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, 2016 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, 2016 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.
52 * This file contains support for diagnostic functions. It is accessed by
53 * opening the hfi1_diag device, normally minor number 129. Diagnostic use
54 * of the chip may render the chip or board unusable until the driver
55 * is unloaded, or in some cases, until the system is rebooted.
57 * Accesses to the chip through this interface are not similar to going
58 * through the /sys/bus/pci resource mmap interface.
62 #include <linux/pci.h>
63 #include <linux/poll.h>
64 #include <linux/vmalloc.h>
65 #include <linux/export.h>
67 #include <linux/uaccess.h>
68 #include <linux/module.h>
69 #include <rdma/ib_smi.h>
73 #include "verbs_txreq.h"
77 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
78 #define snoop_dbg(fmt, ...) \
79 hfi1_cdbg(SNOOP, fmt, ##__VA_ARGS__)
81 /* Snoop option mask */
82 #define SNOOP_DROP_SEND BIT(0)
83 #define SNOOP_USE_METADATA BIT(1)
84 #define SNOOP_SET_VL0TOVL15 BIT(2)
86 static u8 snoop_flags;
89 * Extract packet length from LRH header.
90 * This is in Dwords so multiply by 4 to get size in bytes
92 #define HFI1_GET_PKT_LEN(x) (((be16_to_cpu((x)->lrh[2]) & 0xFFF)) << 2)
94 enum hfi1_filter_status {
100 /* snoop processing functions */
101 rhf_rcv_function_ptr snoop_rhf_rcv_functions[8] = {
102 [RHF_RCV_TYPE_EXPECTED] = snoop_recv_handler,
103 [RHF_RCV_TYPE_EAGER] = snoop_recv_handler,
104 [RHF_RCV_TYPE_IB] = snoop_recv_handler,
105 [RHF_RCV_TYPE_ERROR] = snoop_recv_handler,
106 [RHF_RCV_TYPE_BYPASS] = snoop_recv_handler,
107 [RHF_RCV_TYPE_INVALID5] = process_receive_invalid,
108 [RHF_RCV_TYPE_INVALID6] = process_receive_invalid,
109 [RHF_RCV_TYPE_INVALID7] = process_receive_invalid
112 /* Snoop packet structure */
113 struct snoop_packet {
114 struct list_head list;
119 /* Do not make these an enum or it will blow up the capture_md */
120 #define PKT_DIR_EGRESS 0x0
121 #define PKT_DIR_INGRESS 0x1
123 /* Packet capture metadata returned to the user with the packet. */
134 static atomic_t diagpkt_count = ATOMIC_INIT(0);
135 static struct cdev diagpkt_cdev;
136 static struct device *diagpkt_device;
138 static ssize_t diagpkt_write(struct file *fp, const char __user *data,
139 size_t count, loff_t *off);
141 static const struct file_operations diagpkt_file_ops = {
142 .owner = THIS_MODULE,
143 .write = diagpkt_write,
144 .llseek = noop_llseek,
148 * This is used for communication with user space for snoop extended IOCTLs
150 struct hfi1_link_info {
154 u16 link_speed_active;
155 u16 link_width_active;
159 * Add padding to make this a full IB SMP payload. Note: changing the
160 * size of this structure will make the IOCTLs created with _IOWR
162 * Be sure to run tests on all IOCTLs when making changes to this
169 * This starts our ioctl sequence numbers *way* off from the ones
170 * defined in ib_core.
172 #define SNOOP_CAPTURE_VERSION 0x1
174 #define IB_IOCTL_MAGIC 0x1b /* See Documentation/ioctl-number.txt */
175 #define HFI1_SNOOP_IOC_MAGIC IB_IOCTL_MAGIC
176 #define HFI1_SNOOP_IOC_BASE_SEQ 0x80
178 #define HFI1_SNOOP_IOCGETLINKSTATE \
179 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ)
180 #define HFI1_SNOOP_IOCSETLINKSTATE \
181 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 1)
182 #define HFI1_SNOOP_IOCCLEARQUEUE \
183 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 2)
184 #define HFI1_SNOOP_IOCCLEARFILTER \
185 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 3)
186 #define HFI1_SNOOP_IOCSETFILTER \
187 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 4)
188 #define HFI1_SNOOP_IOCGETVERSION \
189 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 5)
190 #define HFI1_SNOOP_IOCSET_OPTS \
191 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 6)
194 * These offsets +6/+7 could change, but these are already known and used
195 * IOCTL numbers so don't change them without a good reason.
197 #define HFI1_SNOOP_IOCGETLINKSTATE_EXTRA \
198 _IOWR(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 6, \
199 struct hfi1_link_info)
200 #define HFI1_SNOOP_IOCSETLINKSTATE_EXTRA \
201 _IOWR(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 7, \
202 struct hfi1_link_info)
204 static int hfi1_snoop_open(struct inode *in, struct file *fp);
205 static ssize_t hfi1_snoop_read(struct file *fp, char __user *data,
206 size_t pkt_len, loff_t *off);
207 static ssize_t hfi1_snoop_write(struct file *fp, const char __user *data,
208 size_t count, loff_t *off);
209 static long hfi1_ioctl(struct file *fp, unsigned int cmd, unsigned long arg);
210 static unsigned int hfi1_snoop_poll(struct file *fp,
211 struct poll_table_struct *wait);
212 static int hfi1_snoop_release(struct inode *in, struct file *fp);
214 struct hfi1_packet_filter_command {
220 /* Can't re-use PKT_DIR_*GRESS here because 0 means no packets for this */
221 #define HFI1_SNOOP_INGRESS 0x1
222 #define HFI1_SNOOP_EGRESS 0x2
224 enum hfi1_packet_filter_opcodes {
227 FILTER_BY_MAD_MGMT_CLASS,
230 FILTER_BY_SERVICE_LEVEL,
235 static const struct file_operations snoop_file_ops = {
236 .owner = THIS_MODULE,
237 .open = hfi1_snoop_open,
238 .read = hfi1_snoop_read,
239 .unlocked_ioctl = hfi1_ioctl,
240 .poll = hfi1_snoop_poll,
241 .write = hfi1_snoop_write,
242 .release = hfi1_snoop_release
245 struct hfi1_filter_array {
246 int (*filter)(void *, void *, void *);
249 static int hfi1_filter_lid(void *ibhdr, void *packet_data, void *value);
250 static int hfi1_filter_dlid(void *ibhdr, void *packet_data, void *value);
251 static int hfi1_filter_mad_mgmt_class(void *ibhdr, void *packet_data,
253 static int hfi1_filter_qp_number(void *ibhdr, void *packet_data, void *value);
254 static int hfi1_filter_ibpacket_type(void *ibhdr, void *packet_data,
256 static int hfi1_filter_ib_service_level(void *ibhdr, void *packet_data,
