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/pci.h>
52 #include <linux/netdevice.h>
53 #include <linux/vmalloc.h>
54 #include <linux/delay.h>
55 #include <linux/idr.h>
56 #include <linux/module.h>
57 #include <linux/printk.h>
58 #include <linux/hrtimer.h>
69 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
72 * min buffers we want to have per context, after driver
74 #define HFI1_MIN_USER_CTXT_BUFCNT 7
76 #define HFI1_MIN_HDRQ_EGRBUF_CNT 2
77 #define HFI1_MIN_EAGER_BUFFER_SIZE (4 * 1024) /* 4KB */
78 #define HFI1_MAX_EAGER_BUFFER_SIZE (256 * 1024) /* 256KB */
81 * Number of user receive contexts we are configured to use (to allow for more
82 * pio buffers per ctxt, etc.) Zero means use one user context per CPU.
84 uint num_rcv_contexts;
85 module_param_named(num_rcv_contexts, num_rcv_contexts, uint, S_IRUGO);
87 num_rcv_contexts, "Set max number of user receive contexts to use");
89 u8 krcvqs[RXE_NUM_DATA_VL];
91 module_param_array(krcvqs, byte, &krcvqsset, S_IRUGO);
92 MODULE_PARM_DESC(krcvqs, "Array of the number of kernel receive queues by VL");
94 /* computed based on above array */
97 static unsigned hfi1_rcvarr_split = 25;
98 module_param_named(rcvarr_split, hfi1_rcvarr_split, uint, S_IRUGO);
99 MODULE_PARM_DESC(rcvarr_split, "Percent of context's RcvArray entries used for Eager buffers");
101 static uint eager_buffer_size = (2 << 20); /* 2MB */
102 module_param(eager_buffer_size, uint, S_IRUGO);
103 MODULE_PARM_DESC(eager_buffer_size, "Size of the eager buffers, default: 2MB");
105 static uint rcvhdrcnt = 2048; /* 2x the max eager buffer count */
106 module_param_named(rcvhdrcnt, rcvhdrcnt, uint, S_IRUGO);
107 MODULE_PARM_DESC(rcvhdrcnt, "Receive header queue count (default 2048)");
109 static uint hfi1_hdrq_entsize = 32;
110 module_param_named(hdrq_entsize, hfi1_hdrq_entsize, uint, S_IRUGO);
111 MODULE_PARM_DESC(hdrq_entsize, "Size of header queue entries: 2 - 8B, 16 - 64B (default), 32 - 128B");
113 unsigned int user_credit_return_threshold = 33; /* default is 33% */
114 module_param(user_credit_return_threshold, uint, S_IRUGO);
115 MODULE_PARM_DESC(user_credit_return_theshold, "Credit return threshold for user send contexts, return when unreturned credits passes this many blocks (in percent of allocated blocks, 0 is off)");
117 static inline u64 encode_rcv_header_entry_size(u16);
119 static struct idr hfi1_unit_table;
120 u32 hfi1_cpulist_count;
121 unsigned long *hfi1_cpulist;
124 * Common code for creating the receive context array.
126 int hfi1_create_ctxts(struct hfi1_devdata *dd)
130 int local_node_id = pcibus_to_node(dd->pcidev->bus);
132 if (local_node_id < 0)
133 local_node_id = numa_node_id();
134 dd->assigned_node_id = local_node_id;
136 dd->rcd = kcalloc(dd->num_rcv_contexts, sizeof(*dd->rcd), GFP_KERNEL);
140 /* create one or more kernel contexts */
141 for (i = 0; i < dd->first_user_ctxt; ++i) {
142 struct hfi1_pportdata *ppd;
143 struct hfi1_ctxtdata *rcd;
145 ppd = dd->pport + (i % dd->num_pports);
146 rcd = hfi1_create_ctxtdata(ppd, i);
149 "Unable to allocate kernel receive context, failing\n");
153 * Set up the kernel context flags here and now because they
154 * use default values for all receive side memories. User
155 * contexts will be handled as they are created.
157 rcd->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
158 HFI1_CAP_KGET(NODROP_RHQ_FULL) |
159 HFI1_CAP_KGET(NODROP_EGR_FULL) |
160 HFI1_CAP_KGET(DMA_RTAIL);
163 rcd->sc = sc_alloc(dd, SC_ACK, rcd->rcvhdrqentsize, dd->node);
166 "Unable to allocate kernel send context, failing\n");
167 dd->rcd[rcd->ctxt] = NULL;
168 hfi1_free_ctxtdata(dd, rcd);
172 ret = hfi1_init_ctxt(rcd->sc);
175 "Failed to setup kernel receive context, failing\n");
177 dd->rcd[rcd->ctxt] = NULL;
178 hfi1_free_ctxtdata(dd, rcd);
194 * Common code for user and kernel context setup.
196 struct hfi1_ctxtdata *hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, u32 ctxt)
198 struct hfi1_devdata *dd = ppd->dd;
199 struct hfi1_ctxtdata *rcd;
200 unsigned kctxt_ngroups = 0;
203 if (dd->rcv_entries.nctxt_extra >
204 dd->num_rcv_contexts - dd->first_user_ctxt)
205 kctxt_ngroups = (dd->rcv_entries.nctxt_extra -
206 (dd->num_rcv_contexts - dd->first_user_ctxt));
207 rcd = kzalloc(sizeof(*rcd), GFP_KERNEL);
209 u32 rcvtids, max_entries;
211 dd_dev_info(dd, "%s: setting up context %u\n", __func__, ctxt);
213 INIT_LIST_HEAD(&rcd->qp_wait_list);
219 rcd->numa_id = numa_node_id();
220 rcd->rcv_array_groups = dd->rcv_entries.ngroups;
222 spin_lock_init(&rcd->exp_lock);
225 * Calculate the context's RcvArray entry starting point.
226 * We do this here because we have to take into account all
227 * the RcvArray entries that previous context would have
228 * taken and we have to account for any extra groups
229 * assigned to the kernel or user contexts.
