xref: /dpdk/drivers/common/mlx5/mlx5_common.c (revision cf8a8a8f4896c0885d3996716f73513c4317e545)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2019 Mellanox Technologies, Ltd
3  */
4 
5 #include <unistd.h>
6 #include <string.h>
7 #include <stdio.h>
8 
9 #include <rte_errno.h>
10 #include <rte_mempool.h>
11 #include <rte_class.h>
12 #include <rte_malloc.h>
13 
14 #include "mlx5_common.h"
15 #include "mlx5_common_os.h"
16 #include "mlx5_common_log.h"
17 #include "mlx5_common_pci.h"
18 #include "mlx5_common_private.h"
19 
20 uint8_t haswell_broadwell_cpu;
21 
22 /* In case this is an x86_64 intel processor to check if
23  * we should use relaxed ordering.
24  */
25 #ifdef RTE_ARCH_X86_64
26 /**
27  * This function returns processor identification and feature information
28  * into the registers.
29  *
30  * @param eax, ebx, ecx, edx
31  *		Pointers to the registers that will hold cpu information.
32  * @param level
33  *		The main category of information returned.
34  */
35 static inline void mlx5_cpu_id(unsigned int level,
36 				unsigned int *eax, unsigned int *ebx,
37 				unsigned int *ecx, unsigned int *edx)
38 {
39 	__asm__("cpuid\n\t"
40 		: "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx)
41 		: "0" (level));
42 }
43 #endif
44 
45 RTE_LOG_REGISTER_DEFAULT(mlx5_common_logtype, NOTICE)
46 
47 /* Head of list of drivers. */
48 static TAILQ_HEAD(mlx5_drivers, mlx5_class_driver) drivers_list =
49 				TAILQ_HEAD_INITIALIZER(drivers_list);
50 
51 /* Head of devices. */
52 static TAILQ_HEAD(mlx5_devices, mlx5_common_device) devices_list =
53 				TAILQ_HEAD_INITIALIZER(devices_list);
54 
55 static const struct {
56 	const char *name;
57 	unsigned int drv_class;
58 } mlx5_classes[] = {
59 	{ .name = "vdpa", .drv_class = MLX5_CLASS_VDPA },
60 	{ .name = "eth", .drv_class = MLX5_CLASS_ETH },
61 	/* Keep class "net" for backward compatibility. */
62 	{ .name = "net", .drv_class = MLX5_CLASS_ETH },
63 	{ .name = "regex", .drv_class = MLX5_CLASS_REGEX },
64 	{ .name = "compress", .drv_class = MLX5_CLASS_COMPRESS },
65 	{ .name = "crypto", .drv_class = MLX5_CLASS_CRYPTO },
66 };
67 
68 static int
69 class_name_to_value(const char *class_name)
70 {
71 	unsigned int i;
72 
73 	for (i = 0; i < RTE_DIM(mlx5_classes); i++) {
74 		if (strcmp(class_name, mlx5_classes[i].name) == 0)
75 			return mlx5_classes[i].drv_class;
76 	}
77 	return -EINVAL;
78 }
79 
80 static struct mlx5_class_driver *
81 driver_get(uint32_t class)
82 {
83 	struct mlx5_class_driver *driver;
84 
85 	TAILQ_FOREACH(driver, &drivers_list, next) {
86 		if ((uint32_t)driver->drv_class == class)
87 			return driver;
88 	}
89 	return NULL;
90 }
91 
92 static int
93 devargs_class_handler(__rte_unused const char *key,
94 		      const char *class_names, void *opaque)
95 {
96 	int *ret = opaque;
97 	int class_val;
98 	char *scratch;
99 	char *found;
100 	char *refstr = NULL;
101 
102 	*ret = 0;
103 	scratch = strdup(class_names);
104 	if (scratch == NULL) {
105 		*ret = -ENOMEM;
106 		return *ret;
107 	}
108 	found = strtok_r(scratch, ":", &refstr);
109 	if (found == NULL)
110 		/* Empty string. */
111 		goto err;
112 	do {
113 		/* Extract each individual class name. Multiple
114 		 * classes can be supplied as class=net:regex:foo:bar.