258 static int hfi1_filter_ib_pkey(void *ibhdr, void *packet_data, void *value);
259 static int hfi1_filter_direction(void *ibhdr, void *packet_data, void *value);
261 static struct hfi1_filter_array hfi1_filters[] = {
263 { hfi1_filter_dlid },
264 { hfi1_filter_mad_mgmt_class },
265 { hfi1_filter_qp_number },
266 { hfi1_filter_ibpacket_type },
267 { hfi1_filter_ib_service_level },
268 { hfi1_filter_ib_pkey },
269 { hfi1_filter_direction },
272 #define HFI1_MAX_FILTERS ARRAY_SIZE(hfi1_filters)
273 #define HFI1_DIAG_MINOR_BASE 129
275 static int hfi1_snoop_add(struct hfi1_devdata *dd, const char *name);
277 int hfi1_diag_add(struct hfi1_devdata *dd)
282 snprintf(name, sizeof(name), "%s_diagpkt%d", class_name(),
285 * Do this for each device as opposed to the normal diagpkt
286 * interface which is one per host
288 ret = hfi1_snoop_add(dd, name);
290 dd_dev_err(dd, "Unable to init snoop/capture device");
292 snprintf(name, sizeof(name), "%s_diagpkt", class_name());
293 if (atomic_inc_return(&diagpkt_count) == 1) {
294 ret = hfi1_cdev_init(HFI1_DIAGPKT_MINOR, name,
295 &diagpkt_file_ops, &diagpkt_cdev,
296 &diagpkt_device, false);
302 /* this must be called w/ dd->snoop_in_lock held */
303 static void drain_snoop_list(struct list_head *queue)
305 struct list_head *pos, *q;
306 struct snoop_packet *packet;
308 list_for_each_safe(pos, q, queue) {
309 packet = list_entry(pos, struct snoop_packet, list);
315 static void hfi1_snoop_remove(struct hfi1_devdata *dd)
317 unsigned long flags = 0;
319 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
320 drain_snoop_list(&dd->hfi1_snoop.queue);
321 hfi1_cdev_cleanup(&dd->hfi1_snoop.cdev, &dd->hfi1_snoop.class_dev);
322 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
325 void hfi1_diag_remove(struct hfi1_devdata *dd)
327 hfi1_snoop_remove(dd);
328 if (atomic_dec_and_test(&diagpkt_count))
329 hfi1_cdev_cleanup(&diagpkt_cdev, &diagpkt_device);
330 hfi1_cdev_cleanup(&dd->diag_cdev, &dd->diag_device);
334 * Allocated structure shared between the credit return mechanism and
337 struct diagpkt_wait {
338 struct completion credits_returned;
344 * When each side is finished with the structure, they call this.
345 * The last user frees the structure.
347 static void put_diagpkt_wait(struct diagpkt_wait *wait)
349 if (atomic_dec_and_test(&wait->count))
354 * Callback from the credit return code. Set the complete, which
355 * will let diapkt_send() continue.
357 static void diagpkt_complete(void *arg, int code)
359 struct diagpkt_wait *wait = (struct diagpkt_wait *)arg;
362 complete(&wait->credits_returned);
363 put_diagpkt_wait(wait); /* finished with the structure */
367 * diagpkt_send - send a packet
368 * @dp: diag packet descriptor
370 static ssize_t diagpkt_send(struct diag_pkt *dp)
372 struct hfi1_devdata *dd;
373 struct send_context *sc;
374 struct pio_buf *pbuf;
377 u32 pkt_len, total_len;
378 pio_release_cb credit_cb = NULL;
379 void *credit_arg = NULL;
380 struct diagpkt_wait *wait = NULL;
383 dd = hfi1_lookup(dp->unit);
384 if (!dd || !(dd->flags & HFI1_PRESENT) || !dd->kregbase) {
388 if (!(dd->flags & HFI1_INITTED)) {
389 /* no hardware, freeze, etc. */
394 if (dp->version != _DIAG_PKT_VERS) {
395 dd_dev_err(dd, "Invalid version %u for diagpkt_write\n",
401 /* send count must be an exact number of dwords */
407 /* there is only port 1 */
413 /* need a valid context */
414 if (dp->sw_index >= dd->num_send_contexts) {
418 /* can only use kernel contexts */
419 if (dd->send_contexts[dp->sw_index].type != SC_KERNEL) {
423 /* must be allocated */
424 sc = dd->send_contexts[dp->sw_index].sc;
429 /* must be enabled */
430 if (!(sc->flags & SCF_ENABLED)) {
435 /* allocate a buffer and copy the data in */
436 tmpbuf = vmalloc(dp->len);
442 if (copy_from_user(tmpbuf,
443 (const void __user *)(unsigned long)dp->data,
450 * pkt_len is how much data we have to write, includes header and data.
451 * total_len is length of the packet in Dwords plus the PBC should not
454 pkt_len = dp->len >> 2;
455 total_len = pkt_len + 2; /* PBC + packet */
457 /* if 0, fill in a default */
459 struct hfi1_pportdata *ppd = dd->pport;
461 hfi1_cdbg(PKT, "Generating PBC");
462 dp->pbc = create_pbc(ppd, 0, 0, 0, total_len);
464 hfi1_cdbg(PKT, "Using passed in PBC");
467 hfi1_cdbg(PKT, "Egress PBC content is 0x%llx", dp->pbc);
470 * The caller wants to wait until the packet is sent and to
471 * check for errors. The best we can do is wait until
472 * the buffer credits are returned and check if any packet
473 * error has occurred. If there are any late errors, this
474 * could miss it. If there are other senders who generate
475 * an error, this may find it. However, in general, it
478 if (dp->flags & F_DIAGPKT_WAIT) {
479 /* always force a credit return */
480 dp->pbc |= PBC_CREDIT_RETURN;
481 /* turn on credit return interrupts */
482 sc_add_credit_return_intr(sc);
483 wait = kmalloc(sizeof(*wait), GFP_KERNEL);
488 init_completion(&wait->credits_returned);
489 atomic_set(&wait->count, 2);
492 credit_cb = diagpkt_complete;
497 pbuf = sc_buffer_alloc(sc, total_len, credit_cb, credit_arg);
500 /* force a credit return and try again */
501 sc_return_credits(sc);
506 * No send buffer means no credit callback. Undo
507 * the wait set-up that was done above. We free wait
508 * because the callback will never be called.
510 if (dp->flags & F_DIAGPKT_WAIT) {
511 sc_del_credit_return_intr(sc);
519 pio_copy(dd, pbuf, dp->pbc, tmpbuf, pkt_len);
520 /* no flush needed as the HW knows the packet size */
524 if (dp->flags & F_DIAGPKT_WAIT) {
525 /* wait for credit return */
526 ret = wait_for_completion_interruptible(
527 &wait->credits_returned);
529 * If the wait returns an error, the wait was interrupted,
530 * e.g. with a ^C in the user program. The callback is
531 * still pending. This is OK as the wait structure is
532 * kmalloc'ed and the structure will free itself when
533 * all users are done with it.
