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 	bool exclusive_vectors;
58 
59 	struct tasklet_struct datagram_tasklet;
60 	struct tasklet_struct bm_tasklet;
61 
62 	void *data_buffer;
63 	void *notification_bitmap;
64 	dma_addr_t notification_base;
65 };
66 
67 static bool use_ppn64;
68 
69 bool vmci_use_ppn64(void)
70 {
71 	return use_ppn64;
72 }
73 
74 /* vmci_dev singleton device and supporting data*/
75 struct pci_dev *vmci_pdev;
76 static struct vmci_guest_device *vmci_dev_g;
77 static DEFINE_SPINLOCK(vmci_dev_spinlock);
78 
79 static atomic_t vmci_num_guest_devices = ATOMIC_INIT(0);
80 
81 bool vmci_guest_code_active(void)
82 {
83 	return atomic_read(&vmci_num_guest_devices) != 0;
84 }
85 
86 u32 vmci_get_vm_context_id(void)
87 {
88 	if (vm_context_id == VMCI_INVALID_ID) {
89 		struct vmci_datagram get_cid_msg;
90 		get_cid_msg.dst =
91 		    vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
92 				     VMCI_GET_CONTEXT_ID);
93 		get_cid_msg.src = VMCI_ANON_SRC_HANDLE;
94 		get_cid_msg.payload_size = 0;
95 		vm_context_id = vmci_send_datagram(&get_cid_msg);
96 	}
97 	return vm_context_id;
98 }
99 
100 /*
101  * VM to hypervisor call mechanism. We use the standard VMware naming
102  * convention since shared code is calling this function as well.
103  */
104 int vmci_send_datagram(struct vmci_datagram *dg)
105 {
106 	unsigned long flags;
107 	int result;
108 
109 	/* Check args. */
110 	if (dg == NULL)
111 		return VMCI_ERROR_INVALID_ARGS;
112 
113 	/*
114 	 * Need to acquire spinlock on the device because the datagram
115 	 * data may be spread over multiple pages and the monitor may
116 	 * interleave device user rpc calls from multiple
117 	 * VCPUs. Acquiring the spinlock precludes that
118 	 * possibility. Disabling interrupts to avoid incoming
119 	 * datagrams during a "rep out" and possibly landing up in
120 	 * this function.
121 	 */
122 	spin_lock_irqsave(&vmci_dev_spinlock, flags);
123 
124 	if (vmci_dev_g) {
125 		iowrite8_rep(vmci_dev_g->iobase + VMCI_DATA_OUT_ADDR,
126 			     dg, VMCI_DG_SIZE(dg));
127 		result = ioread32(vmci_dev_g->iobase + VMCI_RESULT_LOW_ADDR);
128 	} else {
129 		result = VMCI_ERROR_UNAVAILABLE;
130 	}
131 
132 	spin_unlock_irqrestore(&vmci_dev_spinlock, flags);
133 
134 	return result;
135 }
136 EXPORT_SYMBOL_GPL(vmci_send_datagram);
137 
138 /*
139  * Gets called with the new context id if updated or resumed.
140  * Context id.
141  */
142 static void vmci_guest_cid_update(u32 sub_id,
143 				  const struct vmci_event_data *event_data,
144 				  void *client_data)
145 {
146 	const struct vmci_event_payld_ctx *ev_payload =
147 				vmci_event_data_const_payload(event_data);
148 
149 	if (sub_id != ctx_update_sub_id) {
150 		pr_devel("Invalid subscriber (ID=0x%x)\n", sub_id);
151 		return;
152 	}
153 
154 	if (!event_data || ev_payload->context_id == VMCI_INVALID_ID) {
155 		pr_devel("Invalid event data\n");
156 		return;
157 	}
158 
159 	pr_devel("Updating context from (ID=0x%x) to (ID=0x%x) on event (type=%d)\n",
160 		 vm_context_id, ev_payload->context_id, event_data->event);
161 
162 	vm_context_id = ev_payload->context_id;
163 }
164 
165 /*
166  * Verify that the host supports the hypercalls we need. If it does not,
167  * try to find fallback hypercalls and use those instead.  Returns
168  * true if required hypercalls (or fallback hypercalls) are
169  * supported by the host, false otherwise.
