1 /*
2  * VMware VMCI Driver
3  *
4  * Copyright (C) 2012 VMware, Inc. All rights reserved.
5  *
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.
9  *
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
13  * for more details.
14  */
15 
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>
22 #include <linux/mm.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>
29 #include <linux/io.h>
30 #include <linux/vmalloc.h>
31 
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"
37 
38 #define PCI_DEVICE_ID_VMWARE_VMCI	0x0740
39 
40 #define VMCI_UTIL_NUM_RESOURCES 1
41 
42 static bool vmci_disable_msi;
43 module_param_named(disable_msi, vmci_disable_msi, bool, 0);
44 MODULE_PARM_DESC(disable_msi, "Disable MSI use in driver - (default=0)");
45 
46 static bool vmci_disable_msix;
47 module_param_named(disable_msix, vmci_disable_msix, bool, 0);
48 MODULE_PARM_DESC(disable_msix, "Disable MSI-X use in driver - (default=0)");
49 
50 static u32 ctx_update_sub_id = VMCI_INVALID_ID;
51 static u32 vm_context_id = VMCI_INVALID_ID;
52 
53 struct vmci_guest_device {
54 	struct device *dev;	/* PCI device we are attached to */
55 	void __iomem *iobase;
56 
57 	unsigned int irq;
58 	unsigned int intr_type;
59 	bool exclusive_vectors;
60 	struct msix_entry msix_entries[VMCI_MAX_INTRS];
61 
62 	struct tasklet_struct datagram_tasklet;
63 	struct tasklet_struct bm_tasklet;
64 
65 	void *data_buffer;
66 	void *notification_bitmap;
67 	dma_addr_t notification_base;
68 };
69 
70 /* vmci_dev singleton device and supporting data*/
71 struct pci_dev *vmci_pdev;
72 static struct vmci_guest_device *vmci_dev_g;
73 static DEFINE_SPINLOCK(vmci_dev_spinlock);
74 
75 static atomic_t vmci_num_guest_devices = ATOMIC_INIT(0);
76 
77 bool vmci_guest_code_active(void)
78 {
79 	return atomic_read(&vmci_num_guest_devices) != 0;
80 }
81 
82 u32 vmci_get_vm_context_id(void)
83 {
84 	if (vm_context_id == VMCI_INVALID_ID) {
85 		struct vmci_datagram get_cid_msg;
86 		get_cid_msg.dst =
87 		    vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
88 				     VMCI_GET_CONTEXT_ID);
89 		get_cid_msg.src = VMCI_ANON_SRC_HANDLE;
90 		get_cid_msg.payload_size = 0;
91 		vm_context_id = vmci_send_datagram(&get_cid_msg);
92 	}
93 	return vm_context_id;
94 }
95 
96 /*
97  * VM to hypervisor call mechanism. We use the standard VMware naming
98  * convention since shared code is calling this function as well.
99  */
100 int vmci_send_datagram(struct vmci_datagram *dg)
101 {
102 	unsigned long flags;
103 	int result;
104 
105 	/* Check args. */
106 	if (dg == NULL)
107 		return VMCI_ERROR_INVALID_ARGS;
108 
109 	/*
110 	 * Need to acquire spinlock on the device because the datagram
111 	 * data may be spread over multiple pages and the monitor may
112 	 * interleave device user rpc calls from multiple
113 	 * VCPUs. Acquiring the spinlock precludes that
114 	 * possibility. Disabling interrupts to avoid incoming
115 	 * datagrams during a "rep out" and possibly landing up in
116 	 * this function.
117 	 */
118 	spin_lock_irqsave(&vmci_dev_spinlock, flags);
119 
120 	if (vmci_dev_g) {
121 		iowrite8_rep(vmci_dev_g->iobase + VMCI_DATA_OUT_ADDR,
122 			     dg, VMCI_DG_SIZE(dg));
123 		result = ioread32(vmci_dev_g->iobase + VMCI_RESULT_LOW_ADDR);
124 	} else {
125 		result = VMCI_ERROR_UNAVAILABLE;
126 	}
127 
128 	spin_unlock_irqrestore(&vmci_dev_spinlock, flags);
129 
130 	return result;
131 }
132 EXPORT_SYMBOL_GPL(vmci_send_datagram);
133 
134 /*
135  * Gets called with the new context id if updated or resumed.