535 * A context disable occurs on a send context restart, so
536 * include that in the list of errors below to check for.
537 * NOTE: PRC_FILL_ERR is at best informational and cannot
540 if (!ret && (((wait->code & PRC_STATUS_ERR) ||
541 (wait->code & PRC_FILL_ERR) ||
542 (wait->code & PRC_SC_DISABLE))))
545 put_diagpkt_wait(wait); /* finished with the structure */
546 sc_del_credit_return_intr(sc);
554 static ssize_t diagpkt_write(struct file *fp, const char __user *data,
555 size_t count, loff_t *off)
557 struct hfi1_devdata *dd;
558 struct send_context *sc;
563 if (count != sizeof(dp))
566 if (copy_from_user(&dp, data, sizeof(dp)))
570 * The Send Context is derived from the PbcVL value
571 * if PBC is populated
574 dd = hfi1_lookup(dp.unit);
577 vl = (dp.pbc >> PBC_VL_SHIFT) & PBC_VL_MASK;
580 dp.sw_index = sc->sw_index;
583 "Packet sent over VL %d via Send Context %u(%u)",
584 vl, sc->sw_index, sc->hw_context);
588 return diagpkt_send(&dp);
591 static int hfi1_snoop_add(struct hfi1_devdata *dd, const char *name)
595 dd->hfi1_snoop.mode_flag = 0;
596 spin_lock_init(&dd->hfi1_snoop.snoop_lock);
597 INIT_LIST_HEAD(&dd->hfi1_snoop.queue);
598 init_waitqueue_head(&dd->hfi1_snoop.waitq);
600 ret = hfi1_cdev_init(HFI1_SNOOP_CAPTURE_BASE + dd->unit, name,
602 &dd->hfi1_snoop.cdev, &dd->hfi1_snoop.class_dev,
606 dd_dev_err(dd, "Couldn't create %s device: %d", name, ret);
607 hfi1_cdev_cleanup(&dd->hfi1_snoop.cdev,
608 &dd->hfi1_snoop.class_dev);
614 static struct hfi1_devdata *hfi1_dd_from_sc_inode(struct inode *in)
616 int unit = iminor(in) - HFI1_SNOOP_CAPTURE_BASE;
617 struct hfi1_devdata *dd;
619 dd = hfi1_lookup(unit);
623 /* clear or restore send context integrity checks */
624 static void adjust_integrity_checks(struct hfi1_devdata *dd)
626 struct send_context *sc;
627 unsigned long sc_flags;
630 spin_lock_irqsave(&dd->sc_lock, sc_flags);
631 for (i = 0; i < dd->num_send_contexts; i++) {
634 sc = dd->send_contexts[i].sc;
637 continue; /* not allocated */
639 enable = likely(!HFI1_CAP_IS_KSET(NO_INTEGRITY)) &&
640 dd->hfi1_snoop.mode_flag != HFI1_PORT_SNOOP_MODE;
642 set_pio_integrity(sc);
644 if (enable) /* take HFI_CAP_* flags into account */
647 spin_unlock_irqrestore(&dd->sc_lock, sc_flags);
650 static int hfi1_snoop_open(struct inode *in, struct file *fp)
654 unsigned long flags = 0;
655 struct hfi1_devdata *dd;
656 struct list_head *queue;
658 mutex_lock(&hfi1_mutex);
660 dd = hfi1_dd_from_sc_inode(in);
667 * File mode determines snoop or capture. Some existing user
668 * applications expect the capture device to be able to be opened RDWR
669 * because they expect a dedicated capture device. For this reason we
670 * support a module param to force capture mode even if the file open
671 * mode matches snoop.
673 if ((fp->f_flags & O_ACCMODE) == O_RDONLY) {
674 snoop_dbg("Capture Enabled");
675 mode_flag = HFI1_PORT_CAPTURE_MODE;
676 } else if ((fp->f_flags & O_ACCMODE) == O_RDWR) {
677 snoop_dbg("Snoop Enabled");
678 mode_flag = HFI1_PORT_SNOOP_MODE;
680 snoop_dbg("Invalid");
684 queue = &dd->hfi1_snoop.queue;
687 * We are not supporting snoop and capture at the same time.
689 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
690 if (dd->hfi1_snoop.mode_flag) {
692 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
696 dd->hfi1_snoop.mode_flag = mode_flag;
697 drain_snoop_list(queue);
699 dd->hfi1_snoop.filter_callback = NULL;
700 dd->hfi1_snoop.filter_value = NULL;
703 * Send side packet integrity checks are not helpful when snooping so
704 * disable and re-enable when we stop snooping.
706 if (mode_flag == HFI1_PORT_SNOOP_MODE) {
707 /* clear after snoop mode is on */
708 adjust_integrity_checks(dd); /* clear */
711 * We also do not want to be doing the DLID LMC check for
714 dd->hfi1_snoop.dcc_cfg = read_csr(dd, DCC_CFG_PORT_CONFIG1);
715 write_csr(dd, DCC_CFG_PORT_CONFIG1,
716 (dd->hfi1_snoop.dcc_cfg >> 32) << 32);
720 * As soon as we set these function pointers the recv and send handlers
721 * are active. This is a race condition so we must make sure to drain
722 * the queue and init filter values above. Technically we should add
723 * locking here but all that will happen is on recv a packet will get
724 * allocated and get stuck on the snoop_lock before getting added to the
725 * queue. Same goes for send.
727 dd->rhf_rcv_function_map = snoop_rhf_rcv_functions;
728 dd->process_pio_send = snoop_send_pio_handler;
729 dd->process_dma_send = snoop_send_pio_handler;
730 dd->pio_inline_send = snoop_inline_pio_send;
732 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
736 mutex_unlock(&hfi1_mutex);
741 static int hfi1_snoop_release(struct inode *in, struct file *fp)
743 unsigned long flags = 0;
744 struct hfi1_devdata *dd;
747 dd = hfi1_dd_from_sc_inode(in);
751 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
753 /* clear the snoop mode before re-adjusting send context CSRs */
754 mode_flag = dd->hfi1_snoop.mode_flag;
755 dd->hfi1_snoop.mode_flag = 0;
758 * Drain the queue and clear the filters we are done with it. Don't
759 * forget to restore the packet integrity checks
761 drain_snoop_list(&dd->hfi1_snoop.queue);
762 if (mode_flag == HFI1_PORT_SNOOP_MODE) {
763 /* restore after snoop mode is clear */
764 adjust_integrity_checks(dd); /* restore */
767 * Also should probably reset the DCC_CONFIG1 register for DLID
768 * checking on incoming packets again. Use the value saved when
769 * opening the snoop device.