170  */
171 static int vmci_check_host_caps(struct pci_dev *pdev)
172 {
173 	bool result;
174 	struct vmci_resource_query_msg *msg;
175 	u32 msg_size = sizeof(struct vmci_resource_query_hdr) +
176 				VMCI_UTIL_NUM_RESOURCES * sizeof(u32);
177 	struct vmci_datagram *check_msg;
178 
179 	check_msg = kmalloc(msg_size, GFP_KERNEL);
180 	if (!check_msg) {
181 		dev_err(&pdev->dev, "%s: Insufficient memory\n", __func__);
182 		return -ENOMEM;
183 	}
184 
185 	check_msg->dst = vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
186 					  VMCI_RESOURCES_QUERY);
187 	check_msg->src = VMCI_ANON_SRC_HANDLE;
188 	check_msg->payload_size = msg_size - VMCI_DG_HEADERSIZE;
189 	msg = (struct vmci_resource_query_msg *)VMCI_DG_PAYLOAD(check_msg);
190 
191 	msg->num_resources = VMCI_UTIL_NUM_RESOURCES;
192 	msg->resources[0] = VMCI_GET_CONTEXT_ID;
193 
194 	/* Checks that hyper calls are supported */
195 	result = vmci_send_datagram(check_msg) == 0x01;
196 	kfree(check_msg);
197 
198 	dev_dbg(&pdev->dev, "%s: Host capability check: %s\n",
199 		__func__, result ? "PASSED" : "FAILED");
200 
201 	/* We need the vector. There are no fallbacks. */
202 	return result ? 0 : -ENXIO;
203 }
204 
205 /*
206  * Reads datagrams from the data in port and dispatches them. We
207  * always start reading datagrams into only the first page of the
208  * datagram buffer. If the datagrams don't fit into one page, we
209  * use the maximum datagram buffer size for the remainder of the
210  * invocation. This is a simple heuristic for not penalizing
211  * small datagrams.
212  *
213  * This function assumes that it has exclusive access to the data
214  * in port for the duration of the call.
215  */
216 static void vmci_dispatch_dgs(unsigned long data)
217 {
218 	struct vmci_guest_device *vmci_dev = (struct vmci_guest_device *)data;
219 	u8 *dg_in_buffer = vmci_dev->data_buffer;
220 	struct vmci_datagram *dg;
221 	size_t dg_in_buffer_size = VMCI_MAX_DG_SIZE;
222 	size_t current_dg_in_buffer_size = PAGE_SIZE;
223 	size_t remaining_bytes;
224 
225 	BUILD_BUG_ON(VMCI_MAX_DG_SIZE < PAGE_SIZE);
226 
227 	ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
228 		    vmci_dev->data_buffer, current_dg_in_buffer_size);
229 	dg = (struct vmci_datagram *)dg_in_buffer;
230 	remaining_bytes = current_dg_in_buffer_size;
231 
232 	while (dg->dst.resource != VMCI_INVALID_ID ||
233 	       remaining_bytes > PAGE_SIZE) {
234 		unsigned dg_in_size;
235 
236 		/*
237 		 * When the input buffer spans multiple pages, a datagram can
238 		 * start on any page boundary in the buffer.
239 		 */
240 		if (dg->dst.resource == VMCI_INVALID_ID) {
241 			dg = (struct vmci_datagram *)roundup(
242 				(uintptr_t)dg + 1, PAGE_SIZE);
243 			remaining_bytes =
244 				(size_t)(dg_in_buffer +
245 					 current_dg_in_buffer_size -
246 					 (u8 *)dg);
247 			continue;
248 		}
249 
250 		dg_in_size = VMCI_DG_SIZE_ALIGNED(dg);
251 
252 		if (dg_in_size <= dg_in_buffer_size) {
253 			int result;
254 
255 			/*
256 			 * If the remaining bytes in the datagram
257 			 * buffer doesn't contain the complete
258 			 * datagram, we first make sure we have enough
259 			 * room for it and then we read the reminder
260 			 * of the datagram and possibly any following
261 			 * datagrams.
