4 * Copyright (C) 2012 VMware, Inc. All rights reserved.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License as published by the
8 * Free Software Foundation version 2 and no later version.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
12 * or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 #include <linux/vmw_vmci_defs.h>
17 #include <linux/vmw_vmci_api.h>
18 #include <linux/moduleparam.h>
19 #include <linux/interrupt.h>
20 #include <linux/highmem.h>
21 #include <linux/kernel.h>
23 #include <linux/module.h>
24 #include <linux/sched.h>
25 #include <linux/slab.h>
26 #include <linux/init.h>
27 #include <linux/pci.h>
28 #include <linux/smp.h>
30 #include <linux/vmalloc.h>
32 #include "vmci_datagram.h"
33 #include "vmci_doorbell.h"
34 #include "vmci_context.h"
35 #include "vmci_driver.h"
36 #include "vmci_event.h"
38 #define PCI_VENDOR_ID_VMWARE 0x15AD
39 #define PCI_DEVICE_ID_VMWARE_VMCI 0x0740
41 #define VMCI_UTIL_NUM_RESOURCES 1
43 static bool vmci_disable_msi;
44 module_param_named(disable_msi, vmci_disable_msi, bool, 0);
45 MODULE_PARM_DESC(disable_msi, "Disable MSI use in driver - (default=0)");
47 static bool vmci_disable_msix;
48 module_param_named(disable_msix, vmci_disable_msix, bool, 0);
49 MODULE_PARM_DESC(disable_msix, "Disable MSI-X use in driver - (default=0)");
51 static u32 ctx_update_sub_id = VMCI_INVALID_ID;
52 static u32 vm_context_id = VMCI_INVALID_ID;
54 struct vmci_guest_device {
55 struct device *dev; /* PCI device we are attached to */
59 unsigned int intr_type;
60 bool exclusive_vectors;
61 struct msix_entry msix_entries[VMCI_MAX_INTRS];
63 struct tasklet_struct datagram_tasklet;
64 struct tasklet_struct bm_tasklet;
67 void *notification_bitmap;
70 /* vmci_dev singleton device and supporting data*/
71 static struct vmci_guest_device *vmci_dev_g;
72 static DEFINE_SPINLOCK(vmci_dev_spinlock);
74 static atomic_t vmci_num_guest_devices = ATOMIC_INIT(0);
76 bool vmci_guest_code_active(void)
78 return atomic_read(&vmci_num_guest_devices) != 0;
81 u32 vmci_get_vm_context_id(void)
83 if (vm_context_id == VMCI_INVALID_ID) {
84 struct vmci_datagram get_cid_msg;
86 vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
88 get_cid_msg.src = VMCI_ANON_SRC_HANDLE;
89 get_cid_msg.payload_size = 0;
90 vm_context_id = vmci_send_datagram(&get_cid_msg);
96 * VM to hypervisor call mechanism. We use the standard VMware naming
97 * convention since shared code is calling this function as well.
99 int vmci_send_datagram(struct vmci_datagram *dg)
106 return VMCI_ERROR_INVALID_ARGS;
109 * Need to acquire spinlock on the device because the datagram
110 * data may be spread over multiple pages and the monitor may
111 * interleave device user rpc calls from multiple
112 * VCPUs. Acquiring the spinlock precludes that
113 * possibility. Disabling interrupts to avoid incoming
114 * datagrams during a "rep out" and possibly landing up in
117 spin_lock_irqsave(&vmci_dev_spinlock, flags);
120 iowrite8_rep(vmci_dev_g->iobase + VMCI_DATA_OUT_ADDR,
121 dg, VMCI_DG_SIZE(dg));
122 result = ioread32(vmci_dev_g->iobase + VMCI_RESULT_LOW_ADDR);
124 result = VMCI_ERROR_UNAVAILABLE;
127 spin_unlock_irqrestore(&vmci_dev_spinlock, flags);
131 EXPORT_SYMBOL_GPL(vmci_send_datagram);
134 * Gets called with the new context id if updated or resumed.