771 write_csr(dd, DCC_CFG_PORT_CONFIG1, dd->hfi1_snoop.dcc_cfg);
774 dd->hfi1_snoop.filter_callback = NULL;
775 kfree(dd->hfi1_snoop.filter_value);
776 dd->hfi1_snoop.filter_value = NULL;
779 * User is done snooping and capturing, return control to the normal
780 * handler. Re-enable SDMA handling.
782 dd->rhf_rcv_function_map = dd->normal_rhf_rcv_functions;
783 dd->process_pio_send = hfi1_verbs_send_pio;
784 dd->process_dma_send = hfi1_verbs_send_dma;
785 dd->pio_inline_send = pio_copy;
787 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
789 snoop_dbg("snoop/capture device released");
794 static unsigned int hfi1_snoop_poll(struct file *fp,
795 struct poll_table_struct *wait)
798 unsigned long flags = 0;
800 struct hfi1_devdata *dd;
802 dd = hfi1_dd_from_sc_inode(fp->f_inode);
806 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
808 poll_wait(fp, &dd->hfi1_snoop.waitq, wait);
809 if (!list_empty(&dd->hfi1_snoop.queue))
810 ret |= POLLIN | POLLRDNORM;
812 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
816 static ssize_t hfi1_snoop_write(struct file *fp, const char __user *data,
817 size_t count, loff_t *off)
819 struct diag_pkt dpkt;
820 struct hfi1_devdata *dd;
822 u8 byte_two, sl, sc5, sc4, vl, byte_one;
823 struct send_context *sc;
826 struct hfi1_ibport *ibp;
827 struct hfi1_pportdata *ppd;
829 dd = hfi1_dd_from_sc_inode(fp->f_inode);
834 snoop_dbg("received %lu bytes from user", count);
836 memset(&dpkt, 0, sizeof(struct diag_pkt));
837 dpkt.version = _DIAG_PKT_VERS;
838 dpkt.unit = dd->unit;
841 if (likely(!(snoop_flags & SNOOP_USE_METADATA))) {
843 * We need to generate the PBC and not let diagpkt_send do it,
844 * to do this we need the VL and the length in dwords.
845 * The VL can be determined by using the SL and looking up the
846 * SC. Then the SC can be converted into VL. The exception to
847 * this is those packets which are from an SMI queue pair.
848 * Since we can't detect anything about the QP here we have to
849 * rely on the SC. If its 0xF then we assume its SMI and
850 * do not look at the SL.
852 if (copy_from_user(&byte_one, data, 1))
855 if (copy_from_user(&byte_two, data + 1, 1))
858 sc4 = (byte_one >> 4) & 0xf;
860 snoop_dbg("Detected VL15 packet ignoring SL in packet");
863 sl = (byte_two >> 4) & 0xf;
864 ibp = to_iport(&dd->verbs_dev.rdi.ibdev, 1);
865 sc5 = ibp->sl_to_sc[sl];
866 vl = sc_to_vlt(dd, sc5);
868 snoop_dbg("VL %d does not match SC %d of packet",
874 sc = dd->vld[vl].sc; /* Look up the context based on VL */
876 dpkt.sw_index = sc->sw_index;
877 snoop_dbg("Sending on context %u(%u)", sc->sw_index,
880 snoop_dbg("Could not find context for vl %d", vl);
884 len = (count >> 2) + 2; /* Add in PBC */
885 pbc = create_pbc(ppd, 0, 0, vl, len);
887 if (copy_from_user(&pbc, data, sizeof(pbc)))
889 vl = (pbc >> PBC_VL_SHIFT) & PBC_VL_MASK;
890 sc = dd->vld[vl].sc; /* Look up the context based on VL */
892 dpkt.sw_index = sc->sw_index;
894 snoop_dbg("Could not find context for vl %d", vl);
898 count -= sizeof(pbc);
901 dpkt.data = (unsigned long)data;
903 snoop_dbg("PBC: vl=0x%llx Length=0x%llx",
908 ret = diagpkt_send(&dpkt);
910 * diagpkt_send only returns number of bytes in the diagpkt so patch
911 * that up here before returning.
913 if (ret == sizeof(dpkt))
919 static ssize_t hfi1_snoop_read(struct file *fp, char __user *data,
920 size_t pkt_len, loff_t *off)
923 unsigned long flags = 0;
924 struct snoop_packet *packet = NULL;
925 struct hfi1_devdata *dd;
927 dd = hfi1_dd_from_sc_inode(fp->f_inode);
931 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
933 while (list_empty(&dd->hfi1_snoop.queue)) {
934 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
936 if (fp->f_flags & O_NONBLOCK)
939 if (wait_event_interruptible(
940 dd->hfi1_snoop.waitq,
941 !list_empty(&dd->hfi1_snoop.queue)))
944 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
947 if (!list_empty(&dd->hfi1_snoop.queue)) {
948 packet = list_entry(dd->hfi1_snoop.queue.next,
949 struct snoop_packet, list);
950 list_del(&packet->list);
951 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
952 if (pkt_len >= packet->total_len) {
953 if (copy_to_user(data, packet->data,
957 ret = packet->total_len;
964 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
971 * hfi1_assign_snoop_link_credits -- Set up credits for VL15 and others
972 * @ppd : ptr to hfi1 port data
973 * @value : options from user space
975 * Assumes the rest of the CM credit registers are zero from a
976 * previous global or credit reset.
977 * Leave shared count at zero for both global and all vls.
978 * In snoop mode ideally we don't use shared credits
979 * Reserve 8.5k for VL15
980 * If total credits less than 8.5kbytes return error.
981 * Divide the rest of the credits across VL0 to VL7 and if
982 * each of these levels has less than 34 credits (at least 2048 + 128 bytes)
983 * return with an error.
984 * The credit registers will be reset to zero on link negotiation or link up
985 * so this function should be activated from user space only if the port has
986 * gone past link negotiation and link up.