115 		 */
116 		class_val = class_name_to_value(found);
117 		/* Check if its a valid class. */
118 		if (class_val < 0) {
119 			*ret = -EINVAL;
120 			goto err;
121 		}
122 		*ret |= class_val;
123 		found = strtok_r(NULL, ":", &refstr);
124 	} while (found != NULL);
125 err:
126 	free(scratch);
127 	if (*ret < 0)
128 		DRV_LOG(ERR, "Invalid mlx5 class options: %s.\n", class_names);
129 	return *ret;
130 }
131 
132 static int
133 parse_class_options(const struct rte_devargs *devargs)
134 {
135 	struct rte_kvargs *kvlist;
136 	int ret = 0;
137 
138 	if (devargs == NULL)
139 		return 0;
140 	if (devargs->cls != NULL && devargs->cls->name != NULL)
141 		/* Global syntax, only one class type. */
142 		return class_name_to_value(devargs->cls->name);
143 	/* Legacy devargs support multiple classes. */
144 	kvlist = rte_kvargs_parse(devargs->args, NULL);
145 	if (kvlist == NULL)
146 		return 0;
147 	rte_kvargs_process(kvlist, RTE_DEVARGS_KEY_CLASS,
148 			   devargs_class_handler, &ret);
149 	rte_kvargs_free(kvlist);
150 	return ret;
151 }
152 
153 static const unsigned int mlx5_class_invalid_combinations[] = {
154 	MLX5_CLASS_ETH | MLX5_CLASS_VDPA,
155 	/* New class combination should be added here. */
156 };
157 
158 static int
159 is_valid_class_combination(uint32_t user_classes)
160 {
161 	unsigned int i;
162 
163 	/* Verify if user specified unsupported combination. */
164 	for (i = 0; i < RTE_DIM(mlx5_class_invalid_combinations); i++) {
165 		if ((mlx5_class_invalid_combinations[i] & user_classes) ==
166 		    mlx5_class_invalid_combinations[i])
167 			return -EINVAL;
168 	}
169 	/* Not found any invalid class combination. */
170 	return 0;
171 }
172 
173 static bool
174 device_class_enabled(const struct mlx5_common_device *device, uint32_t class)
175 {
176 	return (device->classes_loaded & class) > 0;
177 }
178 
179 static bool
180 mlx5_bus_match(const struct mlx5_class_driver *drv,
181 	       const struct rte_device *dev)
182 {
183 	if (mlx5_dev_is_pci(dev))
184 		return mlx5_dev_pci_match(drv, dev);
185 	return true;
186 }
187 
188 static struct mlx5_common_device *
189 to_mlx5_device(const struct rte_device *rte_dev)
190 {
191 	struct mlx5_common_device *dev;
192 
193 	TAILQ_FOREACH(dev, &devices_list, next) {
194 		if (rte_dev == dev->dev)
195 			return dev;
196 	}
197 	return NULL;
198 }
199 
200 int
201 mlx5_dev_to_pci_str(const struct rte_device *dev, char *addr, size_t size)
202 {
203 	struct rte_pci_addr pci_addr = { 0 };
204 	int ret;
205 
206 	if (mlx5_dev_is_pci(dev)) {
207 		/* Input might be <BDF>, format PCI address to <DBDF>. */
208 		ret = rte_pci_addr_parse(dev->name, &pci_addr);
209 		if (ret != 0)
210 			return -ENODEV;
211 		rte_pci_device_name(&pci_addr, addr, size);
212 		return 0;
213 	}
214 #ifdef RTE_EXEC_ENV_LINUX
215 	return mlx5_auxiliary_get_pci_str(RTE_DEV_TO_AUXILIARY_CONST(dev),
216 			addr, size);
217 #else
218 	rte_errno = ENODEV;
219 	return -rte_errno;
220 #endif
221 }
222 
223 static void
224 dev_release(struct mlx5_common_device *dev)
225 {
226 	TAILQ_REMOVE(&devices_list, dev, next);
227 	rte_free(dev);
228 }
229 
230 static int
231 drivers_remove(struct mlx5_common_device *dev, uint32_t enabled_classes)
232 {
233 	struct mlx5_class_driver *driver;
234 	int local_ret = -ENODEV;
235 	unsigned int i = 0;
236 	int ret = 0;
237 
238 	enabled_classes &= dev->classes_loaded;
239 	while (enabled_classes) {
240 		driver = driver_get(RTE_BIT64(i));
241 		if (driver != NULL) {
242 			local_ret = driver->remove(dev->dev);
243 			if (local_ret == 0)
244 				dev->classes_loaded &= ~RTE_BIT64(i);
245 			else if (ret == 0)
246 				ret = local_ret;
247 		}
