2 * Copyright(c) 2015, 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
49 * This file contains support for diagnostic functions. It is accessed by
50 * opening the hfi1_diag device, normally minor number 129. Diagnostic use
51 * of the chip may render the chip or board unusable until the driver
52 * is unloaded, or in some cases, until the system is rebooted.
54 * Accesses to the chip through this interface are not similar to going
55 * through the /sys/bus/pci resource mmap interface.
59 #include <linux/pci.h>
60 #include <linux/poll.h>
61 #include <linux/vmalloc.h>
62 #include <linux/export.h>
64 #include <linux/uaccess.h>
65 #include <linux/module.h>
66 #include <rdma/ib_smi.h>
70 #include "verbs_txreq.h"
74 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
75 #define snoop_dbg(fmt, ...) \
76 hfi1_cdbg(SNOOP, fmt, ##__VA_ARGS__)
78 /* Snoop option mask */
79 #define SNOOP_DROP_SEND BIT(0)
80 #define SNOOP_USE_METADATA BIT(1)
81 #define SNOOP_SET_VL0TOVL15 BIT(2)
83 static u8 snoop_flags;
86 * Extract packet length from LRH header.
87 * This is in Dwords so multiply by 4 to get size in bytes
89 #define HFI1_GET_PKT_LEN(x) (((be16_to_cpu((x)->lrh[2]) & 0xFFF)) << 2)
91 enum hfi1_filter_status {
97 /* snoop processing functions */
98 rhf_rcv_function_ptr snoop_rhf_rcv_functions[8] = {
99 [RHF_RCV_TYPE_EXPECTED] = snoop_recv_handler,
100 [RHF_RCV_TYPE_EAGER] = snoop_recv_handler,
101 [RHF_RCV_TYPE_IB] = snoop_recv_handler,
102 [RHF_RCV_TYPE_ERROR] = snoop_recv_handler,
103 [RHF_RCV_TYPE_BYPASS] = snoop_recv_handler,
104 [RHF_RCV_TYPE_INVALID5] = process_receive_invalid,
105 [RHF_RCV_TYPE_INVALID6] = process_receive_invalid,
106 [RHF_RCV_TYPE_INVALID7] = process_receive_invalid
109 /* Snoop packet structure */
110 struct snoop_packet {
111 struct list_head list;
116 /* Do not make these an enum or it will blow up the capture_md */
117 #define PKT_DIR_EGRESS 0x0
118 #define PKT_DIR_INGRESS 0x1
120 /* Packet capture metadata returned to the user with the packet. */
131 static atomic_t diagpkt_count = ATOMIC_INIT(0);
132 static struct cdev diagpkt_cdev;
133 static struct device *diagpkt_device;
135 static ssize_t diagpkt_write(struct file *fp, const char __user *data,
136 size_t count, loff_t *off);
138 static const struct file_operations diagpkt_file_ops = {
139 .owner = THIS_MODULE,
140 .write = diagpkt_write,
141 .llseek = noop_llseek,
145 * This is used for communication with user space for snoop extended IOCTLs
147 struct hfi1_link_info {
151 u16 link_speed_active;
152 u16 link_width_active;
156 * Add padding to make this a full IB SMP payload. Note: changing the
157 * size of this structure will make the IOCTLs created with _IOWR
159 * Be sure to run tests on all IOCTLs when making changes to this
166 * This starts our ioctl sequence numbers *way* off from the ones
167 * defined in ib_core.
169 #define SNOOP_CAPTURE_VERSION 0x1
171 #define IB_IOCTL_MAGIC 0x1b /* See Documentation/ioctl-number.txt */
172 #define HFI1_SNOOP_IOC_MAGIC IB_IOCTL_MAGIC
173 #define HFI1_SNOOP_IOC_BASE_SEQ 0x80
175 #define HFI1_SNOOP_IOCGETLINKSTATE \
176 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ)
177 #define HFI1_SNOOP_IOCSETLINKSTATE \
178 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 1)
179 #define HFI1_SNOOP_IOCCLEARQUEUE \
180 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 2)
181 #define HFI1_SNOOP_IOCCLEARFILTER \
182 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 3)
183 #define HFI1_SNOOP_IOCSETFILTER \
184 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 4)
185 #define HFI1_SNOOP_IOCGETVERSION \
186 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 5)
187 #define HFI1_SNOOP_IOCSET_OPTS \
188 _IO(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 6)
191 * These offsets +6/+7 could change, but these are already known and used
192 * IOCTL numbers so don't change them without a good reason.
194 #define HFI1_SNOOP_IOCGETLINKSTATE_EXTRA \
195 _IOWR(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 6, \
196 struct hfi1_link_info)
197 #define HFI1_SNOOP_IOCSETLINKSTATE_EXTRA \
198 _IOWR(HFI1_SNOOP_IOC_MAGIC, HFI1_SNOOP_IOC_BASE_SEQ + 7, \
199 struct hfi1_link_info)
201 static int hfi1_snoop_open(struct inode *in, struct file *fp);
202 static ssize_t hfi1_snoop_read(struct file *fp, char __user *data,
203 size_t pkt_len, loff_t *off);
204 static ssize_t hfi1_snoop_write(struct file *fp, const char __user *data,
205 size_t count, loff_t *off);
206 static long hfi1_ioctl(struct file *fp, unsigned int cmd, unsigned long arg);
207 static unsigned int hfi1_snoop_poll(struct file *fp,
208 struct poll_table_struct *wait);
209 static int hfi1_snoop_release(struct inode *in, struct file *fp);
211 struct hfi1_packet_filter_command {
217 /* Can't re-use PKT_DIR_*GRESS here because 0 means no packets for this */
218 #define HFI1_SNOOP_INGRESS 0x1
219 #define HFI1_SNOOP_EGRESS 0x2
221 enum hfi1_packet_filter_opcodes {
224 FILTER_BY_MAD_MGMT_CLASS,
227 FILTER_BY_SERVICE_LEVEL,
232 static const struct file_operations snoop_file_ops = {
233 .owner = THIS_MODULE,
234 .open = hfi1_snoop_open,
235 .read = hfi1_snoop_read,
236 .unlocked_ioctl = hfi1_ioctl,
237 .poll = hfi1_snoop_poll,
238 .write = hfi1_snoop_write,
239 .release = hfi1_snoop_release
242 struct hfi1_filter_array {
243 int (*filter)(void *, void *, void *);
246 static int hfi1_filter_lid(void *ibhdr, void *packet_data, void *value);
247 static int hfi1_filter_dlid(void *ibhdr, void *packet_data, void *value);
248 static int hfi1_filter_mad_mgmt_class(void *ibhdr, void *packet_data,
250 static int hfi1_filter_qp_number(void *ibhdr, void *packet_data, void *value);
251 static int hfi1_filter_ibpacket_type(void *ibhdr, void *packet_data,
253 static int hfi1_filter_ib_service_level(void *ibhdr, void *packet_data,
255 static int hfi1_filter_ib_pkey(void *ibhdr, void *packet_data, void *value);
