1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
15 #include <sys/epoll.h>
16 #include <sys/signalfd.h>
17 #include <sys/ioctl.h>
18 #include <sys/eventfd.h>
22 #include <rte_common.h>
23 #include <rte_interrupts.h>
24 #include <rte_memory.h>
25 #include <rte_launch.h>
27 #include <rte_per_lcore.h>
28 #include <rte_lcore.h>
29 #include <rte_atomic.h>
30 #include <rte_branch_prediction.h>
31 #include <rte_debug.h>
33 #include <rte_errno.h>
34 #include <rte_spinlock.h>
35 #include <rte_pause.h>
38 #include "eal_private.h"
40 #include "eal_thread.h"
42 #define EAL_INTR_EPOLL_WAIT_FOREVER (-1)
43 #define NB_OTHER_INTR 1
45 static RTE_DEFINE_PER_LCORE(int, _epfd) = -1; /**< epoll fd per thread */
61 * union buffer for reading on different devices
63 union rte_intr_read_buffer {
64 int uio_intr_count; /* for uio device */
66 uint64_t vfio_intr_count; /* for vfio device */
68 uint64_t timerfd_num; /* for timerfd */
69 char charbuf[16]; /* for others */
72 TAILQ_HEAD(rte_intr_cb_list, rte_intr_callback);
73 TAILQ_HEAD(rte_intr_source_list, rte_intr_source);
75 struct rte_intr_callback {
76 TAILQ_ENTRY(rte_intr_callback) next;
77 rte_intr_callback_fn cb_fn; /**< callback address */
78 void *cb_arg; /**< parameter for callback */
79 uint8_t pending_delete; /**< delete after callback is called */
80 rte_intr_unregister_callback_fn ucb_fn; /**< fn to call before cb is deleted */
83 struct rte_intr_source {
84 TAILQ_ENTRY(rte_intr_source) next;
85 struct rte_intr_handle intr_handle; /**< interrupt handle */
86 struct rte_intr_cb_list callbacks; /**< user callbacks */
90 /* global spinlock for interrupt data operation */
91 static rte_spinlock_t intr_lock = RTE_SPINLOCK_INITIALIZER;
93 /* union buffer for pipe read/write */
94 static union intr_pipefds intr_pipe;
96 /* interrupt sources list */
97 static struct rte_intr_source_list intr_sources;
99 /* interrupt handling thread */
100 static pthread_t intr_thread;
102 /* VFIO interrupts */
105 #define IRQ_SET_BUF_LEN (sizeof(struct vfio_irq_set) + sizeof(int))
106 /* irq set buffer length for queue interrupts and LSC interrupt */
107 #define MSIX_IRQ_SET_BUF_LEN (sizeof(struct vfio_irq_set) + \
108 sizeof(int) * (RTE_MAX_RXTX_INTR_VEC_ID + 1))
110 /* enable legacy (INTx) interrupts */
112 vfio_enable_intx(const struct rte_intr_handle *intr_handle) {
113 struct vfio_irq_set *irq_set;
114 char irq_set_buf[IRQ_SET_BUF_LEN];
118 len = sizeof(irq_set_buf);
121 irq_set = (struct vfio_irq_set *) irq_set_buf;
122 irq_set->argsz = len;
124 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
125 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
127 fd_ptr = (int *) &irq_set->data;
128 *fd_ptr = intr_handle->fd;
130 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
133 RTE_LOG(ERR, EAL, "Error enabling INTx interrupts for fd %d\n",
138 /* unmask INTx after enabling */
139 memset(irq_set, 0, len);
140 len = sizeof(struct vfio_irq_set);
141 irq_set->argsz = len;
143 irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
144 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
147 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
150 RTE_LOG(ERR, EAL, "Error unmasking INTx interrupts for fd %d\n",
157 /* disable legacy (INTx) interrupts */
159 vfio_disable_intx(const struct rte_intr_handle *intr_handle) {
160 struct vfio_irq_set *irq_set;
161 char irq_set_buf[IRQ_SET_BUF_LEN];
164 len = sizeof(struct vfio_irq_set);
166 /* mask interrupts before disabling */
167 irq_set = (struct vfio_irq_set *) irq_set_buf;
168 irq_set->argsz = len;
170 irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK;
171 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
174 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
177 RTE_LOG(ERR, EAL, "Error masking INTx interrupts for fd %d\n",
