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(v: &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(dg: &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(addr: 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, addr: 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(port: vmci_dev->iobase + VMCI_DATA_IN_ADDR, |
122 | buf: dest, count: 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 * = 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(dev: 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 * = 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(dev: vmci_dev_g, VMCI_RESULT_LOW_ADDR); |
177 | if (result == VMCI_SUCCESS) |
178 | result = (int)buffer_header->result; |
179 | } else { |
180 | iowrite8_rep(port: dev->iobase + VMCI_DATA_OUT_ADDR, |
181 | buf: dg, VMCI_DG_SIZE(dg)); |
182 | result = vmci_read_reg(dev: 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(dev: vmci_dev_g, dg); |
214 | result = vmci_read_reg(dev: vmci_dev_g, VMCI_RESULT_LOW_ADDR); |
215 | } else { |
216 | result = VMCI_ERROR_UNAVAILABLE; |
217 | } |
218 | |
219 | spin_unlock_irqrestore(lock: &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(ev_data: 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(size: 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(objp: 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, dest: dg_in_buffer, size: 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 | dest: dg_in_buffer + |
396 | remaining_bytes, |
397 | size: 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(msg: 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, dest: dg_in_buffer, |
436 | size: 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, dest: dg_in_buffer, |
454 | size: 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(bitmap: 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(vmci_dev: 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(vmci_dev: 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(dev: vmci_dev->dev, |
558 | VMCI_DMA_DG_BUFFER_SIZE, |
559 | cpu_addr: vmci_dev->tx_buffer, |
560 | dma_handle: vmci_dev->tx_buffer_base); |
561 | if (vmci_dev->data_buffer != NULL) |
562 | dma_free_coherent(dev: vmci_dev->dev, |
563 | VMCI_DMA_DG_BUFFER_SIZE, |
564 | cpu_addr: vmci_dev->data_buffer, |
565 | dma_handle: vmci_dev->data_buffer_base); |
566 | } else { |
567 | vfree(addr: 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(dev: pdev, bar: 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(dev: &pdev->dev, size: sizeof(*vmci_dev), GFP_KERNEL); |
625 | if (!vmci_dev) { |
626 | dev_err(&pdev->dev, |
627 | "Can't allocate memory for VMCI device\n" ); |
628 | return -ENOMEM; |
629 | } |
630 | |
631 | vmci_dev->dev = &pdev->dev; |
632 | vmci_dev->exclusive_vectors = false; |
633 | vmci_dev->iobase = iobase; |
634 | vmci_dev->mmio_base = mmio_base; |
635 | |
636 | init_waitqueue_head(&vmci_dev->inout_wq); |
637 | |
638 | if (mmio_base != NULL) { |
639 | vmci_dev->tx_buffer = dma_alloc_coherent(dev: &pdev->dev, VMCI_DMA_DG_BUFFER_SIZE, |