988 * Return -- 0 if successful else error condition
991 static long hfi1_assign_snoop_link_credits(struct hfi1_pportdata *ppd,
994 #define OPA_MIN_PER_VL_CREDITS 34 /* 2048 + 128 bytes */
995 struct buffer_control t;
997 struct hfi1_devdata *dd = ppd->dd;
998 u16 total_credits = (value >> 16) & 0xffff;
999 u16 vl15_credits = dd->vl15_init / 2;
1001 __be16 be_per_vl_credits;
1003 if (!(ppd->host_link_state & HLS_UP))
1005 if (total_credits < vl15_credits)
1008 per_vl_credits = (total_credits - vl15_credits) / TXE_NUM_DATA_VL;
1010 if (per_vl_credits < OPA_MIN_PER_VL_CREDITS)
1013 memset(&t, 0, sizeof(t));
1014 be_per_vl_credits = cpu_to_be16(per_vl_credits);
1016 for (i = 0; i < TXE_NUM_DATA_VL; i++)
1017 t.vl[i].dedicated = be_per_vl_credits;
1019 t.vl[15].dedicated = cpu_to_be16(vl15_credits);
1020 return set_buffer_control(ppd, &t);
1023 snoop_dbg("port_state = 0x%x, total_credits = %d, vl15_credits = %d",
1024 ppd->host_link_state, total_credits, vl15_credits);
1029 static long hfi1_ioctl(struct file *fp, unsigned int cmd, unsigned long arg)
1031 struct hfi1_devdata *dd;
1032 void *filter_value = NULL;
1038 unsigned long flags = 0;
1039 unsigned long *argp = NULL;
1040 struct hfi1_packet_filter_command filter_cmd = {0};
1042 struct hfi1_pportdata *ppd = NULL;
1044 struct hfi1_link_info link_info;
1045 int read_cmd, write_cmd, read_ok, write_ok;
1047 dd = hfi1_dd_from_sc_inode(fp->f_inode);
1051 mode_flag = dd->hfi1_snoop.mode_flag;
1052 read_cmd = _IOC_DIR(cmd) & _IOC_READ;
1053 write_cmd = _IOC_DIR(cmd) & _IOC_WRITE;
1054 write_ok = access_ok(VERIFY_WRITE, (void __user *)arg, _IOC_SIZE(cmd));
1055 read_ok = access_ok(VERIFY_READ, (void __user *)arg, _IOC_SIZE(cmd));
1057 if ((read_cmd && !write_ok) || (write_cmd && !read_ok))
1060 if (!capable(CAP_SYS_ADMIN))
1063 if ((mode_flag & HFI1_PORT_CAPTURE_MODE) &&
1064 (cmd != HFI1_SNOOP_IOCCLEARQUEUE) &&
1065 (cmd != HFI1_SNOOP_IOCCLEARFILTER) &&
1066 (cmd != HFI1_SNOOP_IOCSETFILTER))
1067 /* Capture devices are allowed only 3 operations
1068 * 1.Clear capture queue
1069 * 2.Clear capture filter
1070 * 3.Set capture filter
1071 * Other are invalid.
1076 case HFI1_SNOOP_IOCSETLINKSTATE_EXTRA:
1077 memset(&link_info, 0, sizeof(link_info));
1079 if (copy_from_user(&link_info,
1080 (struct hfi1_link_info __user *)arg,
1084 value = link_info.port_state;
1085 index = link_info.port_number;
1086 if (index > dd->num_pports - 1)
1089 ppd = &dd->pport[index];
1093 /* What we want to transition to */
1094 phys_state = (value >> 4) & 0xF;
1095 link_state = value & 0xF;
1096 snoop_dbg("Setting link state 0x%x", value);
1098 switch (link_state) {
1100 if (phys_state == 0)
1104 switch (phys_state) {
1106 dev_state = HLS_DN_DOWNDEF;
1109 dev_state = HLS_DN_POLL;
1112 dev_state = HLS_DN_DISABLE;
1117 ret = set_link_state(ppd, dev_state);
1120 ret = set_link_state(ppd, HLS_UP_ARMED);
1122 send_idle_sma(dd, SMA_IDLE_ARM);
1124 case IB_PORT_ACTIVE:
1125 ret = set_link_state(ppd, HLS_UP_ACTIVE);
1127 send_idle_sma(dd, SMA_IDLE_ACTIVE);
1136 case HFI1_SNOOP_IOCGETLINKSTATE:
1137 case HFI1_SNOOP_IOCGETLINKSTATE_EXTRA:
1138 if (cmd == HFI1_SNOOP_IOCGETLINKSTATE_EXTRA) {
1139 memset(&link_info, 0, sizeof(link_info));
1140 if (copy_from_user(&link_info,
1141 (struct hfi1_link_info __user *)arg,
1144 index = link_info.port_number;
1146 ret = __get_user(index, (int __user *)arg);
1151 if (index > dd->num_pports - 1)
1154 ppd = &dd->pport[index];
1158 value = hfi1_ibphys_portstate(ppd);
1160 value |= driver_lstate(ppd);
1162 snoop_dbg("Link port | Link State: %d", value);
1164 if ((cmd == HFI1_SNOOP_IOCGETLINKSTATE_EXTRA) ||
1165 (cmd == HFI1_SNOOP_IOCSETLINKSTATE_EXTRA)) {
1166 link_info.port_state = value;
1167 link_info.node_guid = cpu_to_be64(ppd->guid);
1168 link_info.link_speed_active =
1169 ppd->link_speed_active;
1170 link_info.link_width_active =
1171 ppd->link_width_active;
1172 if (copy_to_user((struct hfi1_link_info __user *)arg,
1173 &link_info, sizeof(link_info)))
1176 ret = __put_user(value, (int __user *)arg);
1180 case HFI1_SNOOP_IOCCLEARQUEUE:
1181 snoop_dbg("Clearing snoop queue");
1182 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1183 drain_snoop_list(&dd->hfi1_snoop.queue);
1184 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1187 case HFI1_SNOOP_IOCCLEARFILTER:
1188 snoop_dbg("Clearing filter");
1189 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1190 if (dd->hfi1_snoop.filter_callback) {
1191 /* Drain packets first */
1192 drain_snoop_list(&dd->hfi1_snoop.queue);
1193 dd->hfi1_snoop.filter_callback = NULL;
1195 kfree(dd->hfi1_snoop.filter_value);
1196 dd->hfi1_snoop.filter_value = NULL;
1197 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1200 case HFI1_SNOOP_IOCSETFILTER:
1201 snoop_dbg("Setting filter");
1202 /* just copy command structure */
1203 argp = (unsigned long *)arg;
1204 if (copy_from_user(&filter_cmd, (void __user *)argp,
1205 sizeof(filter_cmd)))
1208 if (filter_cmd.opcode >= HFI1_MAX_FILTERS) {
1209 pr_alert("Invalid opcode in request\n");
1213 snoop_dbg("Opcode %d Len %d Ptr %p",
1214 filter_cmd.opcode, filter_cmd.length,
1215 filter_cmd.value_ptr);
1217 filter_value = kcalloc(filter_cmd.length, sizeof(u8),
1222 /* copy remaining data from userspace */
1223 if (copy_from_user((u8 *)filter_value,
1224 (void __user *)filter_cmd.value_ptr,
1225 filter_cmd.length)) {
1226 kfree(filter_value);
1229 /* Drain packets first */
1230 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1231 drain_snoop_list(&dd->hfi1_snoop.queue);
1232 dd->hfi1_snoop.filter_callback =
1233 hfi1_filters[filter_cmd.opcode].filter;
1234 /* just in case we see back to back sets */
1235 kfree(dd->hfi1_snoop.filter_value);
1236 dd->hfi1_snoop.filter_value = filter_value;
1237 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1239 case HFI1_SNOOP_IOCGETVERSION:
1240 value = SNOOP_CAPTURE_VERSION;
1241 snoop_dbg("Getting version: %d", value);
1242 ret = __put_user(value, (int __user *)arg);
1244 case HFI1_SNOOP_IOCSET_OPTS:
1246 ret = __get_user(value, (int __user *)arg);
1250 snoop_dbg("Setting snoop option %d", value);
1251 if (value & SNOOP_DROP_SEND)
1252 snoop_flags |= SNOOP_DROP_SEND;
1253 if (value & SNOOP_USE_METADATA)
1254 snoop_flags |= SNOOP_USE_METADATA;
1255 if (value & (SNOOP_SET_VL0TOVL15)) {
1256 ppd = &dd->pport[0]; /* first port will do */
1257 ret = hfi1_assign_snoop_link_credits(ppd, value);
1267 static void snoop_list_add_tail(struct snoop_packet *packet,
1268 struct hfi1_devdata *dd)
1270 unsigned long flags = 0;
1272 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1273 if (likely((dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE) ||
1274 (dd->hfi1_snoop.mode_flag & HFI1_PORT_CAPTURE_MODE))) {
1275 list_add_tail(&packet->list, &dd->hfi1_snoop.queue);
1276 snoop_dbg("Added packet to list");
1280 * Technically we can could have closed the snoop device while waiting
1281 * on the above lock and it is gone now. The snoop mode_flag will
1282 * prevent us from adding the packet to the queue though.