262 			 */
263 			if (dg_in_size > remaining_bytes) {
264 				if (remaining_bytes !=
265 				    current_dg_in_buffer_size) {
266 
267 					/*
268 					 * We move the partial
269 					 * datagram to the front and
270 					 * read the reminder of the
271 					 * datagram and possibly
272 					 * following calls into the
273 					 * following bytes.
274 					 */
275 					memmove(dg_in_buffer, dg_in_buffer +
276 						current_dg_in_buffer_size -
277 						remaining_bytes,
278 						remaining_bytes);
279 					dg = (struct vmci_datagram *)
280 					    dg_in_buffer;
281 				}
282 
283 				if (current_dg_in_buffer_size !=
284 				    dg_in_buffer_size)
285 					current_dg_in_buffer_size =
286 					    dg_in_buffer_size;
287 
288 				ioread8_rep(vmci_dev->iobase +
289 						VMCI_DATA_IN_ADDR,
290 					vmci_dev->data_buffer +
291 						remaining_bytes,
292 					current_dg_in_buffer_size -
293 						remaining_bytes);
294 			}
295 
296 			/*
297 			 * We special case event datagrams from the
298 			 * hypervisor.
299 			 */
300 			if (dg->src.context == VMCI_HYPERVISOR_CONTEXT_ID &&
301 			    dg->dst.resource == VMCI_EVENT_HANDLER) {
302 				result = vmci_event_dispatch(dg);
303 			} else {
304 				result = vmci_datagram_invoke_guest_handler(dg);
305 			}
306 			if (result < VMCI_SUCCESS)
307 				dev_dbg(vmci_dev->dev,
308 					"Datagram with resource (ID=0x%x) failed (err=%d)\n",
309 					 dg->dst.resource, result);
310 
311 			/* On to the next datagram. */
312 			dg = (struct vmci_datagram *)((u8 *)dg +
313 						      dg_in_size);
314 		} else {
315 			size_t bytes_to_skip;
316 
317 			/*
318 			 * Datagram doesn't fit in datagram buffer of maximal
319 			 * size. We drop it.
320 			 */
321 			dev_dbg(vmci_dev->dev,
322 				"Failed to receive datagram (size=%u bytes)\n",
323 				 dg_in_size);
324 
325 			bytes_to_skip = dg_in_size - remaining_bytes;
326 			if (current_dg_in_buffer_size != dg_in_buffer_size)
327 				current_dg_in_buffer_size = dg_in_buffer_size;
328 
329 			for (;;) {
330 				ioread8_rep(vmci_dev->iobase +
331 						VMCI_DATA_IN_ADDR,
332 					vmci_dev->data_buffer,
333 					current_dg_in_buffer_size);
334 				if (bytes_to_skip <= current_dg_in_buffer_size)
335 					break;
336 
337 				bytes_to_skip -= current_dg_in_buffer_size;
338 			}
339 			dg = (struct vmci_datagram *)(dg_in_buffer +
340 						      bytes_to_skip);
341 		}
342 
343 		remaining_bytes =
344 		    (size_t) (dg_in_buffer + current_dg_in_buffer_size -
345 			      (u8 *)dg);
346 
347 		if (remaining_bytes < VMCI_DG_HEADERSIZE) {
348 			/* Get the next batch of datagrams. */
349 
350 			ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
351 				    vmci_dev->data_buffer,
352 				    current_dg_in_buffer_size);
353 			dg = (struct vmci_datagram *)dg_in_buffer;
354 			remaining_bytes = current_dg_in_buffer_size;
355 		}
356 	}
357 }
358 
359 /*
360  * Scans the notification bitmap for raised flags, clears them
361  * and handles the notifications.
