xref: /linux/drivers/misc/vmw_vmci/vmci_guest.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * VMware VMCI Driver
4  *
5  * Copyright (C) 2012 VMware, Inc. All rights reserved.
6  */
7 
8 #include <linux/vmw_vmci_defs.h>
9 #include <linux/vmw_vmci_api.h>
10 #include <linux/moduleparam.h>
11 #include <linux/interrupt.h>
12 #include <linux/highmem.h>
13 #include <linux/kernel.h>
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/processor.h>
17 #include <linux/sched.h>
18 #include <linux/slab.h>
19 #include <linux/init.h>
20 #include <linux/pci.h>
21 #include <linux/smp.h>
22 #include <linux/io.h>
23 #include <linux/vmalloc.h>
24 
25 #include "vmci_datagram.h"
26 #include "vmci_doorbell.h"
27 #include "vmci_context.h"
28 #include "vmci_driver.h"
29 #include "vmci_event.h"
30 
31 #define PCI_DEVICE_ID_VMWARE_VMCI	0x0740
32 
33 #define VMCI_UTIL_NUM_RESOURCES 1
34 
35 /*
36  * Datagram buffers for DMA send/receive must accommodate at least
37  * a maximum sized datagram and the header.
38  */
39 #define VMCI_DMA_DG_BUFFER_SIZE (VMCI_MAX_DG_SIZE + PAGE_SIZE)
40 
41 static bool vmci_disable_msi;
42 module_param_named(disable_msi, vmci_disable_msi, bool, 0);
43 MODULE_PARM_DESC(disable_msi, "Disable MSI use in driver - (default=0)");
44 
45 static bool vmci_disable_msix;
46 module_param_named(disable_msix, vmci_disable_msix, bool, 0);
47 MODULE_PARM_DESC(disable_msix, "Disable MSI-X use in driver - (default=0)");
48 
49 static u32 ctx_update_sub_id = VMCI_INVALID_ID;
50 static u32 vm_context_id = VMCI_INVALID_ID;
51 
52 struct vmci_guest_device {
53 	struct device *dev;	/* PCI device we are attached to */
54 	void __iomem *iobase;
55 	void __iomem *mmio_base;
56 
57 	bool exclusive_vectors;
58 
59 	struct wait_queue_head inout_wq;
60 
61 	void *data_buffer;
62 	dma_addr_t data_buffer_base;
63 	void *tx_buffer;
64 	dma_addr_t tx_buffer_base;
65 	void *notification_bitmap;
66 	dma_addr_t notification_base;
67 };
68 
69 static bool use_ppn64;
70 
71 bool vmci_use_ppn64(void)
72 {
73 	return use_ppn64;
74 }
75 
76 /* vmci_dev singleton device and supporting data*/
77 struct pci_dev *vmci_pdev;
78 static struct vmci_guest_device *vmci_dev_g;
79 static DEFINE_SPINLOCK(vmci_dev_spinlock);
80 
81 static atomic_t vmci_num_guest_devices = ATOMIC_INIT(0);
82 
83 bool vmci_guest_code_active(void)
84 {
85 	return atomic_read(&vmci_num_guest_devices) != 0;
86 }
87 
88 u32 vmci_get_vm_context_id(void)
89 {
90 	if (vm_context_id == VMCI_INVALID_ID) {
91 		struct vmci_datagram get_cid_msg;
92 		get_cid_msg.dst =
93 		    vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
94 				     VMCI_GET_CONTEXT_ID);
95 		get_cid_msg.src = VMCI_ANON_SRC_HANDLE;
96 		get_cid_msg.payload_size = 0;
97 		vm_context_id = vmci_send_datagram(&get_cid_msg);
98 	}
99 	return vm_context_id;
100 }
101 
102 static unsigned int vmci_read_reg(struct vmci_guest_device *dev, u32 reg)
103 {
104 	if (dev->mmio_base != NULL)
105 		return readl(dev->mmio_base + reg);
106 	return ioread32(dev->iobase + reg);
107 }
108 
109 static void vmci_write_reg(struct vmci_guest_device *dev, u32 val, u32 reg)
110 {
111 	if (dev->mmio_base != NULL)
112 		writel(val, dev->mmio_base + reg);
113 	else
114 		iowrite32(val, dev->iobase + reg);
115 }
116 
117 static void vmci_read_data(struct vmci_guest_device *vmci_dev,
118 			   void *dest, size_t size)
119 {
120 	if (vmci_dev->mmio_base == NULL)
121 		ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
122 			    dest, size);
123 	else {
124 		/*
125 		 * For DMA datagrams, the data_buffer will contain the header on the
126 		 * first page, followed by the incoming datagram(s) on the following
127 		 * pages. The header uses an S/G element immediately following the
128 		 * header on the first page to point to the data area.
