xref: /linux/drivers/char/ipmi/ipmi_si_intf.c (revision 26fbb4c8c7c3ee9a4c3b4de555a8587b5a19154e)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * ipmi_si.c
4  *
5  * The interface to the IPMI driver for the system interfaces (KCS, SMIC,
6  * BT).
7  *
8  * Author: MontaVista Software, Inc.
9  *         Corey Minyard <minyard@mvista.com>
10  *         source@mvista.com
11  *
12  * Copyright 2002 MontaVista Software Inc.
13  * Copyright 2006 IBM Corp., Christian Krafft <krafft@de.ibm.com>
14  */
15 
16 /*
17  * This file holds the "policy" for the interface to the SMI state
18  * machine.  It does the configuration, handles timers and interrupts,
19  * and drives the real SMI state machine.
20  */
21 
22 #define pr_fmt(fmt) "ipmi_si: " fmt
23 
24 #include <linux/module.h>
25 #include <linux/moduleparam.h>
26 #include <linux/sched.h>
27 #include <linux/seq_file.h>
28 #include <linux/timer.h>
29 #include <linux/errno.h>
30 #include <linux/spinlock.h>
31 #include <linux/slab.h>
32 #include <linux/delay.h>
33 #include <linux/list.h>
34 #include <linux/notifier.h>
35 #include <linux/mutex.h>
36 #include <linux/kthread.h>
37 #include <asm/irq.h>
38 #include <linux/interrupt.h>
39 #include <linux/rcupdate.h>
40 #include <linux/ipmi.h>
41 #include <linux/ipmi_smi.h>
42 #include "ipmi_si.h"
43 #include "ipmi_si_sm.h"
44 #include <linux/string.h>
45 #include <linux/ctype.h>
46 
47 /* Measure times between events in the driver. */
48 #undef DEBUG_TIMING
49 
50 /* Call every 10 ms. */
51 #define SI_TIMEOUT_TIME_USEC	10000
52 #define SI_USEC_PER_JIFFY	(1000000/HZ)
53 #define SI_TIMEOUT_JIFFIES	(SI_TIMEOUT_TIME_USEC/SI_USEC_PER_JIFFY)
54 #define SI_SHORT_TIMEOUT_USEC  250 /* .25ms when the SM request a
55 				      short timeout */
56 
57 enum si_intf_state {
58 	SI_NORMAL,
59 	SI_GETTING_FLAGS,
60 	SI_GETTING_EVENTS,
61 	SI_CLEARING_FLAGS,
62 	SI_GETTING_MESSAGES,
63 	SI_CHECKING_ENABLES,
64 	SI_SETTING_ENABLES
65 	/* FIXME - add watchdog stuff. */
66 };
67 
68 /* Some BT-specific defines we need here. */
69 #define IPMI_BT_INTMASK_REG		2
70 #define IPMI_BT_INTMASK_CLEAR_IRQ_BIT	2
71 #define IPMI_BT_INTMASK_ENABLE_IRQ_BIT	1
72 
73 static const char * const si_to_str[] = { "invalid", "kcs", "smic", "bt" };
74 
75 static bool initialized;
76 
77 /*
78  * Indexes into stats[] in smi_info below.
79  */
80 enum si_stat_indexes {
81 	/*
82 	 * Number of times the driver requested a timer while an operation
83 	 * was in progress.
84 	 */
85 	SI_STAT_short_timeouts = 0,
86 
87 	/*
88 	 * Number of times the driver requested a timer while nothing was in
89 	 * progress.
90 	 */
91 	SI_STAT_long_timeouts,
92 
93 	/* Number of times the interface was idle while being polled. */
94 	SI_STAT_idles,
95 
96 	/* Number of interrupts the driver handled. */
97 	SI_STAT_interrupts,
98 
99 	/* Number of time the driver got an ATTN from the hardware. */
100 	SI_STAT_attentions,
101 
102 	/* Number of times the driver requested flags from the hardware. */
103 	SI_STAT_flag_fetches,
104 
105 	/* Number of times the hardware didn't follow the state machine. */
106 	SI_STAT_hosed_count,
107 
108 	/* Number of completed messages. */
109 	SI_STAT_complete_transactions,
110 
111 	/* Number of IPMI events received from the hardware. */
112 	SI_STAT_events,
113 
114 	/* Number of watchdog pretimeouts. */
115 	SI_STAT_watchdog_pretimeouts,
116 
117 	/* Number of asynchronous messages received. */
118 	SI_STAT_incoming_messages,
119 
120 
121 	/* This *must* remain last, add new values above this. */
122 	SI_NUM_STATS
123 };
124 
125 struct smi_info {
126 	int                    si_num;
127 	struct ipmi_smi        *intf;
128 	struct si_sm_data      *si_sm;
129 	const struct si_sm_handlers *handlers;
130 	spinlock_t             si_lock;
131 	struct ipmi_smi_msg    *waiting_msg;
132 	struct ipmi_smi_msg    *curr_msg;
133 	enum si_intf_state     si_state;
134 
135 	/*
136 	 * Used to handle the various types of I/O that can occur with
137 	 * IPMI
138 	 */
139 	struct si_sm_io io;
140 
141 	/*
142 	 * Per-OEM handler, called from handle_flags().  Returns 1
143 	 * when handle_flags() needs to be re-run or 0 indicating it
144 	 * set si_state itself.
145 	 */
146 	int (*oem_data_avail_handler)(struct smi_info *smi_info);
147 
148 	/*
149 	 * Flags from the last GET_MSG_FLAGS command, used when an ATTN
150 	 * is set to hold the flags until we are done handling everything
151 	 * from the flags.
152 	 */
153 #define RECEIVE_MSG_AVAIL	0x01
154 #define EVENT_MSG_BUFFER_FULL	0x02
155 #define WDT_PRE_TIMEOUT_INT	0x08
156 #define OEM0_DATA_AVAIL     0x20
157 #define OEM1_DATA_AVAIL     0x40
158 #define OEM2_DATA_AVAIL     0x80
159 #define OEM_DATA_AVAIL      (OEM0_DATA_AVAIL | \
160 			     OEM1_DATA_AVAIL | \
161 			     OEM2_DATA_AVAIL)
162 	unsigned char       msg_flags;
163 
164 	/* Does the BMC have an event buffer? */
165 	bool		    has_event_buffer;
166 
167 	/*
168 	 * If set to true, this will request events the next time the
169 	 * state machine is idle.
170 	 */
171 	atomic_t            req_events;
172 
173 	/*
174 	 * If true, run the state machine to completion on every send
175 	 * call.  Generally used after a panic to make sure stuff goes
176 	 * out.
177 	 */
178 	bool                run_to_completion;
179 
180 	/* The timer for this si. */
181 	struct timer_list   si_timer;
182 
183 	/* This flag is set, if the timer can be set */
184 	bool		    timer_can_start;
185 
186 	/* This flag is set, if the timer is running (timer_pending() isn't enough) */
187 	bool		    timer_running;
188 
189 	/* The time (in jiffies) the last timeout occurred at. */
190 	unsigned long       last_timeout_jiffies;
191 
192 	/* Are we waiting for the events, pretimeouts, received msgs? */
193 	atomic_t            need_watch;
194 
195 	/*
196 	 * The driver will disable interrupts when it gets into a
197 	 * situation where it cannot handle messages due to lack of
198 	 * memory.  Once that situation clears up, it will re-enable
199 	 * interrupts.
200 	 */
201 	bool interrupt_disabled;
202 
203 	/*
204 	 * Does the BMC support events?
205 	 */
206 	bool supports_event_msg_buff;
207 
208 	/*
209 	 * Can we disable interrupts the global enables receive irq
210 	 * bit?  There are currently two forms of brokenness, some
211 	 * systems cannot disable the bit (which is technically within
212 	 * the spec but a bad idea) and some systems have the bit
213 	 * forced to zero even though interrupts work (which is
214 	 * clearly outside the spec).  The next bool tells which form
215 	 * of brokenness is present.
216 	 */
217 	bool cannot_disable_irq;
218 
219 	/*
220 	 * Some systems are broken and cannot set the irq enable
221 	 * bit, even if they support interrupts.
222 	 */
223 	bool irq_enable_broken;
224 
225 	/* Is the driver in maintenance mode? */
226 	bool in_maintenance_mode;
227 
228 	/*
229 	 * Did we get an attention that we did not handle?
230 	 */
231 	bool got_attn;
232 
233 	/* From the get device id response... */
234 	struct ipmi_device_id device_id;
235 
236 	/* Have we added the device group to the device? */
237 	bool dev_group_added;
238 
239 	/* Counters and things for the proc filesystem. */
240 	atomic_t stats[SI_NUM_STATS];
241 
242 	struct task_struct *thread;
243 
244 	struct list_head link;
245 };
246 
247 #define smi_inc_stat(smi, stat) \
248 	atomic_inc(&(smi)->stats[SI_STAT_ ## stat])
249 #define smi_get_stat(smi, stat) \
250 	((unsigned int) atomic_read(&(smi)->stats[SI_STAT_ ## stat]))
251 
252 #define IPMI_MAX_INTFS 4
253 static int force_kipmid[IPMI_MAX_INTFS];
254 static int num_force_kipmid;
255 
256 static unsigned int kipmid_max_busy_us[IPMI_MAX_INTFS];
257 static int num_max_busy_us;
258 
259 static bool unload_when_empty = true;
260 
261 static int try_smi_init(struct smi_info *smi);
262 static void cleanup_one_si(struct smi_info *smi_info);
263 static void cleanup_ipmi_si(void);
264 
265 #ifdef DEBUG_TIMING
266 void debug_timestamp(char *msg)
267 {
268 	struct timespec64 t;
269 
270 	ktime_get_ts64(&t);
271 	pr_debug("**%s: %lld.%9.9ld\n", msg, t.tv_sec, t.tv_nsec);
272 }
273 #else
274 #define debug_timestamp(x)
275 #endif
276 
277 static ATOMIC_NOTIFIER_HEAD(xaction_notifier_list);
278 static int register_xaction_notifier(struct notifier_block *nb)
279 {
280 	return atomic_notifier_chain_register(&xaction_notifier_list, nb);
281 }
282 
283 static void deliver_recv_msg(struct smi_info *smi_info,
284 			     struct ipmi_smi_msg *msg)
285 {
286 	/* Deliver the message to the upper layer. */
287 	ipmi_smi_msg_received(smi_info->intf, msg);
288 }
289 
290 static void return_hosed_msg(struct smi_info *smi_info, int cCode)
291 {
292 	struct ipmi_smi_msg *msg = smi_info->curr_msg;
293 
294 	if (cCode < 0 || cCode > IPMI_ERR_UNSPECIFIED)
295 		cCode = IPMI_ERR_UNSPECIFIED;
296 	/* else use it as is */
297 
298 	/* Make it a response */
299 	msg->rsp[0] = msg->data[0] | 4;
300 	msg->rsp[1] = msg->data[1];
301 	msg->rsp[2] = cCode;
302 	msg->rsp_size = 3;
303 
304 	smi_info->curr_msg = NULL;
305 	deliver_recv_msg(smi_info, msg);
306 }
307 
308 static enum si_sm_result start_next_msg(struct smi_info *smi_info)
309 {
310 	int              rv;
311 
312 	if (!smi_info->waiting_msg) {
313 		smi_info->curr_msg = NULL;
314 		rv = SI_SM_IDLE;
315 	} else {
316 		int err;
317 
318 		smi_info->curr_msg = smi_info->waiting_msg;
319 		smi_info->waiting_msg = NULL;
320 		debug_timestamp("Start2");
321 		err = atomic_notifier_call_chain(&xaction_notifier_list,
322 				0, smi_info);
323 		if (err & NOTIFY_STOP_MASK) {
324 			rv = SI_SM_CALL_WITHOUT_DELAY;
325 			goto out;
326 		}
327 		err = smi_info->handlers->start_transaction(
328 			smi_info->si_sm,
329 			smi_info->curr_msg->data,
330 			smi_info->curr_msg->data_size);
331 		if (err)
332 			return_hosed_msg(smi_info, err);
333 
334 		rv = SI_SM_CALL_WITHOUT_DELAY;
335 	}
336 out:
337 	return rv;
338 }
339 
340 static void smi_mod_timer(struct smi_info *smi_info, unsigned long new_val)
341 {
342 	if (!smi_info->timer_can_start)
343 		return;
344 	smi_info->last_timeout_jiffies = jiffies;
345 	mod_timer(&smi_info->si_timer, new_val);
346 	smi_info->timer_running = true;
347 }
348 
349 /*
350  * Start a new message and (re)start the timer and thread.
