2 * Copyright(c) 2015 - 2017 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>
67 #include "user_sdma.h"
68 #include "verbs.h" /* for the headers */
69 #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 /* The maximum number of Data io vectors per message/request */
78 #define MAX_VECTORS_PER_REQ 8
80 * Maximum number of packet to send from each message/request
81 * before moving to the next one.
83 #define MAX_PKTS_PER_QUEUE 16
85 #define num_pages(x) (1 + ((((x) - 1) & PAGE_MASK) >> PAGE_SHIFT))
87 #define req_opcode(x) \
88 (((x) >> HFI1_SDMA_REQ_OPCODE_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
89 #define req_version(x) \
90 (((x) >> HFI1_SDMA_REQ_VERSION_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
91 #define req_iovcnt(x) \
92 (((x) >> HFI1_SDMA_REQ_IOVCNT_SHIFT) & HFI1_SDMA_REQ_IOVCNT_MASK)
94 /* Number of BTH.PSN bits used for sequence number in expected rcvs */
95 #define BTH_SEQ_MASK 0x7ffull
98 * Define fields in the KDETH header so we can update the header
101 #define KDETH_OFFSET_SHIFT 0
102 #define KDETH_OFFSET_MASK 0x7fff
103 #define KDETH_OM_SHIFT 15
104 #define KDETH_OM_MASK 0x1
105 #define KDETH_TID_SHIFT 16
106 #define KDETH_TID_MASK 0x3ff
107 #define KDETH_TIDCTRL_SHIFT 26
108 #define KDETH_TIDCTRL_MASK 0x3
109 #define KDETH_INTR_SHIFT 28
110 #define KDETH_INTR_MASK 0x1
111 #define KDETH_SH_SHIFT 29
112 #define KDETH_SH_MASK 0x1
113 #define KDETH_HCRC_UPPER_SHIFT 16
114 #define KDETH_HCRC_UPPER_MASK 0xff
115 #define KDETH_HCRC_LOWER_SHIFT 24
116 #define KDETH_HCRC_LOWER_MASK 0xff
118 #define AHG_KDETH_INTR_SHIFT 12
119 #define AHG_KDETH_SH_SHIFT 13
120 #define AHG_KDETH_ARRAY_SIZE 9
122 #define PBC2LRH(x) ((((x) & 0xfff) << 2) - 4)
123 #define LRH2PBC(x) ((((x) >> 2) + 1) & 0xfff)
125 #define KDETH_GET(val, field) \
126 (((le32_to_cpu((val))) >> KDETH_##field##_SHIFT) & KDETH_##field##_MASK)
127 #define KDETH_SET(dw, field, val) do { \
128 u32 dwval = le32_to_cpu(dw); \
129 dwval &= ~(KDETH_##field##_MASK << KDETH_##field##_SHIFT); \
130 dwval |= (((val) & KDETH_##field##_MASK) << \
131 KDETH_##field##_SHIFT); \
132 dw = cpu_to_le32(dwval); \
135 #define AHG_HEADER_SET(arr, idx, dw, bit, width, value) \
137 if ((idx) < ARRAY_SIZE((arr))) \
138 (arr)[(idx++)] = sdma_build_ahg_descriptor( \
139 (__force u16)(value), (dw), (bit), \
145 /* KDETH OM multipliers and switch over point */
146 #define KDETH_OM_SMALL 4
147 #define KDETH_OM_SMALL_SHIFT 2
148 #define KDETH_OM_LARGE 64
149 #define KDETH_OM_LARGE_SHIFT 6
150 #define KDETH_OM_MAX_SIZE (1 << ((KDETH_OM_LARGE / KDETH_OM_SMALL) + 1))
152 /* Tx request flag bits */
153 #define TXREQ_FLAGS_REQ_ACK BIT(0) /* Set the ACK bit in the header */
154 #define TXREQ_FLAGS_REQ_DISABLE_SH BIT(1) /* Disable header suppression */
156 #define SDMA_PKT_Q_INACTIVE BIT(0)
157 #define SDMA_PKT_Q_ACTIVE BIT(1)
158 #define SDMA_PKT_Q_DEFERRED BIT(2)
161 * Maximum retry attempts to submit a TX request
162 * before putting the process to sleep.
164 #define MAX_DEFER_RETRY_COUNT 1
166 static unsigned initial_pkt_count = 8;
168 #define SDMA_IOWAIT_TIMEOUT 1000 /* in milliseconds */
170 struct sdma_mmu_node;
172 struct user_sdma_iovec {
173 struct list_head list;
175 /* number of pages in this vector */
177 /* array of pinned pages for this vector */
180 * offset into the virtual address space of the vector at
181 * which we last left off.
184 struct sdma_mmu_node *node;
187 struct sdma_mmu_node {
188 struct mmu_rb_node rb;
189 struct hfi1_user_sdma_pkt_q *pq;
195 /* evict operation argument */
197 u32 cleared; /* count evicted so far */
198 u32 target; /* target count to evict */
201 struct user_sdma_request {
202 /* This is the original header from user space */
203 struct hfi1_pkt_header hdr;
205 /* Read mostly fields */
206 struct hfi1_user_sdma_pkt_q *pq ____cacheline_aligned_in_smp;
207 struct hfi1_user_sdma_comp_q *cq;
209 * Pointer to the SDMA engine for this request.