137 static void vmci_guest_cid_update(u32 sub_id,
138 const struct vmci_event_data *event_data,
141 const struct vmci_event_payld_ctx *ev_payload =
142 vmci_event_data_const_payload(event_data);
144 if (sub_id != ctx_update_sub_id) {
145 pr_devel("Invalid subscriber (ID=0x%x)\n", sub_id);
149 if (!event_data || ev_payload->context_id == VMCI_INVALID_ID) {
150 pr_devel("Invalid event data\n");
154 pr_devel("Updating context from (ID=0x%x) to (ID=0x%x) on event (type=%d)\n",
155 vm_context_id, ev_payload->context_id, event_data->event);
157 vm_context_id = ev_payload->context_id;
161 * Verify that the host supports the hypercalls we need. If it does not,
162 * try to find fallback hypercalls and use those instead. Returns
163 * true if required hypercalls (or fallback hypercalls) are
164 * supported by the host, false otherwise.
166 static bool vmci_check_host_caps(struct pci_dev *pdev)
169 struct vmci_resource_query_msg *msg;
170 u32 msg_size = sizeof(struct vmci_resource_query_hdr) +
171 VMCI_UTIL_NUM_RESOURCES * sizeof(u32);
172 struct vmci_datagram *check_msg;
174 check_msg = kmalloc(msg_size, GFP_KERNEL);
176 dev_err(&pdev->dev, "%s: Insufficient memory\n", __func__);
180 check_msg->dst = vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
181 VMCI_RESOURCES_QUERY);
182 check_msg->src = VMCI_ANON_SRC_HANDLE;
183 check_msg->payload_size = msg_size - VMCI_DG_HEADERSIZE;
184 msg = (struct vmci_resource_query_msg *)VMCI_DG_PAYLOAD(check_msg);
186 msg->num_resources = VMCI_UTIL_NUM_RESOURCES;
187 msg->resources[0] = VMCI_GET_CONTEXT_ID;
189 /* Checks that hyper calls are supported */
190 result = vmci_send_datagram(check_msg) == 0x01;
193 dev_dbg(&pdev->dev, "%s: Host capability check: %s\n",
194 __func__, result ? "PASSED" : "FAILED");
196 /* We need the vector. There are no fallbacks. */
201 * Reads datagrams from the data in port and dispatches them. We
202 * always start reading datagrams into only the first page of the
203 * datagram buffer. If the datagrams don't fit into one page, we
204 * use the maximum datagram buffer size for the remainder of the
205 * invocation. This is a simple heuristic for not penalizing
208 * This function assumes that it has exclusive access to the data
209 * in port for the duration of the call.
211 static void vmci_dispatch_dgs(unsigned long data)
213 struct vmci_guest_device *vmci_dev = (struct vmci_guest_device *)data;
214 u8 *dg_in_buffer = vmci_dev->data_buffer;
215 struct vmci_datagram *dg;
216 size_t dg_in_buffer_size = VMCI_MAX_DG_SIZE;
217 size_t current_dg_in_buffer_size = PAGE_SIZE;
218 size_t remaining_bytes;
220 BUILD_BUG_ON(VMCI_MAX_DG_SIZE < PAGE_SIZE);
222 ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
223 vmci_dev->data_buffer, current_dg_in_buffer_size);
224 dg = (struct vmci_datagram *)dg_in_buffer;
225 remaining_bytes = current_dg_in_buffer_size;
227 while (dg->dst.resource != VMCI_INVALID_ID ||
228 remaining_bytes > PAGE_SIZE) {
232 * When the input buffer spans multiple pages, a datagram can
233 * start on any page boundary in the buffer.