231 if (ctxt < dd->first_user_ctxt) {
232 if (ctxt < kctxt_ngroups) {
233 base = ctxt * (dd->rcv_entries.ngroups + 1);
234 rcd->rcv_array_groups++;
236 base = kctxt_ngroups +
237 (ctxt * dd->rcv_entries.ngroups);
239 u16 ct = ctxt - dd->first_user_ctxt;
241 base = ((dd->n_krcv_queues * dd->rcv_entries.ngroups) +
243 if (ct < dd->rcv_entries.nctxt_extra) {
244 base += ct * (dd->rcv_entries.ngroups + 1);
245 rcd->rcv_array_groups++;
247 base += dd->rcv_entries.nctxt_extra +
248 (ct * dd->rcv_entries.ngroups);
250 rcd->eager_base = base * dd->rcv_entries.group_size;
252 /* Validate and initialize Rcv Hdr Q variables */
253 if (rcvhdrcnt % HDRQ_INCREMENT) {
255 "ctxt%u: header queue count %d must be divisible by %d\n",
256 rcd->ctxt, rcvhdrcnt, HDRQ_INCREMENT);
259 rcd->rcvhdrq_cnt = rcvhdrcnt;
260 rcd->rcvhdrqentsize = hfi1_hdrq_entsize;
262 * Simple Eager buffer allocation: we have already pre-allocated
263 * the number of RcvArray entry groups. Each ctxtdata structure
264 * holds the number of groups for that context.
266 * To follow CSR requirements and maintain cacheline alignment,
267 * make sure all sizes and bases are multiples of group_size.
269 * The expected entry count is what is left after assigning
272 max_entries = rcd->rcv_array_groups *
273 dd->rcv_entries.group_size;
274 rcvtids = ((max_entries * hfi1_rcvarr_split) / 100);
275 rcd->egrbufs.count = round_down(rcvtids,
276 dd->rcv_entries.group_size);
277 if (rcd->egrbufs.count > MAX_EAGER_ENTRIES) {
278 dd_dev_err(dd, "ctxt%u: requested too many RcvArray entries.\n",
280 rcd->egrbufs.count = MAX_EAGER_ENTRIES;
282 dd_dev_info(dd, "ctxt%u: max Eager buffer RcvArray entries: %u\n",
283 rcd->ctxt, rcd->egrbufs.count);
286 * Allocate array that will hold the eager buffer accounting
288 * This will allocate the maximum possible buffer count based
289 * on the value of the RcvArray split parameter.
290 * The resulting value will be rounded down to the closest
291 * multiple of dd->rcv_entries.group_size.
293 rcd->egrbufs.buffers = kcalloc(rcd->egrbufs.count,
294 sizeof(*rcd->egrbufs.buffers),
296 if (!rcd->egrbufs.buffers)
298 rcd->egrbufs.rcvtids = kcalloc(rcd->egrbufs.count,
299 sizeof(*rcd->egrbufs.rcvtids),
301 if (!rcd->egrbufs.rcvtids)
303 rcd->egrbufs.size = eager_buffer_size;
305 * The size of the buffers programmed into the RcvArray
306 * entries needs to be big enough to handle the highest
309 if (rcd->egrbufs.size < hfi1_max_mtu) {
310 rcd->egrbufs.size = __roundup_pow_of_two(hfi1_max_mtu);
312 "ctxt%u: eager bufs size too small. Adjusting to %zu\n",
313 rcd->ctxt, rcd->egrbufs.size);
315 rcd->egrbufs.rcvtid_size = HFI1_MAX_EAGER_BUFFER_SIZE;
317 if (ctxt < dd->first_user_ctxt) { /* N/A for PSM contexts */
318 rcd->opstats = kzalloc(sizeof(*rcd->opstats),
327 kfree(rcd->egrbufs.rcvtids);
328 kfree(rcd->egrbufs.buffers);
334 * Convert a receive header entry size that to the encoding used in the CSR.
336 * Return a zero if the given size is invalid.
338 static inline u64 encode_rcv_header_entry_size(u16 size)
340 /* there are only 3 valid receive header entry sizes */
347 return 0; /* invalid */
351 * Select the largest ccti value over all SLs to determine the intra-
352 * packet gap for the link.
354 * called with cca_timer_lock held (to protect access to cca_timer
355 * array), and rcu_read_lock() (to protect access to cc_state).
357 void set_link_ipg(struct hfi1_pportdata *ppd)
359 struct hfi1_devdata *dd = ppd->dd;
360 struct cc_state *cc_state;
362 u16 cce, ccti_limit, max_ccti = 0;
365 u32 current_egress_rate; /* Mbits /sec */
368 * max_pkt_time is the maximum packet egress time in units
369 * of the fabric clock period 1/(805 MHz).
372 cc_state = get_cc_state(ppd);
374 if (cc_state == NULL)
376 * This should _never_ happen - rcu_read_lock() is held,
377 * and set_link_ipg() should not be called if cc_state
382 for (i = 0; i < OPA_MAX_SLS; i++) {
383 u16 ccti = ppd->cca_timer[i].ccti;
389 ccti_limit = cc_state->cct.ccti_limit;
390 if (max_ccti > ccti_limit)
391 max_ccti = ccti_limit;
393 cce = cc_state->cct.entries[max_ccti].entry;
394 shift = (cce & 0xc000) >> 14;
395 mult = (cce & 0x3fff);
397 current_egress_rate = active_egress_rate(ppd);
399 max_pkt_time = egress_cycles(ppd->ibmaxlen, current_egress_rate);
401 src = (max_pkt_time >> shift) * mult;
403 src &= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SMASK;
404 src <<= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SHIFT;
406 write_csr(dd, SEND_STATIC_RATE_CONTROL, src);
409 static enum hrtimer_restart cca_timer_fn(struct hrtimer *t)
411 struct cca_timer *cca_timer;
412 struct hfi1_pportdata *ppd;
414 u16 ccti, ccti_timer, ccti_min;
415 struct cc_state *cc_state;
418 cca_timer = container_of(t, struct cca_timer, hrtimer);
419 ppd = cca_timer->ppd;
424 cc_state = get_cc_state(ppd);
426 if (cc_state == NULL) {
428 return HRTIMER_NORESTART;
432 * 1) decrement ccti for SL
433 * 2) calculate IPG for link (set_link_ipg())
434 * 3) restart timer, unless ccti is at min value
437 ccti_min = cc_state->cong_setting.entries[sl].ccti_min;
438 ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
440 spin_lock_irqsave(&ppd->cca_timer_lock, flags);
442 ccti = cca_timer->ccti;
444 if (ccti > ccti_min) {
449 spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
453 if (ccti > ccti_min) {
454 unsigned long nsec = 1024 * ccti_timer;
455 /* ccti_timer is in units of 1.024 usec */
456 hrtimer_forward_now(t, ns_to_ktime(nsec));
457 return HRTIMER_RESTART;
459 return HRTIMER_NORESTART;
463 * Common code for initializing the physical port structure.