362  */
363 static void vmci_process_bitmap(unsigned long data)
364 {
365 	struct vmci_guest_device *dev = (struct vmci_guest_device *)data;
366 
367 	if (!dev->notification_bitmap) {
368 		dev_dbg(dev->dev, "No bitmap present in %s\n", __func__);
369 		return;
370 	}
371 
372 	vmci_dbell_scan_notification_entries(dev->notification_bitmap);
373 }
374 
375 /*
376  * Interrupt handler for legacy or MSI interrupt, or for first MSI-X
377  * interrupt (vector VMCI_INTR_DATAGRAM).
378  */
379 static irqreturn_t vmci_interrupt(int irq, void *_dev)
380 {
381 	struct vmci_guest_device *dev = _dev;
382 
383 	/*
384 	 * If we are using MSI-X with exclusive vectors then we simply schedule
385 	 * the datagram tasklet, since we know the interrupt was meant for us.
386 	 * Otherwise we must read the ICR to determine what to do.
387 	 */
388 
389 	if (dev->exclusive_vectors) {
390 		tasklet_schedule(&dev->datagram_tasklet);
391 	} else {
392 		unsigned int icr;
393 
394 		/* Acknowledge interrupt and determine what needs doing. */
395 		icr = ioread32(dev->iobase + VMCI_ICR_ADDR);
396 		if (icr == 0 || icr == ~0)
397 			return IRQ_NONE;
398 
399 		if (icr & VMCI_ICR_DATAGRAM) {
400 			tasklet_schedule(&dev->datagram_tasklet);
401 			icr &= ~VMCI_ICR_DATAGRAM;
402 		}
403 
404 		if (icr & VMCI_ICR_NOTIFICATION) {
405 			tasklet_schedule(&dev->bm_tasklet);
406 			icr &= ~VMCI_ICR_NOTIFICATION;
407 		}
408 
409 		if (icr != 0)
410 			dev_warn(dev->dev,
411 				 "Ignoring unknown interrupt cause (%d)\n",
412 				 icr);
413 	}
414 
415 	return IRQ_HANDLED;
416 }
417 
418 /*
419  * Interrupt handler for MSI-X interrupt vector VMCI_INTR_NOTIFICATION,
420  * which is for the notification bitmap.  Will only get called if we are
421  * using MSI-X with exclusive vectors.
422  */
423 static irqreturn_t vmci_interrupt_bm(int irq, void *_dev)
424 {
425 	struct vmci_guest_device *dev = _dev;
426 
427 	/* For MSI-X we can just assume it was meant for us. */
428 	tasklet_schedule(&dev->bm_tasklet);
429 
430 	return IRQ_HANDLED;
431 }
432 
433 /*
434  * Most of the initialization at module load time is done here.
435  */
436 static int vmci_guest_probe_device(struct pci_dev *pdev,
437 				   const struct pci_device_id *id)
438 {
439 	struct vmci_guest_device *vmci_dev;
440 	void __iomem *iobase;
441 	unsigned int capabilities;
442 	unsigned int caps_in_use;
443 	unsigned long cmd;
444 	int vmci_err;
445 	int error;
446 
447 	dev_dbg(&pdev->dev, "Probing for vmci/PCI guest device\n");
448 
449 	error = pcim_enable_device(pdev);
450 	if (error) {
451 		dev_err(&pdev->dev,
452 			"Failed to enable VMCI device: %d\n", error);
453 		return error;
454 	}
455 
456 	error = pcim_iomap_regions(pdev, 1 << 0, KBUILD_MODNAME);
457 	if (error) {
458 		dev_err(&pdev->dev, "Failed to reserve/map IO regions\n");
459 		return error;
460 	}
461 
462 	iobase = pcim_iomap_table(pdev)[0];
463 
464 	dev_info(&pdev->dev, "Found VMCI PCI device at %#lx, irq %u\n",
465 		 (unsigned long)iobase, pdev->irq);
466 
467 	vmci_dev = devm_kzalloc(&pdev->dev, sizeof(*vmci_dev), GFP_KERNEL);