235 if (dg->dst.resource == VMCI_INVALID_ID) {
236 dg = (struct vmci_datagram *)roundup(
237 (uintptr_t)dg + 1, PAGE_SIZE);
239 (size_t)(dg_in_buffer +
240 current_dg_in_buffer_size -
245 dg_in_size = VMCI_DG_SIZE_ALIGNED(dg);
247 if (dg_in_size <= dg_in_buffer_size) {
251 * If the remaining bytes in the datagram
252 * buffer doesn't contain the complete
253 * datagram, we first make sure we have enough
254 * room for it and then we read the reminder
255 * of the datagram and possibly any following
258 if (dg_in_size > remaining_bytes) {
259 if (remaining_bytes !=
260 current_dg_in_buffer_size) {
263 * We move the partial
264 * datagram to the front and
265 * read the reminder of the
266 * datagram and possibly
267 * following calls into the
270 memmove(dg_in_buffer, dg_in_buffer +
271 current_dg_in_buffer_size -
274 dg = (struct vmci_datagram *)
278 if (current_dg_in_buffer_size !=
280 current_dg_in_buffer_size =
283 ioread8_rep(vmci_dev->iobase +
285 vmci_dev->data_buffer +
287 current_dg_in_buffer_size -
292 * We special case event datagrams from the
295 if (dg->src.context == VMCI_HYPERVISOR_CONTEXT_ID &&
296 dg->dst.resource == VMCI_EVENT_HANDLER) {
297 result = vmci_event_dispatch(dg);
299 result = vmci_datagram_invoke_guest_handler(dg);
301 if (result < VMCI_SUCCESS)
302 dev_dbg(vmci_dev->dev,
303 "Datagram with resource (ID=0x%x) failed (err=%d)\n",
304 dg->dst.resource, result);
306 /* On to the next datagram. */
307 dg = (struct vmci_datagram *)((u8 *)dg +
310 size_t bytes_to_skip;
313 * Datagram doesn't fit in datagram buffer of maximal
316 dev_dbg(vmci_dev->dev,
317 "Failed to receive datagram (size=%u bytes)\n",
320 bytes_to_skip = dg_in_size - remaining_bytes;
321 if (current_dg_in_buffer_size != dg_in_buffer_size)
322 current_dg_in_buffer_size = dg_in_buffer_size;
325 ioread8_rep(vmci_dev->iobase +
327 vmci_dev->data_buffer,
328 current_dg_in_buffer_size);
329 if (bytes_to_skip <= current_dg_in_buffer_size)
332 bytes_to_skip -= current_dg_in_buffer_size;
334 dg = (struct vmci_datagram *)(dg_in_buffer +
339 (size_t) (dg_in_buffer + current_dg_in_buffer_size -
342 if (remaining_bytes < VMCI_DG_HEADERSIZE) {
343 /* Get the next batch of datagrams. */
345 ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
346 vmci_dev->data_buffer,
347 current_dg_in_buffer_size);
348 dg = (struct vmci_datagram *)dg_in_buffer;
349 remaining_bytes = current_dg_in_buffer_size;
355 * Scans the notification bitmap for raised flags, clears them
356 * and handles the notifications.
358 static void vmci_process_bitmap(unsigned long data)
360 struct vmci_guest_device *dev = (struct vmci_guest_device *)data;
362 if (!dev->notification_bitmap) {
363 dev_dbg(dev->dev, "No bitmap present in %s\n", __func__);
367 vmci_dbell_scan_notification_entries(dev->notification_bitmap);
371 * Enable MSI-X. Try exclusive vectors first, then shared vectors.
373 static int vmci_enable_msix(struct pci_dev *pdev,
374 struct vmci_guest_device *vmci_dev)
379 for (i = 0; i < VMCI_MAX_INTRS; ++i) {
380 vmci_dev->msix_entries[i].entry = i;
381 vmci_dev->msix_entries[i].vector = i;
384 result = pci_enable_msix(pdev, vmci_dev->msix_entries, VMCI_MAX_INTRS);
386 vmci_dev->exclusive_vectors = true;
388 result = pci_enable_msix(pdev, vmci_dev->msix_entries, 1);
394 * Interrupt handler for legacy or MSI interrupt, or for first MSI-X
395 * interrupt (vector VMCI_INTR_DATAGRAM).