465 void hfi1_init_pportdata(struct pci_dev *pdev, struct hfi1_pportdata *ppd,
466 struct hfi1_devdata *dd, u8 hw_pidx, u8 port)
469 uint default_pkey_idx;
472 ppd->hw_pidx = hw_pidx;
473 ppd->port = port; /* IB port number, not index */
475 default_pkey_idx = 1;
477 ppd->pkeys[default_pkey_idx] = DEFAULT_P_KEY;
479 hfi1_early_err(&pdev->dev,
480 "Faking data partition 0x8001 in idx %u\n",
482 ppd->pkeys[!default_pkey_idx] = 0x8001;
485 INIT_WORK(&ppd->link_vc_work, handle_verify_cap);
486 INIT_WORK(&ppd->link_up_work, handle_link_up);
487 INIT_WORK(&ppd->link_down_work, handle_link_down);
488 INIT_WORK(&ppd->freeze_work, handle_freeze);
489 INIT_WORK(&ppd->link_downgrade_work, handle_link_downgrade);
490 INIT_WORK(&ppd->sma_message_work, handle_sma_message);
491 INIT_WORK(&ppd->link_bounce_work, handle_link_bounce);
492 mutex_init(&ppd->hls_lock);
493 spin_lock_init(&ppd->sdma_alllock);
494 spin_lock_init(&ppd->qsfp_info.qsfp_lock);
496 ppd->sm_trap_qp = 0x0;
501 spin_lock_init(&ppd->cca_timer_lock);
503 for (i = 0; i < OPA_MAX_SLS; i++) {
504 hrtimer_init(&ppd->cca_timer[i].hrtimer, CLOCK_MONOTONIC,
506 ppd->cca_timer[i].ppd = ppd;
507 ppd->cca_timer[i].sl = i;
508 ppd->cca_timer[i].ccti = 0;
509 ppd->cca_timer[i].hrtimer.function = cca_timer_fn;
512 ppd->cc_max_table_entries = IB_CC_TABLE_CAP_DEFAULT;
514 spin_lock_init(&ppd->cc_state_lock);
515 spin_lock_init(&ppd->cc_log_lock);
516 size = sizeof(struct cc_state);
517 RCU_INIT_POINTER(ppd->cc_state, kzalloc(size, GFP_KERNEL));
518 if (!rcu_dereference(ppd->cc_state))
524 hfi1_early_err(&pdev->dev,
525 "Congestion Control Agent disabled for port %d\n", port);
529 * Do initialization for device that is only needed on
530 * first detect, not on resets.
532 static int loadtime_init(struct hfi1_devdata *dd)
538 * init_after_reset - re-initialize after a reset
539 * @dd: the hfi1_ib device
541 * sanity check at least some of the values after reset, and
542 * ensure no receive or transmit (explicitly, in case reset
545 static int init_after_reset(struct hfi1_devdata *dd)
550 * Ensure chip does no sends or receives, tail updates, or
551 * pioavail updates while we re-initialize. This is mostly
552 * for the driver data structures, not chip registers.
554 for (i = 0; i < dd->num_rcv_contexts; i++)
555 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
556 HFI1_RCVCTRL_INTRAVAIL_DIS |
557 HFI1_RCVCTRL_TAILUPD_DIS, i);
558 pio_send_control(dd, PSC_GLOBAL_DISABLE);
559 for (i = 0; i < dd->num_send_contexts; i++)
560 sc_disable(dd->send_contexts[i].sc);
565 static void enable_chip(struct hfi1_devdata *dd)
570 /* enable PIO send */
571 pio_send_control(dd, PSC_GLOBAL_ENABLE);
574 * Enable kernel ctxts' receive and receive interrupt.
575 * Other ctxts done as user opens and initializes them.
577 rcvmask = HFI1_RCVCTRL_CTXT_ENB | HFI1_RCVCTRL_INTRAVAIL_ENB;
578 for (i = 0; i < dd->first_user_ctxt; ++i) {
579 rcvmask |= HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, DMA_RTAIL) ?
580 HFI1_RCVCTRL_TAILUPD_ENB : HFI1_RCVCTRL_TAILUPD_DIS;
581 if (!HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, MULTI_PKT_EGR))
582 rcvmask |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
583 if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_RHQ_FULL))
584 rcvmask |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
585 if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_EGR_FULL))
586 rcvmask |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
587 hfi1_rcvctrl(dd, rcvmask, i);
588 sc_enable(dd->rcd[i]->sc);
593 * create_workqueues - create per port workqueues
594 * @dd: the hfi1_ib device
596 static int create_workqueues(struct hfi1_devdata *dd)
599 struct hfi1_pportdata *ppd;
601 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
602 ppd = dd->pport + pidx;
604 char wq_name[8]; /* 3 + 2 + 1 + 1 + 1 */
606 snprintf(wq_name, sizeof(wq_name), "hfi%d_%d",
609 create_singlethread_workqueue(wq_name);
616 pr_err("create_singlethread_workqueue failed for port %d\n",
618 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
619 ppd = dd->pport + pidx;
621 destroy_workqueue(ppd->hfi1_wq);
629 * hfi1_init - do the actual initialization sequence on the chip
630 * @dd: the hfi1_ib device
631 * @reinit: re-initializing, so don't allocate new memory
633 * Do the actual initialization sequence on the chip. This is done
634 * both from the init routine called from the PCI infrastructure, and
635 * when we reset the chip, or detect that it was reset internally,
636 * or it's administratively re-enabled.
638 * Memory allocation here and in called routines is only done in
639 * the first case (reinit == 0). We have to be careful, because even
640 * without memory allocation, we need to re-write all the chip registers
641 * TIDs, etc. after the reset or enable has completed.