256 static int hfi1_filter_direction(void *ibhdr, void *packet_data, void *value);
258 static const struct hfi1_filter_array hfi1_filters[] = {
260 { hfi1_filter_dlid },
261 { hfi1_filter_mad_mgmt_class },
262 { hfi1_filter_qp_number },
263 { hfi1_filter_ibpacket_type },
264 { hfi1_filter_ib_service_level },
265 { hfi1_filter_ib_pkey },
266 { hfi1_filter_direction },
269 #define HFI1_MAX_FILTERS ARRAY_SIZE(hfi1_filters)
270 #define HFI1_DIAG_MINOR_BASE 129
272 static int hfi1_snoop_add(struct hfi1_devdata *dd, const char *name);
274 int hfi1_diag_add(struct hfi1_devdata *dd)
279 snprintf(name, sizeof(name), "%s_diagpkt%d", class_name(),
282 * Do this for each device as opposed to the normal diagpkt
283 * interface which is one per host
285 ret = hfi1_snoop_add(dd, name);
287 dd_dev_err(dd, "Unable to init snoop/capture device");
289 snprintf(name, sizeof(name), "%s_diagpkt", class_name());
290 if (atomic_inc_return(&diagpkt_count) == 1) {
291 ret = hfi1_cdev_init(HFI1_DIAGPKT_MINOR, name,
292 &diagpkt_file_ops, &diagpkt_cdev,
293 &diagpkt_device, false);
299 /* this must be called w/ dd->snoop_in_lock held */
300 static void drain_snoop_list(struct list_head *queue)
302 struct list_head *pos, *q;
303 struct snoop_packet *packet;
305 list_for_each_safe(pos, q, queue) {
306 packet = list_entry(pos, struct snoop_packet, list);
312 static void hfi1_snoop_remove(struct hfi1_devdata *dd)
314 unsigned long flags = 0;
316 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
317 drain_snoop_list(&dd->hfi1_snoop.queue);
318 hfi1_cdev_cleanup(&dd->hfi1_snoop.cdev, &dd->hfi1_snoop.class_dev);
319 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
322 void hfi1_diag_remove(struct hfi1_devdata *dd)
324 hfi1_snoop_remove(dd);
325 if (atomic_dec_and_test(&diagpkt_count))
326 hfi1_cdev_cleanup(&diagpkt_cdev, &diagpkt_device);
327 hfi1_cdev_cleanup(&dd->diag_cdev, &dd->diag_device);
331 * Allocated structure shared between the credit return mechanism and
334 struct diagpkt_wait {
335 struct completion credits_returned;
341 * When each side is finished with the structure, they call this.
342 * The last user frees the structure.
344 static void put_diagpkt_wait(struct diagpkt_wait *wait)
346 if (atomic_dec_and_test(&wait->count))
351 * Callback from the credit return code. Set the complete, which
352 * will let diapkt_send() continue.
354 static void diagpkt_complete(void *arg, int code)
356 struct diagpkt_wait *wait = (struct diagpkt_wait *)arg;
359 complete(&wait->credits_returned);
360 put_diagpkt_wait(wait); /* finished with the structure */
364 * diagpkt_send - send a packet
365 * @dp: diag packet descriptor
367 static ssize_t diagpkt_send(struct diag_pkt *dp)
369 struct hfi1_devdata *dd;
370 struct send_context *sc;
371 struct pio_buf *pbuf;
374 u32 pkt_len, total_len;
375 pio_release_cb credit_cb = NULL;
376 void *credit_arg = NULL;
377 struct diagpkt_wait *wait = NULL;
380 dd = hfi1_lookup(dp->unit);
381 if (!dd || !(dd->flags & HFI1_PRESENT) || !dd->kregbase) {
385 if (!(dd->flags & HFI1_INITTED)) {
386 /* no hardware, freeze, etc. */
391 if (dp->version != _DIAG_PKT_VERS) {
392 dd_dev_err(dd, "Invalid version %u for diagpkt_write\n",
398 /* send count must be an exact number of dwords */
404 /* there is only port 1 */
410 /* need a valid context */
411 if (dp->sw_index >= dd->num_send_contexts) {
415 /* can only use kernel contexts */
416 if (dd->send_contexts[dp->sw_index].type != SC_KERNEL) {
420 /* must be allocated */
421 sc = dd->send_contexts[dp->sw_index].sc;
426 /* must be enabled */
427 if (!(sc->flags & SCF_ENABLED)) {
432 /* allocate a buffer and copy the data in */
433 tmpbuf = vmalloc(dp->len);
439 if (copy_from_user(tmpbuf,
440 (const void __user *)(unsigned long)dp->data,
447 * pkt_len is how much data we have to write, includes header and data.
448 * total_len is length of the packet in Dwords plus the PBC should not
451 pkt_len = dp->len >> 2;
452 total_len = pkt_len + 2; /* PBC + packet */
454 /* if 0, fill in a default */
456 struct hfi1_pportdata *ppd = dd->pport;
458 hfi1_cdbg(PKT, "Generating PBC");
459 dp->pbc = create_pbc(ppd, 0, 0, 0, total_len);
461 hfi1_cdbg(PKT, "Using passed in PBC");
464 hfi1_cdbg(PKT, "Egress PBC content is 0x%llx", dp->pbc);
467 * The caller wants to wait until the packet is sent and to
468 * check for errors. The best we can do is wait until
469 * the buffer credits are returned and check if any packet
470 * error has occurred. If there are any late errors, this
471 * could miss it. If there are other senders who generate
472 * an error, this may find it. However, in general, it
475 if (dp->flags & F_DIAGPKT_WAIT) {
476 /* always force a credit return */
477 dp->pbc |= PBC_CREDIT_RETURN;
478 /* turn on credit return interrupts */
479 sc_add_credit_return_intr(sc);
480 wait = kmalloc(sizeof(*wait), GFP_KERNEL);
485 init_completion(&wait->credits_returned);
486 atomic_set(&wait->count, 2);
489 credit_cb = diagpkt_complete;
494 pbuf = sc_buffer_alloc(sc, total_len, credit_cb, credit_arg);
497 /* force a credit return and try again */
498 sc_return_credits(sc);
503 * No send buffer means no credit callback. Undo
504 * the wait set-up that was done above. We free wait
505 * because the callback will never be called.
507 if (dp->flags & F_DIAGPKT_WAIT) {
508 sc_del_credit_return_intr(sc);
516 pio_copy(dd, pbuf, dp->pbc, tmpbuf, pkt_len);
517 /* no flush needed as the HW knows the packet size */
521 if (dp->flags & F_DIAGPKT_WAIT) {
522 /* wait for credit return */
523 ret = wait_for_completion_interruptible(
524 &wait->credits_returned);
526 * If the wait returns an error, the wait was interrupted,
527 * e.g. with a ^C in the user program. The callback is
528 * still pending. This is OK as the wait structure is
529 * kmalloc'ed and the structure will free itself when
530 * all users are done with it.