1285 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1286 wake_up_interruptible(&dd->hfi1_snoop.waitq);
1289 static inline int hfi1_filter_check(void *val, const char *msg)
1292 snoop_dbg("Error invalid %s value for filter", msg);
1293 return HFI1_FILTER_ERR;
1298 static int hfi1_filter_lid(void *ibhdr, void *packet_data, void *value)
1300 struct hfi1_ib_header *hdr;
1303 ret = hfi1_filter_check(ibhdr, "header");
1306 ret = hfi1_filter_check(value, "user");
1309 hdr = (struct hfi1_ib_header *)ibhdr;
1311 if (*((u16 *)value) == be16_to_cpu(hdr->lrh[3])) /* matches slid */
1312 return HFI1_FILTER_HIT; /* matched */
1314 return HFI1_FILTER_MISS; /* Not matched */
1317 static int hfi1_filter_dlid(void *ibhdr, void *packet_data, void *value)
1319 struct hfi1_ib_header *hdr;
1322 ret = hfi1_filter_check(ibhdr, "header");
1325 ret = hfi1_filter_check(value, "user");
1329 hdr = (struct hfi1_ib_header *)ibhdr;
1331 if (*((u16 *)value) == be16_to_cpu(hdr->lrh[1]))
1332 return HFI1_FILTER_HIT;
1334 return HFI1_FILTER_MISS;
1337 /* Not valid for outgoing packets, send handler passes null for data*/
1338 static int hfi1_filter_mad_mgmt_class(void *ibhdr, void *packet_data,
1341 struct hfi1_ib_header *hdr;
1342 struct hfi1_other_headers *ohdr = NULL;
1343 struct ib_smp *smp = NULL;
1347 ret = hfi1_filter_check(ibhdr, "header");
1350 ret = hfi1_filter_check(packet_data, "packet_data");
1353 ret = hfi1_filter_check(value, "user");
1357 hdr = (struct hfi1_ib_header *)ibhdr;
1360 if ((be16_to_cpu(hdr->lrh[0]) & 3) == HFI1_LRH_BTH)
1361 ohdr = &hdr->u.oth; /* LRH + BTH + DETH */
1363 ohdr = &hdr->u.l.oth; /* LRH + GRH + BTH + DETH */
1365 qpn = be32_to_cpu(ohdr->bth[1]) & 0x00FFFFFF;
1367 smp = (struct ib_smp *)packet_data;
1368 if (*((u8 *)value) == smp->mgmt_class)
1369 return HFI1_FILTER_HIT;
1371 return HFI1_FILTER_MISS;
1373 return HFI1_FILTER_ERR;
1376 static int hfi1_filter_qp_number(void *ibhdr, void *packet_data, void *value)
1378 struct hfi1_ib_header *hdr;
1379 struct hfi1_other_headers *ohdr = NULL;
1382 ret = hfi1_filter_check(ibhdr, "header");
1385 ret = hfi1_filter_check(value, "user");
1389 hdr = (struct hfi1_ib_header *)ibhdr;
1392 if ((be16_to_cpu(hdr->lrh[0]) & 3) == HFI1_LRH_BTH)
1393 ohdr = &hdr->u.oth; /* LRH + BTH + DETH */
1395 ohdr = &hdr->u.l.oth; /* LRH + GRH + BTH + DETH */
1396 if (*((u32 *)value) == (be32_to_cpu(ohdr->bth[1]) & 0x00FFFFFF))
1397 return HFI1_FILTER_HIT;
1399 return HFI1_FILTER_MISS;
1402 static int hfi1_filter_ibpacket_type(void *ibhdr, void *packet_data,
1407 struct hfi1_ib_header *hdr;
1408 struct hfi1_other_headers *ohdr = NULL;
1411 ret = hfi1_filter_check(ibhdr, "header");
1414 ret = hfi1_filter_check(value, "user");
1418 hdr = (struct hfi1_ib_header *)ibhdr;
1420 lnh = (be16_to_cpu(hdr->lrh[0]) & 3);
1422 if (lnh == HFI1_LRH_BTH)
1424 else if (lnh == HFI1_LRH_GRH)
1425 ohdr = &hdr->u.l.oth;
1427 return HFI1_FILTER_ERR;
1429 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1431 if (*((u8 *)value) == ((opcode >> 5) & 0x7))
1432 return HFI1_FILTER_HIT;
1434 return HFI1_FILTER_MISS;
1437 static int hfi1_filter_ib_service_level(void *ibhdr, void *packet_data,
1440 struct hfi1_ib_header *hdr;
1443 ret = hfi1_filter_check(ibhdr, "header");
1446 ret = hfi1_filter_check(value, "user");
1450 hdr = (struct hfi1_ib_header *)ibhdr;
1452 if ((*((u8 *)value)) == ((be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF))
1453 return HFI1_FILTER_HIT;
1455 return HFI1_FILTER_MISS;
1458 static int hfi1_filter_ib_pkey(void *ibhdr, void *packet_data, void *value)
1461 struct hfi1_ib_header *hdr;
1462 struct hfi1_other_headers *ohdr = NULL;
1465 ret = hfi1_filter_check(ibhdr, "header");
1468 ret = hfi1_filter_check(value, "user");
1472 hdr = (struct hfi1_ib_header *)ibhdr;
1474 lnh = (be16_to_cpu(hdr->lrh[0]) & 3);
1475 if (lnh == HFI1_LRH_BTH)
1477 else if (lnh == HFI1_LRH_GRH)
1478 ohdr = &hdr->u.l.oth;
1480 return HFI1_FILTER_ERR;
1482 /* P_key is 16-bit entity, however top most bit indicates
1483 * type of membership. 0 for limited and 1 for Full.