468 	if (!vmci_dev) {
469 		dev_err(&pdev->dev,
470 			"Can't allocate memory for VMCI device\n");
471 		return -ENOMEM;
472 	}
473 
474 	vmci_dev->dev = &pdev->dev;
475 	vmci_dev->exclusive_vectors = false;
476 	vmci_dev->iobase = iobase;
477 
478 	tasklet_init(&vmci_dev->datagram_tasklet,
479 		     vmci_dispatch_dgs, (unsigned long)vmci_dev);
480 	tasklet_init(&vmci_dev->bm_tasklet,
481 		     vmci_process_bitmap, (unsigned long)vmci_dev);
482 
483 	vmci_dev->data_buffer = vmalloc(VMCI_MAX_DG_SIZE);
484 	if (!vmci_dev->data_buffer) {
485 		dev_err(&pdev->dev,
486 			"Can't allocate memory for datagram buffer\n");
487 		return -ENOMEM;
488 	}
489 
490 	pci_set_master(pdev);	/* To enable queue_pair functionality. */
491 
492 	/*
493 	 * Verify that the VMCI Device supports the capabilities that
494 	 * we need. If the device is missing capabilities that we would
495 	 * like to use, check for fallback capabilities and use those
496 	 * instead (so we can run a new VM on old hosts). Fail the load if
497 	 * a required capability is missing and there is no fallback.
498 	 *
499 	 * Right now, we need datagrams. There are no fallbacks.
500 	 */
501 	capabilities = ioread32(vmci_dev->iobase + VMCI_CAPS_ADDR);
502 	if (!(capabilities & VMCI_CAPS_DATAGRAM)) {
503 		dev_err(&pdev->dev, "Device does not support datagrams\n");
504 		error = -ENXIO;
505 		goto err_free_data_buffer;
506 	}
507 	caps_in_use = VMCI_CAPS_DATAGRAM;
508 
509 	/*
510 	 * Use 64-bit PPNs if the device supports.
511 	 *
512 	 * There is no check for the return value of dma_set_mask_and_coherent
513 	 * since this driver can handle the default mask values if
514 	 * dma_set_mask_and_coherent fails.
515 	 */
516 	if (capabilities & VMCI_CAPS_PPN64) {
517 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
518 		use_ppn64 = true;
519 		caps_in_use |= VMCI_CAPS_PPN64;
520 	} else {
521 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(44));
522 		use_ppn64 = false;
523 	}
524 
525 	/*
526 	 * If the hardware supports notifications, we will use that as
527 	 * well.
528 	 */
529 	if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
530 		vmci_dev->notification_bitmap = dma_alloc_coherent(
531 			&pdev->dev, PAGE_SIZE, &vmci_dev->notification_base,
532 			GFP_KERNEL);
533 		if (!vmci_dev->notification_bitmap) {
534 			dev_warn(&pdev->dev,
535 				 "Unable to allocate notification bitmap\n");
536 		} else {
537 			memset(vmci_dev->notification_bitmap, 0, PAGE_SIZE);
538 			caps_in_use |= VMCI_CAPS_NOTIFICATIONS;
539 		}
540 	}
541 
542 	dev_info(&pdev->dev, "Using capabilities 0x%x\n", caps_in_use);
543 
544 	/* Let the host know which capabilities we intend to use. */
545 	iowrite32(caps_in_use, vmci_dev->iobase + VMCI_CAPS_ADDR);
546 
547 	/* Set up global device so that we can start sending datagrams */
548 	spin_lock_irq(&vmci_dev_spinlock);
549 	vmci_dev_g = vmci_dev;
550 	vmci_pdev = pdev;
551 	spin_unlock_irq(&vmci_dev_spinlock);
552 
553 	/*
554 	 * Register notification bitmap with device if that capability is
555 	 * used.