129 		 */
130 		struct vmci_data_in_out_header *buffer_header = vmci_dev->data_buffer;
131 		struct vmci_sg_elem *sg_array = (struct vmci_sg_elem *)(buffer_header + 1);
132 		size_t buffer_offset = dest - vmci_dev->data_buffer;
133 
134 		buffer_header->opcode = 1;
135 		buffer_header->size = 1;
136 		buffer_header->busy = 0;
137 		sg_array[0].addr = vmci_dev->data_buffer_base + buffer_offset;
138 		sg_array[0].size = size;
139 
140 		vmci_write_reg(vmci_dev, lower_32_bits(vmci_dev->data_buffer_base),
141 			       VMCI_DATA_IN_LOW_ADDR);
142 
143 		wait_event(vmci_dev->inout_wq, buffer_header->busy == 1);
144 	}
145 }
146 
147 static int vmci_write_data(struct vmci_guest_device *dev,
148 			   struct vmci_datagram *dg)
149 {
150 	int result;
151 
152 	if (dev->mmio_base != NULL) {
153 		struct vmci_data_in_out_header *buffer_header = dev->tx_buffer;
154 		u8 *dg_out_buffer = (u8 *)(buffer_header + 1);
155 
156 		if (VMCI_DG_SIZE(dg) > VMCI_MAX_DG_SIZE)
157 			return VMCI_ERROR_INVALID_ARGS;
158 
159 		/*
160 		 * Initialize send buffer with outgoing datagram
161 		 * and set up header for inline data. Device will
162 		 * not access buffer asynchronously - only after
163 		 * the write to VMCI_DATA_OUT_LOW_ADDR.
164 		 */
165 		memcpy(dg_out_buffer, dg, VMCI_DG_SIZE(dg));
166 		buffer_header->opcode = 0;
167 		buffer_header->size = VMCI_DG_SIZE(dg);
168 		buffer_header->busy = 1;
169 
170 		vmci_write_reg(dev, lower_32_bits(dev->tx_buffer_base),
171 			       VMCI_DATA_OUT_LOW_ADDR);
172 
173 		/* Caller holds a spinlock, so cannot block. */
174 		spin_until_cond(buffer_header->busy == 0);
175 
176 		result = vmci_read_reg(vmci_dev_g, VMCI_RESULT_LOW_ADDR);
177 		if (result == VMCI_SUCCESS)
178 			result = (int)buffer_header->result;
179 	} else {
180 		iowrite8_rep(dev->iobase + VMCI_DATA_OUT_ADDR,
181 			     dg, VMCI_DG_SIZE(dg));
182 		result = vmci_read_reg(vmci_dev_g, VMCI_RESULT_LOW_ADDR);
183 	}
184 
185 	return result;
186 }
187 
188 /*
189  * VM to hypervisor call mechanism. We use the standard VMware naming
190  * convention since shared code is calling this function as well.
191  */
192 int vmci_send_datagram(struct vmci_datagram *dg)
193 {
194 	unsigned long flags;
195 	int result;
196 
197 	/* Check args. */
198 	if (dg == NULL)
199 		return VMCI_ERROR_INVALID_ARGS;
200 
201 	/*
202 	 * Need to acquire spinlock on the device because the datagram
203 	 * data may be spread over multiple pages and the monitor may
204 	 * interleave device user rpc calls from multiple
205 	 * VCPUs. Acquiring the spinlock precludes that
206 	 * possibility. Disabling interrupts to avoid incoming
207 	 * datagrams during a "rep out" and possibly landing up in
208 	 * this function.
209 	 */
210 	spin_lock_irqsave(&vmci_dev_spinlock, flags);
211 
212 	if (vmci_dev_g) {
213 		vmci_write_data(vmci_dev_g, dg);
214 		result = vmci_read_reg(vmci_dev_g, VMCI_RESULT_LOW_ADDR);
215 	} else {
216 		result = VMCI_ERROR_UNAVAILABLE;
217 	}
218 
219 	spin_unlock_irqrestore(&vmci_dev_spinlock, flags);
220 
221 	return result;
222 }
223 EXPORT_SYMBOL_GPL(vmci_send_datagram);
224 
225 /*
226  * Gets called with the new context id if updated or resumed.
227  * Context id.
228  */
229 static void vmci_guest_cid_update(u32 sub_id,
230 				  const struct vmci_event_data *event_data,
231 				  void *client_data)
232 {
233 	const struct vmci_event_payld_ctx *ev_payload =
234 				vmci_event_data_const_payload(event_data);
235 
236 	if (sub_id != ctx_update_sub_id) {
237 		pr_devel("Invalid subscriber (ID=0x%x)\n", sub_id);
238 		return;
239 	}
240 
241 	if (!event_data || ev_payload->context_id == VMCI_INVALID_ID) {
242 		pr_devel("Invalid event data\n");
243 		return;
244 	}
245 
246 	pr_devel("Updating context from (ID=0x%x) to (ID=0x%x) on event (type=%d)\n",
247 		 vm_context_id, ev_payload->context_id, event_data->event);
248 
249 	vm_context_id = ev_payload->context_id;
250 }
251 
252 /*
253  * Verify that the host supports the hypercalls we need. If it does not,
254  * try to find fallback hypercalls and use those instead.  Returns 0 if
255  * required hypercalls (or fallback hypercalls) are supported by the host,
256  * an error code otherwise.