351  */
352 static void start_new_msg(struct smi_info *smi_info, unsigned char *msg,
353 			  unsigned int size)
354 {
355 	smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
356 
357 	if (smi_info->thread)
358 		wake_up_process(smi_info->thread);
359 
360 	smi_info->handlers->start_transaction(smi_info->si_sm, msg, size);
361 }
362 
363 static void start_check_enables(struct smi_info *smi_info)
364 {
365 	unsigned char msg[2];
366 
367 	msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
368 	msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
369 
370 	start_new_msg(smi_info, msg, 2);
371 	smi_info->si_state = SI_CHECKING_ENABLES;
372 }
373 
374 static void start_clear_flags(struct smi_info *smi_info)
375 {
376 	unsigned char msg[3];
377 
378 	/* Make sure the watchdog pre-timeout flag is not set at startup. */
379 	msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
380 	msg[1] = IPMI_CLEAR_MSG_FLAGS_CMD;
381 	msg[2] = WDT_PRE_TIMEOUT_INT;
382 
383 	start_new_msg(smi_info, msg, 3);
384 	smi_info->si_state = SI_CLEARING_FLAGS;
385 }
386 
387 static void start_getting_msg_queue(struct smi_info *smi_info)
388 {
389 	smi_info->curr_msg->data[0] = (IPMI_NETFN_APP_REQUEST << 2);
390 	smi_info->curr_msg->data[1] = IPMI_GET_MSG_CMD;
391 	smi_info->curr_msg->data_size = 2;
392 
393 	start_new_msg(smi_info, smi_info->curr_msg->data,
394 		      smi_info->curr_msg->data_size);
395 	smi_info->si_state = SI_GETTING_MESSAGES;
396 }
397 
398 static void start_getting_events(struct smi_info *smi_info)
399 {
400 	smi_info->curr_msg->data[0] = (IPMI_NETFN_APP_REQUEST << 2);
401 	smi_info->curr_msg->data[1] = IPMI_READ_EVENT_MSG_BUFFER_CMD;
402 	smi_info->curr_msg->data_size = 2;
403 
404 	start_new_msg(smi_info, smi_info->curr_msg->data,
405 		      smi_info->curr_msg->data_size);
406 	smi_info->si_state = SI_GETTING_EVENTS;
407 }
408 
409 /*
410  * When we have a situtaion where we run out of memory and cannot
411  * allocate messages, we just leave them in the BMC and run the system
412  * polled until we can allocate some memory.  Once we have some
413  * memory, we will re-enable the interrupt.
414  *
415  * Note that we cannot just use disable_irq(), since the interrupt may
416  * be shared.
417  */
418 static inline bool disable_si_irq(struct smi_info *smi_info)
419 {
420 	if ((smi_info->io.irq) && (!smi_info->interrupt_disabled)) {
421 		smi_info->interrupt_disabled = true;
422 		start_check_enables(smi_info);
423 		return true;
424 	}
425 	return false;
426 }
427 
428 static inline bool enable_si_irq(struct smi_info *smi_info)
429 {
430 	if ((smi_info->io.irq) && (smi_info->interrupt_disabled)) {
431 		smi_info->interrupt_disabled = false;
432 		start_check_enables(smi_info);
433 		return true;
434 	}
435 	return false;
436 }
437 
438 /*
439  * Allocate a message.  If unable to allocate, start the interrupt
440  * disable process and return NULL.  If able to allocate but
441  * interrupts are disabled, free the message and return NULL after
442  * starting the interrupt enable process.
443  */
444 static struct ipmi_smi_msg *alloc_msg_handle_irq(struct smi_info *smi_info)
445 {
446 	struct ipmi_smi_msg *msg;
447 
448 	msg = ipmi_alloc_smi_msg();
449 	if (!msg) {
450 		if (!disable_si_irq(smi_info))
451 			smi_info->si_state = SI_NORMAL;
452 	} else if (enable_si_irq(smi_info)) {
453 		ipmi_free_smi_msg(msg);
454 		msg = NULL;
455 	}
456 	return msg;
457 }
458 
459 static void handle_flags(struct smi_info *smi_info)
460 {
461 retry:
462 	if (smi_info->msg_flags & WDT_PRE_TIMEOUT_INT) {
463 		/* Watchdog pre-timeout */
464 		smi_inc_stat(smi_info, watchdog_pretimeouts);
465 
466 		start_clear_flags(smi_info);
467 		smi_info->msg_flags &= ~WDT_PRE_TIMEOUT_INT;
468 		ipmi_smi_watchdog_pretimeout(smi_info->intf);
469 	} else if (smi_info->msg_flags & RECEIVE_MSG_AVAIL) {
470 		/* Messages available. */
471 		smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
472 		if (!smi_info->curr_msg)
473 			return;
474 
475 		start_getting_msg_queue(smi_info);
476 	} else if (smi_info->msg_flags & EVENT_MSG_BUFFER_FULL) {
477 		/* Events available. */
478 		smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
479 		if (!smi_info->curr_msg)
480 			return;
481 
482 		start_getting_events(smi_info);
483 	} else if (smi_info->msg_flags & OEM_DATA_AVAIL &&
484 		   smi_info->oem_data_avail_handler) {
485 		if (smi_info->oem_data_avail_handler(smi_info))
486 			goto retry;
487 	} else
488 		smi_info->si_state = SI_NORMAL;
489 }
490 
491 /*
492  * Global enables we care about.
493  */
494 #define GLOBAL_ENABLES_MASK (IPMI_BMC_EVT_MSG_BUFF | IPMI_BMC_RCV_MSG_INTR | \
495 			     IPMI_BMC_EVT_MSG_INTR)
496 
497 static u8 current_global_enables(struct smi_info *smi_info, u8 base,
498 				 bool *irq_on)
499 {
500 	u8 enables = 0;
501 
502 	if (smi_info->supports_event_msg_buff)
503 		enables |= IPMI_BMC_EVT_MSG_BUFF;
504 
505 	if (((smi_info->io.irq && !smi_info->interrupt_disabled) ||
506 	     smi_info->cannot_disable_irq) &&
507 	    !smi_info->irq_enable_broken)
508 		enables |= IPMI_BMC_RCV_MSG_INTR;
509 
510 	if (smi_info->supports_event_msg_buff &&
511 	    smi_info->io.irq && !smi_info->interrupt_disabled &&
512 	    !smi_info->irq_enable_broken)
513 		enables |= IPMI_BMC_EVT_MSG_INTR;
514 
515 	*irq_on = enables & (IPMI_BMC_EVT_MSG_INTR | IPMI_BMC_RCV_MSG_INTR);
516 
517 	return enables;
518 }
519 
520 static void check_bt_irq(struct smi_info *smi_info, bool irq_on)
521 {
522 	u8 irqstate = smi_info->io.inputb(&smi_info->io, IPMI_BT_INTMASK_REG);
523 
524 	irqstate &= IPMI_BT_INTMASK_ENABLE_IRQ_BIT;
525 
526 	if ((bool)irqstate == irq_on)
527 		return;
528 
529 	if (irq_on)
530 		smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG,
531 				     IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
532 	else
533 		smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG, 0);
534 }
535 
536 static void handle_transaction_done(struct smi_info *smi_info)
537 {
538 	struct ipmi_smi_msg *msg;
539 
540 	debug_timestamp("Done");
541 	switch (smi_info->si_state) {
542 	case SI_NORMAL:
543 		if (!smi_info->curr_msg)
544 			break;
545 
546 		smi_info->curr_msg->rsp_size
547 			= smi_info->handlers->get_result(
548 				smi_info->si_sm,
549 				smi_info->curr_msg->rsp,
550 				IPMI_MAX_MSG_LENGTH);
551 
552 		/*
553 		 * Do this here becase deliver_recv_msg() releases the
554 		 * lock, and a new message can be put in during the
555 		 * time the lock is released.
556 		 */
557 		msg = smi_info->curr_msg;
558 		smi_info->curr_msg = NULL;
559 		deliver_recv_msg(smi_info, msg);
560 		break;
561 
562 	case SI_GETTING_FLAGS:
563 	{
564 		unsigned char msg[4];
565 		unsigned int  len;
566 
567 		/* We got the flags from the SMI, now handle them. */
568 		len = smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
569 		if (msg[2] != 0) {
570 			/* Error fetching flags, just give up for now. */
571 			smi_info->si_state = SI_NORMAL;
572 		} else if (len < 4) {
573 			/*
574 			 * Hmm, no flags.  That's technically illegal, but
575 			 * don't use uninitialized data.
576 			 */
577 			smi_info->si_state = SI_NORMAL;
578 		} else {
579 			smi_info->msg_flags = msg[3];
580 			handle_flags(smi_info);
581 		}
582 		break;
583 	}
584 
585 	case SI_CLEARING_FLAGS:
586 	{
587 		unsigned char msg[3];
588 
589 		/* We cleared the flags. */
590 		smi_info->handlers->get_result(smi_info->si_sm, msg, 3);
591 		if (msg[2] != 0) {
592 			/* Error clearing flags */
593 			dev_warn(smi_info->io.dev,
594 				 "Error clearing flags: %2.2x\n", msg[2]);
595 		}
596 		smi_info->si_state = SI_NORMAL;
597 		break;
598 	}
599 
600 	case SI_GETTING_EVENTS:
601 	{
602 		smi_info->curr_msg->rsp_size
603 			= smi_info->handlers->get_result(
604 				smi_info->si_sm,
605 				smi_info->curr_msg->rsp,
606 				IPMI_MAX_MSG_LENGTH);
607 
608 		/*
609 		 * Do this here becase deliver_recv_msg() releases the
610 		 * lock, and a new message can be put in during the
611 		 * time the lock is released.