556 	 */
557 	if (caps_in_use & VMCI_CAPS_NOTIFICATIONS) {
558 		unsigned long bitmap_ppn =
559 			vmci_dev->notification_base >> PAGE_SHIFT;
560 		if (!vmci_dbell_register_notification_bitmap(bitmap_ppn)) {
561 			dev_warn(&pdev->dev,
562 				 "VMCI device unable to register notification bitmap with PPN 0x%lx\n",
563 				 bitmap_ppn);
564 			error = -ENXIO;
565 			goto err_remove_vmci_dev_g;
566 		}
567 	}
568 
569 	/* Check host capabilities. */
570 	error = vmci_check_host_caps(pdev);
571 	if (error)
572 		goto err_remove_bitmap;
573 
574 	/* Enable device. */
575 
576 	/*
577 	 * We subscribe to the VMCI_EVENT_CTX_ID_UPDATE here so we can
578 	 * update the internal context id when needed.
579 	 */
580 	vmci_err = vmci_event_subscribe(VMCI_EVENT_CTX_ID_UPDATE,
581 					vmci_guest_cid_update, NULL,
582 					&ctx_update_sub_id);
583 	if (vmci_err < VMCI_SUCCESS)
584 		dev_warn(&pdev->dev,
585 			 "Failed to subscribe to event (type=%d): %d\n",
586 			 VMCI_EVENT_CTX_ID_UPDATE, vmci_err);
587 
588 	/*
589 	 * Enable interrupts.  Try MSI-X first, then MSI, and then fallback on
590 	 * legacy interrupts.
591 	 */
592 	error = pci_alloc_irq_vectors(pdev, VMCI_MAX_INTRS, VMCI_MAX_INTRS,
593 			PCI_IRQ_MSIX);
594 	if (error < 0) {
595 		error = pci_alloc_irq_vectors(pdev, 1, 1,
596 				PCI_IRQ_MSIX | PCI_IRQ_MSI | PCI_IRQ_LEGACY);
597 		if (error < 0)
598 			goto err_remove_bitmap;
599 	} else {
600 		vmci_dev->exclusive_vectors = true;
601 	}
602 
603 	/*
604 	 * Request IRQ for legacy or MSI interrupts, or for first
605 	 * MSI-X vector.
606 	 */
607 	error = request_irq(pci_irq_vector(pdev, 0), vmci_interrupt,
608 			    IRQF_SHARED, KBUILD_MODNAME, vmci_dev);
609 	if (error) {
610 		dev_err(&pdev->dev, "Irq %u in use: %d\n",
611 			pci_irq_vector(pdev, 0), error);
612 		goto err_disable_msi;
613 	}
614 
615 	/*
616 	 * For MSI-X with exclusive vectors we need to request an
617 	 * interrupt for each vector so that we get a separate
618 	 * interrupt handler routine.  This allows us to distinguish
619 	 * between the vectors.
620 	 */
621 	if (vmci_dev->exclusive_vectors) {
622 		error = request_irq(pci_irq_vector(pdev, 1),
623 				    vmci_interrupt_bm, 0, KBUILD_MODNAME,
624 				    vmci_dev);
625 		if (error) {
626 			dev_err(&pdev->dev,
627 				"Failed to allocate irq %u: %d\n",
628 				pci_irq_vector(pdev, 1), error);
629 			goto err_free_irq;
630 		}
631 	}
632 
633 	dev_dbg(&pdev->dev, "Registered device\n");
634 
635 	atomic_inc(&vmci_num_guest_devices);
636 
637 	/* Enable specific interrupt bits. */
638 	cmd = VMCI_IMR_DATAGRAM;
639 	if (caps_in_use & VMCI_CAPS_NOTIFICATIONS)
640 		cmd |= VMCI_IMR_NOTIFICATION;
641 	iowrite32(cmd, vmci_dev->iobase + VMCI_IMR_ADDR);
642 
643 	/* Enable interrupts. */
644 	iowrite32(VMCI_CONTROL_INT_ENABLE,
645 		  vmci_dev->iobase + VMCI_CONTROL_ADDR);
646 
647 	pci_set_drvdata(pdev, vmci_dev);
648 	return 0;
649 
650 err_free_irq:
651 	free_irq(pci_irq_vector(pdev, 0), vmci_dev);
652 	tasklet_kill(&vmci_dev->datagram_tasklet);
653 	tasklet_kill(&vmci_dev->bm_tasklet);
654 
655 err_disable_msi:
656 	pci_free_irq_vectors(pdev);
657 