532 * A context disable occurs on a send context restart, so
533 * include that in the list of errors below to check for.
534 * NOTE: PRC_FILL_ERR is at best informational and cannot
537 if (!ret && (((wait->code & PRC_STATUS_ERR) ||
538 (wait->code & PRC_FILL_ERR) ||
539 (wait->code & PRC_SC_DISABLE))))
542 put_diagpkt_wait(wait); /* finished with the structure */
543 sc_del_credit_return_intr(sc);
551 static ssize_t diagpkt_write(struct file *fp, const char __user *data,
552 size_t count, loff_t *off)
554 struct hfi1_devdata *dd;
555 struct send_context *sc;
560 if (count != sizeof(dp))
563 if (copy_from_user(&dp, data, sizeof(dp)))
567 * The Send Context is derived from the PbcVL value
568 * if PBC is populated
571 dd = hfi1_lookup(dp.unit);
574 vl = (dp.pbc >> PBC_VL_SHIFT) & PBC_VL_MASK;
577 dp.sw_index = sc->sw_index;
580 "Packet sent over VL %d via Send Context %u(%u)",
581 vl, sc->sw_index, sc->hw_context);
585 return diagpkt_send(&dp);
588 static int hfi1_snoop_add(struct hfi1_devdata *dd, const char *name)
592 dd->hfi1_snoop.mode_flag = 0;
593 spin_lock_init(&dd->hfi1_snoop.snoop_lock);
594 INIT_LIST_HEAD(&dd->hfi1_snoop.queue);
595 init_waitqueue_head(&dd->hfi1_snoop.waitq);
597 ret = hfi1_cdev_init(HFI1_SNOOP_CAPTURE_BASE + dd->unit, name,
599 &dd->hfi1_snoop.cdev, &dd->hfi1_snoop.class_dev,
603 dd_dev_err(dd, "Couldn't create %s device: %d", name, ret);
604 hfi1_cdev_cleanup(&dd->hfi1_snoop.cdev,
605 &dd->hfi1_snoop.class_dev);
611 static struct hfi1_devdata *hfi1_dd_from_sc_inode(struct inode *in)
613 int unit = iminor(in) - HFI1_SNOOP_CAPTURE_BASE;
614 struct hfi1_devdata *dd;
616 dd = hfi1_lookup(unit);
620 /* clear or restore send context integrity checks */
621 static void adjust_integrity_checks(struct hfi1_devdata *dd)
623 struct send_context *sc;
624 unsigned long sc_flags;
627 spin_lock_irqsave(&dd->sc_lock, sc_flags);
628 for (i = 0; i < dd->num_send_contexts; i++) {
631 sc = dd->send_contexts[i].sc;
634 continue; /* not allocated */
636 enable = likely(!HFI1_CAP_IS_KSET(NO_INTEGRITY)) &&
637 dd->hfi1_snoop.mode_flag != HFI1_PORT_SNOOP_MODE;
639 set_pio_integrity(sc);
641 if (enable) /* take HFI_CAP_* flags into account */
644 spin_unlock_irqrestore(&dd->sc_lock, sc_flags);
647 static int hfi1_snoop_open(struct inode *in, struct file *fp)
651 unsigned long flags = 0;
652 struct hfi1_devdata *dd;
653 struct list_head *queue;
655 mutex_lock(&hfi1_mutex);
657 dd = hfi1_dd_from_sc_inode(in);
664 * File mode determines snoop or capture. Some existing user
665 * applications expect the capture device to be able to be opened RDWR
666 * because they expect a dedicated capture device. For this reason we
667 * support a module param to force capture mode even if the file open
668 * mode matches snoop.
670 if ((fp->f_flags & O_ACCMODE) == O_RDONLY) {
671 snoop_dbg("Capture Enabled");
672 mode_flag = HFI1_PORT_CAPTURE_MODE;
673 } else if ((fp->f_flags & O_ACCMODE) == O_RDWR) {
674 snoop_dbg("Snoop Enabled");
675 mode_flag = HFI1_PORT_SNOOP_MODE;
677 snoop_dbg("Invalid");
681 queue = &dd->hfi1_snoop.queue;
684 * We are not supporting snoop and capture at the same time.
686 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
687 if (dd->hfi1_snoop.mode_flag) {
689 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
693 dd->hfi1_snoop.mode_flag = mode_flag;
694 drain_snoop_list(queue);
696 dd->hfi1_snoop.filter_callback = NULL;
697 dd->hfi1_snoop.filter_value = NULL;
700 * Send side packet integrity checks are not helpful when snooping so
701 * disable and re-enable when we stop snooping.
703 if (mode_flag == HFI1_PORT_SNOOP_MODE) {
704 /* clear after snoop mode is on */
705 adjust_integrity_checks(dd); /* clear */
708 * We also do not want to be doing the DLID LMC check for
711 dd->hfi1_snoop.dcc_cfg = read_csr(dd, DCC_CFG_PORT_CONFIG1);
712 write_csr(dd, DCC_CFG_PORT_CONFIG1,
713 (dd->hfi1_snoop.dcc_cfg >> 32) << 32);
717 * As soon as we set these function pointers the recv and send handlers
718 * are active. This is a race condition so we must make sure to drain
719 * the queue and init filter values above. Technically we should add
720 * locking here but all that will happen is on recv a packet will get
721 * allocated and get stuck on the snoop_lock before getting added to the
722 * queue. Same goes for send.
724 dd->rhf_rcv_function_map = snoop_rhf_rcv_functions;
725 dd->process_pio_send = snoop_send_pio_handler;
726 dd->process_dma_send = snoop_send_pio_handler;
727 dd->pio_inline_send = snoop_inline_pio_send;
729 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
733 mutex_unlock(&hfi1_mutex);
738 static int hfi1_snoop_release(struct inode *in, struct file *fp)
740 unsigned long flags = 0;
741 struct hfi1_devdata *dd;
744 dd = hfi1_dd_from_sc_inode(in);
748 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
750 /* clear the snoop mode before re-adjusting send context CSRs */
751 mode_flag = dd->hfi1_snoop.mode_flag;
752 dd->hfi1_snoop.mode_flag = 0;
755 * Drain the queue and clear the filters we are done with it. Don't
756 * forget to restore the packet integrity checks
758 drain_snoop_list(&dd->hfi1_snoop.queue);
759 if (mode_flag == HFI1_PORT_SNOOP_MODE) {
760 /* restore after snoop mode is clear */
761 adjust_integrity_checks(dd); /* restore */
764 * Also should probably reset the DCC_CONFIG1 register for DLID
765 * checking on incoming packets again. Use the value saved when
766 * opening the snoop device.