210 * Since different request could be on different VLs,
211 * each request will need it's own engine pointer.
213 struct sdma_engine *sde;
214 struct sdma_req_info info;
215 /* TID array values copied from the tid_iov vector */
217 /* total length of the data in the request */
219 /* number of elements copied to the tids array */
222 * We copy the iovs for this request (based on
223 * info.iovcnt). These are only the data vectors
228 /* Writeable fields shared with interrupt */
229 u64 seqcomp ____cacheline_aligned_in_smp;
231 /* status of the last txreq completed */
234 /* Send side fields */
235 struct list_head txps ____cacheline_aligned_in_smp;
238 * KDETH.OFFSET (TID) field
239 * The offset can cover multiple packets, depending on the
240 * size of the TID entry.
244 * KDETH.Offset (Eager) field
245 * We need to remember the initial value so the headers
246 * can be updated properly.
250 /* TID index copied from the tid_iov vector */
252 /* progress index moving along the iovs array */
257 struct user_sdma_iovec iovs[MAX_VECTORS_PER_REQ];
258 } ____cacheline_aligned_in_smp;
261 * A single txreq could span up to 3 physical pages when the MTU
262 * is sufficiently large (> 4K). Each of the IOV pointers also
263 * needs it's own set of flags so the vector has been handled
264 * independently of each other.
266 struct user_sdma_txreq {
267 /* Packet header for the txreq */
268 struct hfi1_pkt_header hdr;
269 struct sdma_txreq txreq;
270 struct list_head list;
271 struct user_sdma_request *req;
277 #define SDMA_DBG(req, fmt, ...) \
278 hfi1_cdbg(SDMA, "[%u:%u:%u:%u] " fmt, (req)->pq->dd->unit, \
279 (req)->pq->ctxt, (req)->pq->subctxt, (req)->info.comp_idx, \
281 #define SDMA_Q_DBG(pq, fmt, ...) \
282 hfi1_cdbg(SDMA, "[%u:%u:%u] " fmt, (pq)->dd->unit, (pq)->ctxt, \
283 (pq)->subctxt, ##__VA_ARGS__)
285 static int user_sdma_send_pkts(struct user_sdma_request *req,
287 static int num_user_pages(const struct iovec *iov);
288 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status);
289 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq);
290 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin);
291 static int pin_vector_pages(struct user_sdma_request *req,
292 struct user_sdma_iovec *iovec);
293 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
294 unsigned start, unsigned npages);
295 static int check_header_template(struct user_sdma_request *req,
296 struct hfi1_pkt_header *hdr, u32 lrhlen,
298 static int set_txreq_header(struct user_sdma_request *req,
299 struct user_sdma_txreq *tx, u32 datalen);
300 static int set_txreq_header_ahg(struct user_sdma_request *req,
301 struct user_sdma_txreq *tx, u32 len);
302 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
303 struct hfi1_user_sdma_comp_q *cq,
304 u16 idx, enum hfi1_sdma_comp_state state,
306 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags);
307 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
309 static int defer_packet_queue(
310 struct sdma_engine *sde,
312 struct sdma_txreq *txreq,
314 static void activate_packet_queue(struct iowait *wait, int reason);
315 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
317 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode);
318 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
319 void *arg2, bool *stop);
320 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode);
321 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
323 static struct mmu_rb_ops sdma_rb_ops = {
324 .filter = sdma_rb_filter,
325 .insert = sdma_rb_insert,
326 .evict = sdma_rb_evict,
327 .remove = sdma_rb_remove,
328 .invalidate = sdma_rb_invalidate
331 static int defer_packet_queue(
332 struct sdma_engine *sde,
334 struct sdma_txreq *txreq,
337 struct hfi1_user_sdma_pkt_q *pq =
338 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
339 struct hfi1_ibdev *dev = &pq->dd->verbs_dev;
340 struct user_sdma_txreq *tx =
341 container_of(txreq, struct user_sdma_txreq, txreq);
343 if (sdma_progress(sde, seq, txreq)) {
344 if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
348 * We are assuming that if the list is enqueued somewhere, it
349 * is to the dmawait list since that is the only place where
350 * it is supposed to be enqueued.