183 memset(irq_set, 0, len);
184 irq_set->argsz = len;
186 irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
187 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
190 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
194 "Error disabling INTx interrupts for fd %d\n", intr_handle->fd);
200 /* unmask/ack legacy (INTx) interrupts */
202 vfio_ack_intx(const struct rte_intr_handle *intr_handle)
204 struct vfio_irq_set irq_set;
207 memset(&irq_set, 0, sizeof(irq_set));
208 irq_set.argsz = sizeof(irq_set);
210 irq_set.flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
211 irq_set.index = VFIO_PCI_INTX_IRQ_INDEX;
214 if (ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, &irq_set)) {
215 RTE_LOG(ERR, EAL, "Error unmasking INTx interrupts for fd %d\n",
222 /* enable MSI interrupts */
224 vfio_enable_msi(const struct rte_intr_handle *intr_handle) {
226 char irq_set_buf[IRQ_SET_BUF_LEN];
227 struct vfio_irq_set *irq_set;
230 len = sizeof(irq_set_buf);
232 irq_set = (struct vfio_irq_set *) irq_set_buf;
233 irq_set->argsz = len;
235 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
236 irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
238 fd_ptr = (int *) &irq_set->data;
239 *fd_ptr = intr_handle->fd;
241 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
244 RTE_LOG(ERR, EAL, "Error enabling MSI interrupts for fd %d\n",
251 /* disable MSI interrupts */
253 vfio_disable_msi(const struct rte_intr_handle *intr_handle) {
254 struct vfio_irq_set *irq_set;
255 char irq_set_buf[IRQ_SET_BUF_LEN];
258 len = sizeof(struct vfio_irq_set);
260 irq_set = (struct vfio_irq_set *) irq_set_buf;
261 irq_set->argsz = len;
263 irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
264 irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
267 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
271 "Error disabling MSI interrupts for fd %d\n", intr_handle->fd);
276 /* enable MSI-X interrupts */
278 vfio_enable_msix(const struct rte_intr_handle *intr_handle) {
280 char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
281 struct vfio_irq_set *irq_set;
284 len = sizeof(irq_set_buf);
286 irq_set = (struct vfio_irq_set *) irq_set_buf;
287 irq_set->argsz = len;
288 /* 0 < irq_set->count < RTE_MAX_RXTX_INTR_VEC_ID + 1 */
289 irq_set->count = intr_handle->max_intr ?
290 (intr_handle->max_intr > RTE_MAX_RXTX_INTR_VEC_ID + 1 ?
291 RTE_MAX_RXTX_INTR_VEC_ID + 1 : intr_handle->max_intr) : 1;
292 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
293 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
295 fd_ptr = (int *) &irq_set->data;
296 /* INTR vector offset 0 reserve for non-efds mapping */
297 fd_ptr[RTE_INTR_VEC_ZERO_OFFSET] = intr_handle->fd;
298 memcpy(&fd_ptr[RTE_INTR_VEC_RXTX_OFFSET], intr_handle->efds,
299 sizeof(*intr_handle->efds) * intr_handle->nb_efd);
301 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
304 RTE_LOG(ERR, EAL, "Error enabling MSI-X interrupts for fd %d\n",
312 /* disable MSI-X interrupts */
314 vfio_disable_msix(const struct rte_intr_handle *intr_handle) {
315 struct vfio_irq_set *irq_set;
316 char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
319 len = sizeof(struct vfio_irq_set);
321 irq_set = (struct vfio_irq_set *) irq_set_buf;
322 irq_set->argsz = len;
324 irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
325 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
328 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
332 "Error disabling MSI-X interrupts for fd %d\n", intr_handle->fd);
337 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
338 /* enable req notifier */
340 vfio_enable_req(const struct rte_intr_handle *intr_handle)
343 char irq_set_buf[IRQ_SET_BUF_LEN];
344 struct vfio_irq_set *irq_set;
347 len = sizeof(irq_set_buf);
349 irq_set = (struct vfio_irq_set *) irq_set_buf;
