DONE_ERROR 5
158 #define SDMA_PKT_Q_INACTIVE (1 << 0)
159 #define SDMA_PKT_Q_ACTIVE (1 << 1)
160 #define SDMA_PKT_Q_DEFERRED (1 << 2)
163 * Maximum retry attempts to submit a TX request
164 * before putting the process to sleep.
166 #define MAX_DEFER_RETRY_COUNT 1
168 static unsigned initial_pkt_count = 8;
170 #define SDMA_IOWAIT_TIMEOUT 1000 /* in milliseconds */
172 struct user_sdma_iovec {
174 /* number of pages in this vector */
176 /* array of pinned pages for this vector */
178 /* offset into the virtual address space of the vector at
179 * which we last left off. */
183 struct user_sdma_request {
184 struct sdma_req_info info;
185 struct hfi1_user_sdma_pkt_q *pq;
186 struct hfi1_user_sdma_comp_q *cq;
187 /* This is the original header from user space */
188 struct hfi1_pkt_header hdr;
190 * Pointer to the SDMA engine for this request.
191 * Since different request could be on different VLs,
192 * each request will need it's own engine pointer.
194 struct sdma_engine *sde;
198 * KDETH.Offset (Eager) field
199 * We need to remember the initial value so the headers
200 * can be updated properly.
204 * KDETH.OFFSET (TID) field
205 * The offset can cover multiple packets, depending on the
206 * size of the TID entry.
211 * Remember this because the header template always sets it
216 * pointer to the user's task_struct. We are going to
217 * get a reference to it so we can process io vectors
220 struct task_struct *user_proc;
222 * pointer to the user's mm_struct. We are going to
223 * get a reference to it so it doesn't get freed
224 * since we might not be in process context when we
225 * are processing the iov's.
226 * Using this mm_struct, we can get vma based on the
227 * iov's address (find_vma()).
229 struct mm_struct *user_mm;
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 */
247 spinlock_t list_lock;
248 struct list_head txps;
252 struct user_sdma_txreq {
253 /* Packet header for the txreq */
254 struct hfi1_pkt_header hdr;
255 struct sdma_txreq txreq;
256 struct user_sdma_request *req;
257 struct user_sdma_iovec *iovec1;
258 struct user_sdma_iovec *iovec2;
264 #define SDMA_DBG(req, fmt, ...) \
265 hfi1_cdbg(SDMA, "[%u:%u:%u:%u] " fmt, (req)->pq->dd->unit, \
266 (req)->pq->ctxt, (req)->pq->subctxt, (req)->info.comp_idx, \
268 #define SDMA_Q_DBG(pq, fmt, ...) \
269 hfi1_cdbg(SDMA, "[%u:%u:%u] " fmt, (pq)->dd->unit, (pq)->ctxt, \
270 (pq)->subctxt, ##__VA_ARGS__)
272 static int user_sdma_send_pkts(struct user_sdma_request *, unsigned);
273 static int num_user_pages(const struct iovec *);
274 static void user_sdma_txreq_cb(struct sdma_txreq *, int, int);
275 static void user_sdma_free_request(struct user_sdma_request *);
276 static int pin_vector_pages(struct user_sdma_request *,
277 struct user_sdma_iovec *);
278 static void unpin_vector_pages(struct user_sdma_iovec *);
279 static int check_header_template(struct user_sdma_request *,
280 struct hfi1_pkt_header *, u32, u32);
281 static int set_txreq_header(struct user_sdma_request *,
282 struct user_sdma_txreq *, u32);
283 static int set_txreq_header_ahg(struct user_sdma_request *,
284 struct user_sdma_txreq *, u32);
285 static inline void set_comp_state(struct user_sdma_request *,
286 enum hfi1_sdma_comp_state, int);
287 static inline u32 set_pkt_bth_psn(__be32, u8, u32);
288 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
290 static int defer_packet_queue(
291 struct sdma_engine *,
295 static void activate_packet_queue(struct iowait *, int);
297 static inline int iovec_may_free(struct user_sdma_iovec *iovec,
298 void (*free)(struct user_sdma_iovec *))
300 if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
307 static inline void iovec_set_complete(struct user_sdma_iovec *iovec)
309 iovec->offset = iovec->iov.iov_len;
312 static int defer_packet_queue(
313 struct sdma_engine *sde,
315 struct sdma_txreq *txreq,
318 struct hfi1_user_sdma_pkt_q *pq =
319 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
320 struct hfi1_ibdev *dev = &pq->dd->verbs_dev;
321 struct user_sdma_txreq *tx =
322 container_of(txreq, struct user_sdma_txreq, txreq);
324 if (sdma_progress(sde, seq, txreq)) {
325 if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
329 * We are assuming that if the list is enqueued somewhere, it
330 * is to the dmawait list since that is the only place where
331 * it is supposed to be enqueued.