352 xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
353 write_seqlock(&dev->iowait_lock);
354 if (list_empty(&pq->busy.list))
355 list_add_tail(&pq->busy.list, &sde->dmawait);
356 write_sequnlock(&dev->iowait_lock);
362 static void activate_packet_queue(struct iowait *wait, int reason)
364 struct hfi1_user_sdma_pkt_q *pq =
365 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
366 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
367 wake_up(&wait->wait_dma);
370 static void sdma_kmem_cache_ctor(void *obj)
372 struct user_sdma_txreq *tx = obj;
374 memset(tx, 0, sizeof(*tx));
377 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
378 struct hfi1_filedata *fd)
382 struct hfi1_devdata *dd;
383 struct hfi1_user_sdma_comp_q *cq;
384 struct hfi1_user_sdma_pkt_q *pq;
390 if (!hfi1_sdma_comp_ring_size)
395 pq = kzalloc(sizeof(*pq), GFP_KERNEL);
399 INIT_LIST_HEAD(&pq->list);
401 pq->ctxt = uctxt->ctxt;
402 pq->subctxt = fd->subctxt;
403 pq->n_max_reqs = hfi1_sdma_comp_ring_size;
404 pq->state = SDMA_PKT_Q_INACTIVE;
405 atomic_set(&pq->n_reqs, 0);
406 init_waitqueue_head(&pq->wait);
407 atomic_set(&pq->n_locked, 0);
410 iowait_init(&pq->busy, 0, NULL, defer_packet_queue,
411 activate_packet_queue, NULL);
414 pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
420 pq->req_in_use = kcalloc(BITS_TO_LONGS(hfi1_sdma_comp_ring_size),
421 sizeof(*pq->req_in_use),
424 goto pq_reqs_no_in_use;
426 snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
428 pq->txreq_cache = kmem_cache_create(buf,
429 sizeof(struct user_sdma_txreq),
432 sdma_kmem_cache_ctor);
433 if (!pq->txreq_cache) {
434 dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
439 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
443 cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
444 * hfi1_sdma_comp_ring_size));
448 cq->nentries = hfi1_sdma_comp_ring_size;
450 ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
453 dd_dev_err(dd, "Failed to register with MMU %d", ret);
460 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
461 list_add(&pq->list, &uctxt->sdma_queues);
462 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
471 kmem_cache_destroy(pq->txreq_cache);
473 kfree(pq->req_in_use);
482 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd)
484 struct hfi1_ctxtdata *uctxt = fd->uctxt;
485 struct hfi1_user_sdma_pkt_q *pq;
488 hfi1_cdbg(SDMA, "[%u:%u:%u] Freeing user SDMA queues", uctxt->dd->unit,
489 uctxt->ctxt, fd->subctxt);
493 hfi1_mmu_rb_unregister(pq->handler);
494 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
495 if (!list_empty(&pq->list))
496 list_del_init(&pq->list);
497 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
498 iowait_sdma_drain(&pq->busy);
499 /* Wait until all requests have been freed. */
500 wait_event_interruptible(
502 (ACCESS_ONCE(pq->state) == SDMA_PKT_Q_INACTIVE));
504 kfree(pq->req_in_use);
505 kmem_cache_destroy(pq->txreq_cache);
510 vfree(fd->cq->comps);
517 static u8 dlid_to_selector(u16 dlid)
519 static u8 mapping[256];
520 static int initialized;
525 memset(mapping, 0xFF, 256);
529 hash = ((dlid >> 8) ^ dlid) & 0xFF;
530 if (mapping[hash] == 0xFF) {
531 mapping[hash] = next;
532 next = (next + 1) & 0x7F;
535 return mapping[hash];
538 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
539 struct iovec *iovec, unsigned long dim,
540 unsigned long *count)
543 struct hfi1_ctxtdata *uctxt = fd->uctxt;
544 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
545 struct hfi1_user_sdma_comp_q *cq = fd->cq;
546 struct hfi1_devdata *dd = pq->dd;
547 unsigned long idx = 0;
548 u8 pcount = initial_pkt_count;
549 struct sdma_req_info info;
550 struct user_sdma_request *req;
556 if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
559 "[%u:%u:%u] First vector not big enough for header %lu/%lu",
560 dd->unit, uctxt->ctxt, fd->subctxt,
561 iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
564 ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
566 hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
567 dd->unit, uctxt->ctxt, fd->subctxt, ret);
571 trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
574 if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
576 "[%u:%u:%u:%u] Invalid comp index",
577 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
582 * Sanity check the header io vector count. Need at least 1 vector
583 * (header) and cannot be larger than the actual io vector count.
585 if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
587 "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
588 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
589 req_iovcnt(info.ctrl), dim);
593 if (!info.fragsize) {
595 "[%u:%u:%u:%u] Request does not specify fragsize",
596 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
600 /* Try to claim the request. */
601 if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
602 hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
603 dd->unit, uctxt->ctxt, fd->subctxt,
608 * All safety checks have been done and this request has been claimed.
610 hfi1_cdbg(SDMA, "[%u:%u:%u] Using req/comp entry %u\n", dd->unit,
611 uctxt->ctxt, fd->subctxt, info.comp_idx);
612 req = pq->reqs + info.comp_idx;
613 req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
623 req->seqsubmitted = 0;
627 INIT_LIST_HEAD(&req->txps);
629 memcpy(&req->info, &info, sizeof(info));
631 if (req_opcode(info.ctrl) == EXPECTED) {
632 /* expected must have a TID info and at least one data vector */
633 if (req->data_iovs < 2) {
635 "Not enough vectors for expected request");
642 if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
643 SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
644 MAX_VECTORS_PER_REQ);
648 /* Copy the header from the user buffer */
649 ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
652 SDMA_DBG(req, "Failed to copy header template (%d)", ret);
657 /* If Static rate control is not enabled, sanitize the header. */
658 if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
661 /* Validate the opcode. Do not trust packets from user space blindly. */
662 opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
663 if ((opcode & USER_OPCODE_CHECK_MASK) !=
664 USER_OPCODE_CHECK_VAL) {
665 SDMA_DBG(req, "Invalid opcode (%d)", opcode);
670 * Validate the vl. Do not trust packets from user space blindly.
671 * VL comes from PBC, SC comes from LRH, and the VL needs to
672 * match the SC look up.