136  * Context id.
137  */
138 static void vmci_guest_cid_update(u32 sub_id,
139 				  const struct vmci_event_data *event_data,
140 				  void *client_data)
141 {
142 	const struct vmci_event_payld_ctx *ev_payload =
143 				vmci_event_data_const_payload(event_data);
144 
145 	if (sub_id != ctx_update_sub_id) {
146 		pr_devel("Invalid subscriber (ID=0x%x)\n", sub_id);
147 		return;
148 	}
149 
150 	if (!event_data || ev_payload->context_id == VMCI_INVALID_ID) {
151 		pr_devel("Invalid event data\n");
152 		return;
153 	}
154 
155 	pr_devel("Updating context from (ID=0x%x) to (ID=0x%x) on event (type=%d)\n",
156 		 vm_context_id, ev_payload->context_id, event_data->event);
157 
158 	vm_context_id = ev_payload->context_id;
159 }
160 
161 /*
162  * Verify that the host supports the hypercalls we need. If it does not,
163  * try to find fallback hypercalls and use those instead.  Returns
164  * true if required hypercalls (or fallback hypercalls) are
165  * supported by the host, false otherwise.
166  */
167 static int vmci_check_host_caps(struct pci_dev *pdev)
168 {
169 	bool result;
170 	struct vmci_resource_query_msg *msg;
171 	u32 msg_size = sizeof(struct vmci_resource_query_hdr) +
172 				VMCI_UTIL_NUM_RESOURCES * sizeof(u32);
173 	struct vmci_datagram *check_msg;
174 
175 	check_msg = kmalloc(msg_size, GFP_KERNEL);
176 	if (!check_msg) {
177 		dev_err(&pdev->dev, "%s: Insufficient memory\n", __func__);
178 		return -ENOMEM;
179 	}
180 
181 	check_msg->dst = vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
182 					  VMCI_RESOURCES_QUERY);
183 	check_msg->src = VMCI_ANON_SRC_HANDLE;
184 	check_msg->payload_size = msg_size - VMCI_DG_HEADERSIZE;
185 	msg = (struct vmci_resource_query_msg *)VMCI_DG_PAYLOAD(check_msg);
186 
187 	msg->num_resources = VMCI_UTIL_NUM_RESOURCES;
188 	msg->resources[0] = VMCI_GET_CONTEXT_ID;
189 
190 	/* Checks that hyper calls are supported */
191 	result = vmci_send_datagram(check_msg) == 0x01;
192 	kfree(check_msg);
193 
194 	dev_dbg(&pdev->dev, "%s: Host capability check: %s\n",
195 		__func__, result ? "PASSED" : "FAILED");
196 
197 	/* We need the vector. There are no fallbacks. */
198 	return result ? 0 : -ENXIO;
199 }
200 
201 /*
202  * Reads datagrams from the data in port and dispatches them. We
203  * always start reading datagrams into only the first page of the
204  * datagram buffer. If the datagrams don't fit into one page, we
205  * use the maximum datagram buffer size for the remainder of the
206  * invocation. This is a simple heuristic for not penalizing
207  * small datagrams.
208  *
209  * This function assumes that it has exclusive access to the data
210  * in port for the duration of the call.
211  */
212 static void vmci_dispatch_dgs(unsigned long data)
213 {
214 	struct vmci_guest_device *vmci_dev = (struct vmci_guest_device *)data;
215 	u8 *dg_in_buffer = vmci_dev->data_buffer;
216 	struct vmci_datagram *dg;
217 	size_t dg_in_buffer_size = VMCI_MAX_DG_SIZE;
218 	size_t current_dg_in_buffer_size = PAGE_SIZE;
219 	size_t remaining_bytes;
220 
221 	BUILD_BUG_ON(VMCI_MAX_DG_SIZE < PAGE_SIZE);
222 
223 	ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
224 		    vmci_dev->data_buffer, current_dg_in_buffer_size);
225 	dg = (struct vmci_datagram *)dg_in_buffer;
226 	remaining_bytes = current_dg_in_buffer_size;
227 
228 	while (dg->dst.resource != VMCI_INVALID_ID ||
229 	       remaining_bytes > PAGE_SIZE) {
230 		unsigned dg_in_size;
231 
232 		/*
233 		 * When the input buffer spans multiple pages, a datagram can
234 		 * start on any page boundary in the buffer.