612 		 */
613 		msg = smi_info->curr_msg;
614 		smi_info->curr_msg = NULL;
615 		if (msg->rsp[2] != 0) {
616 			/* Error getting event, probably done. */
617 			msg->done(msg);
618 
619 			/* Take off the event flag. */
620 			smi_info->msg_flags &= ~EVENT_MSG_BUFFER_FULL;
621 			handle_flags(smi_info);
622 		} else {
623 			smi_inc_stat(smi_info, events);
624 
625 			/*
626 			 * Do this before we deliver the message
627 			 * because delivering the message releases the
628 			 * lock and something else can mess with the
629 			 * state.
630 			 */
631 			handle_flags(smi_info);
632 
633 			deliver_recv_msg(smi_info, msg);
634 		}
635 		break;
636 	}
637 
638 	case SI_GETTING_MESSAGES:
639 	{
640 		smi_info->curr_msg->rsp_size
641 			= smi_info->handlers->get_result(
642 				smi_info->si_sm,
643 				smi_info->curr_msg->rsp,
644 				IPMI_MAX_MSG_LENGTH);
645 
646 		/*
647 		 * Do this here becase deliver_recv_msg() releases the
648 		 * lock, and a new message can be put in during the
649 		 * time the lock is released.
650 		 */
651 		msg = smi_info->curr_msg;
652 		smi_info->curr_msg = NULL;
653 		if (msg->rsp[2] != 0) {
654 			/* Error getting event, probably done. */
655 			msg->done(msg);
656 
657 			/* Take off the msg flag. */
658 			smi_info->msg_flags &= ~RECEIVE_MSG_AVAIL;
659 			handle_flags(smi_info);
660 		} else {
661 			smi_inc_stat(smi_info, incoming_messages);
662 
663 			/*
664 			 * Do this before we deliver the message
665 			 * because delivering the message releases the
666 			 * lock and something else can mess with the
667 			 * state.
668 			 */
669 			handle_flags(smi_info);
670 
671 			deliver_recv_msg(smi_info, msg);
672 		}
673 		break;
674 	}
675 
676 	case SI_CHECKING_ENABLES:
677 	{
678 		unsigned char msg[4];
679 		u8 enables;
680 		bool irq_on;
681 
682 		/* We got the flags from the SMI, now handle them. */
683 		smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
684 		if (msg[2] != 0) {
685 			dev_warn(smi_info->io.dev,
686 				 "Couldn't get irq info: %x.\n", msg[2]);
687 			dev_warn(smi_info->io.dev,
688 				 "Maybe ok, but ipmi might run very slowly.\n");
689 			smi_info->si_state = SI_NORMAL;
690 			break;
691 		}
692 		enables = current_global_enables(smi_info, 0, &irq_on);
693 		if (smi_info->io.si_type == SI_BT)
694 			/* BT has its own interrupt enable bit. */
695 			check_bt_irq(smi_info, irq_on);
696 		if (enables != (msg[3] & GLOBAL_ENABLES_MASK)) {
697 			/* Enables are not correct, fix them. */
698 			msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
699 			msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
700 			msg[2] = enables | (msg[3] & ~GLOBAL_ENABLES_MASK);
701 			smi_info->handlers->start_transaction(
702 				smi_info->si_sm, msg, 3);
703 			smi_info->si_state = SI_SETTING_ENABLES;
704 		} else if (smi_info->supports_event_msg_buff) {
705 			smi_info->curr_msg = ipmi_alloc_smi_msg();
706 			if (!smi_info->curr_msg) {
707 				smi_info->si_state = SI_NORMAL;
708 				break;
709 			}
710 			start_getting_events(smi_info);
711 		} else {
712 			smi_info->si_state = SI_NORMAL;
713 		}
714 		break;
715 	}
716 
717 	case SI_SETTING_ENABLES:
718 	{
719 		unsigned char msg[4];
720 
721 		smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
722 		if (msg[2] != 0)
723 			dev_warn(smi_info->io.dev,
724 				 "Could not set the global enables: 0x%x.\n",
725 				 msg[2]);
726 
727 		if (smi_info->supports_event_msg_buff) {
728 			smi_info->curr_msg = ipmi_alloc_smi_msg();
729 			if (!smi_info->curr_msg) {
730 				smi_info->si_state = SI_NORMAL;
731 				break;
732 			}
733 			start_getting_events(smi_info);
734 		} else {
735 			smi_info->si_state = SI_NORMAL;
736 		}
737 		break;
738 	}
739 	}
740 }
741 
742 /*
743  * Called on timeouts and events.  Timeouts should pass the elapsed
744  * time, interrupts should pass in zero.  Must be called with
745  * si_lock held and interrupts disabled.
746  */
747 static enum si_sm_result smi_event_handler(struct smi_info *smi_info,
748 					   int time)
749 {
750 	enum si_sm_result si_sm_result;
751 
752 restart:
753 	/*
754 	 * There used to be a loop here that waited a little while
755 	 * (around 25us) before giving up.  That turned out to be
756 	 * pointless, the minimum delays I was seeing were in the 300us
757 	 * range, which is far too long to wait in an interrupt.  So
758 	 * we just run until the state machine tells us something
759 	 * happened or it needs a delay.
760 	 */
761 	si_sm_result = smi_info->handlers->event(smi_info->si_sm, time);
762 	time = 0;
763 	while (si_sm_result == SI_SM_CALL_WITHOUT_DELAY)
764 		si_sm_result = smi_info->handlers->event(smi_info->si_sm, 0);
765 
766 	if (si_sm_result == SI_SM_TRANSACTION_COMPLETE) {
767 		smi_inc_stat(smi_info, complete_transactions);
768 
769 		handle_transaction_done(smi_info);
770 		goto restart;
771 	} else if (si_sm_result == SI_SM_HOSED) {
772 		smi_inc_stat(smi_info, hosed_count);
773 
774 		/*
775 		 * Do the before return_hosed_msg, because that
776 		 * releases the lock.
777 		 */
778 		smi_info->si_state = SI_NORMAL;
779 		if (smi_info->curr_msg != NULL) {
780 			/*
781 			 * If we were handling a user message, format
782 			 * a response to send to the upper layer to
783 			 * tell it about the error.
784 			 */
785 			return_hosed_msg(smi_info, IPMI_ERR_UNSPECIFIED);
786 		}
787 		goto restart;
788 	}
789 
790 	/*
791 	 * We prefer handling attn over new messages.  But don't do
792 	 * this if there is not yet an upper layer to handle anything.
793 	 */
794 	if (si_sm_result == SI_SM_ATTN || smi_info->got_attn) {
795 		unsigned char msg[2];
796 
797 		if (smi_info->si_state != SI_NORMAL) {
798 			/*
799 			 * We got an ATTN, but we are doing something else.
800 			 * Handle the ATTN later.
801 			 */
802 			smi_info->got_attn = true;
803 		} else {
804 			smi_info->got_attn = false;
805 			smi_inc_stat(smi_info, attentions);
806 
807 			/*
808 			 * Got a attn, send down a get message flags to see
809 			 * what's causing it.  It would be better to handle
810 			 * this in the upper layer, but due to the way
811 			 * interrupts work with the SMI, that's not really
812 			 * possible.
813 			 */
814 			msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
815 			msg[1] = IPMI_GET_MSG_FLAGS_CMD;
816 
817 			start_new_msg(smi_info, msg, 2);
818 			smi_info->si_state = SI_GETTING_FLAGS;
819 			goto restart;
820 		}
821 	}
822 
823 	/* If we are currently idle, try to start the next message. */
824 	if (si_sm_result == SI_SM_IDLE) {
825 		smi_inc_stat(smi_info, idles);
826 
827 		si_sm_result = start_next_msg(smi_info);
828 		if (si_sm_result != SI_SM_IDLE)
829 			goto restart;
830 	}
831 
832 	if ((si_sm_result == SI_SM_IDLE)
833 	    && (atomic_read(&smi_info->req_events))) {
834 		/*
835 		 * We are idle and the upper layer requested that I fetch
836 		 * events, so do so.
837 		 */
838 		atomic_set(&smi_info->req_events, 0);
839 
840 		/*
841 		 * Take this opportunity to check the interrupt and
842 		 * message enable state for the BMC.  The BMC can be
843 		 * asynchronously reset, and may thus get interrupts
844 		 * disable and messages disabled.
845 		 */
846 		if (smi_info->supports_event_msg_buff || smi_info->io.irq) {
847 			start_check_enables(smi_info);
848 		} else {
849 			smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
850 			if (!smi_info->curr_msg)
851 				goto out;
852 
853 			start_getting_events(smi_info);
854 		}
855 		goto restart;
856 	}
857 
858 	if (si_sm_result == SI_SM_IDLE && smi_info->timer_running) {
859 		/* Ok it if fails, the timer will just go off. */
860 		if (del_timer(&smi_info->si_timer))
861 			smi_info->timer_running = false;
862 	}
863 
864 out:
865 	return si_sm_result;
866 }
867 
868 static void check_start_timer_thread(struct smi_info *smi_info)
869 {
870 	if (smi_info->si_state == SI_NORMAL && smi_info->curr_msg == NULL) {
871 		smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
872 
873 		if (smi_info->thread)
874 			wake_up_process(smi_info->thread);
875 
876 		start_next_msg(smi_info);
877 		smi_event_handler(smi_info, 0);
878 	}
879 }
880 
881 static void flush_messages(void *send_info)
882 {
883 	struct smi_info *smi_info = send_info;
884 	enum si_sm_result result;
885 
886 	/*
887 	 * Currently, this function is called only in run-to-completion
888 	 * mode.  This means we are single-threaded, no need for locks.
889 	 */
890 	result = smi_event_handler(smi_info, 0);
891 	while (result != SI_SM_IDLE) {
892 		udelay(SI_SHORT_TIMEOUT_USEC);
893 		result = smi_event_handler(smi_info, SI_SHORT_TIMEOUT_USEC);
894 	}
895 }
896 
897 static void sender(void                *send_info,
898 		   struct ipmi_smi_msg *msg)
899 {
900 	struct smi_info   *smi_info = send_info;
901 	unsigned long     flags;
902 
903 	debug_timestamp("Enqueue");
904 
905 	if (smi_info->run_to_completion) {
906 		/*
907 		 * If we are running to completion, start it.  Upper
908 		 * layer will call flush_messages to clear it out.