257  */
258 static int vmci_check_host_caps(struct pci_dev *pdev)
259 {
260 	bool result;
261 	struct vmci_resource_query_msg *msg;
262 	u32 msg_size = sizeof(struct vmci_resource_query_hdr) +
263 				VMCI_UTIL_NUM_RESOURCES * sizeof(u32);
264 	struct vmci_datagram *check_msg;
265 
266 	check_msg = kzalloc(msg_size, GFP_KERNEL);
267 	if (!check_msg) {
268 		dev_err(&pdev->dev, "%s: Insufficient memory\n", __func__);
269 		return -ENOMEM;
270 	}
271 
272 	check_msg->dst = vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
273 					  VMCI_RESOURCES_QUERY);
274 	check_msg->src = VMCI_ANON_SRC_HANDLE;
275 	check_msg->payload_size = msg_size - VMCI_DG_HEADERSIZE;
276 	msg = (struct vmci_resource_query_msg *)VMCI_DG_PAYLOAD(check_msg);
277 
278 	msg->num_resources = VMCI_UTIL_NUM_RESOURCES;
279 	msg->resources[0] = VMCI_GET_CONTEXT_ID;
280 
281 	/* Checks that hyper calls are supported */
282 	result = vmci_send_datagram(check_msg) == 0x01;
283 	kfree(check_msg);
284 
285 	dev_dbg(&pdev->dev, "%s: Host capability check: %s\n",
286 		__func__, result ? "PASSED" : "FAILED");
287 
288 	/* We need the vector. There are no fallbacks. */
289 	return result ? 0 : -ENXIO;
290 }
291 
292 /*
293  * Reads datagrams from the device and dispatches them. For IO port
294  * based access to the device, we always start reading datagrams into
295  * only the first page of the datagram buffer. If the datagrams don't
296  * fit into one page, we use the maximum datagram buffer size for the
297  * remainder of the invocation. This is a simple heuristic for not
298  * penalizing small datagrams. For DMA-based datagrams, we always
299  * use the maximum datagram buffer size, since there is no performance
300  * penalty for doing so.
301  *
302  * This function assumes that it has exclusive access to the data
303  * in register(s) for the duration of the call.
304  */
305 static void vmci_dispatch_dgs(struct vmci_guest_device *vmci_dev)
306 {
307 	u8 *dg_in_buffer = vmci_dev->data_buffer;
308 	struct vmci_datagram *dg;
309 	size_t dg_in_buffer_size = VMCI_MAX_DG_SIZE;
310 	size_t current_dg_in_buffer_size;
311 	size_t remaining_bytes;
312 	bool is_io_port = vmci_dev->mmio_base == NULL;
313 
314 	BUILD_BUG_ON(VMCI_MAX_DG_SIZE < PAGE_SIZE);
315 
316 	if (!is_io_port) {
317 		/* For mmio, the first page is used for the header. */
318 		dg_in_buffer += PAGE_SIZE;
319 
320 		/*
321 		 * For DMA-based datagram operations, there is no performance
322 		 * penalty for reading the maximum buffer size.
323 		 */
324 		current_dg_in_buffer_size = VMCI_MAX_DG_SIZE;
325 	} else {
326 		current_dg_in_buffer_size = PAGE_SIZE;
327 	}
328 	vmci_read_data(vmci_dev, dg_in_buffer, current_dg_in_buffer_size);
329 	dg = (struct vmci_datagram *)dg_in_buffer;
330 	remaining_bytes = current_dg_in_buffer_size;
331 
332 	/*
333 	 * Read through the buffer until an invalid datagram header is
334 	 * encountered. The exit condition for datagrams read through
335 	 * VMCI_DATA_IN_ADDR is a bit more complicated, since a datagram
336 	 * can start on any page boundary in the buffer.
337 	 */
338 	while (dg->dst.resource != VMCI_INVALID_ID ||
339 	       (is_io_port && remaining_bytes > PAGE_SIZE)) {
340 		unsigned dg_in_size;
341 
342 		/*
343 		 * If using VMCI_DATA_IN_ADDR, skip to the next page
344 		 * as a datagram can start on any page boundary.
345 		 */
346 		if (dg->dst.resource == VMCI_INVALID_ID) {
347 			dg = (struct vmci_datagram *)roundup(
348 				(uintptr_t)dg + 1, PAGE_SIZE);
349 			remaining_bytes =
350 				(size_t)(dg_in_buffer +
351 					 current_dg_in_buffer_size -
352 					 (u8 *)dg);
353 			continue;
354 		}
355 
356 		dg_in_size = VMCI_DG_SIZE_ALIGNED(dg);
357 
358 		if (dg_in_size <= dg_in_buffer_size) {
359 			int result;
360 
361 			/*
362 			 * If the remaining bytes in the datagram
363 			 * buffer doesn't contain the complete
364 			 * datagram, we first make sure we have enough
365 			 * room for it and then we read the reminder
366 			 * of the datagram and possibly any following
367 			 * datagrams.