640 | dma_handle: &vmci_dev->tx_buffer_base, |
641 | GFP_KERNEL); |
642 | if (!vmci_dev->tx_buffer) { |
643 | dev_err(&pdev->dev, |
644 | "Can't allocate memory for datagram tx buffer\n" ); |
645 | return -ENOMEM; |
646 | } |
647 | |
648 | vmci_dev->data_buffer = dma_alloc_coherent(dev: &pdev->dev, VMCI_DMA_DG_BUFFER_SIZE, |
649 | dma_handle: &vmci_dev->data_buffer_base, |
650 | GFP_KERNEL); |
651 | } else { |
652 | vmci_dev->data_buffer = vmalloc(VMCI_MAX_DG_SIZE); |
653 | } |
654 | if (!vmci_dev->data_buffer) { |
655 | dev_err(&pdev->dev, |
656 | "Can't allocate memory for datagram buffer\n" ); |
657 | error = -ENOMEM; |
658 | goto err_free_data_buffers; |
659 | } |
660 | |
661 | pci_set_master(dev: pdev); /* To enable queue_pair functionality. */ |
662 | |
663 | /* |
664 | * Verify that the VMCI Device supports the capabilities that |
665 | * we need. If the device is missing capabilities that we would |
666 | * like to use, check for fallback capabilities and use those |
667 | * instead (so we can run a new VM on old hosts). Fail the load if |
668 | * a required capability is missing and there is no fallback. |
669 | * |
670 | * Right now, we need datagrams. There are no fallbacks. |
671 | */ |
672 | capabilities = vmci_read_reg(dev: vmci_dev, VMCI_CAPS_ADDR); |
673 | if (!(capabilities & VMCI_CAPS_DATAGRAM)) { |
674 | dev_err(&pdev->dev, "Device does not support datagrams\n" ); |
675 | error = -ENXIO; |
676 | goto err_free_data_buffers; |
677 | } |
678 | caps_in_use = VMCI_CAPS_DATAGRAM; |
679 | |
680 | /* |
681 | * Use 64-bit PPNs if the device supports. |
682 | * |
683 | * There is no check for the return value of dma_set_mask_and_coherent |
684 | * since this driver can handle the default mask values if |
685 | * dma_set_mask_and_coherent fails. |
686 | */ |
687 | if (capabilities & VMCI_CAPS_PPN64) { |
688 | dma_set_mask_and_coherent(dev: &pdev->dev, DMA_BIT_MASK(64)); |
689 | use_ppn64 = true; |
690 | caps_in_use |= VMCI_CAPS_PPN64; |
691 | } else { |
692 | dma_set_mask_and_coherent(dev: &pdev->dev, DMA_BIT_MASK(44)); |
693 | use_ppn64 = false; |
694 | } |
695 | |
696 | /* |
697 | * If the hardware supports notifications, we will use that as |
698 | * well. |
699 | */ |
700 | if (capabilities & VMCI_CAPS_NOTIFICATIONS) { |
701 | vmci_dev->notification_bitmap = dma_alloc_coherent( |
702 | dev: &pdev->dev, PAGE_SIZE, dma_handle: &vmci_dev->notification_base, |
703 | GFP_KERNEL); |
704 | if (!vmci_dev->notification_bitmap) |
705 | dev_warn(&pdev->dev, |
706 | "Unable to allocate notification bitmap\n" ); |
707 | else |
708 | caps_in_use |= VMCI_CAPS_NOTIFICATIONS; |
709 | } |
710 | |
711 | if (mmio_base != NULL) { |
712 | if (capabilities & VMCI_CAPS_DMA_DATAGRAM) { |
713 | caps_in_use |= VMCI_CAPS_DMA_DATAGRAM; |
714 | } else { |
715 | dev_err(&pdev->dev, |
716 | "Missing capability: VMCI_CAPS_DMA_DATAGRAM\n" ); |
717 | error = -ENXIO; |
718 | goto err_free_notification_bitmap; |
719 | } |
720 | } |
721 | |
722 | dev_info(&pdev->dev, "Using capabilities 0x%x\n" , caps_in_use); |