909 		 */
910 		smi_info->waiting_msg = msg;
911 		return;
912 	}
913 
914 	spin_lock_irqsave(&smi_info->si_lock, flags);
915 	/*
916 	 * The following two lines don't need to be under the lock for
917 	 * the lock's sake, but they do need SMP memory barriers to
918 	 * avoid getting things out of order.  We are already claiming
919 	 * the lock, anyway, so just do it under the lock to avoid the
920 	 * ordering problem.
921 	 */
922 	BUG_ON(smi_info->waiting_msg);
923 	smi_info->waiting_msg = msg;
924 	check_start_timer_thread(smi_info);
925 	spin_unlock_irqrestore(&smi_info->si_lock, flags);
926 }
927 
928 static void set_run_to_completion(void *send_info, bool i_run_to_completion)
929 {
930 	struct smi_info   *smi_info = send_info;
931 
932 	smi_info->run_to_completion = i_run_to_completion;
933 	if (i_run_to_completion)
934 		flush_messages(smi_info);
935 }
936 
937 /*
938  * Use -1 as a special constant to tell that we are spinning in kipmid
939  * looking for something and not delaying between checks
940  */
941 #define IPMI_TIME_NOT_BUSY ns_to_ktime(-1ull)
942 static inline bool ipmi_thread_busy_wait(enum si_sm_result smi_result,
943 					 const struct smi_info *smi_info,
944 					 ktime_t *busy_until)
945 {
946 	unsigned int max_busy_us = 0;
947 
948 	if (smi_info->si_num < num_max_busy_us)
949 		max_busy_us = kipmid_max_busy_us[smi_info->si_num];
950 	if (max_busy_us == 0 || smi_result != SI_SM_CALL_WITH_DELAY)
951 		*busy_until = IPMI_TIME_NOT_BUSY;
952 	else if (*busy_until == IPMI_TIME_NOT_BUSY) {
953 		*busy_until = ktime_get() + max_busy_us * NSEC_PER_USEC;
954 	} else {
955 		if (unlikely(ktime_get() > *busy_until)) {
956 			*busy_until = IPMI_TIME_NOT_BUSY;
957 			return false;
958 		}
959 	}
960 	return true;
961 }
962 
963 
964 /*
965  * A busy-waiting loop for speeding up IPMI operation.
966  *
967  * Lousy hardware makes this hard.  This is only enabled for systems
968  * that are not BT and do not have interrupts.  It starts spinning
969  * when an operation is complete or until max_busy tells it to stop
970  * (if that is enabled).  See the paragraph on kimid_max_busy_us in
971  * Documentation/driver-api/ipmi.rst for details.
972  */
973 static int ipmi_thread(void *data)
974 {
975 	struct smi_info *smi_info = data;
976 	unsigned long flags;
977 	enum si_sm_result smi_result;
978 	ktime_t busy_until = IPMI_TIME_NOT_BUSY;
979 
980 	set_user_nice(current, MAX_NICE);
981 	while (!kthread_should_stop()) {
982 		int busy_wait;
983 
984 		spin_lock_irqsave(&(smi_info->si_lock), flags);
985 		smi_result = smi_event_handler(smi_info, 0);
986 
987 		/*
988 		 * If the driver is doing something, there is a possible
989 		 * race with the timer.  If the timer handler see idle,
990 		 * and the thread here sees something else, the timer
991 		 * handler won't restart the timer even though it is
992 		 * required.  So start it here if necessary.
993 		 */
994 		if (smi_result != SI_SM_IDLE && !smi_info->timer_running)
995 			smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
996 
997 		spin_unlock_irqrestore(&(smi_info->si_lock), flags);
998 		busy_wait = ipmi_thread_busy_wait(smi_result, smi_info,
999 						  &busy_until);
1000 		if (smi_result == SI_SM_CALL_WITHOUT_DELAY) {
1001 			; /* do nothing */
1002 		} else if (smi_result == SI_SM_CALL_WITH_DELAY && busy_wait) {
1003 			/*
1004 			 * In maintenance mode we run as fast as
1005 			 * possible to allow firmware updates to
1006 			 * complete as fast as possible, but normally
1007 			 * don't bang on the scheduler.
1008 			 */
1009 			if (smi_info->in_maintenance_mode)
1010 				schedule();
1011 			else
1012 				usleep_range(100, 200);
1013 		} else if (smi_result == SI_SM_IDLE) {
1014 			if (atomic_read(&smi_info->need_watch)) {
1015 				schedule_timeout_interruptible(100);
1016 			} else {
1017 				/* Wait to be woken up when we are needed. */
1018 				__set_current_state(TASK_INTERRUPTIBLE);
1019 				schedule();
1020 			}
1021 		} else {
1022 			schedule_timeout_interruptible(1);
1023 		}
1024 	}
1025 	return 0;
1026 }
1027 
1028 
1029 static void poll(void *send_info)
1030 {
1031 	struct smi_info *smi_info = send_info;
1032 	unsigned long flags = 0;
1033 	bool run_to_completion = smi_info->run_to_completion;
1034 
1035 	/*
1036 	 * Make sure there is some delay in the poll loop so we can
1037 	 * drive time forward and timeout things.
1038 	 */
1039 	udelay(10);
1040 	if (!run_to_completion)
1041 		spin_lock_irqsave(&smi_info->si_lock, flags);
1042 	smi_event_handler(smi_info, 10);
1043 	if (!run_to_completion)
1044 		spin_unlock_irqrestore(&smi_info->si_lock, flags);
1045 }
1046 
1047 static void request_events(void *send_info)
1048 {
1049 	struct smi_info *smi_info = send_info;
1050 
1051 	if (!smi_info->has_event_buffer)
1052 		return;
1053 
1054 	atomic_set(&smi_info->req_events, 1);
1055 }
1056 
1057 static void set_need_watch(void *send_info, unsigned int watch_mask)
1058 {
1059 	struct smi_info *smi_info = send_info;
1060 	unsigned long flags;
1061 	int enable;
1062 
1063 	enable = !!watch_mask;
1064 
1065 	atomic_set(&smi_info->need_watch, enable);
1066 	spin_lock_irqsave(&smi_info->si_lock, flags);
1067 	check_start_timer_thread(smi_info);
1068 	spin_unlock_irqrestore(&smi_info->si_lock, flags);
1069 }
1070 
1071 static void smi_timeout(struct timer_list *t)
1072 {
1073 	struct smi_info   *smi_info = from_timer(smi_info, t, si_timer);
1074 	enum si_sm_result smi_result;
1075 	unsigned long     flags;
1076 	unsigned long     jiffies_now;
1077 	long              time_diff;
1078 	long		  timeout;
1079 
1080 	spin_lock_irqsave(&(smi_info->si_lock), flags);
1081 	debug_timestamp("Timer");
1082 
1083 	jiffies_now = jiffies;
1084 	time_diff = (((long)jiffies_now - (long)smi_info->last_timeout_jiffies)
1085 		     * SI_USEC_PER_JIFFY);
1086 	smi_result = smi_event_handler(smi_info, time_diff);
1087 
1088 	if ((smi_info->io.irq) && (!smi_info->interrupt_disabled)) {
1089 		/* Running with interrupts, only do long timeouts. */
1090 		timeout = jiffies + SI_TIMEOUT_JIFFIES;
1091 		smi_inc_stat(smi_info, long_timeouts);
1092 		goto do_mod_timer;
1093 	}
1094 
1095 	/*
1096 	 * If the state machine asks for a short delay, then shorten
1097 	 * the timer timeout.
1098 	 */
1099 	if (smi_result == SI_SM_CALL_WITH_DELAY) {
1100 		smi_inc_stat(smi_info, short_timeouts);
1101 		timeout = jiffies + 1;
1102 	} else {
1103 		smi_inc_stat(smi_info, long_timeouts);
1104 		timeout = jiffies + SI_TIMEOUT_JIFFIES;
1105 	}
1106 
1107 do_mod_timer:
1108 	if (smi_result != SI_SM_IDLE)
1109 		smi_mod_timer(smi_info, timeout);
1110 	else
1111 		smi_info->timer_running = false;
1112 	spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1113 }
1114 
1115 irqreturn_t ipmi_si_irq_handler(int irq, void *data)
1116 {
1117 	struct smi_info *smi_info = data;
1118 	unsigned long   flags;
1119 
1120 	if (smi_info->io.si_type == SI_BT)
1121 		/* We need to clear the IRQ flag for the BT interface. */
1122 		smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG,
1123 				     IPMI_BT_INTMASK_CLEAR_IRQ_BIT
1124 				     | IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
1125 
1126 	spin_lock_irqsave(&(smi_info->si_lock), flags);
1127 
1128 	smi_inc_stat(smi_info, interrupts);
1129 
1130 	debug_timestamp("Interrupt");
1131 
1132 	smi_event_handler(smi_info, 0);
1133 	spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1134 	return IRQ_HANDLED;
1135 }
1136 
1137 static int smi_start_processing(void            *send_info,
1138 				struct ipmi_smi *intf)
1139 {
1140 	struct smi_info *new_smi = send_info;
1141 	int             enable = 0;
1142 
1143 	new_smi->intf = intf;
1144 
1145 	/* Set up the timer that drives the interface. */
1146 	timer_setup(&new_smi->si_timer, smi_timeout, 0);
1147 	new_smi->timer_can_start = true;
1148 	smi_mod_timer(new_smi, jiffies + SI_TIMEOUT_JIFFIES);
1149 
1150 	/* Try to claim any interrupts. */
1151 	if (new_smi->io.irq_setup) {
1152 		new_smi->io.irq_handler_data = new_smi;
1153 		new_smi->io.irq_setup(&new_smi->io);
1154 	}
1155 
1156 	/*
1157 	 * Check if the user forcefully enabled the daemon.
1158 	 */
1159 	if (new_smi->si_num < num_force_kipmid)
1160 		enable = force_kipmid[new_smi->si_num];
1161 	/*
1162 	 * The BT interface is efficient enough to not need a thread,
1163 	 * and there is no need for a thread if we have interrupts.
