drivers/char/ipmi/ipmi_si_intf.c

   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 <linux/string.h>
  44 #include <linux/ctype.h>
  45
  46 /* Measure times between events in the driver. */
  47 #undef DEBUG_TIMING
  48
  49 /* Call every 10 ms. */
  50 #define SI_TIMEOUT_TIME_USEC    10000
  51 #define SI_USEC_PER_JIFFY       (1000000/HZ)
  52 #define SI_TIMEOUT_JIFFIES      (SI_TIMEOUT_TIME_USEC/SI_USEC_PER_JIFFY)
  53 #define SI_SHORT_TIMEOUT_USEC  250 /* .25ms when the SM request a
  54                                       short timeout */
  55
  56 enum si_intf_state {
  57         SI_NORMAL,
  58         SI_GETTING_FLAGS,
  59         SI_GETTING_EVENTS,
  60         SI_CLEARING_FLAGS,
  61         SI_GETTING_MESSAGES,
  62         SI_CHECKING_ENABLES,
  63         SI_SETTING_ENABLES
  64         /* FIXME - add watchdog stuff. */
  65 };
  66
  67 /* Some BT-specific defines we need here. */
  68 #define IPMI_BT_INTMASK_REG             2
  69 #define IPMI_BT_INTMASK_CLEAR_IRQ_BIT   2
  70 #define IPMI_BT_INTMASK_ENABLE_IRQ_BIT  1
  71
  72 static const char * const si_to_str[] = { "invalid", "kcs", "smic", "bt" };
  73
  74 static int initialized;
  75
  76 /*
  77  * Indexes into stats[] in smi_info below.
  78  */
  79 enum si_stat_indexes {
  80         /*
  81          * Number of times the driver requested a timer while an operation
  82          * was in progress.
  83          */
  84         SI_STAT_short_timeouts = 0,
  85
  86         /*
  87          * Number of times the driver requested a timer while nothing was in
  88          * progress.
  89          */
  90         SI_STAT_long_timeouts,
  91
  92         /* Number of times the interface was idle while being polled. */
  93         SI_STAT_idles,
  94
  95         /* Number of interrupts the driver handled. */
  96         SI_STAT_interrupts,
  97
  98         /* Number of time the driver got an ATTN from the hardware. */
  99         SI_STAT_attentions,
 100
 101         /* Number of times the driver requested flags from the hardware. */
 102         SI_STAT_flag_fetches,
 103
 104         /* Number of times the hardware didn't follow the state machine. */
 105         SI_STAT_hosed_count,
 106
 107         /* Number of completed messages. */
 108         SI_STAT_complete_transactions,
 109
 110         /* Number of IPMI events received from the hardware. */
 111         SI_STAT_events,
 112
 113         /* Number of watchdog pretimeouts. */
 114         SI_STAT_watchdog_pretimeouts,
 115
 116         /* Number of asynchronous messages received. */
 117         SI_STAT_incoming_messages,
 118
 119
 120         /* This *must* remain last, add new values above this. */
 121         SI_NUM_STATS
 122 };
 123
 124 struct smi_info {
 125         int                    si_num;
 126         struct ipmi_smi        *intf;
 127         struct si_sm_data      *si_sm;
 128         const struct si_sm_handlers *handlers;
 129         spinlock_t             si_lock;
 130         struct ipmi_smi_msg    *waiting_msg;
 131         struct ipmi_smi_msg    *curr_msg;
 132         enum si_intf_state     si_state;
 133
 134         /*
 135          * Used to handle the various types of I/O that can occur with
 136          * IPMI
 137          */
 138         struct si_sm_io io;
 139
 140         /*
 141          * Per-OEM handler, called from handle_flags().  Returns 1
 142          * when handle_flags() needs to be re-run or 0 indicating it
 143          * set si_state itself.
 144          */
 145         int (*oem_data_avail_handler)(struct smi_info *smi_info);
 146
 147         /*
 148          * Flags from the last GET_MSG_FLAGS command, used when an ATTN
 149          * is set to hold the flags until we are done handling everything
 150          * from the flags.