643 int hfi1_init(struct hfi1_devdata *dd, int reinit)
645 int ret = 0, pidx, lastfail = 0;
647 struct hfi1_ctxtdata *rcd;
648 struct hfi1_pportdata *ppd;
650 /* Set up recv low level handlers */
651 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_EXPECTED] =
652 kdeth_process_expected;
653 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_EAGER] =
655 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_IB] = process_receive_ib;
656 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_ERROR] =
657 process_receive_error;
658 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_BYPASS] =
659 process_receive_bypass;
660 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID5] =
661 process_receive_invalid;
662 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID6] =
663 process_receive_invalid;
664 dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID7] =
665 process_receive_invalid;
666 dd->rhf_rcv_function_map = dd->normal_rhf_rcv_functions;
668 /* Set up send low level handlers */
669 dd->process_pio_send = hfi1_verbs_send_pio;
670 dd->process_dma_send = hfi1_verbs_send_dma;
671 dd->pio_inline_send = pio_copy;
674 atomic_set(&dd->drop_packet, DROP_PACKET_ON);
677 atomic_set(&dd->drop_packet, DROP_PACKET_OFF);
681 /* make sure the link is not "up" */
682 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
683 ppd = dd->pport + pidx;
688 ret = init_after_reset(dd);
690 ret = loadtime_init(dd);
694 /* dd->rcd can be NULL if early initialization failed */
695 for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
697 * Set up the (kernel) rcvhdr queue and egr TIDs. If doing
698 * re-init, the simplest way to handle this is to free
699 * existing, and re-allocate.
700 * Need to re-create rest of ctxt 0 ctxtdata as well.
706 rcd->do_interrupt = &handle_receive_interrupt;
708 lastfail = hfi1_create_rcvhdrq(dd, rcd);
710 lastfail = hfi1_setup_eagerbufs(rcd);
713 "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
718 /* Allocate enough memory for user event notification. */
719 len = ALIGN(dd->chip_rcv_contexts * HFI1_MAX_SHARED_CTXTS *
720 sizeof(*dd->events), PAGE_SIZE);
721 dd->events = vmalloc_user(len);
723 dd_dev_err(dd, "Failed to allocate user events page\n");
725 * Allocate a page for device and port status.
726 * Page will be shared amongst all user processes.
728 dd->status = vmalloc_user(PAGE_SIZE);
730 dd_dev_err(dd, "Failed to allocate dev status page\n");
732 dd->freezelen = PAGE_SIZE - (sizeof(*dd->status) -
733 sizeof(dd->status->freezemsg));
734 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
735 ppd = dd->pport + pidx;
737 /* Currently, we only have one port */
738 ppd->statusp = &dd->status->port;
743 /* enable chip even if we have an error, so we can debug cause */
746 ret = hfi1_cq_init(dd);
749 * Set status even if port serdes is not initialized
750 * so that diags will work.
753 dd->status->dev |= HFI1_STATUS_CHIP_PRESENT |
756 /* enable all interrupts from the chip */
757 set_intr_state(dd, 1);
759 /* chip is OK for user apps; mark it as initialized */
760 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
761 ppd = dd->pport + pidx;
763 /* initialize the qsfp if it exists
764 * Requires interrupts to be enabled so we are notified
765 * when the QSFP completes reset, and has
766 * to be done before bringing up the SERDES
770 /* start the serdes - must be after interrupts are
771 enabled so we are notified when the link goes up */
772 lastfail = bringup_serdes(ppd);
775 "Failed to bring up port %u\n",
779 * Set status even if port serdes is not initialized
780 * so that diags will work.
783 *ppd->statusp |= HFI1_STATUS_CHIP_PRESENT |
785 if (!ppd->link_speed_enabled)
790 /* if ret is non-zero, we probably should do some cleanup here... */
794 static inline struct hfi1_devdata *__hfi1_lookup(int unit)
796 return idr_find(&hfi1_unit_table, unit);
799 struct hfi1_devdata *hfi1_lookup(int unit)
801 struct hfi1_devdata *dd;
804 spin_lock_irqsave(&hfi1_devs_lock, flags);
805 dd = __hfi1_lookup(unit);
806 spin_unlock_irqrestore(&hfi1_devs_lock, flags);
812 * Stop the timers during unit shutdown, or after an error late
815 static void stop_timers(struct hfi1_devdata *dd)
817 struct hfi1_pportdata *ppd;
820 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
821 ppd = dd->pport + pidx;
822 if (ppd->led_override_timer.data) {
823 del_timer_sync(&ppd->led_override_timer);
824 atomic_set(&ppd->led_override_timer_active, 0);
830 * shutdown_device - shut down a device
831 * @dd: the hfi1_ib device
833 * This is called to make the device quiet when we are about to
834 * unload the driver, and also when the device is administratively
835 * disabled. It does not free any data structures.
836 * Everything it does has to be setup again by hfi1_init(dd, 1)
838 static void shutdown_device(struct hfi1_devdata *dd)
840 struct hfi1_pportdata *ppd;
844 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
845 ppd = dd->pport + pidx;
849 *ppd->statusp &= ~(HFI1_STATUS_IB_CONF |
850 HFI1_STATUS_IB_READY);
852 dd->flags &= ~HFI1_INITTED;
854 /* mask interrupts, but not errors */
855 set_intr_state(dd, 0);
857 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
858 ppd = dd->pport + pidx;
859 for (i = 0; i < dd->num_rcv_contexts; i++)
860 hfi1_rcvctrl(dd, HFI1_RCVCTRL_TAILUPD_DIS |
861 HFI1_RCVCTRL_CTXT_DIS |
862 HFI1_RCVCTRL_INTRAVAIL_DIS |
863 HFI1_RCVCTRL_PKEY_DIS |
864 HFI1_RCVCTRL_ONE_PKT_EGR_DIS, i);
866 * Gracefully stop all sends allowing any in progress to
869 for (i = 0; i < dd->num_send_contexts; i++)
870 sc_flush(dd->send_contexts[i].sc);
874 * Enough for anything that's going to trickle out to have actually
879 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
880 ppd = dd->pport + pidx;
882 /* disable all contexts */
883 for (i = 0; i < dd->num_send_contexts; i++)
884 sc_disable(dd->send_contexts[i].sc);
885 /* disable the send device */
886 pio_send_control(dd, PSC_GLOBAL_DISABLE);
889 * Clear SerdesEnable.
890 * We can't count on interrupts since we are stopping.
892 hfi1_quiet_serdes(ppd);
895 destroy_workqueue(ppd->hfi1_wq);
903 * hfi1_free_ctxtdata - free a context's allocated data
904 * @dd: the hfi1_ib device
905 * @rcd: the ctxtdata structure
907 * free up any allocated data for a context
908 * This should not touch anything that would affect a simultaneous
909 * re-allocation of context data, because it is called after hfi1_mutex
910 * is released (and can be called from reinit as well).
911 * It should never change any chip state, or global driver state.