1164 	 */
1165 	else if ((new_smi->io.si_type != SI_BT) && (!new_smi->io.irq))
1166 		enable = 1;
1167 
1168 	if (enable) {
1169 		new_smi->thread = kthread_run(ipmi_thread, new_smi,
1170 					      "kipmi%d", new_smi->si_num);
1171 		if (IS_ERR(new_smi->thread)) {
1172 			dev_notice(new_smi->io.dev, "Could not start"
1173 				   " kernel thread due to error %ld, only using"
1174 				   " timers to drive the interface\n",
1175 				   PTR_ERR(new_smi->thread));
1176 			new_smi->thread = NULL;
1177 		}
1178 	}
1179 
1180 	return 0;
1181 }
1182 
1183 static int get_smi_info(void *send_info, struct ipmi_smi_info *data)
1184 {
1185 	struct smi_info *smi = send_info;
1186 
1187 	data->addr_src = smi->io.addr_source;
1188 	data->dev = smi->io.dev;
1189 	data->addr_info = smi->io.addr_info;
1190 	get_device(smi->io.dev);
1191 
1192 	return 0;
1193 }
1194 
1195 static void set_maintenance_mode(void *send_info, bool enable)
1196 {
1197 	struct smi_info   *smi_info = send_info;
1198 
1199 	if (!enable)
1200 		atomic_set(&smi_info->req_events, 0);
1201 	smi_info->in_maintenance_mode = enable;
1202 }
1203 
1204 static void shutdown_smi(void *send_info);
1205 static const struct ipmi_smi_handlers handlers = {
1206 	.owner                  = THIS_MODULE,
1207 	.start_processing       = smi_start_processing,
1208 	.shutdown               = shutdown_smi,
1209 	.get_smi_info		= get_smi_info,
1210 	.sender			= sender,
1211 	.request_events		= request_events,
1212 	.set_need_watch		= set_need_watch,
1213 	.set_maintenance_mode   = set_maintenance_mode,
1214 	.set_run_to_completion  = set_run_to_completion,
1215 	.flush_messages		= flush_messages,
1216 	.poll			= poll,
1217 };
1218 
1219 static LIST_HEAD(smi_infos);
1220 static DEFINE_MUTEX(smi_infos_lock);
1221 static int smi_num; /* Used to sequence the SMIs */
1222 
1223 static const char * const addr_space_to_str[] = { "i/o", "mem" };
1224 
1225 module_param_array(force_kipmid, int, &num_force_kipmid, 0);
1226 MODULE_PARM_DESC(force_kipmid, "Force the kipmi daemon to be enabled (1) or"
1227 		 " disabled(0).  Normally the IPMI driver auto-detects"
1228 		 " this, but the value may be overridden by this parm.");
1229 module_param(unload_when_empty, bool, 0);
1230 MODULE_PARM_DESC(unload_when_empty, "Unload the module if no interfaces are"
1231 		 " specified or found, default is 1.  Setting to 0"
1232 		 " is useful for hot add of devices using hotmod.");
1233 module_param_array(kipmid_max_busy_us, uint, &num_max_busy_us, 0644);
1234 MODULE_PARM_DESC(kipmid_max_busy_us,
1235 		 "Max time (in microseconds) to busy-wait for IPMI data before"
1236 		 " sleeping. 0 (default) means to wait forever. Set to 100-500"
1237 		 " if kipmid is using up a lot of CPU time.");
1238 
1239 void ipmi_irq_finish_setup(struct si_sm_io *io)
1240 {
1241 	if (io->si_type == SI_BT)
1242 		/* Enable the interrupt in the BT interface. */
1243 		io->outputb(io, IPMI_BT_INTMASK_REG,
1244 			    IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
1245 }
1246 
1247 void ipmi_irq_start_cleanup(struct si_sm_io *io)
1248 {
1249 	if (io->si_type == SI_BT)
1250 		/* Disable the interrupt in the BT interface. */
1251 		io->outputb(io, IPMI_BT_INTMASK_REG, 0);
1252 }
1253 
1254 static void std_irq_cleanup(struct si_sm_io *io)
1255 {
1256 	ipmi_irq_start_cleanup(io);
1257 	free_irq(io->irq, io->irq_handler_data);
1258 }
1259 
1260 int ipmi_std_irq_setup(struct si_sm_io *io)
1261 {
1262 	int rv;
1263 
1264 	if (!io->irq)
1265 		return 0;
1266 
1267 	rv = request_irq(io->irq,
1268 			 ipmi_si_irq_handler,
1269 			 IRQF_SHARED,
1270 			 SI_DEVICE_NAME,
1271 			 io->irq_handler_data);
1272 	if (rv) {
1273 		dev_warn(io->dev, "%s unable to claim interrupt %d,"
1274 			 " running polled\n",
1275 			 SI_DEVICE_NAME, io->irq);
1276 		io->irq = 0;
1277 	} else {
1278 		io->irq_cleanup = std_irq_cleanup;
1279 		ipmi_irq_finish_setup(io);
1280 		dev_info(io->dev, "Using irq %d\n", io->irq);
1281 	}
1282 
1283 	return rv;
1284 }
1285 
1286 static int wait_for_msg_done(struct smi_info *smi_info)
1287 {
1288 	enum si_sm_result     smi_result;
1289 
1290 	smi_result = smi_info->handlers->event(smi_info->si_sm, 0);
1291 	for (;;) {
1292 		if (smi_result == SI_SM_CALL_WITH_DELAY ||
1293 		    smi_result == SI_SM_CALL_WITH_TICK_DELAY) {
1294 			schedule_timeout_uninterruptible(1);
1295 			smi_result = smi_info->handlers->event(
1296 				smi_info->si_sm, jiffies_to_usecs(1));
1297 		} else if (smi_result == SI_SM_CALL_WITHOUT_DELAY) {
1298 			smi_result = smi_info->handlers->event(
1299 				smi_info->si_sm, 0);
1300 		} else
1301 			break;
1302 	}
1303 	if (smi_result == SI_SM_HOSED)
1304 		/*
1305 		 * We couldn't get the state machine to run, so whatever's at
1306 		 * the port is probably not an IPMI SMI interface.
1307 		 */
1308 		return -ENODEV;
1309 
1310 	return 0;
1311 }
1312 
1313 static int try_get_dev_id(struct smi_info *smi_info)
1314 {
1315 	unsigned char         msg[2];
1316 	unsigned char         *resp;
1317 	unsigned long         resp_len;
1318 	int                   rv = 0;
1319 	unsigned int          retry_count = 0;
1320 
1321 	resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1322 	if (!resp)
1323 		return -ENOMEM;
1324 
1325 	/*
1326 	 * Do a Get Device ID command, since it comes back with some
1327 	 * useful info.
1328 	 */
1329 	msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1330 	msg[1] = IPMI_GET_DEVICE_ID_CMD;
1331 
1332 retry:
1333 	smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1334 
1335 	rv = wait_for_msg_done(smi_info);
1336 	if (rv)
1337 		goto out;
1338 
1339 	resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1340 						  resp, IPMI_MAX_MSG_LENGTH);
1341 
1342 	/* Check and record info from the get device id, in case we need it. */
1343 	rv = ipmi_demangle_device_id(resp[0] >> 2, resp[1],
1344 			resp + 2, resp_len - 2, &smi_info->device_id);
1345 	if (rv) {
1346 		/* record completion code */
1347 		unsigned char cc = *(resp + 2);
1348 
1349 		if ((cc == IPMI_DEVICE_IN_FW_UPDATE_ERR
1350 		    || cc == IPMI_DEVICE_IN_INIT_ERR
1351 		    || cc == IPMI_NOT_IN_MY_STATE_ERR)
1352 		    && ++retry_count <= GET_DEVICE_ID_MAX_RETRY) {
1353 			dev_warn(smi_info->io.dev,
1354 			    "BMC returned 0x%2.2x, retry get bmc device id\n",
1355 			    cc);
1356 			goto retry;
1357 		}
1358 	}
1359 
1360 out:
1361 	kfree(resp);
1362 	return rv;
1363 }
1364 
1365 static int get_global_enables(struct smi_info *smi_info, u8 *enables)
1366 {
1367 	unsigned char         msg[3];
1368 	unsigned char         *resp;
1369 	unsigned long         resp_len;
1370 	int                   rv;
1371 
1372 	resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1373 	if (!resp)
1374 		return -ENOMEM;
1375 
1376 	msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1377 	msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
1378 	smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1379 
1380 	rv = wait_for_msg_done(smi_info);
1381 	if (rv) {
1382 		dev_warn(smi_info->io.dev,
1383 			 "Error getting response from get global enables command: %d\n",
1384 			 rv);
1385 		goto out;
1386 	}
1387 
1388 	resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1389 						  resp, IPMI_MAX_MSG_LENGTH);
1390 
1391 	if (resp_len < 4 ||
1392 			resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1393 			resp[1] != IPMI_GET_BMC_GLOBAL_ENABLES_CMD   ||
1394 			resp[2] != 0) {
1395 		dev_warn(smi_info->io.dev,
1396 			 "Invalid return from get global enables command: %ld %x %x %x\n",
1397 			 resp_len, resp[0], resp[1], resp[2]);
1398 		rv = -EINVAL;
1399 		goto out;
1400 	} else {
1401 		*enables = resp[3];
1402 	}
1403 
1404 out:
1405 	kfree(resp);
1406 	return rv;
1407 }
1408 
1409 /*
1410  * Returns 1 if it gets an error from the command.
1411  */
1412 static int set_global_enables(struct smi_info *smi_info, u8 enables)
1413 {
1414 	unsigned char         msg[3];
1415 	unsigned char         *resp;
1416 	unsigned long         resp_len;
1417 	int                   rv;
1418 
1419 	resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1420 	if (!resp)
1421 		return -ENOMEM;
1422 
1423 	msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1424 	msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
1425 	msg[2] = enables;
1426 	smi_info->handlers->start_transaction(smi_info->si_sm, msg, 3);
1427 
1428 	rv = wait_for_msg_done(smi_info);
1429 	if (rv) {
1430 		dev_warn(smi_info->io.dev,
1431 			 "Error getting response from set global enables command: %d\n",
1432 			 rv);
1433 		goto out;
1434 	}
1435 
1436 	resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1437 						  resp, IPMI_MAX_MSG_LENGTH);
1438 
1439 	if (resp_len < 3 ||
1440 			resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1441 			resp[1] != IPMI_SET_BMC_GLOBAL_ENABLES_CMD) {
1442 		dev_warn(smi_info->io.dev,
1443 			 "Invalid return from set global enables command: %ld %x %x\n",
1444 			 resp_len, resp[0], resp[1]);
1445 		rv = -EINVAL;
1446 		goto out;
1447 	}
1448 
1449 	if (resp[2] != 0)
1450 		rv = 1;
1451 
1452 out:
1453 	kfree(resp);
1454 	return rv;
1455 }
1456 
1457 /*
1458  * Some BMCs do not support clearing the receive irq bit in the global
1459  * enables (even if they don't support interrupts on the BMC).  Check
1460  * for this and handle it properly.
1461  */
1462 static void check_clr_rcv_irq(struct smi_info *smi_info)
1463 {
1464 	u8 enables = 0;
1465 	int rv;
1466 
1467 	rv = get_global_enables(smi_info, &enables);
1468 	if (!rv) {
1469 		if ((enables & IPMI_BMC_RCV_MSG_INTR) == 0)
1470 			/* Already clear, should work ok. */
1471 			return;
1472 
1473 		enables &= ~IPMI_BMC_RCV_MSG_INTR;
1474 		rv = set_global_enables(smi_info, enables);
1475 	}
1476 
1477 	if (rv < 0) {
1478 		dev_err(smi_info->io.dev,
1479 			"Cannot check clearing the rcv irq: %d\n", rv);
1480 		return;
1481 	}
1482 
1483 	if (rv) {
1484 		/*
1485 		 * An error when setting the event buffer bit means
1486 		 * clearing the bit is not supported.