 151          */
 152 #define RECEIVE_MSG_AVAIL       0x01
 153 #define EVENT_MSG_BUFFER_FULL   0x02
 154 #define WDT_PRE_TIMEOUT_INT     0x08
 155 #define OEM0_DATA_AVAIL     0x20
 156 #define OEM1_DATA_AVAIL     0x40
 157 #define OEM2_DATA_AVAIL     0x80
 158 #define OEM_DATA_AVAIL      (OEM0_DATA_AVAIL | \
 159                              OEM1_DATA_AVAIL | \
 160                              OEM2_DATA_AVAIL)
 161         unsigned char       msg_flags;
 162
 163         /* Does the BMC have an event buffer? */
 164         bool                has_event_buffer;
 165
 166         /*
 167          * If set to true, this will request events the next time the
 168          * state machine is idle.
 169          */
 170         atomic_t            req_events;
 171
 172         /*
 173          * If true, run the state machine to completion on every send
 174          * call.  Generally used after a panic to make sure stuff goes
 175          * out.
 176          */
 177         bool                run_to_completion;
 178
 179         /* The timer for this si. */
 180         struct timer_list   si_timer;
 181
 182         /* This flag is set, if the timer can be set */
 183         bool                timer_can_start;
 184
 185         /* This flag is set, if the timer is running (timer_pending() isn't enough) */
 186         bool                timer_running;
 187
 188         /* The time (in jiffies) the last timeout occurred at. */
 189         unsigned long       last_timeout_jiffies;
 190
 191         /* Are we waiting for the events, pretimeouts, received msgs? */
 192         atomic_t            need_watch;
 193
 194         /*
 195          * The driver will disable interrupts when it gets into a
 196          * situation where it cannot handle messages due to lack of
 197          * memory.  Once that situation clears up, it will re-enable
 198          * interrupts.
 199          */
 200         bool interrupt_disabled;
 201
 202         /*
 203          * Does the BMC support events?
 204          */
 205         bool supports_event_msg_buff;
 206
 207         /*
 208          * Can we disable interrupts the global enables receive irq
 209          * bit?  There are currently two forms of brokenness, some
 210          * systems cannot disable the bit (which is technically within
 211          * the spec but a bad idea) and some systems have the bit
 212          * forced to zero even though interrupts work (which is
 213          * clearly outside the spec).  The next bool tells which form
 214          * of brokenness is present.
 215          */
 216         bool cannot_disable_irq;
 217
 218         /*
 219          * Some systems are broken and cannot set the irq enable
 220          * bit, even if they support interrupts.
 221          */
 222         bool irq_enable_broken;
 223
 224         /*
 225          * Did we get an attention that we did not handle?
 226          */
 227         bool got_attn;
 228
 229         /* From the get device id response... */
 230         struct ipmi_device_id device_id;
 231
 232         /* Default driver model device. */
 233         struct platform_device *pdev;
 234
 235         /* Have we added the device group to the device? */
 236         bool dev_group_added;
 237
 238         /* Have we added the platform device? */
 239         bool pdev_registered;
 240
 241         /* Counters and things for the proc filesystem. */
 242         atomic_t stats[SI_NUM_STATS];
 243
 244         struct task_struct *thread;
 245
 246         struct list_head link;
 247 };
 248
 249 #define smi_inc_stat(smi, stat) \
 250         atomic_inc(&(smi)->stats[SI_STAT_ ## stat])
 251 #define smi_get_stat(smi, stat) \
 252         ((unsigned int) atomic_read(&(smi)->stats[SI_STAT_ ## stat]))
 253
 254 #define IPMI_MAX_INTFS 4
 255 static int force_kipmid[IPMI_MAX_INTFS];
 256 static int num_force_kipmid;
 257
 258 static unsigned int kipmid_max_busy_us[IPMI_MAX_INTFS];
 259 static int num_max_busy_us;
 260
 261 static bool unload_when_empty = true;
 262
 263 static int try_smi_init(struct smi_info *smi);
 264 static void cleanup_one_si(struct smi_info *smi_info);
 265 static void cleanup_ipmi_si(void);
 266
 267 #ifdef DEBUG_TIMING
 268 void debug_timestamp(char *msg)
 269 {
 270         struct timespec64 t;
 271
 272         ktime_get_ts64(&t);
 273         pr_debug("**%s: %lld.%9.9ld\n", msg, (long long) t.tv_sec, t.tv_nsec);
 274 }
 275 #else
 276 #define debug_timestamp(x)
 277 #endif
 278
 279 static ATOMIC_NOTIFIER_HEAD(xaction_notifier_list);
 280 static int register_xaction_notifier(struct notifier_block *nb)
 281 {
 282         return atomic_notifier_chain_register(&xaction_notifier_list, nb);
 283 }
 284
 285 static void deliver_recv_msg(struct smi_info *smi_info,