913 void hfi1_free_ctxtdata(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
921 dma_free_coherent(&dd->pcidev->dev, rcd->rcvhdrq_size,
922 rcd->rcvhdrq, rcd->rcvhdrq_phys);
924 if (rcd->rcvhdrtail_kvaddr) {
925 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
926 (void *)rcd->rcvhdrtail_kvaddr,
927 rcd->rcvhdrqtailaddr_phys);
928 rcd->rcvhdrtail_kvaddr = NULL;
932 /* all the RcvArray entries should have been cleared by now */
933 kfree(rcd->egrbufs.rcvtids);
935 for (e = 0; e < rcd->egrbufs.alloced; e++) {
936 if (rcd->egrbufs.buffers[e].phys)
937 dma_free_coherent(&dd->pcidev->dev,
938 rcd->egrbufs.buffers[e].len,
939 rcd->egrbufs.buffers[e].addr,
940 rcd->egrbufs.buffers[e].phys);
942 kfree(rcd->egrbufs.buffers);
945 vfree(rcd->physshadow);
946 vfree(rcd->tid_pg_list);
947 vfree(rcd->user_event_mask);
948 vfree(rcd->subctxt_uregbase);
949 vfree(rcd->subctxt_rcvegrbuf);
950 vfree(rcd->subctxt_rcvhdr_base);
951 kfree(rcd->tidusemap);
956 void hfi1_free_devdata(struct hfi1_devdata *dd)
960 spin_lock_irqsave(&hfi1_devs_lock, flags);
961 idr_remove(&hfi1_unit_table, dd->unit);
963 spin_unlock_irqrestore(&hfi1_devs_lock, flags);
964 hfi1_dbg_ibdev_exit(&dd->verbs_dev);
965 rcu_barrier(); /* wait for rcu callbacks to complete */
966 free_percpu(dd->int_counter);
967 free_percpu(dd->rcv_limit);
968 ib_dealloc_device(&dd->verbs_dev.ibdev);
972 * Allocate our primary per-unit data structure. Must be done via verbs
973 * allocator, because the verbs cleanup process both does cleanup and
974 * free of the data structure.
975 * "extra" is for chip-specific data.
977 * Use the idr mechanism to get a unit number for this unit.
979 struct hfi1_devdata *hfi1_alloc_devdata(struct pci_dev *pdev, size_t extra)
982 struct hfi1_devdata *dd;
985 dd = (struct hfi1_devdata *)ib_alloc_device(sizeof(*dd) + extra);
987 return ERR_PTR(-ENOMEM);
988 /* extra is * number of ports */
989 dd->num_pports = extra / sizeof(struct hfi1_pportdata);
990 dd->pport = (struct hfi1_pportdata *)(dd + 1);
992 INIT_LIST_HEAD(&dd->list);
993 dd->node = dev_to_node(&pdev->dev);
996 idr_preload(GFP_KERNEL);
997 spin_lock_irqsave(&hfi1_devs_lock, flags);
999 ret = idr_alloc(&hfi1_unit_table, dd, 0, 0, GFP_NOWAIT);
1002 list_add(&dd->list, &hfi1_dev_list);
1005 spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1009 hfi1_early_err(&pdev->dev,
1010 "Could not allocate unit ID: error %d\n", -ret);
1014 * Initialize all locks for the device. This needs to be as early as
1015 * possible so locks are usable.
1017 spin_lock_init(&dd->sc_lock);
1018 spin_lock_init(&dd->sendctrl_lock);
1019 spin_lock_init(&dd->rcvctrl_lock);
1020 spin_lock_init(&dd->uctxt_lock);
1021 spin_lock_init(&dd->hfi1_diag_trans_lock);
1022 spin_lock_init(&dd->sc_init_lock);
1023 spin_lock_init(&dd->dc8051_lock);
1024 spin_lock_init(&dd->dc8051_memlock);
1025 mutex_init(&dd->qsfp_i2c_mutex);
1026 seqlock_init(&dd->sc2vl_lock);
1027 spin_lock_init(&dd->sde_map_lock);
1028 init_waitqueue_head(&dd->event_queue);
1030 dd->int_counter = alloc_percpu(u64);
1031 if (!dd->int_counter) {
1033 hfi1_early_err(&pdev->dev,
1034 "Could not allocate per-cpu int_counter\n");
1038 dd->rcv_limit = alloc_percpu(u64);
1039 if (!dd->rcv_limit) {
1041 hfi1_early_err(&pdev->dev,
1042 "Could not allocate per-cpu rcv_limit\n");
1046 if (!hfi1_cpulist_count) {
1047 u32 count = num_online_cpus();
1049 hfi1_cpulist = kcalloc(BITS_TO_LONGS(count), sizeof(long),
1052 hfi1_cpulist_count = count;
1056 "Could not alloc cpulist info, cpu affinity might be wrong\n");
1058 hfi1_dbg_ibdev_init(&dd->verbs_dev);
1062 if (!list_empty(&dd->list))
1063 list_del_init(&dd->list);
1064 ib_dealloc_device(&dd->verbs_dev.ibdev);
1065 return ERR_PTR(ret);
1069 * Called from freeze mode handlers, and from PCI error
1070 * reporting code. Should be paranoid about state of
1071 * system and data structures.
1073 void hfi1_disable_after_error(struct hfi1_devdata *dd)
1075 if (dd->flags & HFI1_INITTED) {
1078 dd->flags &= ~HFI1_INITTED;
1080 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1081 struct hfi1_pportdata *ppd;
1083 ppd = dd->pport + pidx;
1084 if (dd->flags & HFI1_PRESENT)
1085 set_link_state(ppd, HLS_DN_DISABLE);
1088 *ppd->statusp &= ~HFI1_STATUS_IB_READY;
1093 * Mark as having had an error for driver, and also
1094 * for /sys and status word mapped to user programs.
1095 * This marks unit as not usable, until reset.
1098 dd->status->dev |= HFI1_STATUS_HWERROR;
1101 static void remove_one(struct pci_dev *);
1102 static int init_one(struct pci_dev *, const struct pci_device_id *);
1104 #define DRIVER_LOAD_MSG "Intel " DRIVER_NAME " loaded: "
1105 #define PFX DRIVER_NAME ": "
1107 static const struct pci_device_id hfi1_pci_tbl[] = {
1108 { PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL0) },
1109 { PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL1) },
1113 MODULE_DEVICE_TABLE(pci, hfi1_pci_tbl);
1115 static struct pci_driver hfi1_pci_driver = {
1116 .name = DRIVER_NAME,
1118 .remove = remove_one,
1119 .id_table = hfi1_pci_tbl,
1120 .err_handler = &hfi1_pci_err_handler,
1123 static void __init compute_krcvqs(void)
1127 for (i = 0; i < krcvqsset; i++)
1128 n_krcvqs += krcvqs[i];
1132 * Do all the generic driver unit- and chip-independent memory
1133 * allocation and initialization.