1487 		 */
1488 		dev_warn(smi_info->io.dev,
1489 			 "The BMC does not support clearing the recv irq bit, compensating, but the BMC needs to be fixed.\n");
1490 		smi_info->cannot_disable_irq = true;
1491 	}
1492 }
1493 
1494 /*
1495  * Some BMCs do not support setting the interrupt bits in the global
1496  * enables even if they support interrupts.  Clearly bad, but we can
1497  * compensate.
1498  */
1499 static void check_set_rcv_irq(struct smi_info *smi_info)
1500 {
1501 	u8 enables = 0;
1502 	int rv;
1503 
1504 	if (!smi_info->io.irq)
1505 		return;
1506 
1507 	rv = get_global_enables(smi_info, &enables);
1508 	if (!rv) {
1509 		enables |= IPMI_BMC_RCV_MSG_INTR;
1510 		rv = set_global_enables(smi_info, enables);
1511 	}
1512 
1513 	if (rv < 0) {
1514 		dev_err(smi_info->io.dev,
1515 			"Cannot check setting the rcv irq: %d\n", rv);
1516 		return;
1517 	}
1518 
1519 	if (rv) {
1520 		/*
1521 		 * An error when setting the event buffer bit means
1522 		 * setting the bit is not supported.
1523 		 */
1524 		dev_warn(smi_info->io.dev,
1525 			 "The BMC does not support setting the recv irq bit, compensating, but the BMC needs to be fixed.\n");
1526 		smi_info->cannot_disable_irq = true;
1527 		smi_info->irq_enable_broken = true;
1528 	}
1529 }
1530 
1531 static int try_enable_event_buffer(struct smi_info *smi_info)
1532 {
1533 	unsigned char         msg[3];
1534 	unsigned char         *resp;
1535 	unsigned long         resp_len;
1536 	int                   rv = 0;
1537 
1538 	resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1539 	if (!resp)
1540 		return -ENOMEM;
1541 
1542 	msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1543 	msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
1544 	smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1545 
1546 	rv = wait_for_msg_done(smi_info);
1547 	if (rv) {
1548 		pr_warn("Error getting response from get global enables command, the event buffer is not enabled\n");
1549 		goto out;
1550 	}
1551 
1552 	resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1553 						  resp, IPMI_MAX_MSG_LENGTH);
1554 
1555 	if (resp_len < 4 ||
1556 			resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1557 			resp[1] != IPMI_GET_BMC_GLOBAL_ENABLES_CMD   ||
1558 			resp[2] != 0) {
1559 		pr_warn("Invalid return from get global enables command, cannot enable the event buffer\n");
1560 		rv = -EINVAL;
1561 		goto out;
1562 	}
1563 
1564 	if (resp[3] & IPMI_BMC_EVT_MSG_BUFF) {
1565 		/* buffer is already enabled, nothing to do. */
1566 		smi_info->supports_event_msg_buff = true;
1567 		goto out;
1568 	}
1569 
1570 	msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1571 	msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
1572 	msg[2] = resp[3] | IPMI_BMC_EVT_MSG_BUFF;
1573 	smi_info->handlers->start_transaction(smi_info->si_sm, msg, 3);
1574 
1575 	rv = wait_for_msg_done(smi_info);
1576 	if (rv) {
1577 		pr_warn("Error getting response from set global, enables command, the event buffer is not enabled\n");
1578 		goto out;
1579 	}
1580 
1581 	resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1582 						  resp, IPMI_MAX_MSG_LENGTH);
1583 
1584 	if (resp_len < 3 ||
1585 			resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1586 			resp[1] != IPMI_SET_BMC_GLOBAL_ENABLES_CMD) {
1587 		pr_warn("Invalid return from get global, enables command, not enable the event buffer\n");
1588 		rv = -EINVAL;
1589 		goto out;
1590 	}
1591 
1592 	if (resp[2] != 0)
1593 		/*
1594 		 * An error when setting the event buffer bit means
1595 		 * that the event buffer is not supported.
1596 		 */
1597 		rv = -ENOENT;
1598 	else
1599 		smi_info->supports_event_msg_buff = true;
1600 
1601 out:
1602 	kfree(resp);
1603 	return rv;
1604 }
1605 
1606 #define IPMI_SI_ATTR(name) \
1607 static ssize_t name##_show(struct device *dev,			\
1608 			   struct device_attribute *attr,		\
1609 			   char *buf)					\
1610 {									\
1611 	struct smi_info *smi_info = dev_get_drvdata(dev);		\
1612 									\
1613 	return snprintf(buf, 10, "%u\n", smi_get_stat(smi_info, name));	\
1614 }									\
1615 static DEVICE_ATTR(name, 0444, name##_show, NULL)
1616 
1617 static ssize_t type_show(struct device *dev,
1618 			 struct device_attribute *attr,
1619 			 char *buf)
1620 {
1621 	struct smi_info *smi_info = dev_get_drvdata(dev);
1622 
1623 	return snprintf(buf, 10, "%s\n", si_to_str[smi_info->io.si_type]);
1624 }
1625 static DEVICE_ATTR(type, 0444, type_show, NULL);
1626 
1627 static ssize_t interrupts_enabled_show(struct device *dev,
1628 				       struct device_attribute *attr,
1629 				       char *buf)
1630 {
1631 	struct smi_info *smi_info = dev_get_drvdata(dev);
1632 	int enabled = smi_info->io.irq && !smi_info->interrupt_disabled;
1633 
1634 	return snprintf(buf, 10, "%d\n", enabled);
1635 }
1636 static DEVICE_ATTR(interrupts_enabled, 0444,
1637 		   interrupts_enabled_show, NULL);
1638 
1639 IPMI_SI_ATTR(short_timeouts);
1640 IPMI_SI_ATTR(long_timeouts);
1641 IPMI_SI_ATTR(idles);
1642 IPMI_SI_ATTR(interrupts);
1643 IPMI_SI_ATTR(attentions);
1644 IPMI_SI_ATTR(flag_fetches);
1645 IPMI_SI_ATTR(hosed_count);
1646 IPMI_SI_ATTR(complete_transactions);
1647 IPMI_SI_ATTR(events);
1648 IPMI_SI_ATTR(watchdog_pretimeouts);
1649 IPMI_SI_ATTR(incoming_messages);
1650 
1651 static ssize_t params_show(struct device *dev,
1652 			   struct device_attribute *attr,
1653 			   char *buf)
1654 {
1655 	struct smi_info *smi_info = dev_get_drvdata(dev);
1656 
1657 	return snprintf(buf, 200,
1658 			"%s,%s,0x%lx,rsp=%d,rsi=%d,rsh=%d,irq=%d,ipmb=%d\n",
1659 			si_to_str[smi_info->io.si_type],
1660 			addr_space_to_str[smi_info->io.addr_space],
1661 			smi_info->io.addr_data,
1662 			smi_info->io.regspacing,
1663 			smi_info->io.regsize,
1664 			smi_info->io.regshift,
1665 			smi_info->io.irq,
1666 			smi_info->io.slave_addr);
1667 }
1668 static DEVICE_ATTR(params, 0444, params_show, NULL);
1669 
1670 static struct attribute *ipmi_si_dev_attrs[] = {
1671 	&dev_attr_type.attr,
1672 	&dev_attr_interrupts_enabled.attr,
1673 	&dev_attr_short_timeouts.attr,
1674 	&dev_attr_long_timeouts.attr,
1675 	&dev_attr_idles.attr,
1676 	&dev_attr_interrupts.attr,
1677 	&dev_attr_attentions.attr,
1678 	&dev_attr_flag_fetches.attr,
1679 	&dev_attr_hosed_count.attr,
1680 	&dev_attr_complete_transactions.attr,
1681 	&dev_attr_events.attr,
1682 	&dev_attr_watchdog_pretimeouts.attr,
1683 	&dev_attr_incoming_messages.attr,
1684 	&dev_attr_params.attr,
1685 	NULL
1686 };
1687 
1688 static const struct attribute_group ipmi_si_dev_attr_group = {
1689 	.attrs		= ipmi_si_dev_attrs,
1690 };
1691 
1692 /*
1693  * oem_data_avail_to_receive_msg_avail
1694  * @info - smi_info structure with msg_flags set
1695  *
1696  * Converts flags from OEM_DATA_AVAIL to RECEIVE_MSG_AVAIL
1697  * Returns 1 indicating need to re-run handle_flags().
1698  */
1699 static int oem_data_avail_to_receive_msg_avail(struct smi_info *smi_info)
1700 {
1701 	smi_info->msg_flags = ((smi_info->msg_flags & ~OEM_DATA_AVAIL) |
1702 			       RECEIVE_MSG_AVAIL);
1703 	return 1;
1704 }
1705 
1706 /*
1707  * setup_dell_poweredge_oem_data_handler
1708  * @info - smi_info.device_id must be populated
1709  *
1710  * Systems that match, but have firmware version < 1.40 may assert
1711  * OEM0_DATA_AVAIL on their own, without being told via Set Flags that
1712  * it's safe to do so.  Such systems will de-assert OEM1_DATA_AVAIL
1713  * upon receipt of IPMI_GET_MSG_CMD, so we should treat these flags
1714  * as RECEIVE_MSG_AVAIL instead.
1715  *
1716  * As Dell has no plans to release IPMI 1.5 firmware that *ever*
1717  * assert the OEM[012] bits, and if it did, the driver would have to
1718  * change to handle that properly, we don't actually check for the
1719  * firmware version.
1720  * Device ID = 0x20                BMC on PowerEdge 8G servers
1721  * Device Revision = 0x80
1722  * Firmware Revision1 = 0x01       BMC version 1.40
1723  * Firmware Revision2 = 0x40       BCD encoded
1724  * IPMI Version = 0x51             IPMI 1.5
1725  * Manufacturer ID = A2 02 00      Dell IANA
1726  *
1727  * Additionally, PowerEdge systems with IPMI < 1.5 may also assert
1728  * OEM0_DATA_AVAIL and needs to be treated as RECEIVE_MSG_AVAIL.