333 xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
334 write_seqlock(&dev->iowait_lock);
335 if (list_empty(&pq->busy.list))
336 list_add_tail(&pq->busy.list, &sde->dmawait);
337 write_sequnlock(&dev->iowait_lock);
343 static void activate_packet_queue(struct iowait *wait, int reason)
345 struct hfi1_user_sdma_pkt_q *pq =
346 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
347 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
348 wake_up(&wait->wait_dma);
351 static void sdma_kmem_cache_ctor(void *obj)
353 struct user_sdma_txreq *tx = (struct user_sdma_txreq *)obj;
355 memset(tx, 0, sizeof(*tx));
358 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt, struct file *fp)
363 struct hfi1_devdata *dd;
364 struct hfi1_user_sdma_comp_q *cq;
365 struct hfi1_user_sdma_pkt_q *pq;
373 if (!hfi1_sdma_comp_ring_size) {
380 pq = kzalloc(sizeof(*pq), GFP_KERNEL);
384 memsize = sizeof(*pq->reqs) * hfi1_sdma_comp_ring_size;
385 pq->reqs = kmalloc(memsize, GFP_KERNEL);
389 INIT_LIST_HEAD(&pq->list);
391 pq->ctxt = uctxt->ctxt;
392 pq->subctxt = subctxt_fp(fp);
393 pq->n_max_reqs = hfi1_sdma_comp_ring_size;
394 pq->state = SDMA_PKT_Q_INACTIVE;
395 atomic_set(&pq->n_reqs, 0);
397 iowait_init(&pq->busy, 0, NULL, defer_packet_queue,
398 activate_packet_queue);
400 snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
402 pq->txreq_cache = kmem_cache_create(buf,
403 sizeof(struct user_sdma_txreq),
406 sdma_kmem_cache_ctor);
407 if (!pq->txreq_cache) {
408 dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
412 user_sdma_pkt_fp(fp) = pq;
413 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
417 memsize = ALIGN(sizeof(*cq->comps) * hfi1_sdma_comp_ring_size,
419 cq->comps = vmalloc_user(memsize);
423 cq->nentries = hfi1_sdma_comp_ring_size;
424 user_sdma_comp_fp(fp) = cq;
426 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
427 list_add(&pq->list, &uctxt->sdma_queues);
428 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
434 kmem_cache_destroy(pq->txreq_cache);
439 user_sdma_pkt_fp(fp) = NULL;
446 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd)
448 struct hfi1_ctxtdata *uctxt = fd->uctxt;
449 struct hfi1_user_sdma_pkt_q *pq;
452 hfi1_cdbg(SDMA, "[%u:%u:%u] Freeing user SDMA queues", uctxt->dd->unit,
453 uctxt->ctxt, fd->subctxt);
458 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
459 if (!list_empty(&pq->list))
460 list_del_init(&pq->list);
461 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
462 iowait_sdma_drain(&pq->busy);
464 for (i = 0, j = 0; i < atomic_read(&pq->n_reqs) &&
465 j < pq->n_max_reqs; j++) {
466 struct user_sdma_request *req = &pq->reqs[j];
468 if (test_bit(SDMA_REQ_IN_USE, &req->flags)) {
469 set_comp_state(req, ERROR, -ECOMM);
470 user_sdma_free_request(req);
476 kmem_cache_destroy(pq->txreq_cache);
482 vfree(fd->cq->comps);
489 int hfi1_user_sdma_process_request(struct file *fp, struct iovec *iovec,
490 unsigned long dim, unsigned long *count)
492 int ret = 0, i = 0, sent;
493 struct hfi1_ctxtdata *uctxt = ctxt_fp(fp);
494 struct hfi1_user_sdma_pkt_q *pq = user_sdma_pkt_fp(fp);
495 struct hfi1_user_sdma_comp_q *cq = user_sdma_comp_fp(fp);
496 struct hfi1_devdata *dd = pq->dd;
497 unsigned long idx = 0;
498 u8 pcount = initial_pkt_count;
499 struct sdma_req_info info;
500 struct user_sdma_request *req;
503 if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
506 "[%u:%u:%u] First vector not big enough for header %lu/%lu",
507 dd->unit, uctxt->ctxt, subctxt_fp(fp),
508 iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
512 ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
514 hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
515 dd->unit, uctxt->ctxt, subctxt_fp(fp), ret);
519 trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, subctxt_fp(fp),
521 if (cq->comps[info.comp_idx].status == QUEUED) {
522 hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in QUEUED state",
523 dd->unit, uctxt->ctxt, subctxt_fp(fp),
528 if (!info.fragsize) {
530 "[%u:%u:%u:%u] Request does not specify fragsize",
531 dd->unit, uctxt->ctxt, subctxt_fp(fp), info.comp_idx);
536 * We've done all the safety checks that we can up to this point,
537 * "allocate" the request entry.