674 vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
675 sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
676 (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
677 if (vl >= dd->pport->vls_operational ||
678 vl != sc_to_vlt(dd, sc)) {
679 SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
684 /* Checking P_KEY for requests from user-space */
685 if (egress_pkey_check(dd->pport, req->hdr.lrh, req->hdr.bth, sc,
686 PKEY_CHECK_INVALID)) {
692 * Also should check the BTH.lnh. If it says the next header is GRH then
693 * the RXE parsing will be off and will land in the middle of the KDETH
694 * or miss it entirely.
696 if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
697 SDMA_DBG(req, "User tried to pass in a GRH");
702 req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
704 * Calculate the initial TID offset based on the values of
705 * KDETH.OFFSET and KDETH.OM that are passed in.
707 req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
708 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
709 KDETH_OM_LARGE : KDETH_OM_SMALL);
710 SDMA_DBG(req, "Initial TID offset %u", req->tidoffset);
713 /* Save all the IO vector structures */
714 for (i = 0; i < req->data_iovs; i++) {
715 req->iovs[i].offset = 0;
716 INIT_LIST_HEAD(&req->iovs[i].list);
717 memcpy(&req->iovs[i].iov,
719 sizeof(req->iovs[i].iov));
720 ret = pin_vector_pages(req, &req->iovs[i]);
726 req->data_len += req->iovs[i].iov.iov_len;
728 SDMA_DBG(req, "total data length %u", req->data_len);
730 if (pcount > req->info.npkts)
731 pcount = req->info.npkts;
734 * User space will provide the TID info only when the
735 * request type is EXPECTED. This is true even if there is
736 * only one packet in the request and the header is already
737 * setup. The reason for the singular TID case is that the
738 * driver needs to perform safety checks.
740 if (req_opcode(req->info.ctrl) == EXPECTED) {
741 u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
744 if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
750 * We have to copy all of the tids because they may vary
751 * in size and, therefore, the TID count might not be
752 * equal to the pkt count. However, there is no way to
753 * tell at this point.
755 tmp = memdup_user(iovec[idx].iov_base,
756 ntids * sizeof(*req->tids));
759 SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
769 dlid = be16_to_cpu(req->hdr.lrh[1]);
770 selector = dlid_to_selector(dlid);
771 selector += uctxt->ctxt + fd->subctxt;
772 req->sde = sdma_select_user_engine(dd, selector, vl);
774 if (!req->sde || !sdma_running(req->sde)) {
779 /* We don't need an AHG entry if the request contains only one packet */
780 if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
781 req->ahg_idx = sdma_ahg_alloc(req->sde);
783 set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
784 atomic_inc(&pq->n_reqs);
786 /* Send the first N packets in the request to buy us some time */
787 ret = user_sdma_send_pkts(req, pcount);
788 if (unlikely(ret < 0 && ret != -EBUSY)) {
794 * It is possible that the SDMA engine would have processed all the
795 * submitted packets by the time we get here. Therefore, only set
796 * packet queue state to ACTIVE if there are still uncompleted
799 if (atomic_read(&pq->n_reqs))
800 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
803 * This is a somewhat blocking send implementation.
804 * The driver will block the caller until all packets of the
805 * request have been submitted to the SDMA engine. However, it
806 * will not wait for send completions.
808 while (req->seqsubmitted != req->info.npkts) {
809 ret = user_sdma_send_pkts(req, pcount);
813 WRITE_ONCE(req->has_error, 1);
814 if (ACCESS_ONCE(req->seqcomp) ==
815 req->seqsubmitted - 1)
819 wait_event_interruptible_timeout(
821 (pq->state == SDMA_PKT_Q_ACTIVE),
823 SDMA_IOWAIT_TIMEOUT));
829 user_sdma_free_request(req, true);
832 set_comp_state(pq, cq, info.comp_idx, ERROR, req->status);
836 static inline u32 compute_data_length(struct user_sdma_request *req,
837 struct user_sdma_txreq *tx)
840 * Determine the proper size of the packet data.
841 * The size of the data of the first packet is in the header
842 * template. However, it includes the header and ICRC, which need
844 * The minimum representable packet data length in a header is 4 bytes,
845 * therefore, when the data length request is less than 4 bytes, there's
846 * only one packet, and the packet data length is equal to that of the
847 * request data length.
848 * The size of the remaining packets is the minimum of the frag
849 * size (MTU) or remaining data in the request.