350 irq_set->argsz = len;
352 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
353 VFIO_IRQ_SET_ACTION_TRIGGER;
354 irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
356 fd_ptr = (int *) &irq_set->data;
357 *fd_ptr = intr_handle->fd;
359 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
362 RTE_LOG(ERR, EAL, "Error enabling req interrupts for fd %d\n",
370 /* disable req notifier */
372 vfio_disable_req(const struct rte_intr_handle *intr_handle)
374 struct vfio_irq_set *irq_set;
375 char irq_set_buf[IRQ_SET_BUF_LEN];
378 len = sizeof(struct vfio_irq_set);
380 irq_set = (struct vfio_irq_set *) irq_set_buf;
381 irq_set->argsz = len;
383 irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
384 irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
387 ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
390 RTE_LOG(ERR, EAL, "Error disabling req interrupts for fd %d\n",
399 uio_intx_intr_disable(const struct rte_intr_handle *intr_handle)
401 unsigned char command_high;
403 /* use UIO config file descriptor for uio_pci_generic */
404 if (pread(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
406 "Error reading interrupts status for fd %d\n",
407 intr_handle->uio_cfg_fd);
410 /* disable interrupts */
412 if (pwrite(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
414 "Error disabling interrupts for fd %d\n",
415 intr_handle->uio_cfg_fd);
423 uio_intx_intr_enable(const struct rte_intr_handle *intr_handle)
425 unsigned char command_high;
427 /* use UIO config file descriptor for uio_pci_generic */
428 if (pread(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
430 "Error reading interrupts status for fd %d\n",
431 intr_handle->uio_cfg_fd);
434 /* enable interrupts */
435 command_high &= ~0x4;
436 if (pwrite(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
438 "Error enabling interrupts for fd %d\n",
439 intr_handle->uio_cfg_fd);
447 uio_intr_disable(const struct rte_intr_handle *intr_handle)
451 if (write(intr_handle->fd, &value, sizeof(value)) < 0) {
453 "Error disabling interrupts for fd %d (%s)\n",
454 intr_handle->fd, strerror(errno));
461 uio_intr_enable(const struct rte_intr_handle *intr_handle)
465 if (write(intr_handle->fd, &value, sizeof(value)) < 0) {
467 "Error enabling interrupts for fd %d (%s)\n",
468 intr_handle->fd, strerror(errno));
475 rte_intr_callback_register(const struct rte_intr_handle *intr_handle,
476 rte_intr_callback_fn cb, void *cb_arg)
478 int ret, wake_thread;
479 struct rte_intr_source *src;
480 struct rte_intr_callback *callback;
484 /* first do parameter checking */
485 if (intr_handle == NULL || intr_handle->fd < 0 || cb == NULL) {
487 "Registering with invalid input parameter\n");
491 /* allocate a new interrupt callback entity */
492 callback = calloc(1, sizeof(*callback));
493 if (callback == NULL) {
494 RTE_LOG(ERR, EAL, "Can not allocate memory\n");
497 callback->cb_fn = cb;
498 callback->cb_arg = cb_arg;
499 callback->pending_delete = 0;
500 callback->ucb_fn = NULL;
502 rte_spinlock_lock(&intr_lock);
504 /* check if there is at least one callback registered for the fd */
505 TAILQ_FOREACH(src, &intr_sources, next) {
506 if (src->intr_handle.fd == intr_handle->fd) {
507 /* we had no interrupts for this */
508 if (TAILQ_EMPTY(&src->callbacks))
511 TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
517 /* no existing callbacks for this - add new source */
519 src = calloc(1, sizeof(*src));
521 RTE_LOG(ERR, EAL, "Can not allocate memory\n");
525 src->intr_handle = *intr_handle;
526 TAILQ_INIT(&src->callbacks);
527 TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
528 TAILQ_INSERT_TAIL(&intr_sources, src, next);
534 rte_spinlock_unlock(&intr_lock);
537 * check if need to notify the pipe fd waited by epoll_wait to
538 * rebuild the wait list.