1135 static int __init hfi1_mod_init(void)
1143 /* validate max MTU before any devices start */
1144 if (!valid_opa_max_mtu(hfi1_max_mtu)) {
1145 pr_err("Invalid max_mtu 0x%x, using 0x%x instead\n",
1146 hfi1_max_mtu, HFI1_DEFAULT_MAX_MTU);
1147 hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
1149 /* valid CUs run from 1-128 in powers of 2 */
1150 if (hfi1_cu > 128 || !is_power_of_2(hfi1_cu))
1152 /* valid credit return threshold is 0-100, variable is unsigned */
1153 if (user_credit_return_threshold > 100)
1154 user_credit_return_threshold = 100;
1157 /* sanitize receive interrupt count, time must wait until after
1158 the hardware type is known */
1159 if (rcv_intr_count > RCV_HDR_HEAD_COUNTER_MASK)
1160 rcv_intr_count = RCV_HDR_HEAD_COUNTER_MASK;
1161 /* reject invalid combinations */
1162 if (rcv_intr_count == 0 && rcv_intr_timeout == 0) {
1163 pr_err("Invalid mode: both receive interrupt count and available timeout are zero - setting interrupt count to 1\n");
1166 if (rcv_intr_count > 1 && rcv_intr_timeout == 0) {
1168 * Avoid indefinite packet delivery by requiring a timeout
1171 pr_err("Invalid mode: receive interrupt count greater than 1 and available timeout is zero - setting available timeout to 1\n");
1172 rcv_intr_timeout = 1;
1174 if (rcv_intr_dynamic && !(rcv_intr_count > 1 && rcv_intr_timeout > 0)) {
1176 * The dynamic algorithm expects a non-zero timeout
1179 pr_err("Invalid mode: dynamic receive interrupt mitigation with invalid count and timeout - turning dynamic off\n");
1180 rcv_intr_dynamic = 0;
1183 /* sanitize link CRC options */
1184 link_crc_mask &= SUPPORTED_CRCS;
1187 * These must be called before the driver is registered with
1188 * the PCI subsystem.
1190 idr_init(&hfi1_unit_table);
1193 ret = pci_register_driver(&hfi1_pci_driver);
1195 pr_err("Unable to register driver: error %d\n", -ret);
1198 goto bail; /* all OK */
1202 idr_destroy(&hfi1_unit_table);
1208 module_init(hfi1_mod_init);
1211 * Do the non-unit driver cleanup, memory free, etc. at unload.
1213 static void __exit hfi1_mod_cleanup(void)
1215 pci_unregister_driver(&hfi1_pci_driver);
1217 hfi1_cpulist_count = 0;
1218 kfree(hfi1_cpulist);
1220 idr_destroy(&hfi1_unit_table);
1221 dispose_firmware(); /* asymmetric with obtain_firmware() */
1225 module_exit(hfi1_mod_cleanup);
1227 /* this can only be called after a successful initialization */
1228 static void cleanup_device_data(struct hfi1_devdata *dd)
1232 struct hfi1_ctxtdata **tmp;
1233 unsigned long flags;
1235 /* users can't do anything more with chip */
1236 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1237 struct hfi1_pportdata *ppd = &dd->pport[pidx];
1238 struct cc_state *cc_state;
1242 *ppd->statusp &= ~HFI1_STATUS_CHIP_PRESENT;
1244 for (i = 0; i < OPA_MAX_SLS; i++)
1245 hrtimer_cancel(&ppd->cca_timer[i].hrtimer);
1247 spin_lock(&ppd->cc_state_lock);
1248 cc_state = get_cc_state(ppd);
1249 rcu_assign_pointer(ppd->cc_state, NULL);
1250 spin_unlock(&ppd->cc_state_lock);
1253 call_rcu(&cc_state->rcu, cc_state_reclaim);
1256 free_credit_return(dd);
1259 * Free any resources still in use (usually just kernel contexts)
1260 * at unload; we do for ctxtcnt, because that's what we allocate.
1261 * We acquire lock to be really paranoid that rcd isn't being
1262 * accessed from some interrupt-related code (that should not happen,
1263 * but best to be sure).
1265 spin_lock_irqsave(&dd->uctxt_lock, flags);
1268 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1269 for (ctxt = 0; tmp && ctxt < dd->num_rcv_contexts; ctxt++) {
1270 struct hfi1_ctxtdata *rcd = tmp[ctxt];
1272 tmp[ctxt] = NULL; /* debugging paranoia */
1274 hfi1_clear_tids(rcd);
1275 hfi1_free_ctxtdata(dd, rcd);
1279 /* must follow rcv context free - need to remove rcv's hooks */
1280 for (ctxt = 0; ctxt < dd->num_send_contexts; ctxt++)
1281 sc_free(dd->send_contexts[ctxt].sc);
1282 dd->num_send_contexts = 0;
1283 kfree(dd->send_contexts);
1284 dd->send_contexts = NULL;
1285 kfree(dd->boardname);
1292 * Clean up on unit shutdown, or error during unit load after
1293 * successful initialization.
1295 static void postinit_cleanup(struct hfi1_devdata *dd)
1297 hfi1_start_cleanup(dd);
1299 hfi1_pcie_ddcleanup(dd);
1300 hfi1_pcie_cleanup(dd->pcidev);
1302 cleanup_device_data(dd);
1304 hfi1_free_devdata(dd);
1307 static int init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
1309 int ret = 0, j, pidx, initfail;
1310 struct hfi1_devdata *dd = NULL;
1312 /* First, lock the non-writable module parameters */
1315 /* Validate some global module parameters */
1316 if (rcvhdrcnt <= HFI1_MIN_HDRQ_EGRBUF_CNT) {
1317 hfi1_early_err(&pdev->dev, "Header queue count too small\n");
1321 /* use the encoding function as a sanitization check */
1322 if (!encode_rcv_header_entry_size(hfi1_hdrq_entsize)) {
1323 hfi1_early_err(&pdev->dev, "Invalid HdrQ Entry size %u\n",
1328 /* The receive eager buffer size must be set before the receive
1329 * contexts are created.