854 if (req->data_len < sizeof(u32))
857 len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
858 (sizeof(tx->hdr) - 4));
859 } else if (req_opcode(req->info.ctrl) == EXPECTED) {
860 u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
863 * Get the data length based on the remaining space in the
866 len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
867 /* If we've filled up the TID pair, move to the next one. */
868 if (unlikely(!len) && ++req->tididx < req->n_tids &&
869 req->tids[req->tididx]) {
870 tidlen = EXP_TID_GET(req->tids[req->tididx],
873 len = min_t(u32, tidlen, req->info.fragsize);
876 * Since the TID pairs map entire pages, make sure that we
877 * are not going to try to send more data that we have
880 len = min(len, req->data_len - req->sent);
882 len = min(req->data_len - req->sent, (u32)req->info.fragsize);
884 SDMA_DBG(req, "Data Length = %u", len);
888 static inline u32 pad_len(u32 len)
890 if (len & (sizeof(u32) - 1))
891 len += sizeof(u32) - (len & (sizeof(u32) - 1));
895 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
897 /* (Size of complete header - size of PBC) + 4B ICRC + data length */
898 return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
901 static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
905 struct user_sdma_txreq *tx = NULL;
906 struct hfi1_user_sdma_pkt_q *pq = NULL;
907 struct user_sdma_iovec *iovec = NULL;
914 /* If tx completion has reported an error, we are done. */
915 if (READ_ONCE(req->has_error))
919 * Check if we might have sent the entire request already
921 if (unlikely(req->seqnum == req->info.npkts)) {
922 if (!list_empty(&req->txps))
927 if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
928 maxpkts = req->info.npkts - req->seqnum;
930 while (npkts < maxpkts) {
931 u32 datalen = 0, queued = 0, data_sent = 0;
935 * Check whether any of the completions have come back
936 * with errors. If so, we are not going to process any
937 * more packets from this request.
939 if (READ_ONCE(req->has_error))
942 tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
949 INIT_LIST_HEAD(&tx->list);
952 * For the last packet set the ACK request
953 * and disable header suppression.
955 if (req->seqnum == req->info.npkts - 1)
956 tx->flags |= (TXREQ_FLAGS_REQ_ACK |
957 TXREQ_FLAGS_REQ_DISABLE_SH);
960 * Calculate the payload size - this is min of the fragment
961 * (MTU) size or the remaining bytes in the request but only
962 * if we have payload data.
965 iovec = &req->iovs[req->iov_idx];
966 if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
967 if (++req->iov_idx == req->data_iovs) {
971 iovec = &req->iovs[req->iov_idx];
972 WARN_ON(iovec->offset);
975 datalen = compute_data_length(req, tx);
978 * Disable header suppression for the payload <= 8DWS.
979 * If there is an uncorrectable error in the receive
980 * data FIFO when the received payload size is less than
981 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
982 * not reported.There is set RHF.EccErr if the header
987 "Request has data but pkt len is 0");
990 } else if (datalen <= 32) {
991 tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
995 if (req->ahg_idx >= 0) {
997 u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
998 u32 lrhlen = get_lrh_len(req->hdr,
1001 * Copy the request header into the tx header
1002 * because the HW needs a cacheline-aligned
1004 * This copy can be optimized out if the hdr
1005 * member of user_sdma_request were also
1006 * cacheline aligned.
1008 memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
1009 if (PBC2LRH(pbclen) != lrhlen) {
1010 pbclen = (pbclen & 0xf000) |
1012 tx->hdr.pbc[0] = cpu_to_le16(pbclen);
1014 ret = check_header_template(req, &tx->hdr,
1018 ret = sdma_txinit_ahg(&tx->txreq,
1019 SDMA_TXREQ_F_AHG_COPY,
1020 sizeof(tx->hdr) + datalen,
1021 req->ahg_idx, 0, NULL, 0,
1022 user_sdma_txreq_cb);
1025 ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq,
1033 changes = set_txreq_header_ahg(req, tx,
1039 ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
1040 datalen, user_sdma_txreq_cb);
1044 * Modify the header for this packet. This only needs
1045 * to be done if we are not going to use AHG. Otherwise,
1046 * the HW will do it based on the changes we gave it
1047 * during sdma_txinit_ahg().
1049 ret = set_txreq_header(req, tx, datalen);
1055 * If the request contains any data vectors, add up to
1056 * fragsize bytes to the descriptor.
1058 while (queued < datalen &&
1059 (req->sent + data_sent) < req->data_len) {
1060 unsigned long base, offset;
1061 unsigned pageidx, len;
1063 base = (unsigned long)iovec->iov.iov_base;
1064 offset = offset_in_page(base + iovec->offset +
1066 pageidx = (((iovec->offset + iov_offset +
1067 base) - (base & PAGE_MASK)) >> PAGE_SHIFT);
1068 len = offset + req->info.fragsize > PAGE_SIZE ?
1069 PAGE_SIZE - offset : req->info.fragsize;
1070 len = min((datalen - queued), len);
1071 ret = sdma_txadd_page(pq->dd, &tx->txreq,
1072 iovec->pages[pageidx],
1075 SDMA_DBG(req, "SDMA txreq add page failed %d\n",
1082 if (unlikely(queued < datalen &&
1083 pageidx == iovec->npages &&
1084 req->iov_idx < req->data_iovs - 1)) {
1085 iovec->offset += iov_offset;
1086 iovec = &req->iovs[++req->iov_idx];
1091 * The txreq was submitted successfully so we can update
1094 req->koffset += datalen;
1095 if (req_opcode(req->info.ctrl) == EXPECTED)
1096 req->tidoffset += datalen;
1097 req->sent += data_sent;
1099 iovec->offset += iov_offset;
1100 list_add_tail(&tx->txreq.list, &req->txps);
1102 * It is important to increment this here as it is used to
1103 * generate the BTH.PSN and, therefore, can't be bulk-updated
1104 * outside of the loop.
1106 tx->seqnum = req->seqnum++;
1110 ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps, &count);
1111 req->seqsubmitted += count;
1112 if (req->seqsubmitted == req->info.npkts) {
1113 WRITE_ONCE(req->done, 1);
1115 * The txreq has already been submitted to the HW queue
1116 * so we can free the AHG entry now. Corruption will not
1117 * happen due to the sequential manner in which
1118 * descriptors are processed.