248 		enabled_classes &= ~RTE_BIT64(i);
249 		i++;
250 	}
251 	if (local_ret != 0 && ret == 0)
252 		ret = local_ret;
253 	return ret;
254 }
255 
256 static int
257 drivers_probe(struct mlx5_common_device *dev, uint32_t user_classes)
258 {
259 	struct mlx5_class_driver *driver;
260 	uint32_t enabled_classes = 0;
261 	bool already_loaded;
262 	int ret;
263 
264 	TAILQ_FOREACH(driver, &drivers_list, next) {
265 		if ((driver->drv_class & user_classes) == 0)
266 			continue;
267 		if (!mlx5_bus_match(driver, dev->dev))
268 			continue;
269 		already_loaded = dev->classes_loaded & driver->drv_class;
270 		if (already_loaded && driver->probe_again == 0) {
271 			DRV_LOG(ERR, "Device %s is already probed",
272 				dev->dev->name);
273 			ret = -EEXIST;
274 			goto probe_err;
275 		}
276 		ret = driver->probe(dev->dev);
277 		if (ret < 0) {
278 			DRV_LOG(ERR, "Failed to load driver %s",
279 				driver->name);
280 			goto probe_err;
281 		}
282 		enabled_classes |= driver->drv_class;
283 	}
284 	dev->classes_loaded |= enabled_classes;
285 	return 0;
286 probe_err:
287 	/* Only unload drivers which are enabled which were enabled
288 	 * in this probe instance.
289 	 */
290 	drivers_remove(dev, enabled_classes);
291 	return ret;
292 }
293 
294 int
295 mlx5_common_dev_probe(struct rte_device *eal_dev)
296 {
297 	struct mlx5_common_device *dev;
298 	uint32_t classes = 0;
299 	bool new_device = false;
300 	int ret;
301 
302 	DRV_LOG(INFO, "probe device \"%s\".", eal_dev->name);
303 	ret = parse_class_options(eal_dev->devargs);
304 	if (ret < 0) {
305 		DRV_LOG(ERR, "Unsupported mlx5 class type: %s",
306 			eal_dev->devargs->args);
307 		return ret;
308 	}
309 	classes = ret;
310 	if (classes == 0)
311 		/* Default to net class. */
312 		classes = MLX5_CLASS_ETH;
313 	dev = to_mlx5_device(eal_dev);
314 	if (!dev) {
315 		dev = rte_zmalloc("mlx5_common_device", sizeof(*dev), 0);
316 		if (!dev)
317 			return -ENOMEM;
318 		dev->dev = eal_dev;
319 		TAILQ_INSERT_HEAD(&devices_list, dev, next);
320 		new_device = true;
321 	} else {
322 		/* Validate combination here. */
323 		ret = is_valid_class_combination(classes |
324 						 dev->classes_loaded);
325 		if (ret != 0) {
326 			DRV_LOG(ERR, "Unsupported mlx5 classes combination.");
327 			return ret;
328 		}
329 	}
330 	ret = drivers_probe(dev, classes);
331 	if (ret)
332 		goto class_err;
333 	return 0;
334 class_err:
335 	if (new_device)
336 		dev_release(dev);
337 	return ret;
338 }
339 
340 int
341 mlx5_common_dev_remove(struct rte_device *eal_dev)
342 {
343 	struct mlx5_common_device *dev;
344 	int ret;
345 
346 	dev = to_mlx5_device(eal_dev);
347 	if (!dev)
348 		return -ENODEV;
349 	/* Matching device found, cleanup and unload drivers. */
350 	ret = drivers_remove(dev, dev->classes_loaded);
351 	if (ret != 0)
352 		dev_release(dev);
353 	return ret;
354 }
355 
356 int
357 mlx5_common_dev_dma_map(struct rte_device *dev, void *addr, uint64_t iova,
358 			size_t len)
359 {
360 	struct mlx5_class_driver *driver = NULL;
361 	struct mlx5_class_driver *temp;
362 	struct mlx5_common_device *mdev;
363 	int ret = -EINVAL;
364 
365 	mdev = to_mlx5_device(dev);
366 	if (!mdev)
367 		return -ENODEV;
368 	TAILQ_FOREACH(driver, &drivers_list, next) {
369 		if (!device_class_enabled(mdev, driver->drv_class) ||
370 		    driver->dma_map == NULL)
371 			continue;
372 		ret = driver->dma_map(dev, addr, iova, len);
373 		if (ret)
374 			goto map_err;
375 	}
376 	return ret;
377 map_err:
378 	TAILQ_FOREACH(temp, &drivers_list, next) {
379 		if (temp == driver)
380 			break;
381 		if (device_class_enabled(mdev, temp->drv_class) &&
382 		    temp->dma_map && temp->dma_unmap)