1331 * Set the eager buffer size. Validate that it falls in a range
1332 * allowed by the hardware - all powers of 2 between the min and
1333 * max. The maximum valid MTU is within the eager buffer range
1334 * so we do not need to cap the max_mtu by an eager buffer size
1337 if (eager_buffer_size) {
1338 if (!is_power_of_2(eager_buffer_size))
1340 roundup_pow_of_two(eager_buffer_size);
1342 clamp_val(eager_buffer_size,
1343 MIN_EAGER_BUFFER * 8,
1344 MAX_EAGER_BUFFER_TOTAL);
1345 hfi1_early_info(&pdev->dev, "Eager buffer size %u\n",
1348 hfi1_early_err(&pdev->dev, "Invalid Eager buffer size of 0\n");
1353 /* restrict value of hfi1_rcvarr_split */
1354 hfi1_rcvarr_split = clamp_val(hfi1_rcvarr_split, 0, 100);
1356 ret = hfi1_pcie_init(pdev, ent);
1361 * Do device-specific initialization, function table setup, dd
1364 switch (ent->device) {
1365 case PCI_DEVICE_ID_INTEL0:
1366 case PCI_DEVICE_ID_INTEL1:
1367 dd = hfi1_init_dd(pdev, ent);
1370 hfi1_early_err(&pdev->dev,
1371 "Failing on unknown Intel deviceid 0x%x\n",
1379 goto clean_bail; /* error already printed */
1381 ret = create_workqueues(dd);
1385 /* do the generic initialization */
1386 initfail = hfi1_init(dd, 0);
1388 ret = hfi1_register_ib_device(dd);
1391 * Now ready for use. this should be cleared whenever we
1392 * detect a reset, or initiate one. If earlier failure,
1393 * we still create devices, so diags, etc. can be used
1394 * to determine cause of problem.
1396 if (!initfail && !ret)
1397 dd->flags |= HFI1_INITTED;
1399 j = hfi1_device_create(dd);
1401 dd_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1403 if (initfail || ret) {
1405 flush_workqueue(ib_wq);
1406 for (pidx = 0; pidx < dd->num_pports; ++pidx)
1407 hfi1_quiet_serdes(dd->pport + pidx);
1409 hfi1_device_remove(dd);
1411 hfi1_unregister_ib_device(dd);
1412 postinit_cleanup(dd);
1415 goto bail; /* everything already cleaned */
1423 hfi1_pcie_cleanup(pdev);
1428 static void remove_one(struct pci_dev *pdev)
1430 struct hfi1_devdata *dd = pci_get_drvdata(pdev);
1432 /* unregister from IB core */
1433 hfi1_unregister_ib_device(dd);
1436 * Disable the IB link, disable interrupts on the device,
1437 * clear dma engines, etc.
1439 shutdown_device(dd);
1443 /* wait until all of our (qsfp) queue_work() calls complete */
1444 flush_workqueue(ib_wq);
1446 hfi1_device_remove(dd);
1448 postinit_cleanup(dd);
1452 * hfi1_create_rcvhdrq - create a receive header queue
1453 * @dd: the hfi1_ib device
1454 * @rcd: the context data
1456 * This must be contiguous memory (from an i/o perspective), and must be
1457 * DMA'able (which means for some systems, it will go through an IOMMU,
1458 * or be forced into a low address range).
1460 int hfi1_create_rcvhdrq(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
1465 if (!rcd->rcvhdrq) {
1466 dma_addr_t phys_hdrqtail;
1470 * rcvhdrqentsize is in DWs, so we have to convert to bytes
1473 amt = ALIGN(rcd->rcvhdrq_cnt * rcd->rcvhdrqentsize *
1474 sizeof(u32), PAGE_SIZE);
1476 gfp_flags = (rcd->ctxt >= dd->first_user_ctxt) ?
1477 GFP_USER : GFP_KERNEL;
1478 rcd->rcvhdrq = dma_zalloc_coherent(
1479 &dd->pcidev->dev, amt, &rcd->rcvhdrq_phys,
1480 gfp_flags | __GFP_COMP);
1482 if (!rcd->rcvhdrq) {
1484 "attempt to allocate %d bytes for ctxt %u rcvhdrq failed\n",
1489 /* Event mask is per device now and is in hfi1_devdata */
1490 /*if (rcd->ctxt >= dd->first_user_ctxt) {
1491 rcd->user_event_mask = vmalloc_user(PAGE_SIZE);
1492 if (!rcd->user_event_mask)
1493 goto bail_free_hdrq;
1496 if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
1497 rcd->rcvhdrtail_kvaddr = dma_zalloc_coherent(
1498 &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
1500 if (!rcd->rcvhdrtail_kvaddr)
1502 rcd->rcvhdrqtailaddr_phys = phys_hdrqtail;
1505 rcd->rcvhdrq_size = amt;
1508 * These values are per-context:
1513 reg = ((u64)(rcd->rcvhdrq_cnt >> HDRQ_SIZE_SHIFT)
1514 & RCV_HDR_CNT_CNT_MASK)
1515 << RCV_HDR_CNT_CNT_SHIFT;
1516 write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_CNT, reg);
1517 reg = (encode_rcv_header_entry_size(rcd->rcvhdrqentsize)
1518 & RCV_HDR_ENT_SIZE_ENT_SIZE_MASK)
1519 << RCV_HDR_ENT_SIZE_ENT_SIZE_SHIFT;
1520 write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_ENT_SIZE, reg);
1521 reg = (dd->rcvhdrsize & RCV_HDR_SIZE_HDR_SIZE_MASK)
1522 << RCV_HDR_SIZE_HDR_SIZE_SHIFT;
1523 write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_SIZE, reg);
1528 "attempt to allocate 1 page for ctxt %u rcvhdrqtailaddr failed\n",
1530 vfree(rcd->user_event_mask);
1531 rcd->user_event_mask = NULL;
1532 dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1534 rcd->rcvhdrq = NULL;
1540 * allocate eager buffers, both kernel and user contexts.
1541 * @rcd: the context we are setting up.
1543 * Allocate the eager TID buffers and program them into hip.
1544 * They are no longer completely contiguous, we do multiple allocation
1545 * calls. Otherwise we get the OOM code involved, by asking for too
1546 * much per call, with disastrous results on some kernels.