768 write_csr(dd, DCC_CFG_PORT_CONFIG1, dd->hfi1_snoop.dcc_cfg);
771 dd->hfi1_snoop.filter_callback = NULL;
772 kfree(dd->hfi1_snoop.filter_value);
773 dd->hfi1_snoop.filter_value = NULL;
776 * User is done snooping and capturing, return control to the normal
777 * handler. Re-enable SDMA handling.
779 dd->rhf_rcv_function_map = dd->normal_rhf_rcv_functions;
780 dd->process_pio_send = hfi1_verbs_send_pio;
781 dd->process_dma_send = hfi1_verbs_send_dma;
782 dd->pio_inline_send = pio_copy;
784 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
786 snoop_dbg("snoop/capture device released");
791 static unsigned int hfi1_snoop_poll(struct file *fp,
792 struct poll_table_struct *wait)
795 unsigned long flags = 0;
797 struct hfi1_devdata *dd;
799 dd = hfi1_dd_from_sc_inode(fp->f_inode);
803 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
805 poll_wait(fp, &dd->hfi1_snoop.waitq, wait);
806 if (!list_empty(&dd->hfi1_snoop.queue))
807 ret |= POLLIN | POLLRDNORM;
809 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
813 static ssize_t hfi1_snoop_write(struct file *fp, const char __user *data,
814 size_t count, loff_t *off)
816 struct diag_pkt dpkt;
817 struct hfi1_devdata *dd;
819 u8 byte_two, sl, sc5, sc4, vl, byte_one;
820 struct send_context *sc;
823 struct hfi1_ibport *ibp;
824 struct hfi1_pportdata *ppd;
826 dd = hfi1_dd_from_sc_inode(fp->f_inode);
831 snoop_dbg("received %lu bytes from user", count);
833 memset(&dpkt, 0, sizeof(struct diag_pkt));
834 dpkt.version = _DIAG_PKT_VERS;
835 dpkt.unit = dd->unit;
838 if (likely(!(snoop_flags & SNOOP_USE_METADATA))) {
840 * We need to generate the PBC and not let diagpkt_send do it,
841 * to do this we need the VL and the length in dwords.
842 * The VL can be determined by using the SL and looking up the
843 * SC. Then the SC can be converted into VL. The exception to
844 * this is those packets which are from an SMI queue pair.
845 * Since we can't detect anything about the QP here we have to
846 * rely on the SC. If its 0xF then we assume its SMI and
847 * do not look at the SL.
849 if (copy_from_user(&byte_one, data, 1))
852 if (copy_from_user(&byte_two, data + 1, 1))
855 sc4 = (byte_one >> 4) & 0xf;
857 snoop_dbg("Detected VL15 packet ignoring SL in packet");
860 sl = (byte_two >> 4) & 0xf;
861 ibp = to_iport(&dd->verbs_dev.rdi.ibdev, 1);
862 sc5 = ibp->sl_to_sc[sl];
863 vl = sc_to_vlt(dd, sc5);
865 snoop_dbg("VL %d does not match SC %d of packet",
871 sc = dd->vld[vl].sc; /* Look up the context based on VL */
873 dpkt.sw_index = sc->sw_index;
874 snoop_dbg("Sending on context %u(%u)", sc->sw_index,
877 snoop_dbg("Could not find context for vl %d", vl);
881 len = (count >> 2) + 2; /* Add in PBC */
882 pbc = create_pbc(ppd, 0, 0, vl, len);
884 if (copy_from_user(&pbc, data, sizeof(pbc)))
886 vl = (pbc >> PBC_VL_SHIFT) & PBC_VL_MASK;
887 sc = dd->vld[vl].sc; /* Look up the context based on VL */
889 dpkt.sw_index = sc->sw_index;
891 snoop_dbg("Could not find context for vl %d", vl);
895 count -= sizeof(pbc);
898 dpkt.data = (unsigned long)data;
900 snoop_dbg("PBC: vl=0x%llx Length=0x%llx",
905 ret = diagpkt_send(&dpkt);
907 * diagpkt_send only returns number of bytes in the diagpkt so patch
908 * that up here before returning.
910 if (ret == sizeof(dpkt))
916 static ssize_t hfi1_snoop_read(struct file *fp, char __user *data,
917 size_t pkt_len, loff_t *off)
920 unsigned long flags = 0;
921 struct snoop_packet *packet = NULL;
922 struct hfi1_devdata *dd;
924 dd = hfi1_dd_from_sc_inode(fp->f_inode);
928 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
930 while (list_empty(&dd->hfi1_snoop.queue)) {
931 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
933 if (fp->f_flags & O_NONBLOCK)
936 if (wait_event_interruptible(
937 dd->hfi1_snoop.waitq,
938 !list_empty(&dd->hfi1_snoop.queue)))
941 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
944 if (!list_empty(&dd->hfi1_snoop.queue)) {
945 packet = list_entry(dd->hfi1_snoop.queue.next,
946 struct snoop_packet, list);
947 list_del(&packet->list);
948 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
949 if (pkt_len >= packet->total_len) {
950 if (copy_to_user(data, packet->data,
954 ret = packet->total_len;
961 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
968 * hfi1_assign_snoop_link_credits -- Set up credits for VL15 and others
969 * @ppd : ptr to hfi1 port data
970 * @value : options from user space
972 * Assumes the rest of the CM credit registers are zero from a
973 * previous global or credit reset.
974 * Leave shared count at zero for both global and all vls.
975 * In snoop mode ideally we don't use shared credits
976 * Reserve 8.5k for VL15
977 * If total credits less than 8.5kbytes return error.
978 * Divide the rest of the credits across VL0 to VL7 and if
979 * each of these levels has less than 34 credits (at least 2048 + 128 bytes)
980 * return with an error.
981 * The credit registers will be reset to zero on link negotiation or link up
982 * so this function should be activated from user space only if the port has
983 * gone past link negotiation and link up.