383 			temp->dma_unmap(dev, addr, iova, len);
384 	}
385 	return ret;
386 }
387 
388 int
389 mlx5_common_dev_dma_unmap(struct rte_device *dev, void *addr, uint64_t iova,
390 			  size_t len)
391 {
392 	struct mlx5_class_driver *driver;
393 	struct mlx5_common_device *mdev;
394 	int local_ret = -EINVAL;
395 	int ret = 0;
396 
397 	mdev = to_mlx5_device(dev);
398 	if (!mdev)
399 		return -ENODEV;
400 	/* There is no unmap error recovery in current implementation. */
401 	TAILQ_FOREACH_REVERSE(driver, &drivers_list, mlx5_drivers, next) {
402 		if (!device_class_enabled(mdev, driver->drv_class) ||
403 		    driver->dma_unmap == NULL)
404 			continue;
405 		local_ret = driver->dma_unmap(dev, addr, iova, len);
406 		if (local_ret && (ret == 0))
407 			ret = local_ret;
408 	}
409 	if (local_ret)
410 		ret = local_ret;
411 	return ret;
412 }
413 
414 void
415 mlx5_class_driver_register(struct mlx5_class_driver *driver)
416 {
417 	mlx5_common_driver_on_register_pci(driver);
418 	TAILQ_INSERT_TAIL(&drivers_list, driver, next);
419 }
420 
421 static void mlx5_common_driver_init(void)
422 {
423 	mlx5_common_pci_init();
424 #ifdef RTE_EXEC_ENV_LINUX
425 	mlx5_common_auxiliary_init();
426 #endif
427 }
428 
429 static bool mlx5_common_initialized;
430 
431 /**
432  * One time innitialization routine for run-time dependency on glue library
433  * for multiple PMDs. Each mlx5 PMD that depends on mlx5_common module,
434  * must invoke in its constructor.
435  */
436 void
437 mlx5_common_init(void)
438 {
439 	if (mlx5_common_initialized)
440 		return;
441 
442 	mlx5_glue_constructor();
443 	mlx5_common_driver_init();
444 	mlx5_common_initialized = true;
445 }
446 
447 /**
448  * This function is responsible of initializing the variable
449  *  haswell_broadwell_cpu by checking if the cpu is intel
450  *  and reading the data returned from mlx5_cpu_id().
451  *  since haswell and broadwell cpus don't have improved performance
452  *  when using relaxed ordering we want to check the cpu type before
453  *  before deciding whether to enable RO or not.
454  *  if the cpu is haswell or broadwell the variable will be set to 1
455  *  otherwise it will be 0.
456  */
457 RTE_INIT_PRIO(mlx5_is_haswell_broadwell_cpu, LOG)
458 {
459 #ifdef RTE_ARCH_X86_64
460 	unsigned int broadwell_models[4] = {0x3d, 0x47, 0x4F, 0x56};
461 	unsigned int haswell_models[4] = {0x3c, 0x3f, 0x45, 0x46};
462 	unsigned int i, model, family, brand_id, vendor;
463 	unsigned int signature_intel_ebx = 0x756e6547;
464 	unsigned int extended_model;
465 	unsigned int eax = 0;
466 	unsigned int ebx = 0;
467 	unsigned int ecx = 0;
468 	unsigned int edx = 0;
469 	int max_level;
470 
471 	mlx5_cpu_id(0, &eax, &ebx, &ecx, &edx);
472 	vendor = ebx;
473 	max_level = eax;
474 	if (max_level < 1) {
475 		haswell_broadwell_cpu = 0;
476 		return;
477 	}
478 	mlx5_cpu_id(1, &eax, &ebx, &ecx, &edx);
479 	model = (eax >> 4) & 0x0f;
480 	family = (eax >> 8) & 0x0f;
481 	brand_id = ebx & 0xff;
482 	extended_model = (eax >> 12) & 0xf0;
483 	/* Check if the processor is Haswell or Broadwell */
484 	if (vendor == signature_intel_ebx) {
485 		if (family == 0x06)
486 			model += extended_model;
487 		if (brand_id == 0 && family == 0x6) {
488 			for (i = 0; i < RTE_DIM(broadwell_models); i++)
489 				if (model == broadwell_models[i]) {
490 					haswell_broadwell_cpu = 1;
491 					return;
492 				}
493 			for (i = 0; i < RTE_DIM(haswell_models); i++)
494 				if (model == haswell_models[i]) {
495 					haswell_broadwell_cpu = 1;
496 					return;
497 				}
498 		}
499 	}
500 #endif
501 	haswell_broadwell_cpu = 0;
502 }
503 
504 /**
505  * Allocate the User Access Region with DevX on specified device.