 286                              struct ipmi_smi_msg *msg)
 287 {
 288         /* Deliver the message to the upper layer. */
 289         ipmi_smi_msg_received(smi_info->intf, msg);
 290 }
 291
 292 static void return_hosed_msg(struct smi_info *smi_info, int cCode)
 293 {
 294         struct ipmi_smi_msg *msg = smi_info->curr_msg;
 295
 296         if (cCode < 0 || cCode > IPMI_ERR_UNSPECIFIED)
 297                 cCode = IPMI_ERR_UNSPECIFIED;
 298         /* else use it as is */
 299
 300         /* Make it a response */
 301         msg->rsp[0] = msg->data[0] | 4;
 302         msg->rsp[1] = msg->data[1];
 303         msg->rsp[2] = cCode;
 304         msg->rsp_size = 3;
 305
 306         smi_info->curr_msg = NULL;
 307         deliver_recv_msg(smi_info, msg);
 308 }
 309
 310 static enum si_sm_result start_next_msg(struct smi_info *smi_info)
 311 {
 312         int              rv;
 313
 314         if (!smi_info->waiting_msg) {
 315                 smi_info->curr_msg = NULL;
 316                 rv = SI_SM_IDLE;
 317         } else {
 318                 int err;
 319
 320                 smi_info->curr_msg = smi_info->waiting_msg;
 321                 smi_info->waiting_msg = NULL;
 322                 debug_timestamp("Start2");
 323                 err = atomic_notifier_call_chain(&xaction_notifier_list,
 324                                 0, smi_info);
 325                 if (err & NOTIFY_STOP_MASK) {
 326                         rv = SI_SM_CALL_WITHOUT_DELAY;
 327                         goto out;
 328                 }
 329                 err = smi_info->handlers->start_transaction(
 330                         smi_info->si_sm,
 331                         smi_info->curr_msg->data,
 332                         smi_info->curr_msg->data_size);
 333                 if (err)
 334                         return_hosed_msg(smi_info, err);
 335
 336                 rv = SI_SM_CALL_WITHOUT_DELAY;
 337         }
 338 out:
 339         return rv;
 340 }
 341
 342 static void smi_mod_timer(struct smi_info *smi_info, unsigned long new_val)
 343 {
 344         if (!smi_info->timer_can_start)
 345                 return;
 346         smi_info->last_timeout_jiffies = jiffies;
 347         mod_timer(&smi_info->si_timer, new_val);
 348         smi_info->timer_running = true;
 349 }
 350
 351 /*
 352  * Start a new message and (re)start the timer and thread.
 353  */
 354 static void start_new_msg(struct smi_info *smi_info, unsigned char *msg,
 355                           unsigned int size)
 356 {
 357         smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
 358
 359         if (smi_info->thread)
 360                 wake_up_process(smi_info->thread);
 361
 362         smi_info->handlers->start_transaction(smi_info->si_sm, msg, size);
 363 }
 364
 365 static void start_check_enables(struct smi_info *smi_info)
 366 {
 367         unsigned char msg[2];
 368
 369         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 370         msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
 371
 372         start_new_msg(smi_info, msg, 2);
 373         smi_info->si_state = SI_CHECKING_ENABLES;
 374 }
 375
 376 static void start_clear_flags(struct smi_info *smi_info)
 377 {
 378         unsigned char msg[3];
 379
 380         /* Make sure the watchdog pre-timeout flag is not set at startup. */
 381         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 382         msg[1] = IPMI_CLEAR_MSG_FLAGS_CMD;
 383         msg[2] = WDT_PRE_TIMEOUT_INT;
 384
 385         start_new_msg(smi_info, msg, 3);
 386         smi_info->si_state = SI_CLEARING_FLAGS;
 387 }
 388
 389 static void start_getting_msg_queue(struct smi_info *smi_info)
 390 {
 391         smi_info->curr_msg->data[0] = (IPMI_NETFN_APP_REQUEST << 2);
 392         smi_info->curr_msg->data[1] = IPMI_GET_MSG_CMD;
 393         smi_info->curr_msg->data_size = 2;
 394
 395         start_new_msg(smi_info, smi_info->curr_msg->data,
 396                       smi_info->curr_msg->data_size);
 397         smi_info->si_state = SI_GETTING_MESSAGES;
 398 }
 399
 400 static void start_getting_events(struct smi_info *smi_info)
 401 {
 402         smi_info->curr_msg->data[0] = (IPMI_NETFN_APP_REQUEST << 2);
 403         smi_info->curr_msg->data[1] = IPMI_READ_EVENT_MSG_BUFFER_CMD;
 404         smi_info->curr_msg->data_size = 2;
 405
 406         start_new_msg(smi_info, smi_info->curr_msg->data,
 407                       smi_info->curr_msg->data_size);
 408         smi_info->si_state = SI_GETTING_EVENTS;
 409 }
 410
 411 /*
 412  * When we have a situtaion where we run out of memory and cannot
 413  * allocate messages, we just leave them in the BMC and run the system
 414  * polled until we can allocate some memory.  Once we have some
 415  * memory, we will re-enable the interrupt.