541 if (write(intr_pipe.writefd, "1", 1) < 0)
548 rte_intr_callback_unregister_pending(const struct rte_intr_handle *intr_handle,
549 rte_intr_callback_fn cb_fn, void *cb_arg,
550 rte_intr_unregister_callback_fn ucb_fn)
553 struct rte_intr_source *src;
554 struct rte_intr_callback *cb, *next;
556 /* do parameter checking first */
557 if (intr_handle == NULL || intr_handle->fd < 0) {
559 "Unregistering with invalid input parameter\n");
563 rte_spinlock_lock(&intr_lock);
565 /* check if the insterrupt source for the fd is existent */
566 TAILQ_FOREACH(src, &intr_sources, next)
567 if (src->intr_handle.fd == intr_handle->fd)
570 /* No interrupt source registered for the fd */
574 /* only usable if the source is active */
575 } else if (src->active == 0) {
581 /* walk through the callbacks and mark all that match. */
582 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
583 next = TAILQ_NEXT(cb, next);
584 if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
585 cb->cb_arg == cb_arg)) {
586 cb->pending_delete = 1;
593 rte_spinlock_unlock(&intr_lock);
599 rte_intr_callback_unregister(const struct rte_intr_handle *intr_handle,
600 rte_intr_callback_fn cb_fn, void *cb_arg)
603 struct rte_intr_source *src;
604 struct rte_intr_callback *cb, *next;
606 /* do parameter checking first */
607 if (intr_handle == NULL || intr_handle->fd < 0) {
609 "Unregistering with invalid input parameter\n");
613 rte_spinlock_lock(&intr_lock);
615 /* check if the insterrupt source for the fd is existent */
616 TAILQ_FOREACH(src, &intr_sources, next)
617 if (src->intr_handle.fd == intr_handle->fd)
620 /* No interrupt source registered for the fd */
624 /* interrupt source has some active callbacks right now. */
625 } else if (src->active != 0) {
632 /*walk through the callbacks and remove all that match. */
633 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
635 next = TAILQ_NEXT(cb, next);
637 if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
638 cb->cb_arg == cb_arg)) {
639 TAILQ_REMOVE(&src->callbacks, cb, next);
645 /* all callbacks for that source are removed. */
646 if (TAILQ_EMPTY(&src->callbacks)) {
647 TAILQ_REMOVE(&intr_sources, src, next);
652 rte_spinlock_unlock(&intr_lock);
654 /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
655 if (ret >= 0 && write(intr_pipe.writefd, "1", 1) < 0) {
663 rte_intr_enable(const struct rte_intr_handle *intr_handle)
665 if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
668 if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
671 switch (intr_handle->type){
672 /* write to the uio fd to enable the interrupt */
673 case RTE_INTR_HANDLE_UIO:
674 if (uio_intr_enable(intr_handle))
677 case RTE_INTR_HANDLE_UIO_INTX:
678 if (uio_intx_intr_enable(intr_handle))
681 /* not used at this moment */
682 case RTE_INTR_HANDLE_ALARM:
685 case RTE_INTR_HANDLE_VFIO_MSIX:
686 if (vfio_enable_msix(intr_handle))
689 case RTE_INTR_HANDLE_VFIO_MSI:
690 if (vfio_enable_msi(intr_handle))
693 case RTE_INTR_HANDLE_VFIO_LEGACY:
694 if (vfio_enable_intx(intr_handle))
697 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
698 case RTE_INTR_HANDLE_VFIO_REQ:
699 if (vfio_enable_req(intr_handle))
704 /* not used at this moment */
705 case RTE_INTR_HANDLE_DEV_EVENT:
707 /* unknown handle type */
710 "Unknown handle type of fd %d\n",
719 * PMD generally calls this function at the end of its IRQ callback.
720 * Internally, it unmasks the interrupt if possible.
722 * For INTx, unmasking is required as the interrupt is auto-masked prior to
725 * For MSI/MSI-X, unmasking is typically not needed as the interrupt is not
726 * auto-masked. In fact, for interrupt handle types VFIO_MSIX and VFIO_MSI,
727 * this function is no-op.
730 rte_intr_ack(const struct rte_intr_handle *intr_handle)
732 if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
735 if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
738 switch (intr_handle->type) {
739 /* Both acking and enabling are same for UIO */
740 case RTE_INTR_HANDLE_UIO:
741 if (uio_intr_enable(intr_handle))
744 case RTE_INTR_HANDLE_UIO_INTX:
745 if (uio_intx_intr_enable(intr_handle))
748 /* not used at this moment */
749 case RTE_INTR_HANDLE_ALARM:
752 /* VFIO MSI* is implicitly acked unlike INTx, nothing to do */
753 case RTE_INTR_HANDLE_VFIO_MSIX:
754 case RTE_INTR_HANDLE_VFIO_MSI:
756 case RTE_INTR_HANDLE_VFIO_LEGACY:
757 if (vfio_ack_intx(intr_handle))
760 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
761 case RTE_INTR_HANDLE_VFIO_REQ:
765 /* not used at this moment */
766 case RTE_INTR_HANDLE_DEV_EVENT:
768 /* unknown handle type */
770 RTE_LOG(ERR, EAL, "Unknown handle type of fd %d\n",
779 rte_intr_disable(const struct rte_intr_handle *intr_handle)
781 if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
784 if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
787 switch (intr_handle->type){
788 /* write to the uio fd to disable the interrupt */
789 case RTE_INTR_HANDLE_UIO:
790 if (uio_intr_disable(intr_handle))
793 case RTE_INTR_HANDLE_UIO_INTX:
794 if (uio_intx_intr_disable(intr_handle))
797 /* not used at this moment */
798 case RTE_INTR_HANDLE_ALARM:
801 case RTE_INTR_HANDLE_VFIO_MSIX:
802 if (vfio_disable_msix(intr_handle))
805 case RTE_INTR_HANDLE_VFIO_MSI:
806 if (vfio_disable_msi(intr_handle))
809 case RTE_INTR_HANDLE_VFIO_LEGACY:
810 if (vfio_disable_intx(intr_handle))
813 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
814 case RTE_INTR_HANDLE_VFIO_REQ:
815 if (vfio_disable_req(intr_handle))
820 /* not used at this moment */
821 case RTE_INTR_HANDLE_DEV_EVENT:
823 /* unknown handle type */
826 "Unknown handle type of fd %d\n",
835 eal_intr_process_interrupts(struct epoll_event *events, int nfds)
838 int n, bytes_read, rv;
839 struct rte_intr_source *src;
840 struct rte_intr_callback *cb, *next;
841 union rte_intr_read_buffer buf;
842 struct rte_intr_callback active_cb;
844 for (n = 0; n < nfds; n++) {
847 * if the pipe fd is ready to read, return out to
848 * rebuild the wait list.