368 			 */
369 			if (dg_in_size > remaining_bytes) {
370 				if (remaining_bytes !=
371 				    current_dg_in_buffer_size) {
372 
373 					/*
374 					 * We move the partial
375 					 * datagram to the front and
376 					 * read the reminder of the
377 					 * datagram and possibly
378 					 * following calls into the
379 					 * following bytes.
380 					 */
381 					memmove(dg_in_buffer, dg_in_buffer +
382 						current_dg_in_buffer_size -
383 						remaining_bytes,
384 						remaining_bytes);
385 					dg = (struct vmci_datagram *)
386 					    dg_in_buffer;
387 				}
388 
389 				if (current_dg_in_buffer_size !=
390 				    dg_in_buffer_size)
391 					current_dg_in_buffer_size =
392 					    dg_in_buffer_size;
393 
394 				vmci_read_data(vmci_dev,
395 					       dg_in_buffer +
396 						remaining_bytes,
397 					       current_dg_in_buffer_size -
398 						remaining_bytes);
399 			}
400 
401 			/*
402 			 * We special case event datagrams from the
403 			 * hypervisor.
404 			 */
405 			if (dg->src.context == VMCI_HYPERVISOR_CONTEXT_ID &&
406 			    dg->dst.resource == VMCI_EVENT_HANDLER) {
407 				result = vmci_event_dispatch(dg);
408 			} else {
409 				result = vmci_datagram_invoke_guest_handler(dg);
410 			}
411 			if (result < VMCI_SUCCESS)
412 				dev_dbg(vmci_dev->dev,
413 					"Datagram with resource (ID=0x%x) failed (err=%d)\n",
414 					 dg->dst.resource, result);
415 
416 			/* On to the next datagram. */
417 			dg = (struct vmci_datagram *)((u8 *)dg +
418 						      dg_in_size);
419 		} else {
420 			size_t bytes_to_skip;
421 
422 			/*
423 			 * Datagram doesn't fit in datagram buffer of maximal
424 			 * size. We drop it.
425 			 */
426 			dev_dbg(vmci_dev->dev,
427 				"Failed to receive datagram (size=%u bytes)\n",
428 				 dg_in_size);
429 
430 			bytes_to_skip = dg_in_size - remaining_bytes;
431 			if (current_dg_in_buffer_size != dg_in_buffer_size)
432 				current_dg_in_buffer_size = dg_in_buffer_size;
433 
434 			for (;;) {
435 				vmci_read_data(vmci_dev, dg_in_buffer,
436 					       current_dg_in_buffer_size);
437 				if (bytes_to_skip <= current_dg_in_buffer_size)
438 					break;
439 
440 				bytes_to_skip -= current_dg_in_buffer_size;
441 			}
442 			dg = (struct vmci_datagram *)(dg_in_buffer +
443 						      bytes_to_skip);
444 		}
445 
446 		remaining_bytes =
447 		    (size_t) (dg_in_buffer + current_dg_in_buffer_size -
448 			      (u8 *)dg);
449 
450 		if (remaining_bytes < VMCI_DG_HEADERSIZE) {
451 			/* Get the next batch of datagrams. */
452 
453 			vmci_read_data(vmci_dev, dg_in_buffer,
454 				    current_dg_in_buffer_size);
455 			dg = (struct vmci_datagram *)dg_in_buffer;
456 			remaining_bytes = current_dg_in_buffer_size;
457 		}
458 	}
459 }
460 
461 /*
462  * Scans the notification bitmap for raised flags, clears them
463  * and handles the notifications.
464  */
465 static void vmci_process_bitmap(struct vmci_guest_device *dev)
466 {
467 	if (!dev->notification_bitmap) {
468 		dev_dbg(dev->dev, "No bitmap present in %s\n", __func__);
469 		return;
470 	}
471 
472 	vmci_dbell_scan_notification_entries(dev->notification_bitmap);
473 }
474 
475 /*
476  * Interrupt handler for legacy or MSI interrupt, or for first MSI-X
477  * interrupt (vector VMCI_INTR_DATAGRAM).
478  */
479 static irqreturn_t vmci_interrupt(int irq, void *_dev)
480 {
481 	struct vmci_guest_device *dev = _dev;
482 
483 	/*
484 	 * If we are using MSI-X with exclusive vectors then we simply call
485 	 * vmci_dispatch_dgs(), since we know the interrupt was meant for us.
486 	 * Otherwise we must read the ICR to determine what to do.