985 * Return -- 0 if successful else error condition
988 static long hfi1_assign_snoop_link_credits(struct hfi1_pportdata *ppd,
991 #define OPA_MIN_PER_VL_CREDITS 34 /* 2048 + 128 bytes */
992 struct buffer_control t;
994 struct hfi1_devdata *dd = ppd->dd;
995 u16 total_credits = (value >> 16) & 0xffff;
996 u16 vl15_credits = dd->vl15_init / 2;
998 __be16 be_per_vl_credits;
1000 if (!(ppd->host_link_state & HLS_UP))
1002 if (total_credits < vl15_credits)
1005 per_vl_credits = (total_credits - vl15_credits) / TXE_NUM_DATA_VL;
1007 if (per_vl_credits < OPA_MIN_PER_VL_CREDITS)
1010 memset(&t, 0, sizeof(t));
1011 be_per_vl_credits = cpu_to_be16(per_vl_credits);
1013 for (i = 0; i < TXE_NUM_DATA_VL; i++)
1014 t.vl[i].dedicated = be_per_vl_credits;
1016 t.vl[15].dedicated = cpu_to_be16(vl15_credits);
1017 return set_buffer_control(ppd, &t);
1020 snoop_dbg("port_state = 0x%x, total_credits = %d, vl15_credits = %d",
1021 ppd->host_link_state, total_credits, vl15_credits);
1026 static long hfi1_ioctl(struct file *fp, unsigned int cmd, unsigned long arg)
1028 struct hfi1_devdata *dd;
1029 void *filter_value = NULL;
1035 unsigned long flags = 0;
1036 unsigned long *argp = NULL;
1037 struct hfi1_packet_filter_command filter_cmd = {0};
1039 struct hfi1_pportdata *ppd = NULL;
1041 struct hfi1_link_info link_info;
1042 int read_cmd, write_cmd, read_ok, write_ok;
1044 dd = hfi1_dd_from_sc_inode(fp->f_inode);
1048 mode_flag = dd->hfi1_snoop.mode_flag;
1049 read_cmd = _IOC_DIR(cmd) & _IOC_READ;
1050 write_cmd = _IOC_DIR(cmd) & _IOC_WRITE;
1051 write_ok = access_ok(VERIFY_WRITE, (void __user *)arg, _IOC_SIZE(cmd));
1052 read_ok = access_ok(VERIFY_READ, (void __user *)arg, _IOC_SIZE(cmd));
1054 if ((read_cmd && !write_ok) || (write_cmd && !read_ok))
1057 if (!capable(CAP_SYS_ADMIN))
1060 if ((mode_flag & HFI1_PORT_CAPTURE_MODE) &&
1061 (cmd != HFI1_SNOOP_IOCCLEARQUEUE) &&
1062 (cmd != HFI1_SNOOP_IOCCLEARFILTER) &&
1063 (cmd != HFI1_SNOOP_IOCSETFILTER))
1064 /* Capture devices are allowed only 3 operations
1065 * 1.Clear capture queue
1066 * 2.Clear capture filter
1067 * 3.Set capture filter
1068 * Other are invalid.
1073 case HFI1_SNOOP_IOCSETLINKSTATE_EXTRA:
1074 memset(&link_info, 0, sizeof(link_info));
1076 if (copy_from_user(&link_info,
1077 (struct hfi1_link_info __user *)arg,
1081 value = link_info.port_state;
1082 index = link_info.port_number;
1083 if (index > dd->num_pports - 1)
1086 ppd = &dd->pport[index];
1090 /* What we want to transition to */
1091 phys_state = (value >> 4) & 0xF;
1092 link_state = value & 0xF;
1093 snoop_dbg("Setting link state 0x%x", value);
1095 switch (link_state) {
1097 if (phys_state == 0)
1101 switch (phys_state) {
1103 dev_state = HLS_DN_DOWNDEF;
1106 dev_state = HLS_DN_POLL;
1109 dev_state = HLS_DN_DISABLE;
1114 ret = set_link_state(ppd, dev_state);
1117 ret = set_link_state(ppd, HLS_UP_ARMED);
1119 send_idle_sma(dd, SMA_IDLE_ARM);
1121 case IB_PORT_ACTIVE:
1122 ret = set_link_state(ppd, HLS_UP_ACTIVE);
1124 send_idle_sma(dd, SMA_IDLE_ACTIVE);
1133 case HFI1_SNOOP_IOCGETLINKSTATE:
1134 case HFI1_SNOOP_IOCGETLINKSTATE_EXTRA:
1135 if (cmd == HFI1_SNOOP_IOCGETLINKSTATE_EXTRA) {
1136 memset(&link_info, 0, sizeof(link_info));
1137 if (copy_from_user(&link_info,
1138 (struct hfi1_link_info __user *)arg,
1141 index = link_info.port_number;
1143 ret = __get_user(index, (int __user *)arg);
1148 if (index > dd->num_pports - 1)
1151 ppd = &dd->pport[index];
1155 value = hfi1_ibphys_portstate(ppd);
1157 value |= driver_lstate(ppd);
1159 snoop_dbg("Link port | Link State: %d", value);
1161 if ((cmd == HFI1_SNOOP_IOCGETLINKSTATE_EXTRA) ||
1162 (cmd == HFI1_SNOOP_IOCSETLINKSTATE_EXTRA)) {
1163 link_info.port_state = value;
1164 link_info.node_guid = cpu_to_be64(ppd->guid);
1165 link_info.link_speed_active =
1166 ppd->link_speed_active;
1167 link_info.link_width_active =
1168 ppd->link_width_active;
1169 if (copy_to_user((struct hfi1_link_info __user *)arg,
1170 &link_info, sizeof(link_info)))
1173 ret = __put_user(value, (int __user *)arg);
1177 case HFI1_SNOOP_IOCCLEARQUEUE:
1178 snoop_dbg("Clearing snoop queue");
1179 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1180 drain_snoop_list(&dd->hfi1_snoop.queue);
1181 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1184 case HFI1_SNOOP_IOCCLEARFILTER:
1185 snoop_dbg("Clearing filter");
1186 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1187 if (dd->hfi1_snoop.filter_callback) {
1188 /* Drain packets first */
1189 drain_snoop_list(&dd->hfi1_snoop.queue);
1190 dd->hfi1_snoop.filter_callback = NULL;
1192 kfree(dd->hfi1_snoop.filter_value);
1193 dd->hfi1_snoop.filter_value = NULL;
1194 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1197 case HFI1_SNOOP_IOCSETFILTER:
1198 snoop_dbg("Setting filter");
1199 /* just copy command structure */
1200 argp = (unsigned long *)arg;
1201 if (copy_from_user(&filter_cmd, (void __user *)argp,
1202 sizeof(filter_cmd)))
1205 if (filter_cmd.opcode >= HFI1_MAX_FILTERS) {
1206 pr_alert("Invalid opcode in request\n");
1210 snoop_dbg("Opcode %d Len %d Ptr %p",
1211 filter_cmd.opcode, filter_cmd.length,
1212 filter_cmd.value_ptr);
1214 filter_value = kcalloc(filter_cmd.length, sizeof(u8),
1219 /* copy remaining data from userspace */
1220 if (copy_from_user((u8 *)filter_value,
1221 (void __user *)filter_cmd.value_ptr,
1222 filter_cmd.length)) {
1223 kfree(filter_value);
1226 /* Drain packets first */
1227 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1228 drain_snoop_list(&dd->hfi1_snoop.queue);
1229 dd->hfi1_snoop.filter_callback =
1230 hfi1_filters[filter_cmd.opcode].filter;
1231 /* just in case we see back to back sets */
1232 kfree(dd->hfi1_snoop.filter_value);
1233 dd->hfi1_snoop.filter_value = filter_value;
1234 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1236 case HFI1_SNOOP_IOCGETVERSION:
1237 value = SNOOP_CAPTURE_VERSION;
1238 snoop_dbg("Getting version: %d", value);
1239 ret = __put_user(value, (int __user *)arg);
1241 case HFI1_SNOOP_IOCSET_OPTS:
1243 ret = __get_user(value, (int __user *)arg);
1247 snoop_dbg("Setting snoop option %d", value);
1248 if (value & SNOOP_DROP_SEND)
1249 snoop_flags |= SNOOP_DROP_SEND;
1250 if (value & SNOOP_USE_METADATA)
1251 snoop_flags |= SNOOP_USE_METADATA;
1252 if (value & (SNOOP_SET_VL0TOVL15)) {
1253 ppd = &dd->pport[0]; /* first port will do */
1254 ret = hfi1_assign_snoop_link_credits(ppd, value);
1264 static void snoop_list_add_tail(struct snoop_packet *packet,
1265 struct hfi1_devdata *dd)
1267 unsigned long flags = 0;
1269 spin_lock_irqsave(&dd->hfi1_snoop.snoop_lock, flags);
1270 if (likely((dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE) ||
1271 (dd->hfi1_snoop.mode_flag & HFI1_PORT_CAPTURE_MODE))) {
1272 list_add_tail(&packet->list, &dd->hfi1_snoop.queue);
1273 snoop_dbg("Added packet to list");
1277 * Technically we can could have closed the snoop device while waiting
1278 * on the above lock and it is gone now. The snoop mode_flag will
1279 * prevent us from adding the packet to the queue though.