235 		 */
236 		if (dg->dst.resource == VMCI_INVALID_ID) {
237 			dg = (struct vmci_datagram *)roundup(
238 				(uintptr_t)dg + 1, PAGE_SIZE);
239 			remaining_bytes =
240 				(size_t)(dg_in_buffer +
241 					 current_dg_in_buffer_size -
242 					 (u8 *)dg);
243 			continue;
244 		}
245 
246 		dg_in_size = VMCI_DG_SIZE_ALIGNED(dg);
247 
248 		if (dg_in_size <= dg_in_buffer_size) {
249 			int result;
250 
251 			/*
252 			 * If the remaining bytes in the datagram
253 			 * buffer doesn't contain the complete
254 			 * datagram, we first make sure we have enough
255 			 * room for it and then we read the reminder
256 			 * of the datagram and possibly any following
257 			 * datagrams.
258 			 */
259 			if (dg_in_size > remaining_bytes) {
260 				if (remaining_bytes !=
261 				    current_dg_in_buffer_size) {
262 
263 					/*
264 					 * We move the partial
265 					 * datagram to the front and
266 					 * read the reminder of the
267 					 * datagram and possibly
268 					 * following calls into the
269 					 * following bytes.
270 					 */
271 					memmove(dg_in_buffer, dg_in_buffer +
272 						current_dg_in_buffer_size -
273 						remaining_bytes,
274 						remaining_bytes);
275 					dg = (struct vmci_datagram *)
276 					    dg_in_buffer;
277 				}
278 
279 				if (current_dg_in_buffer_size !=
280 				    dg_in_buffer_size)
281 					current_dg_in_buffer_size =
282 					    dg_in_buffer_size;
283 
284 				ioread8_rep(vmci_dev->iobase +
285 						VMCI_DATA_IN_ADDR,
286 					vmci_dev->data_buffer +
287 						remaining_bytes,
288 					current_dg_in_buffer_size -
289 						remaining_bytes);
290 			}
291 
292 			/*
293 			 * We special case event datagrams from the
294 			 * hypervisor.
295 			 */
296 			if (dg->src.context == VMCI_HYPERVISOR_CONTEXT_ID &&
297 			    dg->dst.resource == VMCI_EVENT_HANDLER) {
298 				result = vmci_event_dispatch(dg);
299 			} else {
300 				result = vmci_datagram_invoke_guest_handler(dg);
301 			}
302 			if (result < VMCI_SUCCESS)
303 				dev_dbg(vmci_dev->dev,
304 					"Datagram with resource (ID=0x%x) failed (err=%d)\n",
305 					 dg->dst.resource, result);
306 
307 			/* On to the next datagram. */
308 			dg = (struct vmci_datagram *)((u8 *)dg +
309 						      dg_in_size);
310 		} else {
311 			size_t bytes_to_skip;
312 
313 			/*
314 			 * Datagram doesn't fit in datagram buffer of maximal
315 			 * size. We drop it.
316 			 */
317 			dev_dbg(vmci_dev->dev,
318 				"Failed to receive datagram (size=%u bytes)\n",
319 				 dg_in_size);
320 
321 			bytes_to_skip = dg_in_size - remaining_bytes;
322 			if (current_dg_in_buffer_size != dg_in_buffer_size)
323 				current_dg_in_buffer_size = dg_in_buffer_size;
324 
325 			for (;;) {
326 				ioread8_rep(vmci_dev->iobase +
327 						VMCI_DATA_IN_ADDR,
328 					vmci_dev->data_buffer,
329 					current_dg_in_buffer_size);
330 				if (bytes_to_skip <= current_dg_in_buffer_size)
331 					break;
332 
333 				bytes_to_skip -= current_dg_in_buffer_size;
334 			}
335 			dg = (struct vmci_datagram *)(dg_in_buffer +
336 						      bytes_to_skip);
337 		}
338 
339 		remaining_bytes =
340 		    (size_t) (dg_in_buffer + current_dg_in_buffer_size -
341 			      (u8 *)dg);
342 
343 		if (remaining_bytes < VMCI_DG_HEADERSIZE) {
344 			/* Get the next batch of datagrams. */
345 
346 			ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
347 				    vmci_dev->data_buffer,
348 				    current_dg_in_buffer_size);
349 			dg = (struct vmci_datagram *)dg_in_buffer;
350 			remaining_bytes = current_dg_in_buffer_size;
351 		}
352 	}
353 }
354 
355 /*
356  * Scans the notification bitmap for raised flags, clears them
357  * and handles the notifications.