658 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
659 	if (vmci_err < VMCI_SUCCESS)
660 		dev_warn(&pdev->dev,
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);
663 
664 err_remove_bitmap:
665 	if (vmci_dev->notification_bitmap) {
666 		iowrite32(VMCI_CONTROL_RESET,
667 			  vmci_dev->iobase + VMCI_CONTROL_ADDR);
668 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
669 				  vmci_dev->notification_bitmap,
670 				  vmci_dev->notification_base);
671 	}
672 
673 err_remove_vmci_dev_g:
674 	spin_lock_irq(&vmci_dev_spinlock);
675 	vmci_pdev = NULL;
676 	vmci_dev_g = NULL;
677 	spin_unlock_irq(&vmci_dev_spinlock);
678 
679 err_free_data_buffer:
680 	vfree(vmci_dev->data_buffer);
681 
682 	/* The rest are managed resources and will be freed by PCI core */
683 	return error;
684 }
685 
686 static void vmci_guest_remove_device(struct pci_dev *pdev)
687 {
688 	struct vmci_guest_device *vmci_dev = pci_get_drvdata(pdev);
689 	int vmci_err;
690 
691 	dev_dbg(&pdev->dev, "Removing device\n");
692 
693 	atomic_dec(&vmci_num_guest_devices);
694 
695 	vmci_qp_guest_endpoints_exit();
696 
697 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
698 	if (vmci_err < VMCI_SUCCESS)
699 		dev_warn(&pdev->dev,
700 			 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
701 			 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
702 
703 	spin_lock_irq(&vmci_dev_spinlock);
704 	vmci_dev_g = NULL;
705 	vmci_pdev = NULL;
706 	spin_unlock_irq(&vmci_dev_spinlock);
707 
708 	dev_dbg(&pdev->dev, "Resetting vmci device\n");
709 	iowrite32(VMCI_CONTROL_RESET, vmci_dev->iobase + VMCI_CONTROL_ADDR);
710 
711 	/*
712 	 * Free IRQ and then disable MSI/MSI-X as appropriate.  For
713 	 * MSI-X, we might have multiple vectors, each with their own
714 	 * IRQ, which we must free too.
715 	 */
716 	if (vmci_dev->exclusive_vectors)
717 		free_irq(pci_irq_vector(pdev, 1), vmci_dev);
718 	free_irq(pci_irq_vector(pdev, 0), vmci_dev);
719 	pci_free_irq_vectors(pdev);
720 
721 	tasklet_kill(&vmci_dev->datagram_tasklet);
722 	tasklet_kill(&vmci_dev->bm_tasklet);
723 
724 	if (vmci_dev->notification_bitmap) {
725 		/*
726 		 * The device reset above cleared the bitmap state of the
727 		 * device, so we can safely free it here.
728 		 */
729 
730 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
731 				  vmci_dev->notification_bitmap,
732 				  vmci_dev->notification_base);
733 	}
734 
735 	vfree(vmci_dev->data_buffer);
736 
737 	/* The rest are managed resources and will be freed by PCI core */
738 }
739 
740 static const struct pci_device_id vmci_ids[] = {
741 	{ PCI_DEVICE(PCI_VENDOR_ID_VMWARE, PCI_DEVICE_ID_VMWARE_VMCI), },
742 	{ 0 },
743 };
744 MODULE_DEVICE_TABLE(pci, vmci_ids);
745 
746 static struct pci_driver vmci_guest_driver = {
747 	.name		= KBUILD_MODNAME,
748 	.id_table	= vmci_ids,
749 	.probe		= vmci_guest_probe_device,
750 	.remove		= vmci_guest_remove_device,
751 };
752 
753 int __init vmci_guest_init(void)
754 {
755 	return pci_register_driver(&vmci_guest_driver);
756 }
757 
758 void __exit vmci_guest_exit(void)
759 {
760 	pci_unregister_driver(&vmci_guest_driver);
761 }
762