1120 if (req->ahg_idx >= 0)
1121 sdma_ahg_free(req->sde, req->ahg_idx);
1126 sdma_txclean(pq->dd, &tx->txreq);
1128 kmem_cache_free(pq->txreq_cache, tx);
1133 * How many pages in this iovec element?
1135 static inline int num_user_pages(const struct iovec *iov)
1137 const unsigned long addr = (unsigned long)iov->iov_base;
1138 const unsigned long len = iov->iov_len;
1139 const unsigned long spage = addr & PAGE_MASK;
1140 const unsigned long epage = (addr + len - 1) & PAGE_MASK;
1142 return 1 + ((epage - spage) >> PAGE_SHIFT);
1145 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
1147 struct evict_data evict_data;
1149 evict_data.cleared = 0;
1150 evict_data.target = npages;
1151 hfi1_mmu_rb_evict(pq->handler, &evict_data);
1152 return evict_data.cleared;
1155 static int pin_vector_pages(struct user_sdma_request *req,
1156 struct user_sdma_iovec *iovec)
1158 int ret = 0, pinned, npages, cleared;
1159 struct page **pages;
1160 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1161 struct sdma_mmu_node *node = NULL;
1162 struct mmu_rb_node *rb_node;
1166 hfi1_mmu_rb_remove_unless_exact(pq->handler,
1168 iovec->iov.iov_base,
1169 iovec->iov.iov_len, &rb_node);
1171 node = container_of(rb_node, struct sdma_mmu_node, rb);
1173 atomic_inc(&node->refcount);
1174 iovec->pages = node->pages;
1175 iovec->npages = node->npages;
1182 node = kzalloc(sizeof(*node), GFP_KERNEL);
1186 node->rb.addr = (unsigned long)iovec->iov.iov_base;
1188 atomic_set(&node->refcount, 0);
1191 npages = num_user_pages(&iovec->iov);
1192 if (node->npages < npages) {
1193 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1195 SDMA_DBG(req, "Failed page array alloc");
1199 memcpy(pages, node->pages, node->npages * sizeof(*pages));
1201 npages -= node->npages;
1204 if (!hfi1_can_pin_pages(pq->dd, pq->mm,
1205 atomic_read(&pq->n_locked), npages)) {
1206 cleared = sdma_cache_evict(pq, npages);
1207 if (cleared >= npages)
1210 pinned = hfi1_acquire_user_pages(pq->mm,
1211 ((unsigned long)iovec->iov.iov_base +
1212 (node->npages * PAGE_SIZE)), npages, 0,
1213 pages + node->npages);
1219 if (pinned != npages) {
1220 unpin_vector_pages(pq->mm, pages, node->npages,
1226 node->rb.len = iovec->iov.iov_len;
1227 node->pages = pages;
1228 node->npages += pinned;
1229 npages = node->npages;
1230 atomic_add(pinned, &pq->n_locked);
1232 iovec->pages = node->pages;
1233 iovec->npages = npages;
1236 ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1238 atomic_sub(node->npages, &pq->n_locked);
1245 unpin_vector_pages(pq->mm, node->pages, 0, node->npages);
1250 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1251 unsigned start, unsigned npages)
1253 hfi1_release_user_pages(mm, pages + start, npages, false);
1257 static int check_header_template(struct user_sdma_request *req,
1258 struct hfi1_pkt_header *hdr, u32 lrhlen,
1262 * Perform safety checks for any type of packet:
1263 * - transfer size is multiple of 64bytes
1264 * - packet length is multiple of 4 bytes
1265 * - packet length is not larger than MTU size
1267 * These checks are only done for the first packet of the
1268 * transfer since the header is "given" to us by user space.
1269 * For the remainder of the packets we compute the values.
1271 if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1272 lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1275 if (req_opcode(req->info.ctrl) == EXPECTED) {
1277 * The header is checked only on the first packet. Furthermore,
1278 * we ensure that at least one TID entry is copied when the
1279 * request is submitted. Therefore, we don't have to verify that
1280 * tididx points to something sane.
1282 u32 tidval = req->tids[req->tididx],
1283 tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1284 tididx = EXP_TID_GET(tidval, IDX),
1285 tidctrl = EXP_TID_GET(tidval, CTRL),
1287 __le32 kval = hdr->kdeth.ver_tid_offset;
1289 tidoff = KDETH_GET(kval, OFFSET) *
1290 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1291 KDETH_OM_LARGE : KDETH_OM_SMALL);
1293 * Expected receive packets have the following
1294 * additional checks:
1295 * - offset is not larger than the TID size
1296 * - TIDCtrl values match between header and TID array
1297 * - TID indexes match between header and TID array
1299 if ((tidoff + datalen > tidlen) ||
1300 KDETH_GET(kval, TIDCTRL) != tidctrl ||
1301 KDETH_GET(kval, TID) != tididx)
1308 * Correctly set the BTH.PSN field based on type of
1309 * transfer - eager packets can just increment the PSN but
1310 * expected packets encode generation and sequence in the
1311 * BTH.PSN field so just incrementing will result in errors.