397 static irqreturn_t vmci_interrupt(int irq, void *_dev)
399 struct vmci_guest_device *dev = _dev;
402 * If we are using MSI-X with exclusive vectors then we simply schedule
403 * the datagram tasklet, since we know the interrupt was meant for us.
404 * Otherwise we must read the ICR to determine what to do.
407 if (dev->intr_type == VMCI_INTR_TYPE_MSIX && dev->exclusive_vectors) {
408 tasklet_schedule(&dev->datagram_tasklet);
412 /* Acknowledge interrupt and determine what needs doing. */
413 icr = ioread32(dev->iobase + VMCI_ICR_ADDR);
414 if (icr == 0 || icr == ~0)
417 if (icr & VMCI_ICR_DATAGRAM) {
418 tasklet_schedule(&dev->datagram_tasklet);
419 icr &= ~VMCI_ICR_DATAGRAM;
422 if (icr & VMCI_ICR_NOTIFICATION) {
423 tasklet_schedule(&dev->bm_tasklet);
424 icr &= ~VMCI_ICR_NOTIFICATION;
429 "Ignoring unknown interrupt cause (%d)\n",
437 * Interrupt handler for MSI-X interrupt vector VMCI_INTR_NOTIFICATION,
438 * which is for the notification bitmap. Will only get called if we are
439 * using MSI-X with exclusive vectors.
441 static irqreturn_t vmci_interrupt_bm(int irq, void *_dev)
443 struct vmci_guest_device *dev = _dev;
445 /* For MSI-X we can just assume it was meant for us. */
446 tasklet_schedule(&dev->bm_tasklet);
452 * Most of the initialization at module load time is done here.
454 static int vmci_guest_probe_device(struct pci_dev *pdev,
455 const struct pci_device_id *id)
457 struct vmci_guest_device *vmci_dev;
458 void __iomem *iobase;
459 unsigned int capabilities;
464 dev_dbg(&pdev->dev, "Probing for vmci/PCI guest device\n");
466 error = pcim_enable_device(pdev);
469 "Failed to enable VMCI device: %d\n", error);
473 error = pcim_iomap_regions(pdev, 1 << 0, KBUILD_MODNAME);
475 dev_err(&pdev->dev, "Failed to reserve/map IO regions\n");
479 iobase = pcim_iomap_table(pdev)[0];
481 dev_info(&pdev->dev, "Found VMCI PCI device at %#lx, irq %u\n",
482 (unsigned long)iobase, pdev->irq);
484 vmci_dev = devm_kzalloc(&pdev->dev, sizeof(*vmci_dev), GFP_KERNEL);
487 "Can't allocate memory for VMCI device\n");
491 vmci_dev->dev = &pdev->dev;
492 vmci_dev->intr_type = VMCI_INTR_TYPE_INTX;
493 vmci_dev->exclusive_vectors = false;
494 vmci_dev->iobase = iobase;
496 tasklet_init(&vmci_dev->datagram_tasklet,
497 vmci_dispatch_dgs, (unsigned long)vmci_dev);
498 tasklet_init(&vmci_dev->bm_tasklet,
499 vmci_process_bitmap, (unsigned long)vmci_dev);
501 vmci_dev->data_buffer = vmalloc(VMCI_MAX_DG_SIZE);
502 if (!vmci_dev->data_buffer) {
504 "Can't allocate memory for datagram buffer\n");
508 pci_set_master(pdev); /* To enable queue_pair functionality. */
511 * Verify that the VMCI Device supports the capabilities that
512 * we need. If the device is missing capabilities that we would
513 * like to use, check for fallback capabilities and use those
514 * instead (so we can run a new VM on old hosts). Fail the load if
515 * a required capability is missing and there is no fallback.
517 * Right now, we need datagrams. There are no fallbacks.