539 hfi1_cdbg(SDMA, "[%u:%u:%u] Using req/comp entry %u\n", dd->unit,
540 uctxt->ctxt, subctxt_fp(fp), info.comp_idx);
541 req = pq->reqs + info.comp_idx;
542 memset(req, 0, sizeof(*req));
543 /* Mark the request as IN_USE before we start filling it in. */
544 set_bit(SDMA_REQ_IN_USE, &req->flags);
545 req->data_iovs = req_iovcnt(info.ctrl) - 1;
548 INIT_LIST_HEAD(&req->txps);
549 spin_lock_init(&req->list_lock);
550 memcpy(&req->info, &info, sizeof(info));
552 if (req_opcode(info.ctrl) == EXPECTED)
555 if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
556 SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
557 MAX_VECTORS_PER_REQ);
561 /* Copy the header from the user buffer */
562 ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
565 SDMA_DBG(req, "Failed to copy header template (%d)", ret);
570 /* If Static rate control is not enabled, sanitize the header. */
571 if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
574 /* Validate the opcode. Do not trust packets from user space blindly. */
575 opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
576 if ((opcode & USER_OPCODE_CHECK_MASK) !=
577 USER_OPCODE_CHECK_VAL) {
578 SDMA_DBG(req, "Invalid opcode (%d)", opcode);
583 * Validate the vl. Do not trust packets from user space blindly.
584 * VL comes from PBC, SC comes from LRH, and the VL needs to
585 * match the SC look up.
587 vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
588 sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
589 (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
590 if (vl >= dd->pport->vls_operational ||
591 vl != sc_to_vlt(dd, sc)) {
592 SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
598 * Also should check the BTH.lnh. If it says the next header is GRH then
599 * the RXE parsing will be off and will land in the middle of the KDETH
600 * or miss it entirely.
602 if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
603 SDMA_DBG(req, "User tried to pass in a GRH");
608 req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
609 /* Calculate the initial TID offset based on the values of
610 KDETH.OFFSET and KDETH.OM that are passed in. */
611 req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
612 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
613 KDETH_OM_LARGE : KDETH_OM_SMALL);
614 SDMA_DBG(req, "Initial TID offset %u", req->tidoffset);
617 /* Save all the IO vector structures */
618 while (i < req->data_iovs) {
619 memcpy(&req->iovs[i].iov, iovec + idx++, sizeof(struct iovec));
620 req->iovs[i].offset = 0;
621 req->data_len += req->iovs[i++].iov.iov_len;
623 SDMA_DBG(req, "total data length %u", req->data_len);
625 if (pcount > req->info.npkts)
626 pcount = req->info.npkts;
629 * User space will provide the TID info only when the
630 * request type is EXPECTED. This is true even if there is
631 * only one packet in the request and the header is already
632 * setup. The reason for the singular TID case is that the
633 * driver needs to perform safety checks.
635 if (req_opcode(req->info.ctrl) == EXPECTED) {
636 u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
638 if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
642 req->tids = kcalloc(ntids, sizeof(*req->tids), GFP_KERNEL);
648 * We have to copy all of the tids because they may vary
649 * in size and, therefore, the TID count might not be
650 * equal to the pkt count. However, there is no way to
651 * tell at this point.