1313 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1315 u32 val = be32_to_cpu(bthpsn),
1316 mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1320 psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1326 static int set_txreq_header(struct user_sdma_request *req,
1327 struct user_sdma_txreq *tx, u32 datalen)
1329 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1330 struct hfi1_pkt_header *hdr = &tx->hdr;
1331 u8 omfactor; /* KDETH.OM */
1334 u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1336 /* Copy the header template to the request before modification */
1337 memcpy(hdr, &req->hdr, sizeof(*hdr));
1340 * Check if the PBC and LRH length are mismatched. If so
1341 * adjust both in the header.
1343 pbclen = le16_to_cpu(hdr->pbc[0]);
1344 if (PBC2LRH(pbclen) != lrhlen) {
1345 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1346 hdr->pbc[0] = cpu_to_le16(pbclen);
1347 hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1350 * This is the first packet in the sequence that has
1351 * a "static" size that can be used for the rest of
1352 * the packets (besides the last one).
1354 if (unlikely(req->seqnum == 2)) {
1356 * From this point on the lengths in both the
1357 * PBC and LRH are the same until the last
1359 * Adjust the template so we don't have to update
1362 req->hdr.pbc[0] = hdr->pbc[0];
1363 req->hdr.lrh[2] = hdr->lrh[2];
1367 * We only have to modify the header if this is not the
1368 * first packet in the request. Otherwise, we use the
1369 * header given to us.
1371 if (unlikely(!req->seqnum)) {
1372 ret = check_header_template(req, hdr, lrhlen, datalen);
1378 hdr->bth[2] = cpu_to_be32(
1379 set_pkt_bth_psn(hdr->bth[2],
1380 (req_opcode(req->info.ctrl) == EXPECTED),
1383 /* Set ACK request on last packet */
1384 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1385 hdr->bth[2] |= cpu_to_be32(1UL << 31);
1387 /* Set the new offset */
1388 hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1389 /* Expected packets have to fill in the new TID information */
1390 if (req_opcode(req->info.ctrl) == EXPECTED) {
1391 tidval = req->tids[req->tididx];
1393 * If the offset puts us at the end of the current TID,
1394 * advance everything.
1396 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1400 * Since we don't copy all the TIDs, all at once,
1401 * we have to check again.
1403 if (++req->tididx > req->n_tids - 1 ||
1404 !req->tids[req->tididx]) {
1407 tidval = req->tids[req->tididx];
1409 omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1410 KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
1411 KDETH_OM_SMALL_SHIFT;
1412 /* Set KDETH.TIDCtrl based on value for this TID. */
1413 KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1414 EXP_TID_GET(tidval, CTRL));
1415 /* Set KDETH.TID based on value for this TID */
1416 KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1417 EXP_TID_GET(tidval, IDX));
1418 /* Clear KDETH.SH when DISABLE_SH flag is set */
1419 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1420 KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1422 * Set the KDETH.OFFSET and KDETH.OM based on size of
1425 SDMA_DBG(req, "TID offset %ubytes %uunits om%u",
1426 req->tidoffset, req->tidoffset >> omfactor,
1427 omfactor != KDETH_OM_SMALL_SHIFT);
1428 KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1429 req->tidoffset >> omfactor);
1430 KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1431 omfactor != KDETH_OM_SMALL_SHIFT);
1434 trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1435 req->info.comp_idx, hdr, tidval);
1436 return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1439 static int set_txreq_header_ahg(struct user_sdma_request *req,
1440 struct user_sdma_txreq *tx, u32 datalen)
1442 u32 ahg[AHG_KDETH_ARRAY_SIZE];
1444 u8 omfactor; /* KDETH.OM */
1445 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1446 struct hfi1_pkt_header *hdr = &req->hdr;
1447 u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1448 u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1450 if (PBC2LRH(pbclen) != lrhlen) {
1451 /* PBC.PbcLengthDWs */
1452 AHG_HEADER_SET(ahg, diff, 0, 0, 12,
1453 cpu_to_le16(LRH2PBC(lrhlen)));
1454 /* LRH.PktLen (we need the full 16 bits due to byte swap) */
1455 AHG_HEADER_SET(ahg, diff, 3, 0, 16,
1456 cpu_to_be16(lrhlen >> 2));
1460 * Do the common updates
1462 /* BTH.PSN and BTH.A */
1463 val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1464 (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1465 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1467 AHG_HEADER_SET(ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16));
1468 AHG_HEADER_SET(ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff));
1470 AHG_HEADER_SET(ahg, diff, 15, 0, 16,
1471 cpu_to_le16(req->koffset & 0xffff));
1472 AHG_HEADER_SET(ahg, diff, 15, 16, 16, cpu_to_le16(req->koffset >> 16));
1473 if (req_opcode(req->info.ctrl) == EXPECTED) {
1476 tidval = req->tids[req->tididx];
1479 * If the offset puts us at the end of the current TID,
1480 * advance everything.
1482 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1486 * Since we don't copy all the TIDs, all at once,
1487 * we have to check again.