1729  *
1730  */
1731 #define DELL_POWEREDGE_8G_BMC_DEVICE_ID  0x20
1732 #define DELL_POWEREDGE_8G_BMC_DEVICE_REV 0x80
1733 #define DELL_POWEREDGE_8G_BMC_IPMI_VERSION 0x51
1734 #define DELL_IANA_MFR_ID 0x0002a2
1735 static void setup_dell_poweredge_oem_data_handler(struct smi_info *smi_info)
1736 {
1737 	struct ipmi_device_id *id = &smi_info->device_id;
1738 	if (id->manufacturer_id == DELL_IANA_MFR_ID) {
1739 		if (id->device_id       == DELL_POWEREDGE_8G_BMC_DEVICE_ID  &&
1740 		    id->device_revision == DELL_POWEREDGE_8G_BMC_DEVICE_REV &&
1741 		    id->ipmi_version   == DELL_POWEREDGE_8G_BMC_IPMI_VERSION) {
1742 			smi_info->oem_data_avail_handler =
1743 				oem_data_avail_to_receive_msg_avail;
1744 		} else if (ipmi_version_major(id) < 1 ||
1745 			   (ipmi_version_major(id) == 1 &&
1746 			    ipmi_version_minor(id) < 5)) {
1747 			smi_info->oem_data_avail_handler =
1748 				oem_data_avail_to_receive_msg_avail;
1749 		}
1750 	}
1751 }
1752 
1753 #define CANNOT_RETURN_REQUESTED_LENGTH 0xCA
1754 static void return_hosed_msg_badsize(struct smi_info *smi_info)
1755 {
1756 	struct ipmi_smi_msg *msg = smi_info->curr_msg;
1757 
1758 	/* Make it a response */
1759 	msg->rsp[0] = msg->data[0] | 4;
1760 	msg->rsp[1] = msg->data[1];
1761 	msg->rsp[2] = CANNOT_RETURN_REQUESTED_LENGTH;
1762 	msg->rsp_size = 3;
1763 	smi_info->curr_msg = NULL;
1764 	deliver_recv_msg(smi_info, msg);
1765 }
1766 
1767 /*
1768  * dell_poweredge_bt_xaction_handler
1769  * @info - smi_info.device_id must be populated
1770  *
1771  * Dell PowerEdge servers with the BT interface (x6xx and 1750) will
1772  * not respond to a Get SDR command if the length of the data
1773  * requested is exactly 0x3A, which leads to command timeouts and no
1774  * data returned.  This intercepts such commands, and causes userspace
1775  * callers to try again with a different-sized buffer, which succeeds.
1776  */
1777 
1778 #define STORAGE_NETFN 0x0A
1779 #define STORAGE_CMD_GET_SDR 0x23
1780 static int dell_poweredge_bt_xaction_handler(struct notifier_block *self,
1781 					     unsigned long unused,
1782 					     void *in)
1783 {
1784 	struct smi_info *smi_info = in;
1785 	unsigned char *data = smi_info->curr_msg->data;
1786 	unsigned int size   = smi_info->curr_msg->data_size;
1787 	if (size >= 8 &&
1788 	    (data[0]>>2) == STORAGE_NETFN &&
1789 	    data[1] == STORAGE_CMD_GET_SDR &&
1790 	    data[7] == 0x3A) {
1791 		return_hosed_msg_badsize(smi_info);
1792 		return NOTIFY_STOP;
1793 	}
1794 	return NOTIFY_DONE;
1795 }
1796 
1797 static struct notifier_block dell_poweredge_bt_xaction_notifier = {
1798 	.notifier_call	= dell_poweredge_bt_xaction_handler,
1799 };
1800 
1801 /*
1802  * setup_dell_poweredge_bt_xaction_handler
1803  * @info - smi_info.device_id must be filled in already
1804  *
1805  * Fills in smi_info.device_id.start_transaction_pre_hook
1806  * when we know what function to use there.
1807  */
1808 static void
1809 setup_dell_poweredge_bt_xaction_handler(struct smi_info *smi_info)
1810 {
1811 	struct ipmi_device_id *id = &smi_info->device_id;
1812 	if (id->manufacturer_id == DELL_IANA_MFR_ID &&
1813 	    smi_info->io.si_type == SI_BT)
1814 		register_xaction_notifier(&dell_poweredge_bt_xaction_notifier);
1815 }
1816 
1817 /*
1818  * setup_oem_data_handler
1819  * @info - smi_info.device_id must be filled in already
1820  *
1821  * Fills in smi_info.device_id.oem_data_available_handler
1822  * when we know what function to use there.
1823  */
1824 
1825 static void setup_oem_data_handler(struct smi_info *smi_info)
1826 {
1827 	setup_dell_poweredge_oem_data_handler(smi_info);
1828 }
1829 
1830 static void setup_xaction_handlers(struct smi_info *smi_info)
1831 {
1832 	setup_dell_poweredge_bt_xaction_handler(smi_info);
1833 }
1834 
1835 static void check_for_broken_irqs(struct smi_info *smi_info)
1836 {
1837 	check_clr_rcv_irq(smi_info);
1838 	check_set_rcv_irq(smi_info);
1839 }
1840 
1841 static inline void stop_timer_and_thread(struct smi_info *smi_info)
1842 {
1843 	if (smi_info->thread != NULL) {
1844 		kthread_stop(smi_info->thread);
1845 		smi_info->thread = NULL;
1846 	}
1847 
1848 	smi_info->timer_can_start = false;
1849 	del_timer_sync(&smi_info->si_timer);
1850 }
1851 
1852 static struct smi_info *find_dup_si(struct smi_info *info)
1853 {
1854 	struct smi_info *e;
1855 
1856 	list_for_each_entry(e, &smi_infos, link) {
1857 		if (e->io.addr_space != info->io.addr_space)
1858 			continue;
1859 		if (e->io.addr_data == info->io.addr_data) {
1860 			/*
1861 			 * This is a cheap hack, ACPI doesn't have a defined
1862 			 * slave address but SMBIOS does.  Pick it up from
1863 			 * any source that has it available.
1864 			 */
1865 			if (info->io.slave_addr && !e->io.slave_addr)
1866 				e->io.slave_addr = info->io.slave_addr;
1867 			return e;
1868 		}
1869 	}
1870 
1871 	return NULL;
1872 }
1873 
1874 int ipmi_si_add_smi(struct si_sm_io *io)
1875 {
1876 	int rv = 0;
1877 	struct smi_info *new_smi, *dup;
1878 
1879 	/*
1880 	 * If the user gave us a hard-coded device at the same
1881 	 * address, they presumably want us to use it and not what is
1882 	 * in the firmware.
1883 	 */
1884 	if (io->addr_source != SI_HARDCODED && io->addr_source != SI_HOTMOD &&
1885 	    ipmi_si_hardcode_match(io->addr_space, io->addr_data)) {
1886 		dev_info(io->dev,
1887 			 "Hard-coded device at this address already exists");
1888 		return -ENODEV;
1889 	}
1890 
1891 	if (!io->io_setup) {
1892 		if (io->addr_space == IPMI_IO_ADDR_SPACE) {
1893 			io->io_setup = ipmi_si_port_setup;
1894 		} else if (io->addr_space == IPMI_MEM_ADDR_SPACE) {
1895 			io->io_setup = ipmi_si_mem_setup;
1896 		} else {
1897 			return -EINVAL;
1898 		}
1899 	}
1900 
1901 	new_smi = kzalloc(sizeof(*new_smi), GFP_KERNEL);
1902 	if (!new_smi)
1903 		return -ENOMEM;
1904 	spin_lock_init(&new_smi->si_lock);
1905 
1906 	new_smi->io = *io;
1907 
1908 	mutex_lock(&smi_infos_lock);
1909 	dup = find_dup_si(new_smi);
1910 	if (dup) {
1911 		if (new_smi->io.addr_source == SI_ACPI &&
1912 		    dup->io.addr_source == SI_SMBIOS) {
1913 			/* We prefer ACPI over SMBIOS. */
1914 			dev_info(dup->io.dev,
1915 				 "Removing SMBIOS-specified %s state machine in favor of ACPI\n",
1916 				 si_to_str[new_smi->io.si_type]);
1917 			cleanup_one_si(dup);
1918 		} else {
1919 			dev_info(new_smi->io.dev,
1920 				 "%s-specified %s state machine: duplicate\n",
1921 				 ipmi_addr_src_to_str(new_smi->io.addr_source),
1922 				 si_to_str[new_smi->io.si_type]);
1923 			rv = -EBUSY;
1924 			kfree(new_smi);
1925 			goto out_err;
1926 		}
1927 	}
1928 
1929 	pr_info("Adding %s-specified %s state machine\n",
1930 		ipmi_addr_src_to_str(new_smi->io.addr_source),
1931 		si_to_str[new_smi->io.si_type]);
1932 
1933 	list_add_tail(&new_smi->link, &smi_infos);
1934 
1935 	if (initialized)
1936 		rv = try_smi_init(new_smi);
1937 out_err:
1938 	mutex_unlock(&smi_infos_lock);
1939 	return rv;
1940 }
1941 
1942 /*
1943  * Try to start up an interface.  Must be called with smi_infos_lock
1944  * held, primarily to keep smi_num consistent, we only one to do these
1945  * one at a time.
1946  */
1947 static int try_smi_init(struct smi_info *new_smi)
1948 {
1949 	int rv = 0;
1950 	int i;
1951 
1952 	pr_info("Trying %s-specified %s state machine at %s address 0x%lx, slave address 0x%x, irq %d\n",
1953 		ipmi_addr_src_to_str(new_smi->io.addr_source),
1954 		si_to_str[new_smi->io.si_type],
1955 		addr_space_to_str[new_smi->io.addr_space],
1956 		new_smi->io.addr_data,
1957 		new_smi->io.slave_addr, new_smi->io.irq);
1958 
1959 	switch (new_smi->io.si_type) {
1960 	case SI_KCS:
1961 		new_smi->handlers = &kcs_smi_handlers;
1962 		break;
1963 
1964 	case SI_SMIC:
1965 		new_smi->handlers = &smic_smi_handlers;
1966 		break;
1967 
1968 	case SI_BT:
1969 		new_smi->handlers = &bt_smi_handlers;
1970 		break;
1971 
1972 	default:
1973 		/* No support for anything else yet. */
1974 		rv = -EIO;
1975 		goto out_err;
1976 	}
1977 
1978 	new_smi->si_num = smi_num;
1979 
1980 	/* Do this early so it's available for logs. */
1981 	if (!new_smi->io.dev) {
1982 		pr_err("IPMI interface added with no device\n");
1983 		rv = -EIO;
1984 		goto out_err;
1985 	}
1986 
1987 	/* Allocate the state machine's data and initialize it. */
1988 	new_smi->si_sm = kmalloc(new_smi->handlers->size(), GFP_KERNEL);
1989 	if (!new_smi->si_sm) {
1990 		rv = -ENOMEM;
1991 		goto out_err;
1992 	}
1993 	new_smi->io.io_size = new_smi->handlers->init_data(new_smi->si_sm,
1994 							   &new_smi->io);
1995 
1996 	/* Now that we know the I/O size, we can set up the I/O. */
1997 	rv = new_smi->io.io_setup(&new_smi->io);
1998 	if (rv) {
1999 		dev_err(new_smi->io.dev, "Could not set up I/O space\n");
2000 		goto out_err;
2001 	}
2002 
2003 	/* Do low-level detection first. */
2004 	if (new_smi->handlers->detect(new_smi->si_sm)) {
2005 		if (new_smi->io.addr_source)
2006 			dev_err(new_smi->io.dev,
2007 				"Interface detection failed\n");
2008 		rv = -ENODEV;
2009 		goto out_err;
2010 	}
2011 
2012 	/*
2013 	 * Attempt a get device id command.  If it fails, we probably
2014 	 * don't have a BMC here.