850 if (events[n].data.fd == intr_pipe.readfd){
851 int r = read(intr_pipe.readfd, buf.charbuf,
852 sizeof(buf.charbuf));
856 rte_spinlock_lock(&intr_lock);
857 TAILQ_FOREACH(src, &intr_sources, next)
858 if (src->intr_handle.fd ==
862 rte_spinlock_unlock(&intr_lock);
866 /* mark this interrupt source as active and release the lock. */
868 rte_spinlock_unlock(&intr_lock);
870 /* set the length to be read dor different handle type */
871 switch (src->intr_handle.type) {
872 case RTE_INTR_HANDLE_UIO:
873 case RTE_INTR_HANDLE_UIO_INTX:
874 bytes_read = sizeof(buf.uio_intr_count);
876 case RTE_INTR_HANDLE_ALARM:
877 bytes_read = sizeof(buf.timerfd_num);
880 case RTE_INTR_HANDLE_VFIO_MSIX:
881 case RTE_INTR_HANDLE_VFIO_MSI:
882 case RTE_INTR_HANDLE_VFIO_LEGACY:
883 bytes_read = sizeof(buf.vfio_intr_count);
885 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
886 case RTE_INTR_HANDLE_VFIO_REQ:
892 case RTE_INTR_HANDLE_VDEV:
893 case RTE_INTR_HANDLE_EXT:
897 case RTE_INTR_HANDLE_DEV_EVENT:
906 if (bytes_read > 0) {
908 * read out to clear the ready-to-be-read flag
911 bytes_read = read(events[n].data.fd, &buf, bytes_read);
912 if (bytes_read < 0) {
913 if (errno == EINTR || errno == EWOULDBLOCK)
916 RTE_LOG(ERR, EAL, "Error reading from file "
917 "descriptor %d: %s\n",
921 * The device is unplugged or buggy, remove
922 * it as an interrupt source and return to
923 * force the wait list to be rebuilt.
925 rte_spinlock_lock(&intr_lock);
926 TAILQ_REMOVE(&intr_sources, src, next);
927 rte_spinlock_unlock(&intr_lock);
929 for (cb = TAILQ_FIRST(&src->callbacks); cb;
931 next = TAILQ_NEXT(cb, next);
932 TAILQ_REMOVE(&src->callbacks, cb, next);
937 } else if (bytes_read == 0)
938 RTE_LOG(ERR, EAL, "Read nothing from file "
939 "descriptor %d\n", events[n].data.fd);
944 /* grab a lock, again to call callbacks and update status. */
945 rte_spinlock_lock(&intr_lock);
949 /* Finally, call all callbacks. */
950 TAILQ_FOREACH(cb, &src->callbacks, next) {
952 /* make a copy and unlock. */
954 rte_spinlock_unlock(&intr_lock);
956 /* call the actual callback */
957 active_cb.cb_fn(active_cb.cb_arg);
959 /*get the lock back. */
960 rte_spinlock_lock(&intr_lock);
963 /* we done with that interrupt source, release it. */
968 /* check if any callback are supposed to be removed */
969 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
970 next = TAILQ_NEXT(cb, next);
971 if (cb->pending_delete) {
972 TAILQ_REMOVE(&src->callbacks, cb, next);
974 cb->ucb_fn(&src->intr_handle, cb->cb_arg);
980 /* all callbacks for that source are removed. */
981 if (TAILQ_EMPTY(&src->callbacks)) {
982 TAILQ_REMOVE(&intr_sources, src, next);
986 /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
987 if (rv >= 0 && write(intr_pipe.writefd, "1", 1) < 0) {
988 rte_spinlock_unlock(&intr_lock);
992 rte_spinlock_unlock(&intr_lock);
999 * It handles all the interrupts.