487 	 */
488 
489 	if (dev->exclusive_vectors) {
490 		vmci_dispatch_dgs(dev);
491 	} else {
492 		unsigned int icr;
493 
494 		/* Acknowledge interrupt and determine what needs doing. */
495 		icr = vmci_read_reg(dev, VMCI_ICR_ADDR);
496 		if (icr == 0 || icr == ~0)
497 			return IRQ_NONE;
498 
499 		if (icr & VMCI_ICR_DATAGRAM) {
500 			vmci_dispatch_dgs(dev);
501 			icr &= ~VMCI_ICR_DATAGRAM;
502 		}
503 
504 		if (icr & VMCI_ICR_NOTIFICATION) {
505 			vmci_process_bitmap(dev);
506 			icr &= ~VMCI_ICR_NOTIFICATION;
507 		}
508 
509 
510 		if (icr & VMCI_ICR_DMA_DATAGRAM) {
511 			wake_up_all(&dev->inout_wq);
512 			icr &= ~VMCI_ICR_DMA_DATAGRAM;
513 		}
514 
515 		if (icr != 0)
516 			dev_warn(dev->dev,
517 				 "Ignoring unknown interrupt cause (%d)\n",
518 				 icr);
519 	}
520 
521 	return IRQ_HANDLED;
522 }
523 
524 /*
525  * Interrupt handler for MSI-X interrupt vector VMCI_INTR_NOTIFICATION,
526  * which is for the notification bitmap.  Will only get called if we are
527  * using MSI-X with exclusive vectors.
528  */
529 static irqreturn_t vmci_interrupt_bm(int irq, void *_dev)
530 {
531 	struct vmci_guest_device *dev = _dev;
532 
533 	/* For MSI-X we can just assume it was meant for us. */
534 	vmci_process_bitmap(dev);
535 
536 	return IRQ_HANDLED;
537 }
538 
539 /*
540  * Interrupt handler for MSI-X interrupt vector VMCI_INTR_DMA_DATAGRAM,
541  * which is for the completion of a DMA datagram send or receive operation.
542  * Will only get called if we are using MSI-X with exclusive vectors.
543  */
544 static irqreturn_t vmci_interrupt_dma_datagram(int irq, void *_dev)
545 {
546 	struct vmci_guest_device *dev = _dev;
547 
548 	wake_up_all(&dev->inout_wq);
549 
550 	return IRQ_HANDLED;
551 }
552 
553 static void vmci_free_dg_buffers(struct vmci_guest_device *vmci_dev)
554 {
555 	if (vmci_dev->mmio_base != NULL) {
556 		if (vmci_dev->tx_buffer != NULL)
557 			dma_free_coherent(vmci_dev->dev,
558 					  VMCI_DMA_DG_BUFFER_SIZE,
559 					  vmci_dev->tx_buffer,
560 					  vmci_dev->tx_buffer_base);
561 		if (vmci_dev->data_buffer != NULL)
562 			dma_free_coherent(vmci_dev->dev,
563 					  VMCI_DMA_DG_BUFFER_SIZE,
564 					  vmci_dev->data_buffer,
565 					  vmci_dev->data_buffer_base);
566 	} else {
567 		vfree(vmci_dev->data_buffer);
568 	}
569 }
570 
571 /*
572  * Most of the initialization at module load time is done here.
573  */
574 static int vmci_guest_probe_device(struct pci_dev *pdev,
575 				   const struct pci_device_id *id)
576 {
577 	struct vmci_guest_device *vmci_dev;
578 	void __iomem *iobase = NULL;
579 	void __iomem *mmio_base = NULL;
580 	unsigned int num_irq_vectors;
581 	unsigned int capabilities;
582 	unsigned int caps_in_use;
583 	unsigned long cmd;
584 	int vmci_err;
585 	int error;
586 
587 	dev_dbg(&pdev->dev, "Probing for vmci/PCI guest device\n");
588 
589 	error = pcim_enable_device(pdev);
590 	if (error) {
591 		dev_err(&pdev->dev,
592 			"Failed to enable VMCI device: %d\n", error);
593 		return error;
594 	}
595 
596 	/*
597 	 * The VMCI device with mmio access to registers requests 256KB
598 	 * for BAR1. If present, driver will use new VMCI device
599 	 * functionality for register access and datagram send/recv.