358  */
359 static void vmci_process_bitmap(unsigned long data)
360 {
361 	struct vmci_guest_device *dev = (struct vmci_guest_device *)data;
362 
363 	if (!dev->notification_bitmap) {
364 		dev_dbg(dev->dev, "No bitmap present in %s\n", __func__);
365 		return;
366 	}
367 
368 	vmci_dbell_scan_notification_entries(dev->notification_bitmap);
369 }
370 
371 /*
372  * Enable MSI-X.  Try exclusive vectors first, then shared vectors.
373  */
374 static int vmci_enable_msix(struct pci_dev *pdev,
375 			    struct vmci_guest_device *vmci_dev)
376 {
377 	int i;
378 	int result;
379 
380 	for (i = 0; i < VMCI_MAX_INTRS; ++i) {
381 		vmci_dev->msix_entries[i].entry = i;
382 		vmci_dev->msix_entries[i].vector = i;
383 	}
384 
385 	result = pci_enable_msix_exact(pdev,
386 				       vmci_dev->msix_entries, VMCI_MAX_INTRS);
387 	if (result == 0)
388 		vmci_dev->exclusive_vectors = true;
389 	else if (result == -ENOSPC)
390 		result = pci_enable_msix_exact(pdev, vmci_dev->msix_entries, 1);
391 
392 	return result;
393 }
394 
395 /*
396  * Interrupt handler for legacy or MSI interrupt, or for first MSI-X
397  * interrupt (vector VMCI_INTR_DATAGRAM).
398  */
399 static irqreturn_t vmci_interrupt(int irq, void *_dev)
400 {
401 	struct vmci_guest_device *dev = _dev;
402 
403 	/*
404 	 * If we are using MSI-X with exclusive vectors then we simply schedule
405 	 * the datagram tasklet, since we know the interrupt was meant for us.
406 	 * Otherwise we must read the ICR to determine what to do.
407 	 */
408 
409 	if (dev->intr_type == VMCI_INTR_TYPE_MSIX && dev->exclusive_vectors) {
410 		tasklet_schedule(&dev->datagram_tasklet);
411 	} else {
412 		unsigned int icr;
413 
414 		/* Acknowledge interrupt and determine what needs doing. */
415 		icr = ioread32(dev->iobase + VMCI_ICR_ADDR);
416 		if (icr == 0 || icr == ~0)
417 			return IRQ_NONE;
418 
419 		if (icr & VMCI_ICR_DATAGRAM) {
420 			tasklet_schedule(&dev->datagram_tasklet);
421 			icr &= ~VMCI_ICR_DATAGRAM;
422 		}
423 
424 		if (icr & VMCI_ICR_NOTIFICATION) {
425 			tasklet_schedule(&dev->bm_tasklet);
426 			icr &= ~VMCI_ICR_NOTIFICATION;
427 		}
428 
429 		if (icr != 0)
430 			dev_warn(dev->dev,
431 				 "Ignoring unknown interrupt cause (%d)\n",
432 				 icr);
433 	}
434 
435 	return IRQ_HANDLED;
436 }
437 
438 /*
439  * Interrupt handler for MSI-X interrupt vector VMCI_INTR_NOTIFICATION,
440  * which is for the notification bitmap.  Will only get called if we are
441  * using MSI-X with exclusive vectors.
442  */
443 static irqreturn_t vmci_interrupt_bm(int irq, void *_dev)
444 {
445 	struct vmci_guest_device *dev = _dev;
446 
447 	/* For MSI-X we can just assume it was meant for us. */
448 	tasklet_schedule(&dev->bm_tasklet);
449 
450 	return IRQ_HANDLED;
451 }
452 
453 /*
454  * Most of the initialization at module load time is done here.