653 ret = copy_from_user(req->tids, iovec[idx].iov_base,
654 ntids * sizeof(*req->tids));
656 SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
665 /* Have to select the engine */
666 req->sde = sdma_select_engine_vl(dd,
667 (u32)(uctxt->ctxt + subctxt_fp(fp)),
669 if (!req->sde || !sdma_running(req->sde)) {
674 /* We don't need an AHG entry if the request contains only one packet */
675 if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG)) {
676 int ahg = sdma_ahg_alloc(req->sde);
678 if (likely(ahg >= 0)) {
679 req->ahg_idx = (u8)ahg;
680 set_bit(SDMA_REQ_HAVE_AHG, &req->flags);
684 set_comp_state(req, QUEUED, 0);
685 /* Send the first N packets in the request to buy us some time */
686 sent = user_sdma_send_pkts(req, pcount);
687 if (unlikely(sent < 0)) {
688 if (sent != -EBUSY) {
694 atomic_inc(&pq->n_reqs);
696 if (sent < req->info.npkts) {
697 /* Take the references to the user's task and mm_struct */
698 get_task_struct(current);
699 req->user_proc = current;
702 * This is a somewhat blocking send implementation.
703 * The driver will block the caller until all packets of the
704 * request have been submitted to the SDMA engine. However, it
705 * will not wait for send completions.
707 while (!test_bit(SDMA_REQ_SEND_DONE, &req->flags)) {
708 ret = user_sdma_send_pkts(req, pcount);
712 wait_event_interruptible_timeout(
714 (pq->state == SDMA_PKT_Q_ACTIVE),
716 SDMA_IOWAIT_TIMEOUT));
725 set_comp_state(req, ERROR, ret);
727 user_sdma_free_request(req);
732 static inline u32 compute_data_length(struct user_sdma_request *req,
733 struct user_sdma_txreq *tx)
736 * Determine the proper size of the packet data.
737 * The size of the data of the first packet is in the header
738 * template. However, it includes the header and ICRC, which need
740 * The size of the remaining packets is the minimum of the frag
741 * size (MTU) or remaining data in the request.
746 len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
747 (sizeof(tx->hdr) - 4));
748 } else if (req_opcode(req->info.ctrl) == EXPECTED) {
749 u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
751 /* Get the data length based on the remaining space in the
753 len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
754 /* If we've filled up the TID pair, move to the next one. */
755 if (unlikely(!len) && ++req->tididx < req->n_tids &&
756 req->tids[req->tididx]) {
757 tidlen = EXP_TID_GET(req->tids[req->tididx],
760 len = min_t(u32, tidlen, req->info.fragsize);
762 /* Since the TID pairs map entire pages, make sure that we
763 * are not going to try to send more data that we have
765 len = min(len, req->data_len - req->sent);
767 len = min(req->data_len - req->sent, (u32)req->info.fragsize);
768 SDMA_DBG(req, "Data Length = %u", len);
772 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
774 /* (Size of complete header - size of PBC) + 4B ICRC + data length */
775 return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
778 static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
782 struct user_sdma_txreq *tx = NULL;
783 struct hfi1_user_sdma_pkt_q *pq = NULL;
784 struct user_sdma_iovec *iovec = NULL;
794 * Check if we might have sent the entire request already
796 if (unlikely(req->seqnum == req->info.npkts)) {
797 if (!list_empty(&req->txps))
802 if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
803 maxpkts = req->info.npkts - req->seqnum;
805 while (npkts < maxpkts) {
806 u32 datalen = 0, queued = 0, data_sent = 0;
810 * Check whether any of the completions have come back
811 * with errors. If so, we are not going to process any
812 * more packets from this request.
814 if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) {
815 set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
820 tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
831 if (req->seqnum == req->info.npkts - 1)
832 tx->flags |= USER_SDMA_TXREQ_FLAGS_LAST_PKT;
835 * Calculate the payload size - this is min of the fragment
836 * (MTU) size or the remaining bytes in the request but only
837 * if we have payload data.
840 iovec = &req->iovs[req->iov_idx];
841 if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
842 if (++req->iov_idx == req->data_iovs) {
846 iovec = &req->iovs[req->iov_idx];
847 WARN_ON(iovec->offset);
851 * This request might include only a header and no user
852 * data, so pin pages only if there is data and it the
853 * pages have not been pinned already.