723 | |
724 | /* Let the host know which capabilities we intend to use. */ |
725 | vmci_write_reg(dev: vmci_dev, val: caps_in_use, VMCI_CAPS_ADDR); |
726 | |
727 | if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM) { |
728 | /* Let the device know the size for pages passed down. */ |
729 | vmci_write_reg(dev: vmci_dev, PAGE_SHIFT, VMCI_GUEST_PAGE_SHIFT); |
730 | |
731 | /* Configure the high order parts of the data in/out buffers. */ |
732 | vmci_write_reg(dev: vmci_dev, upper_32_bits(vmci_dev->data_buffer_base), |
733 | VMCI_DATA_IN_HIGH_ADDR); |
734 | vmci_write_reg(dev: vmci_dev, upper_32_bits(vmci_dev->tx_buffer_base), |
735 | VMCI_DATA_OUT_HIGH_ADDR); |
736 | } |
737 | |
738 | /* Set up global device so that we can start sending datagrams */ |
739 | spin_lock_irq(lock: &vmci_dev_spinlock); |
740 | vmci_dev_g = vmci_dev; |
741 | vmci_pdev = pdev; |
742 | spin_unlock_irq(lock: &vmci_dev_spinlock); |
743 | |
744 | /* |
745 | * Register notification bitmap with device if that capability is |
746 | * used. |
747 | */ |
748 | if (caps_in_use & VMCI_CAPS_NOTIFICATIONS) { |
749 | unsigned long bitmap_ppn = |
750 | vmci_dev->notification_base >> PAGE_SHIFT; |
751 | if (!vmci_dbell_register_notification_bitmap(bitmap_ppn)) { |
752 | dev_warn(&pdev->dev, |
753 | "VMCI device unable to register notification bitmap with PPN 0x%lx\n" , |
754 | bitmap_ppn); |
755 | error = -ENXIO; |
756 | goto err_remove_vmci_dev_g; |
757 | } |
758 | } |
759 | |
760 | /* Check host capabilities. */ |
761 | error = vmci_check_host_caps(pdev); |
762 | if (error) |
763 | goto err_remove_vmci_dev_g; |
764 | |
765 | /* Enable device. */ |
766 | |
767 | /* |
768 | * We subscribe to the VMCI_EVENT_CTX_ID_UPDATE here so we can |
769 | * update the internal context id when needed. |
770 | */ |
771 | vmci_err = vmci_event_subscribe(event: VMCI_EVENT_CTX_ID_UPDATE, |
772 | callback: vmci_guest_cid_update, NULL, |
773 | subid: &ctx_update_sub_id); |
774 | if (vmci_err < VMCI_SUCCESS) |
775 | dev_warn(&pdev->dev, |
776 | "Failed to subscribe to event (type=%d): %d\n" , |
777 | VMCI_EVENT_CTX_ID_UPDATE, vmci_err); |
778 | |
779 | /* |
780 | * Enable interrupts. Try MSI-X first, then MSI, and then fallback on |
781 | * legacy interrupts. |
782 | */ |
783 | if (vmci_dev->mmio_base != NULL) |
784 | num_irq_vectors = VMCI_MAX_INTRS; |
785 | else |
786 | num_irq_vectors = VMCI_MAX_INTRS_NOTIFICATION; |
787 | error = pci_alloc_irq_vectors(dev: pdev, min_vecs: num_irq_vectors, max_vecs: num_irq_vectors, |
788 | PCI_IRQ_MSIX); |
789 | if (error < 0) { |
790 | error = pci_alloc_irq_vectors(dev: pdev, min_vecs: 1, max_vecs: 1, |
791 | PCI_IRQ_MSIX | PCI_IRQ_MSI | PCI_IRQ_LEGACY); |
792 | if (error < 0) |
793 | goto err_unsubscribe_event; |
794 | } else { |
795 | vmci_dev->exclusive_vectors = true; |
796 | } |
797 | |
798 | /* |
799 | * Request IRQ for legacy or MSI interrupts, or for first |
800 | * MSI-X vector. |
801 | */ |
802 | error = request_threaded_irq(irq: pci_irq_vector(dev: pdev, nr: 0), NULL, |
803 | thread_fn: vmci_interrupt, IRQF_SHARED, |