1548 int hfi1_setup_eagerbufs(struct hfi1_ctxtdata *rcd)
1550 struct hfi1_devdata *dd = rcd->dd;
1551 u32 max_entries, egrtop, alloced_bytes = 0, idx = 0;
1555 u16 round_mtu = roundup_pow_of_two(hfi1_max_mtu);
1558 * GFP_USER, but without GFP_FS, so buffer cache can be
1559 * coalesced (we hope); otherwise, even at order 4,
1560 * heavy filesystem activity makes these fail, and we can
1561 * use compound pages.
1563 gfp_flags = __GFP_RECLAIM | __GFP_IO | __GFP_COMP;
1566 * The minimum size of the eager buffers is a groups of MTU-sized
1568 * The global eager_buffer_size parameter is checked against the
1569 * theoretical lower limit of the value. Here, we check against the
1572 if (rcd->egrbufs.size < (round_mtu * dd->rcv_entries.group_size))
1573 rcd->egrbufs.size = round_mtu * dd->rcv_entries.group_size;
1575 * If using one-pkt-per-egr-buffer, lower the eager buffer
1576 * size to the max MTU (page-aligned).
1578 if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
1579 rcd->egrbufs.rcvtid_size = round_mtu;
1582 * Eager buffers sizes of 1MB or less require smaller TID sizes
1583 * to satisfy the "multiple of 8 RcvArray entries" requirement.
1585 if (rcd->egrbufs.size <= (1 << 20))
1586 rcd->egrbufs.rcvtid_size = max((unsigned long)round_mtu,
1587 rounddown_pow_of_two(rcd->egrbufs.size / 8));
1589 while (alloced_bytes < rcd->egrbufs.size &&
1590 rcd->egrbufs.alloced < rcd->egrbufs.count) {
1591 rcd->egrbufs.buffers[idx].addr =
1592 dma_zalloc_coherent(&dd->pcidev->dev,
1593 rcd->egrbufs.rcvtid_size,
1594 &rcd->egrbufs.buffers[idx].phys,
1596 if (rcd->egrbufs.buffers[idx].addr) {
1597 rcd->egrbufs.buffers[idx].len =
1598 rcd->egrbufs.rcvtid_size;
1599 rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].addr =
1600 rcd->egrbufs.buffers[idx].addr;
1601 rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].phys =
1602 rcd->egrbufs.buffers[idx].phys;
1603 rcd->egrbufs.alloced++;
1604 alloced_bytes += rcd->egrbufs.rcvtid_size;
1611 * Fail the eager buffer allocation if:
1612 * - we are already using the lowest acceptable size
1613 * - we are using one-pkt-per-egr-buffer (this implies
1614 * that we are accepting only one size)
1616 if (rcd->egrbufs.rcvtid_size == round_mtu ||
1617 !HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR)) {
1618 dd_dev_err(dd, "ctxt%u: Failed to allocate eager buffers\n",
1620 goto bail_rcvegrbuf_phys;
1623 new_size = rcd->egrbufs.rcvtid_size / 2;
1626 * If the first attempt to allocate memory failed, don't
1627 * fail everything but continue with the next lower
1631 rcd->egrbufs.rcvtid_size = new_size;
1636 * Re-partition already allocated buffers to a smaller
1639 rcd->egrbufs.alloced = 0;
1640 for (i = 0, j = 0, offset = 0; j < idx; i++) {
1641 if (i >= rcd->egrbufs.count)
1643 rcd->egrbufs.rcvtids[i].phys =
1644 rcd->egrbufs.buffers[j].phys + offset;
1645 rcd->egrbufs.rcvtids[i].addr =
1646 rcd->egrbufs.buffers[j].addr + offset;
1647 rcd->egrbufs.alloced++;
1648 if ((rcd->egrbufs.buffers[j].phys + offset +
1650 (rcd->egrbufs.buffers[j].phys +
1651 rcd->egrbufs.buffers[j].len)) {
1657 rcd->egrbufs.rcvtid_size = new_size;
1660 rcd->egrbufs.numbufs = idx;
1661 rcd->egrbufs.size = alloced_bytes;
1663 dd_dev_info(dd, "ctxt%u: Alloced %u rcv tid entries @ %uKB, total %zuKB\n",
1664 rcd->ctxt, rcd->egrbufs.alloced, rcd->egrbufs.rcvtid_size,
1668 * Set the contexts rcv array head update threshold to the closest
1669 * power of 2 (so we can use a mask instead of modulo) below half
1670 * the allocated entries.
1672 rcd->egrbufs.threshold =
1673 rounddown_pow_of_two(rcd->egrbufs.alloced / 2);
1675 * Compute the expected RcvArray entry base. This is done after
1676 * allocating the eager buffers in order to maximize the
1677 * expected RcvArray entries for the context.
1679 max_entries = rcd->rcv_array_groups * dd->rcv_entries.group_size;
1680 egrtop = roundup(rcd->egrbufs.alloced, dd->rcv_entries.group_size);
1681 rcd->expected_count = max_entries - egrtop;
1682 if (rcd->expected_count > MAX_TID_PAIR_ENTRIES * 2)
1683 rcd->expected_count = MAX_TID_PAIR_ENTRIES * 2;
1685 rcd->expected_base = rcd->eager_base + egrtop;
1686 dd_dev_info(dd, "ctxt%u: eager:%u, exp:%u, egrbase:%u, expbase:%u\n",
1687 rcd->ctxt, rcd->egrbufs.alloced, rcd->expected_count,
1688 rcd->eager_base, rcd->expected_base);
1690 if (!hfi1_rcvbuf_validate(rcd->egrbufs.rcvtid_size, PT_EAGER, &order)) {
1691 dd_dev_err(dd, "ctxt%u: current Eager buffer size is invalid %u\n",
1692 rcd->ctxt, rcd->egrbufs.rcvtid_size);
1697 for (idx = 0; idx < rcd->egrbufs.alloced; idx++) {
1698 hfi1_put_tid(dd, rcd->eager_base + idx, PT_EAGER,
1699 rcd->egrbufs.rcvtids[idx].phys, order);
1704 bail_rcvegrbuf_phys:
1705 for (idx = 0; idx < rcd->egrbufs.alloced &&
1706 rcd->egrbufs.buffers[idx].addr;
1708 dma_free_coherent(&dd->pcidev->dev,
1709 rcd->egrbufs.buffers[idx].len,
1710 rcd->egrbufs.buffers[idx].addr,
1711 rcd->egrbufs.buffers[idx].phys);
1712 rcd->egrbufs.buffers[idx].addr = NULL;
1713 rcd->egrbufs.buffers[idx].phys = 0;
1714 rcd->egrbufs.buffers[idx].len = 0;