600 	 */
601 
602 	if (pci_resource_len(pdev, 1) == VMCI_WITH_MMIO_ACCESS_BAR_SIZE) {
603 		dev_info(&pdev->dev, "MMIO register access is available\n");
604 		mmio_base = pci_iomap_range(pdev, 1, VMCI_MMIO_ACCESS_OFFSET,
605 					    VMCI_MMIO_ACCESS_SIZE);
606 		/* If the map fails, we fall back to IOIO access. */
607 		if (!mmio_base)
608 			dev_warn(&pdev->dev, "Failed to map MMIO register access\n");
609 	}
610 
611 	if (!mmio_base) {
612 		if (IS_ENABLED(CONFIG_ARM64)) {
613 			dev_err(&pdev->dev, "MMIO base is invalid\n");
614 			return -ENXIO;
615 		}
616 		error = pcim_iomap_regions(pdev, BIT(0), KBUILD_MODNAME);
617 		if (error) {
618 			dev_err(&pdev->dev, "Failed to reserve/map IO regions\n");
619 			return error;
620 		}
621 		iobase = pcim_iomap_table(pdev)[0];
622 	}
623 
624 	vmci_dev = devm_kzalloc(&pdev->dev, sizeof(*vmci_dev), GFP_KERNEL);
625 	if (!vmci_dev) {
626 		dev_err(&pdev->dev,
627 			"Can't allocate memory for VMCI device\n");
628 		error = -ENOMEM;
629 		goto err_unmap_mmio_base;
630 	}
631 
632 	vmci_dev->dev = &pdev->dev;
633 	vmci_dev->exclusive_vectors = false;
634 	vmci_dev->iobase = iobase;
635 	vmci_dev->mmio_base = mmio_base;
636 
637 	init_waitqueue_head(&vmci_dev->inout_wq);
638 
639 	if (mmio_base != NULL) {
640 		vmci_dev->tx_buffer = dma_alloc_coherent(&pdev->dev, VMCI_DMA_DG_BUFFER_SIZE,
641 							 &vmci_dev->tx_buffer_base,
642 							 GFP_KERNEL);
643 		if (!vmci_dev->tx_buffer) {
644 			dev_err(&pdev->dev,
645 				"Can't allocate memory for datagram tx buffer\n");
646 			error = -ENOMEM;
647 			goto err_unmap_mmio_base;
648 		}
649 
650 		vmci_dev->data_buffer = dma_alloc_coherent(&pdev->dev, VMCI_DMA_DG_BUFFER_SIZE,
651 							   &vmci_dev->data_buffer_base,
652 							   GFP_KERNEL);
653 	} else {
654 		vmci_dev->data_buffer = vmalloc(VMCI_MAX_DG_SIZE);
655 	}
656 	if (!vmci_dev->data_buffer) {
657 		dev_err(&pdev->dev,
658 			"Can't allocate memory for datagram buffer\n");
659 		error = -ENOMEM;
660 		goto err_free_data_buffers;
661 	}
662 
663 	pci_set_master(pdev);	/* To enable queue_pair functionality. */
664 
665 	/*
666 	 * Verify that the VMCI Device supports the capabilities that
667 	 * we need. If the device is missing capabilities that we would
668 	 * like to use, check for fallback capabilities and use those
669 	 * instead (so we can run a new VM on old hosts). Fail the load if
670 	 * a required capability is missing and there is no fallback.
671 	 *
672 	 * Right now, we need datagrams. There are no fallbacks.
673 	 */
674 	capabilities = vmci_read_reg(vmci_dev, VMCI_CAPS_ADDR);
675 	if (!(capabilities & VMCI_CAPS_DATAGRAM)) {
676 		dev_err(&pdev->dev, "Device does not support datagrams\n");
677 		error = -ENXIO;
678 		goto err_free_data_buffers;
679 	}
680 	caps_in_use = VMCI_CAPS_DATAGRAM;
681 
682 	/*
683 	 * Use 64-bit PPNs if the device supports.
684 	 *
685 	 * There is no check for the return value of dma_set_mask_and_coherent
686 	 * since this driver can handle the default mask values if
687 	 * dma_set_mask_and_coherent fails.
688 	 */
689 	if (capabilities & VMCI_CAPS_PPN64) {
690 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
691 		use_ppn64 = true;
692 		caps_in_use |= VMCI_CAPS_PPN64;
693 	} else {
694 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(44));
695 		use_ppn64 = false;
696 	}
697 
698 	/*
699 	 * If the hardware supports notifications, we will use that as
700 	 * well.
701 	 */
702 	if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
703 		vmci_dev->notification_bitmap = dma_alloc_coherent(
704 			&pdev->dev, PAGE_SIZE, &vmci_dev->notification_base,
705 			GFP_KERNEL);
706 		if (!vmci_dev->notification_bitmap)
707 			dev_warn(&pdev->dev,
708 				 "Unable to allocate notification bitmap\n");
709 		else
710 			caps_in_use |= VMCI_CAPS_NOTIFICATIONS;
711 	}
712 
713 	if (mmio_base != NULL) {
714 		if (capabilities & VMCI_CAPS_DMA_DATAGRAM) {
715 			caps_in_use |= VMCI_CAPS_DMA_DATAGRAM;
716 		} else {
717 			dev_err(&pdev->dev,
718 				"Missing capability: VMCI_CAPS_DMA_DATAGRAM\n");
719 			error = -ENXIO;
720 			goto err_free_notification_bitmap;
721 		}
722 	}
723 
724 	dev_info(&pdev->dev, "Using capabilities 0x%x\n", caps_in_use);
725 
726 	/* Let the host know which capabilities we intend to use. */
727 	vmci_write_reg(vmci_dev, caps_in_use, VMCI_CAPS_ADDR);
728 
729 	if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM) {
730 		/* Let the device know the size for pages passed down. */
731 		vmci_write_reg(vmci_dev, PAGE_SHIFT, VMCI_GUEST_PAGE_SHIFT);
732 
733 		/* Configure the high order parts of the data in/out buffers. */
734 		vmci_write_reg(vmci_dev, upper_32_bits(vmci_dev->data_buffer_base),
735 			       VMCI_DATA_IN_HIGH_ADDR);
736 		vmci_write_reg(vmci_dev, upper_32_bits(vmci_dev->tx_buffer_base),
737 			       VMCI_DATA_OUT_HIGH_ADDR);
738 	}
739 
740 	/* Set up global device so that we can start sending datagrams */
741 	spin_lock_irq(&vmci_dev_spinlock);
742 	vmci_dev_g = vmci_dev;
743 	vmci_pdev = pdev;
744 	spin_unlock_irq(&vmci_dev_spinlock);
745 
746 	/*
747 	 * Register notification bitmap with device if that capability is
748 	 * used.