804 | KBUILD_MODNAME, dev: vmci_dev); |
805 | if (error) { |
806 | dev_err(&pdev->dev, "Irq %u in use: %d\n" , |
807 | pci_irq_vector(pdev, 0), error); |
808 | goto err_disable_msi; |
809 | } |
810 | |
811 | /* |
812 | * For MSI-X with exclusive vectors we need to request an |
813 | * interrupt for each vector so that we get a separate |
814 | * interrupt handler routine. This allows us to distinguish |
815 | * between the vectors. |
816 | */ |
817 | if (vmci_dev->exclusive_vectors) { |
818 | error = request_threaded_irq(irq: pci_irq_vector(dev: pdev, nr: 1), NULL, |
819 | thread_fn: vmci_interrupt_bm, flags: 0, |
820 | KBUILD_MODNAME, dev: vmci_dev); |
821 | if (error) { |
822 | dev_err(&pdev->dev, |
823 | "Failed to allocate irq %u: %d\n" , |
824 | pci_irq_vector(pdev, 1), error); |
825 | goto err_free_irq; |
826 | } |
827 | if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM) { |
828 | error = request_threaded_irq(irq: pci_irq_vector(dev: pdev, nr: 2), |
829 | NULL, |
830 | thread_fn: vmci_interrupt_dma_datagram, |
831 | flags: 0, KBUILD_MODNAME, |
832 | dev: vmci_dev); |
833 | if (error) { |
834 | dev_err(&pdev->dev, |
835 | "Failed to allocate irq %u: %d\n" , |
836 | pci_irq_vector(pdev, 2), error); |
837 | goto err_free_bm_irq; |
838 | } |
839 | } |
840 | } |
841 | |
842 | dev_dbg(&pdev->dev, "Registered device\n" ); |
843 | |
844 | atomic_inc(v: &vmci_num_guest_devices); |
845 | |
846 | /* Enable specific interrupt bits. */ |
847 | cmd = VMCI_IMR_DATAGRAM; |
848 | if (caps_in_use & VMCI_CAPS_NOTIFICATIONS) |
849 | cmd |= VMCI_IMR_NOTIFICATION; |
850 | if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM) |
851 | cmd |= VMCI_IMR_DMA_DATAGRAM; |
852 | vmci_write_reg(dev: vmci_dev, val: cmd, VMCI_IMR_ADDR); |
853 | |
854 | /* Enable interrupts. */ |
855 | vmci_write_reg(dev: vmci_dev, VMCI_CONTROL_INT_ENABLE, VMCI_CONTROL_ADDR); |
856 | |
857 | pci_set_drvdata(pdev, data: vmci_dev); |
858 | |
859 | vmci_call_vsock_callback(is_host: false); |
860 | return 0; |
861 | |
862 | err_free_bm_irq: |
863 | if (vmci_dev->exclusive_vectors) |
864 | free_irq(pci_irq_vector(dev: pdev, nr: 1), vmci_dev); |
865 | |
866 | err_free_irq: |
867 | free_irq(pci_irq_vector(dev: pdev, nr: 0), vmci_dev); |
868 | |
869 | err_disable_msi: |
870 | pci_free_irq_vectors(dev: pdev); |
871 | |
872 | err_unsubscribe_event: |
873 | vmci_err = vmci_event_unsubscribe(subid: ctx_update_sub_id); |
874 | if (vmci_err < VMCI_SUCCESS) |
875 | dev_warn(&pdev->dev, |
876 | "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n" , |
877 | VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err); |
878 | |
879 | err_remove_vmci_dev_g: |
880 | spin_lock_irq(lock: &vmci_dev_spinlock); |
881 | vmci_pdev = NULL; |
882 | vmci_dev_g = NULL; |
883 | spin_unlock_irq(lock: &vmci_dev_spinlock); |
884 | |
885 | err_free_notification_bitmap: |
886 | if (vmci_dev->notification_bitmap) { |
887 | vmci_write_reg(dev: vmci_dev, VMCI_CONTROL_RESET, VMCI_CONTROL_ADDR); |
888 | dma_free_coherent(dev: &pdev->dev, PAGE_SIZE, |