455  */
456 static int vmci_guest_probe_device(struct pci_dev *pdev,
457 				   const struct pci_device_id *id)
458 {
459 	struct vmci_guest_device *vmci_dev;
460 	void __iomem *iobase;
461 	unsigned int capabilities;
462 	unsigned long cmd;
463 	int vmci_err;
464 	int error;
465 
466 	dev_dbg(&pdev->dev, "Probing for vmci/PCI guest device\n");
467 
468 	error = pcim_enable_device(pdev);
469 	if (error) {
470 		dev_err(&pdev->dev,
471 			"Failed to enable VMCI device: %d\n", error);
472 		return error;
473 	}
474 
475 	error = pcim_iomap_regions(pdev, 1 << 0, KBUILD_MODNAME);
476 	if (error) {
477 		dev_err(&pdev->dev, "Failed to reserve/map IO regions\n");
478 		return error;
479 	}
480 
481 	iobase = pcim_iomap_table(pdev)[0];
482 
483 	dev_info(&pdev->dev, "Found VMCI PCI device at %#lx, irq %u\n",
484 		 (unsigned long)iobase, pdev->irq);
485 
486 	vmci_dev = devm_kzalloc(&pdev->dev, sizeof(*vmci_dev), GFP_KERNEL);
487 	if (!vmci_dev) {
488 		dev_err(&pdev->dev,
489 			"Can't allocate memory for VMCI device\n");
490 		return -ENOMEM;
491 	}
492 
493 	vmci_dev->dev = &pdev->dev;
494 	vmci_dev->intr_type = VMCI_INTR_TYPE_INTX;
495 	vmci_dev->exclusive_vectors = false;
496 	vmci_dev->iobase = iobase;
497 
498 	tasklet_init(&vmci_dev->datagram_tasklet,
499 		     vmci_dispatch_dgs, (unsigned long)vmci_dev);
500 	tasklet_init(&vmci_dev->bm_tasklet,
501 		     vmci_process_bitmap, (unsigned long)vmci_dev);
502 
503 	vmci_dev->data_buffer = vmalloc(VMCI_MAX_DG_SIZE);
504 	if (!vmci_dev->data_buffer) {
505 		dev_err(&pdev->dev,
506 			"Can't allocate memory for datagram buffer\n");
507 		return -ENOMEM;
508 	}
509 
510 	pci_set_master(pdev);	/* To enable queue_pair functionality. */
511 
512 	/*
513 	 * Verify that the VMCI Device supports the capabilities that
514 	 * we need. If the device is missing capabilities that we would
515 	 * like to use, check for fallback capabilities and use those
516 	 * instead (so we can run a new VM on old hosts). Fail the load if
517 	 * a required capability is missing and there is no fallback.
518 	 *
519 	 * Right now, we need datagrams. There are no fallbacks.
520 	 */
521 	capabilities = ioread32(vmci_dev->iobase + VMCI_CAPS_ADDR);
522 	if (!(capabilities & VMCI_CAPS_DATAGRAM)) {
523 		dev_err(&pdev->dev, "Device does not support datagrams\n");
524 		error = -ENXIO;
525 		goto err_free_data_buffer;
526 	}
527 
528 	/*
529 	 * If the hardware supports notifications, we will use that as
530 	 * well.
531 	 */
532 	if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
533 		vmci_dev->notification_bitmap = dma_alloc_coherent(
534 			&pdev->dev, PAGE_SIZE, &vmci_dev->notification_base,
535 			GFP_KERNEL);
536 		if (!vmci_dev->notification_bitmap) {
537 			dev_warn(&pdev->dev,
538 				 "Unable to allocate notification bitmap\n");
539 		} else {
540 			memset(vmci_dev->notification_bitmap, 0, PAGE_SIZE);
541 			capabilities |= VMCI_CAPS_NOTIFICATIONS;
542 		}
543 	}
544 
545 	dev_info(&pdev->dev, "Using capabilities 0x%x\n", capabilities);
546 
547 	/* Let the host know which capabilities we intend to use. */
548 	iowrite32(capabilities, vmci_dev->iobase + VMCI_CAPS_ADDR);
549 
550 	/* Set up global device so that we can start sending datagrams */
551 	spin_lock_irq(&vmci_dev_spinlock);
552 	vmci_dev_g = vmci_dev;
553 	vmci_pdev = pdev;
554 	spin_unlock_irq(&vmci_dev_spinlock);
555 
556 	/*
557 	 * Register notification bitmap with device if that capability is
558 	 * used.