855 if (unlikely(!iovec->pages && iovec->iov.iov_len)) {
856 ret = pin_vector_pages(req, iovec);
862 datalen = compute_data_length(req, tx);
865 "Request has data but pkt len is 0");
871 if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags)) {
873 u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
874 u32 lrhlen = get_lrh_len(req->hdr, datalen);
876 * Copy the request header into the tx header
877 * because the HW needs a cacheline-aligned
879 * This copy can be optimized out if the hdr
880 * member of user_sdma_request were also
883 memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
884 if (PBC2LRH(pbclen) != lrhlen) {
885 pbclen = (pbclen & 0xf000) |
887 tx->hdr.pbc[0] = cpu_to_le16(pbclen);
889 ret = sdma_txinit_ahg(&tx->txreq,
890 SDMA_TXREQ_F_AHG_COPY,
891 sizeof(tx->hdr) + datalen,
892 req->ahg_idx, 0, NULL, 0,
896 ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq,
904 changes = set_txreq_header_ahg(req, tx,
908 sdma_txinit_ahg(&tx->txreq,
909 SDMA_TXREQ_F_USE_AHG,
910 datalen, req->ahg_idx, changes,
911 req->ahg, sizeof(req->hdr),
915 ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
916 datalen, user_sdma_txreq_cb);
920 * Modify the header for this packet. This only needs
921 * to be done if we are not going to use AHG. Otherwise,
922 * the HW will do it based on the changes we gave it
923 * during sdma_txinit_ahg().
925 ret = set_txreq_header(req, tx, datalen);
931 * If the request contains any data vectors, add up to
932 * fragsize bytes to the descriptor.
934 while (queued < datalen &&
935 (req->sent + data_sent) < req->data_len) {
936 unsigned long base, offset;
937 unsigned pageidx, len;
939 base = (unsigned long)iovec->iov.iov_base;
940 offset = ((base + iovec->offset + iov_offset) &
942 pageidx = (((iovec->offset + iov_offset +
943 base) - (base & PAGE_MASK)) >> PAGE_SHIFT);
944 len = offset + req->info.fragsize > PAGE_SIZE ?
945 PAGE_SIZE - offset : req->info.fragsize;
946 len = min((datalen - queued), len);
947 ret = sdma_txadd_page(pq->dd, &tx->txreq,
948 iovec->pages[pageidx],
952 "SDMA txreq add page failed %d\n",
954 iovec_set_complete(iovec);
960 if (unlikely(queued < datalen &&
961 pageidx == iovec->npages &&
962 req->iov_idx < req->data_iovs - 1)) {
963 iovec->offset += iov_offset;
964 iovec = &req->iovs[++req->iov_idx];
966 ret = pin_vector_pages(req, iovec);
976 * The txreq was submitted successfully so we can update
979 req->koffset += datalen;
980 if (req_opcode(req->info.ctrl) == EXPECTED)
981 req->tidoffset += datalen;
982 req->sent += data_sent;
984 if (tx->iovec1 && !tx->iovec2)
985 tx->iovec1->offset += iov_offset;
987 tx->iovec2->offset += iov_offset;
990 * It is important to increment this here as it is used to
991 * generate the BTH.PSN and, therefore, can't be bulk-updated
992 * outside of the loop.
994 tx->seqnum = req->seqnum++;
995 list_add_tail(&tx->txreq.list, &req->txps);
999 ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps);
1000 if (list_empty(&req->txps))
1001 if (req->seqnum == req->info.npkts) {
1002 set_bit(SDMA_REQ_SEND_DONE, &req->flags);
1004 * The txreq has already been submitted to the HW queue
1005 * so we can free the AHG entry now. Corruption will not
1006 * happen due to the sequential manner in which
1007 * descriptors are processed.
1009 if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags))
1010 sdma_ahg_free(req->sde, req->ahg_idx);
1014 sdma_txclean(pq->dd, &tx->txreq);
1016 kmem_cache_free(pq->txreq_cache, tx);
1022 * How many pages in this iovec element?