519 capabilities = ioread32(vmci_dev->iobase + VMCI_CAPS_ADDR);
520 if (!(capabilities & VMCI_CAPS_DATAGRAM)) {
521 dev_err(&pdev->dev, "Device does not support datagrams\n");
523 goto err_free_data_buffer;
527 * If the hardware supports notifications, we will use that as
530 if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
531 vmci_dev->notification_bitmap = vmalloc(PAGE_SIZE);
532 if (!vmci_dev->notification_bitmap) {
534 "Unable to allocate notification bitmap\n");
536 memset(vmci_dev->notification_bitmap, 0, PAGE_SIZE);
537 capabilities |= VMCI_CAPS_NOTIFICATIONS;
541 dev_info(&pdev->dev, "Using capabilities 0x%x\n", capabilities);
543 /* Let the host know which capabilities we intend to use. */
544 iowrite32(capabilities, vmci_dev->iobase + VMCI_CAPS_ADDR);
546 /* Set up global device so that we can start sending datagrams */
547 spin_lock_irq(&vmci_dev_spinlock);
548 vmci_dev_g = vmci_dev;
549 spin_unlock_irq(&vmci_dev_spinlock);
552 * Register notification bitmap with device if that capability is
555 if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
557 vmalloc_to_page(vmci_dev->notification_bitmap);
558 unsigned long bitmap_ppn = page_to_pfn(page);
559 if (!vmci_dbell_register_notification_bitmap(bitmap_ppn)) {
561 "VMCI device unable to register notification bitmap with PPN 0x%x\n",
563 goto err_remove_vmci_dev_g;
567 /* Check host capabilities. */
568 if (!vmci_check_host_caps(pdev))
569 goto err_remove_bitmap;
574 * We subscribe to the VMCI_EVENT_CTX_ID_UPDATE here so we can
575 * update the internal context id when needed.
577 vmci_err = vmci_event_subscribe(VMCI_EVENT_CTX_ID_UPDATE,
578 vmci_guest_cid_update, NULL,
580 if (vmci_err < VMCI_SUCCESS)
582 "Failed to subscribe to event (type=%d): %d\n",
583 VMCI_EVENT_CTX_ID_UPDATE, vmci_err);
586 * Enable interrupts. Try MSI-X first, then MSI, and then fallback on
589 if (!vmci_disable_msix && !vmci_enable_msix(pdev, vmci_dev)) {
590 vmci_dev->intr_type = VMCI_INTR_TYPE_MSIX;
591 vmci_dev->irq = vmci_dev->msix_entries[0].vector;
592 } else if (!vmci_disable_msi && !pci_enable_msi(pdev)) {
593 vmci_dev->intr_type = VMCI_INTR_TYPE_MSI;
594 vmci_dev->irq = pdev->irq;
596 vmci_dev->intr_type = VMCI_INTR_TYPE_INTX;
597 vmci_dev->irq = pdev->irq;
601 * Request IRQ for legacy or MSI interrupts, or for first
604 error = request_irq(vmci_dev->irq, vmci_interrupt, IRQF_SHARED,
605 KBUILD_MODNAME, vmci_dev);
607 dev_err(&pdev->dev, "Irq %u in use: %d\n",
608 vmci_dev->irq, error);
609 goto err_disable_msi;
613 * For MSI-X with exclusive vectors we need to request an
614 * interrupt for each vector so that we get a separate
615 * interrupt handler routine. This allows us to distinguish
616 * between the vectors.