1282 spin_unlock_irqrestore(&dd->hfi1_snoop.snoop_lock, flags);
1283 wake_up_interruptible(&dd->hfi1_snoop.waitq);
1286 static inline int hfi1_filter_check(void *val, const char *msg)
1289 snoop_dbg("Error invalid %s value for filter", msg);
1290 return HFI1_FILTER_ERR;
1295 static int hfi1_filter_lid(void *ibhdr, void *packet_data, void *value)
1297 struct hfi1_ib_header *hdr;
1300 ret = hfi1_filter_check(ibhdr, "header");
1303 ret = hfi1_filter_check(value, "user");
1306 hdr = (struct hfi1_ib_header *)ibhdr;
1308 if (*((u16 *)value) == be16_to_cpu(hdr->lrh[3])) /* matches slid */
1309 return HFI1_FILTER_HIT; /* matched */
1311 return HFI1_FILTER_MISS; /* Not matched */
1314 static int hfi1_filter_dlid(void *ibhdr, void *packet_data, void *value)
1316 struct hfi1_ib_header *hdr;
1319 ret = hfi1_filter_check(ibhdr, "header");
1322 ret = hfi1_filter_check(value, "user");
1326 hdr = (struct hfi1_ib_header *)ibhdr;
1328 if (*((u16 *)value) == be16_to_cpu(hdr->lrh[1]))
1329 return HFI1_FILTER_HIT;
1331 return HFI1_FILTER_MISS;
1334 /* Not valid for outgoing packets, send handler passes null for data*/
1335 static int hfi1_filter_mad_mgmt_class(void *ibhdr, void *packet_data,
1338 struct hfi1_ib_header *hdr;
1339 struct hfi1_other_headers *ohdr = NULL;
1340 struct ib_smp *smp = NULL;
1344 ret = hfi1_filter_check(ibhdr, "header");
1347 ret = hfi1_filter_check(packet_data, "packet_data");
1350 ret = hfi1_filter_check(value, "user");
1354 hdr = (struct hfi1_ib_header *)ibhdr;
1357 if ((be16_to_cpu(hdr->lrh[0]) & 3) == HFI1_LRH_BTH)
1358 ohdr = &hdr->u.oth; /* LRH + BTH + DETH */
1360 ohdr = &hdr->u.l.oth; /* LRH + GRH + BTH + DETH */
1362 qpn = be32_to_cpu(ohdr->bth[1]) & 0x00FFFFFF;
1364 smp = (struct ib_smp *)packet_data;
1365 if (*((u8 *)value) == smp->mgmt_class)
1366 return HFI1_FILTER_HIT;
1368 return HFI1_FILTER_MISS;
1370 return HFI1_FILTER_ERR;
1373 static int hfi1_filter_qp_number(void *ibhdr, void *packet_data, void *value)
1375 struct hfi1_ib_header *hdr;
1376 struct hfi1_other_headers *ohdr = NULL;
1379 ret = hfi1_filter_check(ibhdr, "header");
1382 ret = hfi1_filter_check(value, "user");
1386 hdr = (struct hfi1_ib_header *)ibhdr;
1389 if ((be16_to_cpu(hdr->lrh[0]) & 3) == HFI1_LRH_BTH)
1390 ohdr = &hdr->u.oth; /* LRH + BTH + DETH */
1392 ohdr = &hdr->u.l.oth; /* LRH + GRH + BTH + DETH */
1393 if (*((u32 *)value) == (be32_to_cpu(ohdr->bth[1]) & 0x00FFFFFF))
1394 return HFI1_FILTER_HIT;
1396 return HFI1_FILTER_MISS;
1399 static int hfi1_filter_ibpacket_type(void *ibhdr, void *packet_data,
1404 struct hfi1_ib_header *hdr;
1405 struct hfi1_other_headers *ohdr = NULL;
1408 ret = hfi1_filter_check(ibhdr, "header");
1411 ret = hfi1_filter_check(value, "user");
1415 hdr = (struct hfi1_ib_header *)ibhdr;
1417 lnh = (be16_to_cpu(hdr->lrh[0]) & 3);
1419 if (lnh == HFI1_LRH_BTH)
1421 else if (lnh == HFI1_LRH_GRH)
1422 ohdr = &hdr->u.l.oth;
1424 return HFI1_FILTER_ERR;
1426 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1428 if (*((u8 *)value) == ((opcode >> 5) & 0x7))
1429 return HFI1_FILTER_HIT;
1431 return HFI1_FILTER_MISS;
1434 static int hfi1_filter_ib_service_level(void *ibhdr, void *packet_data,
1437 struct hfi1_ib_header *hdr;
1440 ret = hfi1_filter_check(ibhdr, "header");
1443 ret = hfi1_filter_check(value, "user");
1447 hdr = (struct hfi1_ib_header *)ibhdr;
1449 if ((*((u8 *)value)) == ((be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF))
1450 return HFI1_FILTER_HIT;
1452 return HFI1_FILTER_MISS;
1455 static int hfi1_filter_ib_pkey(void *ibhdr, void *packet_data, void *value)
1458 struct hfi1_ib_header *hdr;
1459 struct hfi1_other_headers *ohdr = NULL;
1462 ret = hfi1_filter_check(ibhdr, "header");
1465 ret = hfi1_filter_check(value, "user");
1469 hdr = (struct hfi1_ib_header *)ibhdr;
1471 lnh = (be16_to_cpu(hdr->lrh[0]) & 3);
1472 if (lnh == HFI1_LRH_BTH)
1474 else if (lnh == HFI1_LRH_GRH)
1475 ohdr = &hdr->u.l.oth;
1477 return HFI1_FILTER_ERR;
1479 /* P_key is 16-bit entity, however top most bit indicates
1480 * type of membership. 0 for limited and 1 for Full.