1484 * Limited members cannot accept information from other
1485 * Limited members, but communication is allowed between
1486 * every other combination of membership.
1487 * Hence we'll omit comparing top-most bit while filtering
1490 if ((*(u16 *)value & 0x7FFF) ==
1491 ((be32_to_cpu(ohdr->bth[0])) & 0x7FFF))
1492 return HFI1_FILTER_HIT;
1494 return HFI1_FILTER_MISS;
1498 * If packet_data is NULL then this is coming from one of the send functions.
1499 * Thus we know if its an ingressed or egressed packet.
1501 static int hfi1_filter_direction(void *ibhdr, void *packet_data, void *value)
1503 u8 user_dir = *(u8 *)value;
1506 ret = hfi1_filter_check(value, "user");
1511 /* Incoming packet */
1512 if (user_dir & HFI1_SNOOP_INGRESS)
1513 return HFI1_FILTER_HIT;
1515 /* Outgoing packet */
1516 if (user_dir & HFI1_SNOOP_EGRESS)
1517 return HFI1_FILTER_HIT;
1520 return HFI1_FILTER_MISS;
1524 * Allocate a snoop packet. The structure that is stored in the ring buffer, not
1525 * to be confused with an hfi packet type.
1527 static struct snoop_packet *allocate_snoop_packet(u32 hdr_len,
1531 struct snoop_packet *packet;
1533 packet = kzalloc(sizeof(*packet) + hdr_len + data_len
1535 GFP_ATOMIC | __GFP_NOWARN);
1537 INIT_LIST_HEAD(&packet->list);
1543 * Instead of having snoop and capture code intermixed with the recv functions,
1544 * both the interrupt handler and hfi1_ib_rcv() we are going to hijack the call
1545 * and land in here for snoop/capture but if not enabled the call will go
1546 * through as before. This gives us a single point to constrain all of the snoop
1547 * snoop recv logic. There is nothing special that needs to happen for bypass
1548 * packets. This routine should not try to look into the packet. It just copied
1549 * it. There is no guarantee for filters when it comes to bypass packets as
1550 * there is no specific support. Bottom line is this routine does now even know
1551 * what a bypass packet is.
1553 int snoop_recv_handler(struct hfi1_packet *packet)
1555 struct hfi1_pportdata *ppd = packet->rcd->ppd;
1556 struct hfi1_ib_header *hdr = packet->hdr;
1557 int header_size = packet->hlen;
1558 void *data = packet->ebuf;
1559 u32 tlen = packet->tlen;
1560 struct snoop_packet *s_packet = NULL;
1564 struct capture_md md;
1566 snoop_dbg("PACKET IN: hdr size %d tlen %d data %p", header_size, tlen,
1569 trace_snoop_capture(ppd->dd, header_size, hdr, tlen - header_size,
1572 if (!ppd->dd->hfi1_snoop.filter_callback) {
1573 snoop_dbg("filter not set");
1574 ret = HFI1_FILTER_HIT;
1576 ret = ppd->dd->hfi1_snoop.filter_callback(hdr, data,
1577 ppd->dd->hfi1_snoop.filter_value);
1581 case HFI1_FILTER_ERR:
1582 snoop_dbg("Error in filter call");
1584 case HFI1_FILTER_MISS:
1585 snoop_dbg("Filter Miss");
1587 case HFI1_FILTER_HIT:
1589 if (ppd->dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE)
1591 if ((snoop_mode == 0) ||
1592 unlikely(snoop_flags & SNOOP_USE_METADATA))
1593 md_len = sizeof(struct capture_md);
1595 s_packet = allocate_snoop_packet(header_size,
1599 if (unlikely(!s_packet)) {
1600 dd_dev_warn_ratelimited(ppd->dd, "Unable to allocate snoop/capture packet\n");
1605 memset(&md, 0, sizeof(struct capture_md));
1607 md.dir = PKT_DIR_INGRESS;
1608 md.u.rhf = packet->rhf;
1609 memcpy(s_packet->data, &md, md_len);
1612 /* We should always have a header */
1614 memcpy(s_packet->data + md_len, hdr, header_size);
1616 dd_dev_err(ppd->dd, "Unable to copy header to snoop/capture packet\n");
1622 * Packets with no data are possible. If there is no data needed
1623 * to take care of the last 4 bytes which are normally included
1624 * with data buffers and are included in tlen. Since we kzalloc
1625 * the buffer we do not need to set any values but if we decide
1626 * not to use kzalloc we should zero them.
1629 memcpy(s_packet->data + header_size + md_len, data,
1630 tlen - header_size);
1632 s_packet->total_len = tlen + md_len;
1633 snoop_list_add_tail(s_packet, ppd->dd);
1636 * If we are snooping the packet not capturing then throw away
1637 * after adding to the list.
1639 snoop_dbg("Capturing packet");
1640 if (ppd->dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE) {
1641 snoop_dbg("Throwing packet away");
1643 * If we are dropping the packet we still may need to
1644 * handle the case where error flags are set, this is
1645 * normally done by the type specific handler but that
1646 * won't be called in this case.
1648 if (unlikely(rhf_err_flags(packet->rhf)))
1649 handle_eflags(packet);
1651 /* throw the packet on the floor */
1652 return RHF_RCV_CONTINUE;
1660 * We do not care what type of packet came in here - just pass it off
1661 * to the normal handler.
1663 return ppd->dd->normal_rhf_rcv_functions[rhf_rcv_type(packet->rhf)]
1668 * Handle snooping and capturing packets when sdma is being used.
1670 int snoop_send_dma_handler(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
1673 pr_alert("Snooping/Capture of Send DMA Packets Is Not Supported!\n");
1674 snoop_dbg("Unsupported Operation");
1675 return hfi1_verbs_send_dma(qp, ps, 0);
1679 * Handle snooping and capturing packets when pio is being used. Does not handle
1680 * bypass packets. The only way to send a bypass packet currently is to use the
1681 * diagpkt interface. When that interface is enable snoop/capture is not.