1024 static inline int num_user_pages(const struct iovec *iov)
1026 const unsigned long addr = (unsigned long) iov->iov_base;
1027 const unsigned long len = iov->iov_len;
1028 const unsigned long spage = addr & PAGE_MASK;
1029 const unsigned long epage = (addr + len - 1) & PAGE_MASK;
1031 return 1 + ((epage - spage) >> PAGE_SHIFT);
1034 static int pin_vector_pages(struct user_sdma_request *req,
1035 struct user_sdma_iovec *iovec) {
1039 iovec->npages = num_user_pages(&iovec->iov);
1040 iovec->pages = kcalloc(iovec->npages, sizeof(*iovec->pages),
1042 if (!iovec->pages) {
1043 SDMA_DBG(req, "Failed page array alloc");
1047 /* If called by the kernel thread, use the user's mm */
1048 if (current->flags & PF_KTHREAD)
1049 use_mm(req->user_proc->mm);
1050 pinned = get_user_pages_fast(
1051 (unsigned long)iovec->iov.iov_base,
1052 iovec->npages, 0, iovec->pages);
1053 /* If called by the kernel thread, unuse the user's mm */
1054 if (current->flags & PF_KTHREAD)
1055 unuse_mm(req->user_proc->mm);
1056 if (pinned != iovec->npages) {
1057 SDMA_DBG(req, "Failed to pin pages (%u/%u)", pinned,
1064 unpin_vector_pages(iovec);
1069 static void unpin_vector_pages(struct user_sdma_iovec *iovec)
1073 if (ACCESS_ONCE(iovec->offset) != iovec->iov.iov_len) {
1075 "the complete vector has not been sent yet %llu %zu",
1076 iovec->offset, iovec->iov.iov_len);
1079 for (i = 0; i < iovec->npages; i++)
1080 if (iovec->pages[i])
1081 put_page(iovec->pages[i]);
1082 kfree(iovec->pages);
1083 iovec->pages = NULL;
1088 static int check_header_template(struct user_sdma_request *req,
1089 struct hfi1_pkt_header *hdr, u32 lrhlen,
1093 * Perform safety checks for any type of packet:
1094 * - transfer size is multiple of 64bytes
1095 * - packet length is multiple of 4bytes
1096 * - entire request length is multiple of 4bytes
1097 * - packet length is not larger than MTU size
1099 * These checks are only done for the first packet of the
1100 * transfer since the header is "given" to us by user space.
1101 * For the remainder of the packets we compute the values.
1103 if (req->info.fragsize % PIO_BLOCK_SIZE ||
1104 lrhlen & 0x3 || req->data_len & 0x3 ||
1105 lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1108 if (req_opcode(req->info.ctrl) == EXPECTED) {
1110 * The header is checked only on the first packet. Furthermore,
1111 * we ensure that at least one TID entry is copied when the
1112 * request is submitted. Therefore, we don't have to verify that
1113 * tididx points to something sane.
1115 u32 tidval = req->tids[req->tididx],
1116 tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1117 tididx = EXP_TID_GET(tidval, IDX),
1118 tidctrl = EXP_TID_GET(tidval, CTRL),
1120 __le32 kval = hdr->kdeth.ver_tid_offset;
1122 tidoff = KDETH_GET(kval, OFFSET) *
1123 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1124 KDETH_OM_LARGE : KDETH_OM_SMALL);
1126 * Expected receive packets have the following
1127 * additional checks:
1128 * - offset is not larger than the TID size
1129 * - TIDCtrl values match between header and TID array
1130 * - TID indexes match between header and TID array
1132 if ((tidoff + datalen > tidlen) ||
1133 KDETH_GET(kval, TIDCTRL) != tidctrl ||
1134 KDETH_GET(kval, TID) != tididx)
1141 * Correctly set the BTH.PSN field based on type of
1142 * transfer - eager packets can just increment the PSN but
1143 * expected packets encode generation and sequence in the
1144 * BTH.PSN field so just incrementing will result in errors.
1146 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1148 u32 val = be32_to_cpu(bthpsn),
1149 mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1153 psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1159 static int set_txreq_header(struct user_sdma_request *req,
1160 struct user_sdma_txreq *tx, u32 datalen)
1162 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1163 struct hfi1_pkt_header *hdr = &tx->hdr;
1166 u32 tidval = 0, lrhlen = get_lrh_len(*hdr, datalen);
1168 /* Copy the header template to the request before modification */
1169 memcpy(hdr, &req->hdr, sizeof(*hdr));
1172 * Check if the PBC and LRH length are mismatched. If so
1173 * adjust both in the header.