1481 * Limited members cannot accept information from other
1482 * Limited members, but communication is allowed between
1483 * every other combination of membership.
1484 * Hence we'll omit comparing top-most bit while filtering
1487 if ((*(u16 *)value & 0x7FFF) ==
1488 ((be32_to_cpu(ohdr->bth[0])) & 0x7FFF))
1489 return HFI1_FILTER_HIT;
1491 return HFI1_FILTER_MISS;
1495 * If packet_data is NULL then this is coming from one of the send functions.
1496 * Thus we know if its an ingressed or egressed packet.
1498 static int hfi1_filter_direction(void *ibhdr, void *packet_data, void *value)
1500 u8 user_dir = *(u8 *)value;
1503 ret = hfi1_filter_check(value, "user");
1508 /* Incoming packet */
1509 if (user_dir & HFI1_SNOOP_INGRESS)
1510 return HFI1_FILTER_HIT;
1512 /* Outgoing packet */
1513 if (user_dir & HFI1_SNOOP_EGRESS)
1514 return HFI1_FILTER_HIT;
1517 return HFI1_FILTER_MISS;
1521 * Allocate a snoop packet. The structure that is stored in the ring buffer, not
1522 * to be confused with an hfi packet type.
1524 static struct snoop_packet *allocate_snoop_packet(u32 hdr_len,
1528 struct snoop_packet *packet;
1530 packet = kzalloc(sizeof(*packet) + hdr_len + data_len
1532 GFP_ATOMIC | __GFP_NOWARN);
1534 INIT_LIST_HEAD(&packet->list);
1540 * Instead of having snoop and capture code intermixed with the recv functions,
1541 * both the interrupt handler and hfi1_ib_rcv() we are going to hijack the call
1542 * and land in here for snoop/capture but if not enabled the call will go
1543 * through as before. This gives us a single point to constrain all of the snoop
1544 * snoop recv logic. There is nothing special that needs to happen for bypass
1545 * packets. This routine should not try to look into the packet. It just copied
1546 * it. There is no guarantee for filters when it comes to bypass packets as
1547 * there is no specific support. Bottom line is this routine does now even know
1548 * what a bypass packet is.
1550 int snoop_recv_handler(struct hfi1_packet *packet)
1552 struct hfi1_pportdata *ppd = packet->rcd->ppd;
1553 struct hfi1_ib_header *hdr = packet->hdr;
1554 int header_size = packet->hlen;
1555 void *data = packet->ebuf;
1556 u32 tlen = packet->tlen;
1557 struct snoop_packet *s_packet = NULL;
1561 struct capture_md md;
1563 snoop_dbg("PACKET IN: hdr size %d tlen %d data %p", header_size, tlen,
1566 trace_snoop_capture(ppd->dd, header_size, hdr, tlen - header_size,
1569 if (!ppd->dd->hfi1_snoop.filter_callback) {
1570 snoop_dbg("filter not set");
1571 ret = HFI1_FILTER_HIT;
1573 ret = ppd->dd->hfi1_snoop.filter_callback(hdr, data,
1574 ppd->dd->hfi1_snoop.filter_value);
1578 case HFI1_FILTER_ERR:
1579 snoop_dbg("Error in filter call");
1581 case HFI1_FILTER_MISS:
1582 snoop_dbg("Filter Miss");
1584 case HFI1_FILTER_HIT:
1586 if (ppd->dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE)
1588 if ((snoop_mode == 0) ||
1589 unlikely(snoop_flags & SNOOP_USE_METADATA))
1590 md_len = sizeof(struct capture_md);
1592 s_packet = allocate_snoop_packet(header_size,
1596 if (unlikely(!s_packet)) {
1597 dd_dev_warn_ratelimited(ppd->dd, "Unable to allocate snoop/capture packet\n");
1602 memset(&md, 0, sizeof(struct capture_md));
1604 md.dir = PKT_DIR_INGRESS;
1605 md.u.rhf = packet->rhf;
1606 memcpy(s_packet->data, &md, md_len);
1609 /* We should always have a header */
1611 memcpy(s_packet->data + md_len, hdr, header_size);
1613 dd_dev_err(ppd->dd, "Unable to copy header to snoop/capture packet\n");
1619 * Packets with no data are possible. If there is no data needed
1620 * to take care of the last 4 bytes which are normally included
1621 * with data buffers and are included in tlen. Since we kzalloc
1622 * the buffer we do not need to set any values but if we decide
1623 * not to use kzalloc we should zero them.
1626 memcpy(s_packet->data + header_size + md_len, data,
1627 tlen - header_size);
1629 s_packet->total_len = tlen + md_len;
1630 snoop_list_add_tail(s_packet, ppd->dd);
1633 * If we are snooping the packet not capturing then throw away
1634 * after adding to the list.
1636 snoop_dbg("Capturing packet");
1637 if (ppd->dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE) {
1638 snoop_dbg("Throwing packet away");
1640 * If we are dropping the packet we still may need to
1641 * handle the case where error flags are set, this is
1642 * normally done by the type specific handler but that
1643 * won't be called in this case.
1645 if (unlikely(rhf_err_flags(packet->rhf)))
1646 handle_eflags(packet);
1648 /* throw the packet on the floor */
1649 return RHF_RCV_CONTINUE;
1657 * We do not care what type of packet came in here - just pass it off
1658 * to the normal handler.
1660 return ppd->dd->normal_rhf_rcv_functions[rhf_rcv_type(packet->rhf)]
1665 * Handle snooping and capturing packets when sdma is being used.
1667 int snoop_send_dma_handler(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
1670 pr_alert("Snooping/Capture of Send DMA Packets Is Not Supported!\n");
1671 snoop_dbg("Unsupported Operation");
1672 return hfi1_verbs_send_dma(qp, ps, 0);
1676 * Handle snooping and capturing packets when pio is being used. Does not handle
1677 * bypass packets. The only way to send a bypass packet currently is to use the
1678 * diagpkt interface. When that interface is enable snoop/capture is not.