749 	 */
750 	if (caps_in_use & VMCI_CAPS_NOTIFICATIONS) {
751 		unsigned long bitmap_ppn =
752 			vmci_dev->notification_base >> PAGE_SHIFT;
753 		if (!vmci_dbell_register_notification_bitmap(bitmap_ppn)) {
754 			dev_warn(&pdev->dev,
755 				 "VMCI device unable to register notification bitmap with PPN 0x%lx\n",
756 				 bitmap_ppn);
757 			error = -ENXIO;
758 			goto err_remove_vmci_dev_g;
759 		}
760 	}
761 
762 	/* Check host capabilities. */
763 	error = vmci_check_host_caps(pdev);
764 	if (error)
765 		goto err_remove_vmci_dev_g;
766 
767 	/* Enable device. */
768 
769 	/*
770 	 * We subscribe to the VMCI_EVENT_CTX_ID_UPDATE here so we can
771 	 * update the internal context id when needed.
772 	 */
773 	vmci_err = vmci_event_subscribe(VMCI_EVENT_CTX_ID_UPDATE,
774 					vmci_guest_cid_update, NULL,
775 					&ctx_update_sub_id);
776 	if (vmci_err < VMCI_SUCCESS)
777 		dev_warn(&pdev->dev,
778 			 "Failed to subscribe to event (type=%d): %d\n",
779 			 VMCI_EVENT_CTX_ID_UPDATE, vmci_err);
780 
781 	/*
782 	 * Enable interrupts.  Try MSI-X first, then MSI, and then fallback on
783 	 * legacy interrupts.
784 	 */
785 	if (vmci_dev->mmio_base != NULL)
786 		num_irq_vectors = VMCI_MAX_INTRS;
787 	else
788 		num_irq_vectors = VMCI_MAX_INTRS_NOTIFICATION;
789 	error = pci_alloc_irq_vectors(pdev, num_irq_vectors, num_irq_vectors,
790 				      PCI_IRQ_MSIX);
791 	if (error < 0) {
792 		error = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
793 		if (error < 0)
794 			goto err_unsubscribe_event;
795 	} else {
796 		vmci_dev->exclusive_vectors = true;
797 	}
798 
799 	/*
800 	 * Request IRQ for legacy or MSI interrupts, or for first
801 	 * MSI-X vector.
802 	 */
803 	error = request_threaded_irq(pci_irq_vector(pdev, 0), NULL,
804 				     vmci_interrupt, IRQF_SHARED,
805 				     KBUILD_MODNAME, vmci_dev);
806 	if (error) {
807 		dev_err(&pdev->dev, "Irq %u in use: %d\n",
808 			pci_irq_vector(pdev, 0), error);
809 		goto err_disable_msi;
810 	}
811 
812 	/*
813 	 * For MSI-X with exclusive vectors we need to request an
814 	 * interrupt for each vector so that we get a separate
815 	 * interrupt handler routine.  This allows us to distinguish
816 	 * between the vectors.
817 	 */
818 	if (vmci_dev->exclusive_vectors) {
819 		error = request_threaded_irq(pci_irq_vector(pdev, 1), NULL,
820 					     vmci_interrupt_bm, 0,
821 					     KBUILD_MODNAME, vmci_dev);
822 		if (error) {
823 			dev_err(&pdev->dev,
824 				"Failed to allocate irq %u: %d\n",
825 				pci_irq_vector(pdev, 1), error);
826 			goto err_free_irq;
827 		}
828 		if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM) {
829 			error = request_threaded_irq(pci_irq_vector(pdev, 2),
830 						     NULL,
831 						    vmci_interrupt_dma_datagram,
832 						     0, KBUILD_MODNAME,
833 						     vmci_dev);
834 			if (error) {
835 				dev_err(&pdev->dev,
836 					"Failed to allocate irq %u: %d\n",
837 					pci_irq_vector(pdev, 2), error);
838 				goto err_free_bm_irq;
839 			}
840 		}
841 	}
842 
843 	dev_dbg(&pdev->dev, "Registered device\n");
844 
845 	atomic_inc(&vmci_num_guest_devices);
846 
847 	/* Enable specific interrupt bits. */
848 	cmd = VMCI_IMR_DATAGRAM;
849 	if (caps_in_use & VMCI_CAPS_NOTIFICATIONS)
850 		cmd |= VMCI_IMR_NOTIFICATION;
851 	if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM)
852 		cmd |= VMCI_IMR_DMA_DATAGRAM;
853 	vmci_write_reg(vmci_dev, cmd, VMCI_IMR_ADDR);
854 
855 	/* Enable interrupts. */
856 	vmci_write_reg(vmci_dev, VMCI_CONTROL_INT_ENABLE, VMCI_CONTROL_ADDR);
857 
858 	pci_set_drvdata(pdev, vmci_dev);
859 
860 	vmci_call_vsock_callback(false);
861 	return 0;
862 