1175 pbclen = le16_to_cpu(hdr->pbc[0]);
1176 if (PBC2LRH(pbclen) != lrhlen) {
1177 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1178 hdr->pbc[0] = cpu_to_le16(pbclen);
1179 hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1182 * This is the first packet in the sequence that has
1183 * a "static" size that can be used for the rest of
1184 * the packets (besides the last one).
1186 if (unlikely(req->seqnum == 2)) {
1188 * From this point on the lengths in both the
1189 * PBC and LRH are the same until the last
1191 * Adjust the template so we don't have to update
1194 req->hdr.pbc[0] = hdr->pbc[0];
1195 req->hdr.lrh[2] = hdr->lrh[2];
1199 * We only have to modify the header if this is not the
1200 * first packet in the request. Otherwise, we use the
1201 * header given to us.
1203 if (unlikely(!req->seqnum)) {
1204 ret = check_header_template(req, hdr, lrhlen, datalen);
1211 hdr->bth[2] = cpu_to_be32(
1212 set_pkt_bth_psn(hdr->bth[2],
1213 (req_opcode(req->info.ctrl) == EXPECTED),
1216 /* Set ACK request on last packet */
1217 if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT))
1218 hdr->bth[2] |= cpu_to_be32(1UL<<31);
1220 /* Set the new offset */
1221 hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1222 /* Expected packets have to fill in the new TID information */
1223 if (req_opcode(req->info.ctrl) == EXPECTED) {
1224 tidval = req->tids[req->tididx];
1226 * If the offset puts us at the end of the current TID,
1227 * advance everything.
1229 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1232 /* Since we don't copy all the TIDs, all at once,
1233 * we have to check again. */
1234 if (++req->tididx > req->n_tids - 1 ||
1235 !req->tids[req->tididx]) {
1238 tidval = req->tids[req->tididx];
1240 req->omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1241 KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE : KDETH_OM_SMALL;
1242 /* Set KDETH.TIDCtrl based on value for this TID. */
1243 KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1244 EXP_TID_GET(tidval, CTRL));
1245 /* Set KDETH.TID based on value for this TID */
1246 KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1247 EXP_TID_GET(tidval, IDX));
1248 /* Clear KDETH.SH only on the last packet */
1249 if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT))
1250 KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1252 * Set the KDETH.OFFSET and KDETH.OM based on size of
1255 SDMA_DBG(req, "TID offset %ubytes %uunits om%u",
1256 req->tidoffset, req->tidoffset / req->omfactor,
1257 !!(req->omfactor - KDETH_OM_SMALL));
1258 KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1259 req->tidoffset / req->omfactor);
1260 KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1261 !!(req->omfactor - KDETH_OM_SMALL));
1264 trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1265 req->info.comp_idx, hdr, tidval);
1266 return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1269 static int set_txreq_header_ahg(struct user_sdma_request *req,
1270 struct user_sdma_txreq *tx, u32 len)
1273 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1274 struct hfi1_pkt_header *hdr = &req->hdr;
1275 u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1276 u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, len);
1278 if (PBC2LRH(pbclen) != lrhlen) {
1279 /* PBC.PbcLengthDWs */
1280 AHG_HEADER_SET(req->ahg, diff, 0, 0, 12,
1281 cpu_to_le16(LRH2PBC(lrhlen)));
1282 /* LRH.PktLen (we need the full 16 bits due to byte swap) */
1283 AHG_HEADER_SET(req->ahg, diff, 3, 0, 16,
1284 cpu_to_be16(lrhlen >> 2));
1288 * Do the common updates
1290 /* BTH.PSN and BTH.A */
1291 val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1292 (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1293 if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT))