559 	 */
560 	if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
561 		unsigned long bitmap_ppn =
562 			vmci_dev->notification_base >> PAGE_SHIFT;
563 		if (!vmci_dbell_register_notification_bitmap(bitmap_ppn)) {
564 			dev_warn(&pdev->dev,
565 				 "VMCI device unable to register notification bitmap with PPN 0x%x\n",
566 				 (u32) bitmap_ppn);
567 			error = -ENXIO;
568 			goto err_remove_vmci_dev_g;
569 		}
570 	}
571 
572 	/* Check host capabilities. */
573 	error = vmci_check_host_caps(pdev);
574 	if (error)
575 		goto err_remove_bitmap;
576 
577 	/* Enable device. */
578 
579 	/*
580 	 * We subscribe to the VMCI_EVENT_CTX_ID_UPDATE here so we can
581 	 * update the internal context id when needed.
582 	 */
583 	vmci_err = vmci_event_subscribe(VMCI_EVENT_CTX_ID_UPDATE,
584 					vmci_guest_cid_update, NULL,
585 					&ctx_update_sub_id);
586 	if (vmci_err < VMCI_SUCCESS)
587 		dev_warn(&pdev->dev,
588 			 "Failed to subscribe to event (type=%d): %d\n",
589 			 VMCI_EVENT_CTX_ID_UPDATE, vmci_err);
590 
591 	/*
592 	 * Enable interrupts.  Try MSI-X first, then MSI, and then fallback on
593 	 * legacy interrupts.
594 	 */
595 	if (!vmci_disable_msix && !vmci_enable_msix(pdev, vmci_dev)) {
596 		vmci_dev->intr_type = VMCI_INTR_TYPE_MSIX;
597 		vmci_dev->irq = vmci_dev->msix_entries[0].vector;
598 	} else if (!vmci_disable_msi && !pci_enable_msi(pdev)) {
599 		vmci_dev->intr_type = VMCI_INTR_TYPE_MSI;
600 		vmci_dev->irq = pdev->irq;
601 	} else {
602 		vmci_dev->intr_type = VMCI_INTR_TYPE_INTX;
603 		vmci_dev->irq = pdev->irq;
604 	}
605 
606 	/*
607 	 * Request IRQ for legacy or MSI interrupts, or for first
608 	 * MSI-X vector.
609 	 */
610 	error = request_irq(vmci_dev->irq, vmci_interrupt, IRQF_SHARED,
611 			    KBUILD_MODNAME, vmci_dev);
612 	if (error) {
613 		dev_err(&pdev->dev, "Irq %u in use: %d\n",
614 			vmci_dev->irq, error);
615 		goto err_disable_msi;
616 	}
617 
618 	/*
619 	 * For MSI-X with exclusive vectors we need to request an
620 	 * interrupt for each vector so that we get a separate
621 	 * interrupt handler routine.  This allows us to distinguish
622 	 * between the vectors.
623 	 */
624 	if (vmci_dev->exclusive_vectors) {
625 		error = request_irq(vmci_dev->msix_entries[1].vector,
626 				    vmci_interrupt_bm, 0, KBUILD_MODNAME,
627 				    vmci_dev);
628 		if (error) {
629 			dev_err(&pdev->dev,
630 				"Failed to allocate irq %u: %d\n",
631 				vmci_dev->msix_entries[1].vector, error);
632 			goto err_free_irq;
633 		}
634 	}
635 
636 	dev_dbg(&pdev->dev, "Registered device\n");
637 
638 	atomic_inc(&vmci_num_guest_devices);
639 
640 	/* Enable specific interrupt bits. */
641 	cmd = VMCI_IMR_DATAGRAM;
642 	if (capabilities & VMCI_CAPS_NOTIFICATIONS)
643 		cmd |= VMCI_IMR_NOTIFICATION;
644 	iowrite32(cmd, vmci_dev->iobase + VMCI_IMR_ADDR);
645 
646 	/* Enable interrupts. */
647 	iowrite32(VMCI_CONTROL_INT_ENABLE,
648 		  vmci_dev->iobase + VMCI_CONTROL_ADDR);
649 
650 	pci_set_drvdata(pdev, vmci_dev);
651 	return 0;
652 
653 err_free_irq:
654 	free_irq(vmci_dev->irq, vmci_dev);
655 	tasklet_kill(&vmci_dev->datagram_tasklet);
656 	tasklet_kill(&vmci_dev->bm_tasklet);
657 
658 err_disable_msi:
659 	if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSIX)
660 		pci_disable_msix(pdev);
661 	else if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSI)
662 		pci_disable_msi(pdev);