863 err_free_bm_irq:
864 	if (vmci_dev->exclusive_vectors)
865 		free_irq(pci_irq_vector(pdev, 1), vmci_dev);
866 
867 err_free_irq:
868 	free_irq(pci_irq_vector(pdev, 0), vmci_dev);
869 
870 err_disable_msi:
871 	pci_free_irq_vectors(pdev);
872 
873 err_unsubscribe_event:
874 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
875 	if (vmci_err < VMCI_SUCCESS)
876 		dev_warn(&pdev->dev,
877 			 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
878 			 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
879 
880 err_remove_vmci_dev_g:
881 	spin_lock_irq(&vmci_dev_spinlock);
882 	vmci_pdev = NULL;
883 	vmci_dev_g = NULL;
884 	spin_unlock_irq(&vmci_dev_spinlock);
885 
886 err_free_notification_bitmap:
887 	if (vmci_dev->notification_bitmap) {
888 		vmci_write_reg(vmci_dev, VMCI_CONTROL_RESET, VMCI_CONTROL_ADDR);
889 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
890 				  vmci_dev->notification_bitmap,
891 				  vmci_dev->notification_base);
892 	}
893 
894 err_free_data_buffers:
895 	vmci_free_dg_buffers(vmci_dev);
896 
897 err_unmap_mmio_base:
898 	if (mmio_base != NULL)
899 		pci_iounmap(pdev, mmio_base);
900 
901 	/* The rest are managed resources and will be freed by PCI core */
902 	return error;
903 }
904 
905 static void vmci_guest_remove_device(struct pci_dev *pdev)
906 {
907 	struct vmci_guest_device *vmci_dev = pci_get_drvdata(pdev);
908 	int vmci_err;
909 
910 	dev_dbg(&pdev->dev, "Removing device\n");
911 
912 	atomic_dec(&vmci_num_guest_devices);
913 
914 	vmci_qp_guest_endpoints_exit();
915 
916 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
917 	if (vmci_err < VMCI_SUCCESS)
918 		dev_warn(&pdev->dev,
919 			 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
920 			 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
921 
922 	spin_lock_irq(&vmci_dev_spinlock);
923 	vmci_dev_g = NULL;
924 	vmci_pdev = NULL;
925 	spin_unlock_irq(&vmci_dev_spinlock);
926 
927 	dev_dbg(&pdev->dev, "Resetting vmci device\n");
928 	vmci_write_reg(vmci_dev, VMCI_CONTROL_RESET, VMCI_CONTROL_ADDR);
929 
930 	/*
931 	 * Free IRQ and then disable MSI/MSI-X as appropriate.  For
932 	 * MSI-X, we might have multiple vectors, each with their own
933 	 * IRQ, which we must free too.
934 	 */
935 	if (vmci_dev->exclusive_vectors) {
936 		free_irq(pci_irq_vector(pdev, 1), vmci_dev);
937 		if (vmci_dev->mmio_base != NULL)
938 			free_irq(pci_irq_vector(pdev, 2), vmci_dev);
939 	}
940 	free_irq(pci_irq_vector(pdev, 0), vmci_dev);
941 	pci_free_irq_vectors(pdev);
942 
943 	if (vmci_dev->notification_bitmap) {
944 		/*
945 		 * The device reset above cleared the bitmap state of the
946 		 * device, so we can safely free it here.
947 		 */
948 
949 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
950 				  vmci_dev->notification_bitmap,
951 				  vmci_dev->notification_base);
952 	}
953 
954 	vmci_free_dg_buffers(vmci_dev);
955 
956 	if (vmci_dev->mmio_base != NULL)
957 		pci_iounmap(pdev, vmci_dev->mmio_base);
958 
959 	/* The rest are managed resources and will be freed by PCI core */
960 }
961 
962 static const struct pci_device_id vmci_ids[] = {
963 	{ PCI_DEVICE(PCI_VENDOR_ID_VMWARE, PCI_DEVICE_ID_VMWARE_VMCI), },
964 	{ 0 },
965 };
966 MODULE_DEVICE_TABLE(pci, vmci_ids);
967 
968 static struct pci_driver vmci_guest_driver = {
969 	.name		= KBUILD_MODNAME,
970 	.id_table	= vmci_ids,
971 	.probe		= vmci_guest_probe_device,
972 	.remove		= vmci_guest_remove_device,
973 };
974 
975 int __init vmci_guest_init(void)
976 {
977 	return pci_register_driver(&vmci_guest_driver);
978 }
979 
980 void __exit vmci_guest_exit(void)
981 {
982 	pci_unregister_driver(&vmci_guest_driver);
983 }
984