1002 * epoll file descriptor.
1004 * The number of file descriptors added in epoll.
1010 eal_intr_handle_interrupts(int pfd, unsigned totalfds)
1012 struct epoll_event events[totalfds];
1016 nfds = epoll_wait(pfd, events, totalfds,
1017 EAL_INTR_EPOLL_WAIT_FOREVER);
1018 /* epoll_wait fail */
1023 "epoll_wait returns with fail\n");
1026 /* epoll_wait timeout, will never happens here */
1029 /* epoll_wait has at least one fd ready to read */
1030 if (eal_intr_process_interrupts(events, nfds) < 0)
1036 * It builds/rebuilds up the epoll file descriptor with all the
1037 * file descriptors being waited on. Then handles the interrupts.
1045 static __attribute__((noreturn)) void *
1046 eal_intr_thread_main(__rte_unused void *arg)
1048 struct epoll_event ev;
1050 /* host thread, never break out */
1052 /* build up the epoll fd with all descriptors we are to
1053 * wait on then pass it to the handle_interrupts function
1055 static struct epoll_event pipe_event = {
1056 .events = EPOLLIN | EPOLLPRI,
1058 struct rte_intr_source *src;
1059 unsigned numfds = 0;
1061 /* create epoll fd */
1062 int pfd = epoll_create(1);
1064 rte_panic("Cannot create epoll instance\n");
1066 pipe_event.data.fd = intr_pipe.readfd;
1068 * add pipe fd into wait list, this pipe is used to
1069 * rebuild the wait list.
1071 if (epoll_ctl(pfd, EPOLL_CTL_ADD, intr_pipe.readfd,
1073 rte_panic("Error adding fd to %d epoll_ctl, %s\n",
1074 intr_pipe.readfd, strerror(errno));
1078 rte_spinlock_lock(&intr_lock);
1080 TAILQ_FOREACH(src, &intr_sources, next) {
1081 if (src->callbacks.tqh_first == NULL)
1082 continue; /* skip those with no callbacks */
1083 ev.events = EPOLLIN | EPOLLPRI | EPOLLRDHUP | EPOLLHUP;
1084 ev.data.fd = src->intr_handle.fd;
1087 * add all the uio device file descriptor
1090 if (epoll_ctl(pfd, EPOLL_CTL_ADD,
1091 src->intr_handle.fd, &ev) < 0){
1092 rte_panic("Error adding fd %d epoll_ctl, %s\n",
1093 src->intr_handle.fd, strerror(errno));
1098 rte_spinlock_unlock(&intr_lock);
1099 /* serve the interrupt */
1100 eal_intr_handle_interrupts(pfd, numfds);
1103 * when we return, we need to rebuild the
1104 * list of fds to monitor.
1111 rte_eal_intr_init(void)
1115 /* init the global interrupt source head */
1116 TAILQ_INIT(&intr_sources);
1119 * create a pipe which will be waited by epoll and notified to
1120 * rebuild the wait list of epoll.