1489 if (++req->tididx > req->n_tids - 1 ||
1490 !req->tids[req->tididx])
1492 tidval = req->tids[req->tididx];
1494 omfactor = ((EXP_TID_GET(tidval, LEN) *
1496 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1497 KDETH_OM_SMALL_SHIFT;
1498 /* KDETH.OM and KDETH.OFFSET (TID) */
1499 AHG_HEADER_SET(ahg, diff, 7, 0, 16,
1500 ((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1501 ((req->tidoffset >> omfactor)
1503 /* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1504 val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1505 (EXP_TID_GET(tidval, IDX) & 0x3ff));
1507 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1508 val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1510 AHG_KDETH_INTR_SHIFT));
1512 val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1513 cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1514 cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1516 AHG_KDETH_INTR_SHIFT));
1519 AHG_HEADER_SET(ahg, diff, 7, 16, 14, val);
1524 trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1525 req->info.comp_idx, req->sde->this_idx,
1526 req->ahg_idx, ahg, diff, tidval);
1527 sdma_txinit_ahg(&tx->txreq,
1528 SDMA_TXREQ_F_USE_AHG,
1529 datalen, req->ahg_idx, diff,
1530 ahg, sizeof(req->hdr),
1531 user_sdma_txreq_cb);
1537 * SDMA tx request completion callback. Called when the SDMA progress
1538 * state machine gets notification that the SDMA descriptors for this
1539 * tx request have been processed by the DMA engine. Called in
1540 * interrupt context.
1542 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1544 struct user_sdma_txreq *tx =
1545 container_of(txreq, struct user_sdma_txreq, txreq);
1546 struct user_sdma_request *req;
1547 struct hfi1_user_sdma_pkt_q *pq;
1548 struct hfi1_user_sdma_comp_q *cq;
1558 if (status != SDMA_TXREQ_S_OK) {
1559 SDMA_DBG(req, "SDMA completion with error %d",
1561 WRITE_ONCE(req->has_error, 1);
1564 req->seqcomp = tx->seqnum;
1565 kmem_cache_free(pq->txreq_cache, tx);
1568 idx = req->info.comp_idx;
1569 if (req->status == -1 && status == SDMA_TXREQ_S_OK) {
1570 if (req->seqcomp == req->info.npkts - 1) {
1572 user_sdma_free_request(req, false);
1574 set_comp_state(pq, cq, idx, COMPLETE, 0);
1577 if (status != SDMA_TXREQ_S_OK)
1578 req->status = status;
1579 if (req->seqcomp == (ACCESS_ONCE(req->seqsubmitted) - 1) &&
1580 (READ_ONCE(req->done) ||
1581 READ_ONCE(req->has_error))) {
1582 user_sdma_free_request(req, false);
1584 set_comp_state(pq, cq, idx, ERROR, req->status);
1589 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1591 if (atomic_dec_and_test(&pq->n_reqs)) {
1592 xchg(&pq->state, SDMA_PKT_Q_INACTIVE);
1597 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1599 if (!list_empty(&req->txps)) {
1600 struct sdma_txreq *t, *p;
1602 list_for_each_entry_safe(t, p, &req->txps, list) {
1603 struct user_sdma_txreq *tx =
1604 container_of(t, struct user_sdma_txreq, txreq);
1605 list_del_init(&t->list);
1606 sdma_txclean(req->pq->dd, t);
1607 kmem_cache_free(req->pq->txreq_cache, tx);
1610 if (req->data_iovs) {
1611 struct sdma_mmu_node *node;
1614 for (i = 0; i < req->data_iovs; i++) {
1615 node = req->iovs[i].node;
1620 hfi1_mmu_rb_remove(req->pq->handler,
1623 atomic_dec(&node->refcount);
1627 clear_bit(req->info.comp_idx, req->pq->req_in_use);
1630 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1631 struct hfi1_user_sdma_comp_q *cq,
1632 u16 idx, enum hfi1_sdma_comp_state state,
1635 hfi1_cdbg(SDMA, "[%u:%u:%u:%u] Setting completion status %u %d",
1636 pq->dd->unit, pq->ctxt, pq->subctxt, idx, state, ret);
1638 cq->comps[idx].errcode = -ret;
1639 smp_wmb(); /* make sure errcode is visible first */
1640 cq->comps[idx].status = state;
1641 trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1645 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1648 return (bool)(node->addr == addr);
1651 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1653 struct sdma_mmu_node *node =
1654 container_of(mnode, struct sdma_mmu_node, rb);
1656 atomic_inc(&node->refcount);
1661 * Return 1 to remove the node from the rb tree and call the remove op.
1663 * Called with the rb tree lock held.
1665 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1666 void *evict_arg, bool *stop)
1668 struct sdma_mmu_node *node =
1669 container_of(mnode, struct sdma_mmu_node, rb);
1670 struct evict_data *evict_data = evict_arg;
1672 /* is this node still being used? */
1673 if (atomic_read(&node->refcount))
1674 return 0; /* keep this node */
1676 /* this node will be evicted, add its pages to our count */
1677 evict_data->cleared += node->npages;
1679 /* have enough pages been cleared? */
1680 if (evict_data->cleared >= evict_data->target)
1683 return 1; /* remove this node */
1686 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1688 struct sdma_mmu_node *node =
1689 container_of(mnode, struct sdma_mmu_node, rb);
1691 atomic_sub(node->npages, &node->pq->n_locked);
1693 unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
1698 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1700 struct sdma_mmu_node *node =
1701 container_of(mnode, struct sdma_mmu_node, rb);
1703 if (!atomic_read(&node->refcount))