618 if (vmci_dev->exclusive_vectors) {
619 error = request_irq(vmci_dev->msix_entries[1].vector,
620 vmci_interrupt_bm, 0, KBUILD_MODNAME,
624 "Failed to allocate irq %u: %d\n",
625 vmci_dev->msix_entries[1].vector, error);
630 dev_dbg(&pdev->dev, "Registered device\n");
632 atomic_inc(&vmci_num_guest_devices);
634 /* Enable specific interrupt bits. */
635 cmd = VMCI_IMR_DATAGRAM;
636 if (capabilities & VMCI_CAPS_NOTIFICATIONS)
637 cmd |= VMCI_IMR_NOTIFICATION;
638 iowrite32(cmd, vmci_dev->iobase + VMCI_IMR_ADDR);
640 /* Enable interrupts. */
641 iowrite32(VMCI_CONTROL_INT_ENABLE,
642 vmci_dev->iobase + VMCI_CONTROL_ADDR);
644 pci_set_drvdata(pdev, vmci_dev);
648 free_irq(vmci_dev->irq, &vmci_dev);
649 tasklet_kill(&vmci_dev->datagram_tasklet);
650 tasklet_kill(&vmci_dev->bm_tasklet);
653 if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSIX)
654 pci_disable_msix(pdev);
655 else if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSI)
656 pci_disable_msi(pdev);
658 vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
659 if (vmci_err < VMCI_SUCCESS)
661 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
662 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
665 if (vmci_dev->notification_bitmap) {
666 iowrite32(VMCI_CONTROL_RESET,
667 vmci_dev->iobase + VMCI_CONTROL_ADDR);
668 vfree(vmci_dev->notification_bitmap);
671 err_remove_vmci_dev_g:
672 spin_lock_irq(&vmci_dev_spinlock);
674 spin_unlock_irq(&vmci_dev_spinlock);
676 err_free_data_buffer:
677 vfree(vmci_dev->data_buffer);
679 /* The rest are managed resources and will be freed by PCI core */
683 static void vmci_guest_remove_device(struct pci_dev *pdev)
685 struct vmci_guest_device *vmci_dev = pci_get_drvdata(pdev);
688 dev_dbg(&pdev->dev, "Removing device\n");
690 atomic_dec(&vmci_num_guest_devices);
692 vmci_qp_guest_endpoints_exit();
694 vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
695 if (vmci_err < VMCI_SUCCESS)
697 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
698 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
700 spin_lock_irq(&vmci_dev_spinlock);
702 spin_unlock_irq(&vmci_dev_spinlock);
704 dev_dbg(&pdev->dev, "Resetting vmci device\n");
705 iowrite32(VMCI_CONTROL_RESET, vmci_dev->iobase + VMCI_CONTROL_ADDR);
708 * Free IRQ and then disable MSI/MSI-X as appropriate. For
709 * MSI-X, we might have multiple vectors, each with their own
710 * IRQ, which we must free too.
712 free_irq(vmci_dev->irq, vmci_dev);
713 if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSIX) {
714 if (vmci_dev->exclusive_vectors)
715 free_irq(vmci_dev->msix_entries[1].vector, vmci_dev);
716 pci_disable_msix(pdev);
717 } else if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSI) {
718 pci_disable_msi(pdev);
721 tasklet_kill(&vmci_dev->datagram_tasklet);
722 tasklet_kill(&vmci_dev->bm_tasklet);
724 if (vmci_dev->notification_bitmap) {
726 * The device reset above cleared the bitmap state of the
727 * device, so we can safely free it here.
730 vfree(vmci_dev->notification_bitmap);
733 vfree(vmci_dev->data_buffer);
735 /* The rest are managed resources and will be freed by PCI core */
738 static DEFINE_PCI_DEVICE_TABLE(vmci_ids) = {
739 { PCI_DEVICE(PCI_VENDOR_ID_VMWARE, PCI_DEVICE_ID_VMWARE_VMCI), },
742 MODULE_DEVICE_TABLE(pci, vmci_ids);
744 static struct pci_driver vmci_guest_driver = {
745 .name = KBUILD_MODNAME,
746 .id_table = vmci_ids,
747 .probe = vmci_guest_probe_device,
748 .remove = vmci_guest_remove_device,
751 int __init vmci_guest_init(void)
753 return pci_register_driver(&vmci_guest_driver);
756 void __exit vmci_guest_exit(void)
758 pci_unregister_driver(&vmci_guest_driver);