663 
664 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
665 	if (vmci_err < VMCI_SUCCESS)
666 		dev_warn(&pdev->dev,
667 			 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
668 			 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
669 
670 err_remove_bitmap:
671 	if (vmci_dev->notification_bitmap) {
672 		iowrite32(VMCI_CONTROL_RESET,
673 			  vmci_dev->iobase + VMCI_CONTROL_ADDR);
674 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
675 				  vmci_dev->notification_bitmap,
676 				  vmci_dev->notification_base);
677 	}
678 
679 err_remove_vmci_dev_g:
680 	spin_lock_irq(&vmci_dev_spinlock);
681 	vmci_pdev = NULL;
682 	vmci_dev_g = NULL;
683 	spin_unlock_irq(&vmci_dev_spinlock);
684 
685 err_free_data_buffer:
686 	vfree(vmci_dev->data_buffer);
687 
688 	/* The rest are managed resources and will be freed by PCI core */
689 	return error;
690 }
691 
692 static void vmci_guest_remove_device(struct pci_dev *pdev)
693 {
694 	struct vmci_guest_device *vmci_dev = pci_get_drvdata(pdev);
695 	int vmci_err;
696 
697 	dev_dbg(&pdev->dev, "Removing device\n");
698 
699 	atomic_dec(&vmci_num_guest_devices);
700 
701 	vmci_qp_guest_endpoints_exit();
702 
703 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
704 	if (vmci_err < VMCI_SUCCESS)
705 		dev_warn(&pdev->dev,
706 			 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
707 			 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
708 
709 	spin_lock_irq(&vmci_dev_spinlock);
710 	vmci_dev_g = NULL;
711 	vmci_pdev = NULL;
712 	spin_unlock_irq(&vmci_dev_spinlock);
713 
714 	dev_dbg(&pdev->dev, "Resetting vmci device\n");
715 	iowrite32(VMCI_CONTROL_RESET, vmci_dev->iobase + VMCI_CONTROL_ADDR);
716 
717 	/*
718 	 * Free IRQ and then disable MSI/MSI-X as appropriate.  For
719 	 * MSI-X, we might have multiple vectors, each with their own
720 	 * IRQ, which we must free too.
721 	 */
722 	free_irq(vmci_dev->irq, vmci_dev);
723 	if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSIX) {
724 		if (vmci_dev->exclusive_vectors)
725 			free_irq(vmci_dev->msix_entries[1].vector, vmci_dev);
726 		pci_disable_msix(pdev);
727 	} else if (vmci_dev->intr_type == VMCI_INTR_TYPE_MSI) {
728 		pci_disable_msi(pdev);
729 	}
730 
731 	tasklet_kill(&vmci_dev->datagram_tasklet);
732 	tasklet_kill(&vmci_dev->bm_tasklet);
733 
734 	if (vmci_dev->notification_bitmap) {
735 		/*
736 		 * The device reset above cleared the bitmap state of the
737 		 * device, so we can safely free it here.
738 		 */
739 
740 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
741 				  vmci_dev->notification_bitmap,
742 				  vmci_dev->notification_base);
743 	}
744 
745 	vfree(vmci_dev->data_buffer);
746 
747 	/* The rest are managed resources and will be freed by PCI core */
748 }
749 
750 static const struct pci_device_id vmci_ids[] = {
751 	{ PCI_DEVICE(PCI_VENDOR_ID_VMWARE, PCI_DEVICE_ID_VMWARE_VMCI), },
752 	{ 0 },
753 };
754 MODULE_DEVICE_TABLE(pci, vmci_ids);
755 
756 static struct pci_driver vmci_guest_driver = {
757 	.name		= KBUILD_MODNAME,
758 	.id_table	= vmci_ids,
759 	.probe		= vmci_guest_probe_device,
760 	.remove		= vmci_guest_remove_device,
761 };
762 
763 int __init vmci_guest_init(void)
764 {
765 	return pci_register_driver(&vmci_guest_driver);
766 }
767 
768 void __exit vmci_guest_exit(void)
769 {
770 	pci_unregister_driver(&vmci_guest_driver);
771 }
772