2015 	 */
2016 	rv = try_get_dev_id(new_smi);
2017 	if (rv) {
2018 		if (new_smi->io.addr_source)
2019 			dev_err(new_smi->io.dev,
2020 			       "There appears to be no BMC at this location\n");
2021 		goto out_err;
2022 	}
2023 
2024 	setup_oem_data_handler(new_smi);
2025 	setup_xaction_handlers(new_smi);
2026 	check_for_broken_irqs(new_smi);
2027 
2028 	new_smi->waiting_msg = NULL;
2029 	new_smi->curr_msg = NULL;
2030 	atomic_set(&new_smi->req_events, 0);
2031 	new_smi->run_to_completion = false;
2032 	for (i = 0; i < SI_NUM_STATS; i++)
2033 		atomic_set(&new_smi->stats[i], 0);
2034 
2035 	new_smi->interrupt_disabled = true;
2036 	atomic_set(&new_smi->need_watch, 0);
2037 
2038 	rv = try_enable_event_buffer(new_smi);
2039 	if (rv == 0)
2040 		new_smi->has_event_buffer = true;
2041 
2042 	/*
2043 	 * Start clearing the flags before we enable interrupts or the
2044 	 * timer to avoid racing with the timer.
2045 	 */
2046 	start_clear_flags(new_smi);
2047 
2048 	/*
2049 	 * IRQ is defined to be set when non-zero.  req_events will
2050 	 * cause a global flags check that will enable interrupts.
2051 	 */
2052 	if (new_smi->io.irq) {
2053 		new_smi->interrupt_disabled = false;
2054 		atomic_set(&new_smi->req_events, 1);
2055 	}
2056 
2057 	dev_set_drvdata(new_smi->io.dev, new_smi);
2058 	rv = device_add_group(new_smi->io.dev, &ipmi_si_dev_attr_group);
2059 	if (rv) {
2060 		dev_err(new_smi->io.dev,
2061 			"Unable to add device attributes: error %d\n",
2062 			rv);
2063 		goto out_err;
2064 	}
2065 	new_smi->dev_group_added = true;
2066 
2067 	rv = ipmi_register_smi(&handlers,
2068 			       new_smi,
2069 			       new_smi->io.dev,
2070 			       new_smi->io.slave_addr);
2071 	if (rv) {
2072 		dev_err(new_smi->io.dev,
2073 			"Unable to register device: error %d\n",
2074 			rv);
2075 		goto out_err;
2076 	}
2077 
2078 	/* Don't increment till we know we have succeeded. */
2079 	smi_num++;
2080 
2081 	dev_info(new_smi->io.dev, "IPMI %s interface initialized\n",
2082 		 si_to_str[new_smi->io.si_type]);
2083 
2084 	WARN_ON(new_smi->io.dev->init_name != NULL);
2085 
2086  out_err:
2087 	if (rv && new_smi->io.io_cleanup) {
2088 		new_smi->io.io_cleanup(&new_smi->io);
2089 		new_smi->io.io_cleanup = NULL;
2090 	}
2091 
2092 	return rv;
2093 }
2094 
2095 static int __init init_ipmi_si(void)
2096 {
2097 	struct smi_info *e;
2098 	enum ipmi_addr_src type = SI_INVALID;
2099 
2100 	if (initialized)
2101 		return 0;
2102 
2103 	ipmi_hardcode_init();
2104 
2105 	pr_info("IPMI System Interface driver\n");
2106 
2107 	ipmi_si_platform_init();
2108 
2109 	ipmi_si_pci_init();
2110 
2111 	ipmi_si_parisc_init();
2112 
2113 	/* We prefer devices with interrupts, but in the case of a machine
2114 	   with multiple BMCs we assume that there will be several instances
2115 	   of a given type so if we succeed in registering a type then also
2116 	   try to register everything else of the same type */
2117 	mutex_lock(&smi_infos_lock);
2118 	list_for_each_entry(e, &smi_infos, link) {
2119 		/* Try to register a device if it has an IRQ and we either
2120 		   haven't successfully registered a device yet or this
2121 		   device has the same type as one we successfully registered */
2122 		if (e->io.irq && (!type || e->io.addr_source == type)) {
2123 			if (!try_smi_init(e)) {
2124 				type = e->io.addr_source;
2125 			}
2126 		}
2127 	}
2128 
2129 	/* type will only have been set if we successfully registered an si */
2130 	if (type)
2131 		goto skip_fallback_noirq;
2132 
2133 	/* Fall back to the preferred device */
2134 
2135 	list_for_each_entry(e, &smi_infos, link) {
2136 		if (!e->io.irq && (!type || e->io.addr_source == type)) {
2137 			if (!try_smi_init(e)) {
2138 				type = e->io.addr_source;
2139 			}
2140 		}
2141 	}
2142 
2143 skip_fallback_noirq:
2144 	initialized = true;
2145 	mutex_unlock(&smi_infos_lock);
2146 
2147 	if (type)
2148 		return 0;
2149 
2150 	mutex_lock(&smi_infos_lock);
2151 	if (unload_when_empty && list_empty(&smi_infos)) {
2152 		mutex_unlock(&smi_infos_lock);
2153 		cleanup_ipmi_si();
2154 		pr_warn("Unable to find any System Interface(s)\n");
2155 		return -ENODEV;
2156 	} else {
2157 		mutex_unlock(&smi_infos_lock);
2158 		return 0;
2159 	}
2160 }
2161 module_init(init_ipmi_si);
2162 
2163 static void shutdown_smi(void *send_info)
2164 {
2165 	struct smi_info *smi_info = send_info;
2166 
2167 	if (smi_info->dev_group_added) {
2168 		device_remove_group(smi_info->io.dev, &ipmi_si_dev_attr_group);
2169 		smi_info->dev_group_added = false;
2170 	}
2171 	if (smi_info->io.dev)
2172 		dev_set_drvdata(smi_info->io.dev, NULL);
2173 
2174 	/*
2175 	 * Make sure that interrupts, the timer and the thread are
2176 	 * stopped and will not run again.
2177 	 */
2178 	smi_info->interrupt_disabled = true;
2179 	if (smi_info->io.irq_cleanup) {
2180 		smi_info->io.irq_cleanup(&smi_info->io);
2181 		smi_info->io.irq_cleanup = NULL;
2182 	}
2183 	stop_timer_and_thread(smi_info);
2184 
2185 	/*
2186 	 * Wait until we know that we are out of any interrupt
2187 	 * handlers might have been running before we freed the
2188 	 * interrupt.
2189 	 */
2190 	synchronize_rcu();
2191 
2192 	/*
2193 	 * Timeouts are stopped, now make sure the interrupts are off
2194 	 * in the BMC.  Note that timers and CPU interrupts are off,
2195 	 * so no need for locks.
2196 	 */
2197 	while (smi_info->curr_msg || (smi_info->si_state != SI_NORMAL)) {
2198 		poll(smi_info);
2199 		schedule_timeout_uninterruptible(1);
2200 	}
2201 	if (smi_info->handlers)
2202 		disable_si_irq(smi_info);
2203 	while (smi_info->curr_msg || (smi_info->si_state != SI_NORMAL)) {
2204 		poll(smi_info);
2205 		schedule_timeout_uninterruptible(1);
2206 	}
2207 	if (smi_info->handlers)
2208 		smi_info->handlers->cleanup(smi_info->si_sm);
2209 
2210 	if (smi_info->io.addr_source_cleanup) {
2211 		smi_info->io.addr_source_cleanup(&smi_info->io);
2212 		smi_info->io.addr_source_cleanup = NULL;
2213 	}
2214 	if (smi_info->io.io_cleanup) {
2215 		smi_info->io.io_cleanup(&smi_info->io);
2216 		smi_info->io.io_cleanup = NULL;
2217 	}
2218 
2219 	kfree(smi_info->si_sm);
2220 	smi_info->si_sm = NULL;
2221 
2222 	smi_info->intf = NULL;
2223 }
2224 
2225 /*
2226  * Must be called with smi_infos_lock held, to serialize the
2227  * smi_info->intf check.
2228  */
2229 static void cleanup_one_si(struct smi_info *smi_info)
2230 {
2231 	if (!smi_info)
2232 		return;
2233 
2234 	list_del(&smi_info->link);
2235 
2236 	if (smi_info->intf)
2237 		ipmi_unregister_smi(smi_info->intf);
2238 
2239 	kfree(smi_info);
2240 }
2241 
2242 int ipmi_si_remove_by_dev(struct device *dev)
2243 {
2244 	struct smi_info *e;
2245 	int rv = -ENOENT;
2246 
2247 	mutex_lock(&smi_infos_lock);
2248 	list_for_each_entry(e, &smi_infos, link) {
2249 		if (e->io.dev == dev) {
2250 			cleanup_one_si(e);
2251 			rv = 0;
2252 			break;
2253 		}
2254 	}
2255 	mutex_unlock(&smi_infos_lock);
2256 
2257 	return rv;
2258 }
2259 
2260 struct device *ipmi_si_remove_by_data(int addr_space, enum si_type si_type,
2261 				      unsigned long addr)
2262 {
2263 	/* remove */
2264 	struct smi_info *e, *tmp_e;
2265 	struct device *dev = NULL;
2266 
2267 	mutex_lock(&smi_infos_lock);
2268 	list_for_each_entry_safe(e, tmp_e, &smi_infos, link) {
2269 		if (e->io.addr_space != addr_space)
2270 			continue;
2271 		if (e->io.si_type != si_type)
2272 			continue;
2273 		if (e->io.addr_data == addr) {
2274 			dev = get_device(e->io.dev);
2275 			cleanup_one_si(e);
2276 		}
2277 	}
2278 	mutex_unlock(&smi_infos_lock);
2279 
2280 	return dev;
2281 }
2282 
2283 static void cleanup_ipmi_si(void)
2284 {
2285 	struct smi_info *e, *tmp_e;
2286 
2287 	if (!initialized)
2288 		return;
2289 
2290 	ipmi_si_pci_shutdown();
2291 
2292 	ipmi_si_parisc_shutdown();
2293 
2294 	ipmi_si_platform_shutdown();
2295 
2296 	mutex_lock(&smi_infos_lock);
2297 	list_for_each_entry_safe(e, tmp_e, &smi_infos, link)
2298 		cleanup_one_si(e);
2299 	mutex_unlock(&smi_infos_lock);
2300 
2301 	ipmi_si_hardcode_exit();
2302 	ipmi_si_hotmod_exit();
2303 }
2304 module_exit(cleanup_ipmi_si);
2305 
2306 MODULE_ALIAS("platform:dmi-ipmi-si");
2307 MODULE_LICENSE("GPL");
2308 MODULE_AUTHOR("Corey Minyard <minyard@mvista.com>");
2309 MODULE_DESCRIPTION("Interface to the IPMI driver for the KCS, SMIC, and BT"
2310 		   " system interfaces.");
2311