1680 int snoop_send_pio_handler(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
1683 u32 hdrwords = qp->s_hdrwords;
1684 struct rvt_sge_state *ss = qp->s_cur_sge;
1685 u32 len = qp->s_cur_size;
1686 u32 dwords = (len + 3) >> 2;
1687 u32 plen = hdrwords + dwords + 2; /* includes pbc */
1688 struct hfi1_pportdata *ppd = ps->ppd;
1689 struct snoop_packet *s_packet = NULL;
1690 u32 *hdr = (u32 *)&ps->s_txreq->phdr.hdr;
1692 struct rvt_sge_state temp_ss;
1694 void *data_start = NULL;
1698 struct capture_md md;
1700 u32 hdr_len = hdrwords << 2;
1701 u32 tlen = HFI1_GET_PKT_LEN(&ps->s_txreq->phdr.hdr);
1705 snoop_dbg("PACKET OUT: hdrword %u len %u plen %u dwords %u tlen %u",
1706 hdrwords, len, plen, dwords, tlen);
1707 if (ppd->dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE)
1709 if ((snoop_mode == 0) ||
1710 unlikely(snoop_flags & SNOOP_USE_METADATA))
1711 md_len = sizeof(struct capture_md);
1713 /* not using ss->total_len as arg 2 b/c that does not count CRC */
1714 s_packet = allocate_snoop_packet(hdr_len, tlen - hdr_len, md_len);
1716 if (unlikely(!s_packet)) {
1717 dd_dev_warn_ratelimited(ppd->dd, "Unable to allocate snoop/capture packet\n");
1721 s_packet->total_len = tlen + md_len;
1724 memset(&md, 0, sizeof(struct capture_md));
1726 md.dir = PKT_DIR_EGRESS;
1727 if (likely(pbc == 0)) {
1728 vl = be16_to_cpu(ps->s_txreq->phdr.hdr.lrh[0]) >> 12;
1729 md.u.pbc = create_pbc(ppd, 0, qp->s_srate, vl, plen);
1733 memcpy(s_packet->data, &md, md_len);
1740 memcpy(s_packet->data + md_len, hdr, hdr_len);
1743 "Unable to copy header to snoop/capture packet\n");
1749 data = s_packet->data + hdr_len + md_len;
1754 * The update_sge() function below will not modify the
1755 * individual SGEs in the array. It will make a copy each time
1756 * and operate on that. So we only need to copy this instance
1757 * and it won't impact PIO.
1762 snoop_dbg("Need to copy %d bytes", length);
1764 void *addr = temp_ss.sge.vaddr;
1765 u32 slen = temp_ss.sge.length;
1767 if (slen > length) {
1769 snoop_dbg("slen %d > len %d", slen, length);
1771 snoop_dbg("copy %d to %p", slen, addr);
1772 memcpy(data, addr, slen);
1773 update_sge(&temp_ss, slen);
1776 snoop_dbg("data is now %p bytes left %d", data, length);
1778 snoop_dbg("Completed SGE copy");
1782 * Why do the filter check down here? Because the event tracing has its
1783 * own filtering and we need to have the walked the SGE list.
1785 if (!ppd->dd->hfi1_snoop.filter_callback) {
1786 snoop_dbg("filter not set\n");
1787 ret = HFI1_FILTER_HIT;
1789 ret = ppd->dd->hfi1_snoop.filter_callback(
1790 &ps->s_txreq->phdr.hdr,
1792 ppd->dd->hfi1_snoop.filter_value);
1796 case HFI1_FILTER_ERR:
1797 snoop_dbg("Error in filter call");
1799 case HFI1_FILTER_MISS:
1800 snoop_dbg("Filter Miss");
1803 case HFI1_FILTER_HIT:
1804 snoop_dbg("Capturing packet");
1805 snoop_list_add_tail(s_packet, ppd->dd);
1807 if (unlikely((snoop_flags & SNOOP_DROP_SEND) &&
1808 (ppd->dd->hfi1_snoop.mode_flag &
1809 HFI1_PORT_SNOOP_MODE))) {
1810 unsigned long flags;
1812 snoop_dbg("Dropping packet");
1814 spin_lock_irqsave(&qp->s_lock, flags);
1819 spin_unlock_irqrestore(&qp->s_lock, flags);
1820 } else if (qp->ibqp.qp_type == IB_QPT_RC) {
1821 spin_lock_irqsave(&qp->s_lock, flags);
1822 hfi1_rc_send_complete(qp,
1823 &ps->s_txreq->phdr.hdr);
1824 spin_unlock_irqrestore(&qp->s_lock, flags);
1828 * If snoop is dropping the packet we need to put the
1829 * txreq back because no one else will.
1831 hfi1_put_txreq(ps->s_txreq);
1840 return hfi1_verbs_send_pio(qp, ps, md.u.pbc);
1844 * Callers of this must pass a hfi1_ib_header type for the from ptr. Currently
1845 * this can be used anywhere, but the intention is for inline ACKs for RC and
1846 * CCA packets. We don't restrict this usage though.
1848 void snoop_inline_pio_send(struct hfi1_devdata *dd, struct pio_buf *pbuf,
1849 u64 pbc, const void *from, size_t count)
1853 struct capture_md md;
1854 struct snoop_packet *s_packet = NULL;
1857 * count is in dwords so we need to convert to bytes.
1858 * We also need to account for CRC which would be tacked on by hardware.
1860 int packet_len = (count << 2) + 4;
1863 snoop_dbg("ACK OUT: len %d", packet_len);
1865 if (!dd->hfi1_snoop.filter_callback) {
1866 snoop_dbg("filter not set");
1867 ret = HFI1_FILTER_HIT;
1869 ret = dd->hfi1_snoop.filter_callback(
1870 (struct hfi1_ib_header *)from,
1872 dd->hfi1_snoop.filter_value);
1876 case HFI1_FILTER_ERR:
1877 snoop_dbg("Error in filter call");
1879 case HFI1_FILTER_MISS:
1880 snoop_dbg("Filter Miss");
1882 case HFI1_FILTER_HIT:
1883 snoop_dbg("Capturing packet");
1884 if (dd->hfi1_snoop.mode_flag & HFI1_PORT_SNOOP_MODE)
1886 if ((snoop_mode == 0) ||
1887 unlikely(snoop_flags & SNOOP_USE_METADATA))
1888 md_len = sizeof(struct capture_md);
1890 s_packet = allocate_snoop_packet(packet_len, 0, md_len);
1892 if (unlikely(!s_packet)) {
1893 dd_dev_warn_ratelimited(dd, "Unable to allocate snoop/capture packet\n");
1894 goto inline_pio_out;
1897 s_packet->total_len = packet_len + md_len;
1899 /* Fill in the metadata for the packet */
1901 memset(&md, 0, sizeof(struct capture_md));
1903 md.dir = PKT_DIR_EGRESS;
1905 memcpy(s_packet->data, &md, md_len);
1908 /* Add the packet data which is a single buffer */
1909 memcpy(s_packet->data + md_len, from, packet_len);
1911 snoop_list_add_tail(s_packet, dd);
1913 if (unlikely((snoop_flags & SNOOP_DROP_SEND) && snoop_mode)) {
1914 snoop_dbg("Dropping packet");
1923 pio_copy(dd, pbuf, pbc, from, count);