889 | cpu_addr: vmci_dev->notification_bitmap, |
890 | dma_handle: vmci_dev->notification_base); |
891 | } |
892 | |
893 | err_free_data_buffers: |
894 | vmci_free_dg_buffers(vmci_dev); |
895 | |
896 | /* The rest are managed resources and will be freed by PCI core */ |
897 | return error; |
898 | } |
899 | |
900 | static void vmci_guest_remove_device(struct pci_dev *pdev) |
901 | { |
902 | struct vmci_guest_device *vmci_dev = pci_get_drvdata(pdev); |
903 | int vmci_err; |
904 | |
905 | dev_dbg(&pdev->dev, "Removing device\n" ); |
906 | |
907 | atomic_dec(v: &vmci_num_guest_devices); |
908 | |
909 | vmci_qp_guest_endpoints_exit(); |
910 | |
911 | vmci_err = vmci_event_unsubscribe(subid: ctx_update_sub_id); |
912 | if (vmci_err < VMCI_SUCCESS) |
913 | dev_warn(&pdev->dev, |
914 | "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n" , |
915 | VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err); |
916 | |
917 | spin_lock_irq(lock: &vmci_dev_spinlock); |
918 | vmci_dev_g = NULL; |
919 | vmci_pdev = NULL; |
920 | spin_unlock_irq(lock: &vmci_dev_spinlock); |
921 | |
922 | dev_dbg(&pdev->dev, "Resetting vmci device\n" ); |
923 | vmci_write_reg(dev: vmci_dev, VMCI_CONTROL_RESET, VMCI_CONTROL_ADDR); |
924 | |
925 | /* |
926 | * Free IRQ and then disable MSI/MSI-X as appropriate. For |
927 | * MSI-X, we might have multiple vectors, each with their own |
928 | * IRQ, which we must free too. |
929 | */ |
930 | if (vmci_dev->exclusive_vectors) { |
931 | free_irq(pci_irq_vector(dev: pdev, nr: 1), vmci_dev); |
932 | if (vmci_dev->mmio_base != NULL) |
933 | free_irq(pci_irq_vector(dev: pdev, nr: 2), vmci_dev); |
934 | } |
935 | free_irq(pci_irq_vector(dev: pdev, nr: 0), vmci_dev); |
936 | pci_free_irq_vectors(dev: pdev); |
937 | |
938 | if (vmci_dev->notification_bitmap) { |
939 | /* |
940 | * The device reset above cleared the bitmap state of the |
941 | * device, so we can safely free it here. |
942 | */ |
943 | |
944 | dma_free_coherent(dev: &pdev->dev, PAGE_SIZE, |
945 | cpu_addr: vmci_dev->notification_bitmap, |
946 | dma_handle: vmci_dev->notification_base); |
947 | } |
948 | |
949 | vmci_free_dg_buffers(vmci_dev); |
950 | |
951 | if (vmci_dev->mmio_base != NULL) |
952 | pci_iounmap(dev: pdev, vmci_dev->mmio_base); |
953 | |
954 | /* The rest are managed resources and will be freed by PCI core */ |
955 | } |
956 | |
957 | static const struct pci_device_id vmci_ids[] = { |
958 | { PCI_DEVICE(PCI_VENDOR_ID_VMWARE, PCI_DEVICE_ID_VMWARE_VMCI), }, |
959 | { 0 }, |
960 | }; |
961 | MODULE_DEVICE_TABLE(pci, vmci_ids); |
962 | |
963 | static struct pci_driver vmci_guest_driver = { |
964 | .name = KBUILD_MODNAME, |
965 | .id_table = vmci_ids, |
966 | .probe = vmci_guest_probe_device, |
967 | .remove = vmci_guest_remove_device, |
968 | }; |
969 | |
970 | int __init vmci_guest_init(void) |
971 | { |
972 | return pci_register_driver(&vmci_guest_driver); |
973 | } |
974 | |
975 | void __exit vmci_guest_exit(void) |
976 | { |
977 | pci_unregister_driver(dev: &vmci_guest_driver); |
978 | } |
979 | |