506  *
507  * @param [in] ctx
508  *   Infiniband device context to perform allocation on.
509  * @param [in] mapping
510  *   MLX5DV_UAR_ALLOC_TYPE_BF - allocate as cached memory with write-combining
511  *				attributes (if supported by the host), the
512  *				writes to the UAR registers must be followed
513  *				by write memory barrier.
514  *   MLX5DV_UAR_ALLOC_TYPE_NC - allocate as non-cached nenory, all writes are
515  *				promoted to the registers immediately, no
516  *				memory barriers needed.
517  *   mapping < 0 - the first attempt is performed with MLX5DV_UAR_ALLOC_TYPE_BF,
518  *		   if this fails the next attempt with MLX5DV_UAR_ALLOC_TYPE_NC
519  *		   is performed. The drivers specifying negative values should
520  *		   always provide the write memory barrier operation after UAR
521  *		   register writings.
522  * If there is no definitions for the MLX5DV_UAR_ALLOC_TYPE_xx (older rdma
523  * library headers), the caller can specify 0.
524  *
525  * @return
526  *   UAR object pointer on success, NULL otherwise and rte_errno is set.
527  */
528 void *
529 mlx5_devx_alloc_uar(void *ctx, int mapping)
530 {
531 	void *uar;
532 	uint32_t retry, uar_mapping;
533 	void *base_addr;
534 
535 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
536 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
537 		/* Control the mapping type according to the settings. */
538 		uar_mapping = (mapping < 0) ?
539 			      MLX5DV_UAR_ALLOC_TYPE_NC : mapping;
540 #else
541 		/*
542 		 * It seems we have no way to control the memory mapping type
543 		 * for the UAR, the default "Write-Combining" type is supposed.
544 		 */
545 		uar_mapping = 0;
546 		RTE_SET_USED(mapping);
547 #endif
548 		uar = mlx5_glue->devx_alloc_uar(ctx, uar_mapping);
549 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
550 		if (!uar &&
551 		    mapping < 0 &&
552 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
553 			/*
554 			 * In some environments like virtual machine the
555 			 * Write Combining mapped might be not supported and
556 			 * UAR allocation fails. We tried "Non-Cached" mapping
557 			 * for the case.
558 			 */
559 			DRV_LOG(WARNING, "Failed to allocate DevX UAR (BF)");
560 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
561 			uar = mlx5_glue->devx_alloc_uar(ctx, uar_mapping);
562 		} else if (!uar &&
563 			   mapping < 0 &&
564 			   uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
565 			/*
566 			 * If Verbs/kernel does not support "Non-Cached"
567 			 * try the "Write-Combining".
568 			 */
569 			DRV_LOG(WARNING, "Failed to allocate DevX UAR (NC)");
570 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
571 			uar = mlx5_glue->devx_alloc_uar(ctx, uar_mapping);
572 		}
573 #endif
574 		if (!uar) {
575 			DRV_LOG(ERR, "Failed to allocate DevX UAR (BF/NC)");
576 			rte_errno = ENOMEM;
577 			goto exit;
578 		}
579 		base_addr = mlx5_os_get_devx_uar_base_addr(uar);
580 		if (base_addr)
581 			break;
582 		/*
583 		 * The UARs are allocated by rdma_core within the
584 		 * IB device context, on context closure all UARs
585 		 * will be freed, should be no memory/object leakage.
586 		 */
587 		DRV_LOG(WARNING, "Retrying to allocate DevX UAR");
588 		uar = NULL;
589 	}
590 	/* Check whether we finally succeeded with valid UAR allocation. */
591 	if (!uar) {
592 		DRV_LOG(ERR, "Failed to allocate DevX UAR (NULL base)");
593 		rte_errno = ENOMEM;
594 	}
595 	/*
596 	 * Return void * instead of struct mlx5dv_devx_uar *
597 	 * is for compatibility with older rdma-core library headers.
598 	 */
599 exit:
600 	return uar;
601 }
602 
603 RTE_PMD_EXPORT_NAME(mlx5_common_driver, __COUNTER__);
604