1122 if (pipe(intr_pipe.pipefd) < 0) {
1127 /* create the host thread to wait/handle the interrupt */
1128 ret = rte_ctrl_thread_create(&intr_thread, "eal-intr-thread", NULL,
1129 eal_intr_thread_main, NULL);
1133 "Failed to create thread for interrupt handling\n");
1140 eal_intr_proc_rxtx_intr(int fd, const struct rte_intr_handle *intr_handle)
1142 union rte_intr_read_buffer buf;
1146 switch (intr_handle->type) {
1147 case RTE_INTR_HANDLE_UIO:
1148 case RTE_INTR_HANDLE_UIO_INTX:
1149 bytes_read = sizeof(buf.uio_intr_count);
1152 case RTE_INTR_HANDLE_VFIO_MSIX:
1153 case RTE_INTR_HANDLE_VFIO_MSI:
1154 case RTE_INTR_HANDLE_VFIO_LEGACY:
1155 bytes_read = sizeof(buf.vfio_intr_count);
1158 case RTE_INTR_HANDLE_VDEV:
1159 bytes_read = intr_handle->efd_counter_size;
1160 /* For vdev, number of bytes to read is set by driver */
1162 case RTE_INTR_HANDLE_EXT:
1166 RTE_LOG(INFO, EAL, "unexpected intr type\n");
1171 * read out to clear the ready-to-be-read flag
1174 if (bytes_read == 0)
1177 nbytes = read(fd, &buf, bytes_read);
1179 if (errno == EINTR || errno == EWOULDBLOCK ||
1183 "Error reading from fd %d: %s\n",
1184 fd, strerror(errno));
1185 } else if (nbytes == 0)
1186 RTE_LOG(ERR, EAL, "Read nothing from fd %d\n", fd);
1192 eal_epoll_process_event(struct epoll_event *evs, unsigned int n,
1193 struct rte_epoll_event *events)
1195 unsigned int i, count = 0;
1196 struct rte_epoll_event *rev;
1198 for (i = 0; i < n; i++) {
1199 rev = evs[i].data.ptr;
1200 if (!rev || !rte_atomic32_cmpset(&rev->status, RTE_EPOLL_VALID,
1204 events[count].status = RTE_EPOLL_VALID;
1205 events[count].fd = rev->fd;
1206 events[count].epfd = rev->epfd;
1207 events[count].epdata.event = rev->epdata.event;
1208 events[count].epdata.data = rev->epdata.data;
1209 if (rev->epdata.cb_fun)
1210 rev->epdata.cb_fun(rev->fd,
1211 rev->epdata.cb_arg);
1213 rte_compiler_barrier();
1214 rev->status = RTE_EPOLL_VALID;
1221 eal_init_tls_epfd(void)
1223 int pfd = epoll_create(255);
1227 "Cannot create epoll instance\n");
1234 rte_intr_tls_epfd(void)
1236 if (RTE_PER_LCORE(_epfd) == -1)
1237 RTE_PER_LCORE(_epfd) = eal_init_tls_epfd();
1239 return RTE_PER_LCORE(_epfd);
1243 rte_epoll_wait(int epfd, struct rte_epoll_event *events,
1244 int maxevents, int timeout)
1246 struct epoll_event evs[maxevents];
1250 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1254 /* using per thread epoll fd */
1255 if (epfd == RTE_EPOLL_PER_THREAD)
1256 epfd = rte_intr_tls_epfd();
1259 rc = epoll_wait(epfd, evs, maxevents, timeout);
1260 if (likely(rc > 0)) {
1261 /* epoll_wait has at least one fd ready to read */
1262 rc = eal_epoll_process_event(evs, rc, events);
1264 } else if (rc < 0) {
1267 /* epoll_wait fail */
1268 RTE_LOG(ERR, EAL, "epoll_wait returns with fail %s\n",
1273 /* rc == 0, epoll_wait timed out */
1282 eal_epoll_data_safe_free(struct rte_epoll_event *ev)
1284 while (!rte_atomic32_cmpset(&ev->status, RTE_EPOLL_VALID,
1286 while (ev->status != RTE_EPOLL_VALID)
1288 memset(&ev->epdata, 0, sizeof(ev->epdata));
1294 rte_epoll_ctl(int epfd, int op, int fd,
1295 struct rte_epoll_event *event)
1297 struct epoll_event ev;
1300 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1304 /* using per thread epoll fd */
1305 if (epfd == RTE_EPOLL_PER_THREAD)
1306 epfd = rte_intr_tls_epfd();
1308 if (op == EPOLL_CTL_ADD) {
1309 event->status = RTE_EPOLL_VALID;
1310 event->fd = fd; /* ignore fd in event */
1312 ev.data.ptr = (void *)event;
1315 ev.events = event->epdata.event;
1316 if (epoll_ctl(epfd, op, fd, &ev) < 0) {
1317 RTE_LOG(ERR, EAL, "Error op %d fd %d epoll_ctl, %s\n",
1318 op, fd, strerror(errno));
1319 if (op == EPOLL_CTL_ADD)
1320 /* rollback status when CTL_ADD fail */
1321 event->status = RTE_EPOLL_INVALID;
1325 if (op == EPOLL_CTL_DEL && event->status != RTE_EPOLL_INVALID)
1326 eal_epoll_data_safe_free(event);
1332 rte_intr_rx_ctl(struct rte_intr_handle *intr_handle, int epfd,
1333 int op, unsigned int vec, void *data)
1335 struct rte_epoll_event *rev;
1336 struct rte_epoll_data *epdata;
1338 unsigned int efd_idx;
1341 efd_idx = (vec >= RTE_INTR_VEC_RXTX_OFFSET) ?