1295 AHG_HEADER_SET(req->ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16));
1296 AHG_HEADER_SET(req->ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff));
1298 AHG_HEADER_SET(req->ahg, diff, 15, 0, 16,
1299 cpu_to_le16(req->koffset & 0xffff));
1300 AHG_HEADER_SET(req->ahg, diff, 15, 16, 16,
1301 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) *
1314 /* Since we don't copy all the TIDs, all at once,
1315 * we have to check again. */
1316 if (++req->tididx > req->n_tids - 1 ||
1317 !req->tids[req->tididx]) {
1320 tidval = req->tids[req->tididx];
1322 req->omfactor = ((EXP_TID_GET(tidval, LEN) *
1324 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE :
1326 /* KDETH.OM and KDETH.OFFSET (TID) */
1327 AHG_HEADER_SET(req->ahg, diff, 7, 0, 16,
1328 ((!!(req->omfactor - KDETH_OM_SMALL)) << 15 |
1329 ((req->tidoffset / req->omfactor) & 0x7fff)));
1330 /* KDETH.TIDCtrl, KDETH.TID */
1331 val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1332 (EXP_TID_GET(tidval, IDX) & 0x3ff));
1333 /* Clear KDETH.SH on last packet */
1334 if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT)) {
1335 val |= cpu_to_le16(KDETH_GET(hdr->kdeth.ver_tid_offset,
1337 val &= cpu_to_le16(~(1U << 13));
1338 AHG_HEADER_SET(req->ahg, diff, 7, 16, 14, val);
1340 AHG_HEADER_SET(req->ahg, diff, 7, 16, 12, val);
1343 trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1344 req->info.comp_idx, req->sde->this_idx,
1345 req->ahg_idx, req->ahg, diff, tidval);
1349 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status,
1352 struct user_sdma_txreq *tx =
1353 container_of(txreq, struct user_sdma_txreq, txreq);
1354 struct user_sdma_request *req = tx->req;
1355 struct hfi1_user_sdma_pkt_q *pq = req ? req->pq : NULL;
1358 if (unlikely(!req || !pq))
1362 iovec_may_free(tx->iovec1, unpin_vector_pages);
1364 iovec_may_free(tx->iovec2, unpin_vector_pages);
1366 tx_seqnum = tx->seqnum;
1367 kmem_cache_free(pq->txreq_cache, tx);
1369 if (status != SDMA_TXREQ_S_OK) {
1370 dd_dev_err(pq->dd, "SDMA completion with error %d", status);
1371 set_comp_state(req, ERROR, status);
1372 set_bit(SDMA_REQ_HAS_ERROR, &req->flags);
1373 /* Do not free the request until the sender loop has ack'ed
1374 * the error and we've seen all txreqs. */
1375 if (tx_seqnum == ACCESS_ONCE(req->seqnum) &&
1376 test_bit(SDMA_REQ_DONE_ERROR, &req->flags)) {
1377 atomic_dec(&pq->n_reqs);
1378 user_sdma_free_request(req);
1381 if (tx_seqnum == req->info.npkts - 1) {
1382 /* We've sent and completed all packets in this
1383 * request. Signal completion to the user */
1384 atomic_dec(&pq->n_reqs);
1385 set_comp_state(req, COMPLETE, 0);
1386 user_sdma_free_request(req);
1389 if (!atomic_read(&pq->n_reqs))
1390 xchg(&pq->state, SDMA_PKT_Q_INACTIVE);
1393 static void user_sdma_free_request(struct user_sdma_request *req)
1395 if (!list_empty(&req->txps)) {
1396 struct sdma_txreq *t, *p;
1398 list_for_each_entry_safe(t, p, &req->txps, list) {
1399 struct user_sdma_txreq *tx =
1400 container_of(t, struct user_sdma_txreq, txreq);
1401 list_del_init(&t->list);
1402 sdma_txclean(req->pq->dd, t);
1403 kmem_cache_free(req->pq->txreq_cache, tx);
1406 if (req->data_iovs) {
1409 for (i = 0; i < req->data_iovs; i++)
1410 if (req->iovs[i].npages && req->iovs[i].pages)
1411 unpin_vector_pages(&req->iovs[i]);
1414 put_task_struct(req->user_proc);
1416 clear_bit(SDMA_REQ_IN_USE, &req->flags);
1419 static inline void set_comp_state(struct user_sdma_request *req,
1420 enum hfi1_sdma_comp_state state,
1423 SDMA_DBG(req, "Setting completion status %u %d", state, ret);
1424 req->cq->comps[req->info.comp_idx].status = state;
1426 req->cq->comps[req->info.comp_idx].errcode = -ret;
1427 trace_hfi1_sdma_user_completion(req->pq->dd, req->pq->ctxt,
1428 req->pq->subctxt, req->info.comp_idx,