1683 int snoop_send_pio_handler(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
1686 u32 hdrwords = qp->s_hdrwords;
1687 struct rvt_sge_state *ss = qp->s_cur_sge;
1688 u32 len = qp->s_cur_size;
1689 u32 dwords = (len + 3) >> 2;
1690 u32 plen = hdrwords + dwords + 2; /* includes pbc */
1691 struct hfi1_pportdata *ppd = ps->ppd;
1692 struct snoop_packet *s_packet = NULL;
1693 u32 *hdr = (u32 *)&ps->s_txreq->phdr.hdr;
1695 struct rvt_sge_state temp_ss;
1697 void *data_start = NULL;
1701 struct capture_md md;
1703 u32 hdr_len = hdrwords << 2;
1704 u32 tlen = HFI1_GET_PKT_LEN(&ps->s_txreq->phdr.hdr);
1708 snoop_dbg("PACKET OUT: hdrword %u len %u plen %u dwords %u tlen %u",
1709 hdrwords, len, plen, dwords, tlen);
1710 if (ppd->dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE)
1712 if ((snoop_mode == 0) ||
1713 unlikely(snoop_flags & SNOOP_USE_METADATA))
1714 md_len = sizeof(struct capture_md);
1716 /* not using ss->total_len as arg 2 b/c that does not count CRC */
1717 s_packet = allocate_snoop_packet(hdr_len, tlen - hdr_len, md_len);
1719 if (unlikely(!s_packet)) {
1720 dd_dev_warn_ratelimited(ppd->dd, "Unable to allocate snoop/capture packet\n");
1724 s_packet->total_len = tlen + md_len;
1727 memset(&md, 0, sizeof(struct capture_md));
1729 md.dir = PKT_DIR_EGRESS;
1730 if (likely(pbc == 0)) {
1731 vl = be16_to_cpu(ps->s_txreq->phdr.hdr.lrh[0]) >> 12;
1732 md.u.pbc = create_pbc(ppd, 0, qp->s_srate, vl, plen);
1736 memcpy(s_packet->data, &md, md_len);
1743 memcpy(s_packet->data + md_len, hdr, hdr_len);
1746 "Unable to copy header to snoop/capture packet\n");
1752 data = s_packet->data + hdr_len + md_len;
1757 * The update_sge() function below will not modify the
1758 * individual SGEs in the array. It will make a copy each time
1759 * and operate on that. So we only need to copy this instance
1760 * and it won't impact PIO.
1765 snoop_dbg("Need to copy %d bytes", length);
1767 void *addr = temp_ss.sge.vaddr;
1768 u32 slen = temp_ss.sge.length;
1770 if (slen > length) {
1772 snoop_dbg("slen %d > len %d", slen, length);
1774 snoop_dbg("copy %d to %p", slen, addr);
1775 memcpy(data, addr, slen);
1776 update_sge(&temp_ss, slen);
1779 snoop_dbg("data is now %p bytes left %d", data, length);
1781 snoop_dbg("Completed SGE copy");
1785 * Why do the filter check down here? Because the event tracing has its
1786 * own filtering and we need to have the walked the SGE list.
1788 if (!ppd->dd->hfi1_snoop.filter_callback) {
1789 snoop_dbg("filter not set\n");
1790 ret = HFI1_FILTER_HIT;
1792 ret = ppd->dd->hfi1_snoop.filter_callback(
1793 &ps->s_txreq->phdr.hdr,
1795 ppd->dd->hfi1_snoop.filter_value);
1799 case HFI1_FILTER_ERR:
1800 snoop_dbg("Error in filter call");
1802 case HFI1_FILTER_MISS:
1803 snoop_dbg("Filter Miss");
1806 case HFI1_FILTER_HIT:
1807 snoop_dbg("Capturing packet");
1808 snoop_list_add_tail(s_packet, ppd->dd);
1810 if (unlikely((snoop_flags & SNOOP_DROP_SEND) &&
1811 (ppd->dd->hfi1_snoop.mode_flag &
1812 HFI1_PORT_SNOOP_MODE))) {
1813 unsigned long flags;
1815 snoop_dbg("Dropping packet");
1817 spin_lock_irqsave(&qp->s_lock, flags);
1822 spin_unlock_irqrestore(&qp->s_lock, flags);
1823 } else if (qp->ibqp.qp_type == IB_QPT_RC) {
1824 spin_lock_irqsave(&qp->s_lock, flags);
1825 hfi1_rc_send_complete(qp,
1826 &ps->s_txreq->phdr.hdr);
1827 spin_unlock_irqrestore(&qp->s_lock, flags);
1831 * If snoop is dropping the packet we need to put the
1832 * txreq back because no one else will.
1834 hfi1_put_txreq(ps->s_txreq);
1843 return hfi1_verbs_send_pio(qp, ps, md.u.pbc);
1847 * Callers of this must pass a hfi1_ib_header type for the from ptr. Currently
1848 * this can be used anywhere, but the intention is for inline ACKs for RC and
1849 * CCA packets. We don't restrict this usage though.
1851 void snoop_inline_pio_send(struct hfi1_devdata *dd, struct pio_buf *pbuf,
1852 u64 pbc, const void *from, size_t count)
1856 struct capture_md md;
1857 struct snoop_packet *s_packet = NULL;
1860 * count is in dwords so we need to convert to bytes.
1861 * We also need to account for CRC which would be tacked on by hardware.
1863 int packet_len = (count << 2) + 4;
1866 snoop_dbg("ACK OUT: len %d", packet_len);
1868 if (!dd->hfi1_snoop.filter_callback) {
1869 snoop_dbg("filter not set");
1870 ret = HFI1_FILTER_HIT;
1872 ret = dd->hfi1_snoop.filter_callback(
1873 (struct hfi1_ib_header *)from,
1875 dd->hfi1_snoop.filter_value);
1879 case HFI1_FILTER_ERR:
1880 snoop_dbg("Error in filter call");
1882 case HFI1_FILTER_MISS:
1883 snoop_dbg("Filter Miss");
1885 case HFI1_FILTER_HIT:
1886 snoop_dbg("Capturing packet");
1887 if (dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE)
1889 if ((snoop_mode == 0) ||
1890 unlikely(snoop_flags & SNOOP_USE_METADATA))
1891 md_len = sizeof(struct capture_md);
1893 s_packet = allocate_snoop_packet(packet_len, 0, md_len);
1895 if (unlikely(!s_packet)) {
1896 dd_dev_warn_ratelimited(dd, "Unable to allocate snoop/capture packet\n");
1897 goto inline_pio_out;
1900 s_packet->total_len = packet_len + md_len;
1902 /* Fill in the metadata for the packet */
1904 memset(&md, 0, sizeof(struct capture_md));
1906 md.dir = PKT_DIR_EGRESS;
1908 memcpy(s_packet->data, &md, md_len);
1911 /* Add the packet data which is a single buffer */
1912 memcpy(s_packet->data + md_len, from, packet_len);
1914 snoop_list_add_tail(s_packet, dd);
1916 if (unlikely((snoop_flags & SNOOP_DROP_SEND) && snoop_mode)) {
1917 snoop_dbg("Dropping packet");
1926 pio_copy(dd, pbuf, pbc, from, count);