1342 (vec - RTE_INTR_VEC_RXTX_OFFSET) : vec;
1344 if (!intr_handle || intr_handle->nb_efd == 0 ||
1345 efd_idx >= intr_handle->nb_efd) {
1346 RTE_LOG(ERR, EAL, "Wrong intr vector number.\n");
1351 case RTE_INTR_EVENT_ADD:
1352 epfd_op = EPOLL_CTL_ADD;
1353 rev = &intr_handle->elist[efd_idx];
1354 if (rev->status != RTE_EPOLL_INVALID) {
1355 RTE_LOG(INFO, EAL, "Event already been added.\n");
1359 /* attach to intr vector fd */
1360 epdata = &rev->epdata;
1361 epdata->event = EPOLLIN | EPOLLPRI | EPOLLET;
1362 epdata->data = data;
1363 epdata->cb_fun = (rte_intr_event_cb_t)eal_intr_proc_rxtx_intr;
1364 epdata->cb_arg = (void *)intr_handle;
1365 rc = rte_epoll_ctl(epfd, epfd_op,
1366 intr_handle->efds[efd_idx], rev);
1369 "efd %d associated with vec %d added on epfd %d"
1370 "\n", rev->fd, vec, epfd);
1374 case RTE_INTR_EVENT_DEL:
1375 epfd_op = EPOLL_CTL_DEL;
1376 rev = &intr_handle->elist[efd_idx];
1377 if (rev->status == RTE_EPOLL_INVALID) {
1378 RTE_LOG(INFO, EAL, "Event does not exist.\n");
1382 rc = rte_epoll_ctl(rev->epfd, epfd_op, rev->fd, rev);
1387 RTE_LOG(ERR, EAL, "event op type mismatch\n");
1395 rte_intr_free_epoll_fd(struct rte_intr_handle *intr_handle)
1398 struct rte_epoll_event *rev;
1400 for (i = 0; i < intr_handle->nb_efd; i++) {
1401 rev = &intr_handle->elist[i];
1402 if (rev->status == RTE_EPOLL_INVALID)
1404 if (rte_epoll_ctl(rev->epfd, EPOLL_CTL_DEL, rev->fd, rev)) {
1405 /* force free if the entry valid */
1406 eal_epoll_data_safe_free(rev);
1407 rev->status = RTE_EPOLL_INVALID;
1413 rte_intr_efd_enable(struct rte_intr_handle *intr_handle, uint32_t nb_efd)
1417 uint32_t n = RTE_MIN(nb_efd, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1419 assert(nb_efd != 0);
1421 if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX) {
1422 for (i = 0; i < n; i++) {
1423 fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
1426 "can't setup eventfd, error %i (%s)\n",
1427 errno, strerror(errno));
1430 intr_handle->efds[i] = fd;
1432 intr_handle->nb_efd = n;
1433 intr_handle->max_intr = NB_OTHER_INTR + n;
1434 } else if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
1435 /* only check, initialization would be done in vdev driver.*/
1436 if (intr_handle->efd_counter_size >
1437 sizeof(union rte_intr_read_buffer)) {
1438 RTE_LOG(ERR, EAL, "the efd_counter_size is oversized");
1442 intr_handle->efds[0] = intr_handle->fd;
1443 intr_handle->nb_efd = RTE_MIN(nb_efd, 1U);
1444 intr_handle->max_intr = NB_OTHER_INTR;
1451 rte_intr_efd_disable(struct rte_intr_handle *intr_handle)
1455 rte_intr_free_epoll_fd(intr_handle);
1456 if (intr_handle->max_intr > intr_handle->nb_efd) {
1457 for (i = 0; i < intr_handle->nb_efd; i++)
1458 close(intr_handle->efds[i]);
1460 intr_handle->nb_efd = 0;
1461 intr_handle->max_intr = 0;
1465 rte_intr_dp_is_en(struct rte_intr_handle *intr_handle)
1467 return !(!intr_handle->nb_efd);
1471 rte_intr_allow_others(struct rte_intr_handle *intr_handle)
1473 if (!rte_intr_dp_is_en(intr_handle))
1476 return !!(intr_handle->max_intr - intr_handle->nb_efd);
1480 rte_intr_cap_multiple(struct rte_intr_handle *intr_handle)
1482 if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX)
1485 if (intr_handle->type == RTE_INTR_HANDLE_VDEV)