 416  *
 417  * Note that we cannot just use disable_irq(), since the interrupt may
 418  * be shared.
 419  */
 420 static inline bool disable_si_irq(struct smi_info *smi_info)
 421 {
 422         if ((smi_info->io.irq) && (!smi_info->interrupt_disabled)) {
 423                 smi_info->interrupt_disabled = true;
 424                 start_check_enables(smi_info);
 425                 return true;
 426         }
 427         return false;
 428 }
 429
 430 static inline bool enable_si_irq(struct smi_info *smi_info)
 431 {
 432         if ((smi_info->io.irq) && (smi_info->interrupt_disabled)) {
 433                 smi_info->interrupt_disabled = false;
 434                 start_check_enables(smi_info);
 435                 return true;
 436         }
 437         return false;
 438 }
 439
 440 /*
 441  * Allocate a message.  If unable to allocate, start the interrupt
 442  * disable process and return NULL.  If able to allocate but
 443  * interrupts are disabled, free the message and return NULL after
 444  * starting the interrupt enable process.
 445  */
 446 static struct ipmi_smi_msg *alloc_msg_handle_irq(struct smi_info *smi_info)
 447 {
 448         struct ipmi_smi_msg *msg;
 449
 450         msg = ipmi_alloc_smi_msg();
 451         if (!msg) {
 452                 if (!disable_si_irq(smi_info))
 453                         smi_info->si_state = SI_NORMAL;
 454         } else if (enable_si_irq(smi_info)) {
 455                 ipmi_free_smi_msg(msg);
 456                 msg = NULL;
 457         }
 458         return msg;
 459 }
 460
 461 static void handle_flags(struct smi_info *smi_info)
 462 {
 463 retry:
 464         if (smi_info->msg_flags & WDT_PRE_TIMEOUT_INT) {
 465                 /* Watchdog pre-timeout */
 466                 smi_inc_stat(smi_info, watchdog_pretimeouts);
 467
 468                 start_clear_flags(smi_info);
 469                 smi_info->msg_flags &= ~WDT_PRE_TIMEOUT_INT;
 470                 ipmi_smi_watchdog_pretimeout(smi_info->intf);
 471         } else if (smi_info->msg_flags & RECEIVE_MSG_AVAIL) {
 472                 /* Messages available. */
 473                 smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
 474                 if (!smi_info->curr_msg)
 475                         return;
 476
 477                 start_getting_msg_queue(smi_info);
 478         } else if (smi_info->msg_flags & EVENT_MSG_BUFFER_FULL) {
 479                 /* Events available. */
 480                 smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
 481                 if (!smi_info->curr_msg)
 482                         return;
 483
 484                 start_getting_events(smi_info);
 485         } else if (smi_info->msg_flags & OEM_DATA_AVAIL &&
 486                    smi_info->oem_data_avail_handler) {
 487                 if (smi_info->oem_data_avail_handler(smi_info))
 488                         goto retry;
 489         } else
 490                 smi_info->si_state = SI_NORMAL;
 491 }
 492
 493 /*
 494  * Global enables we care about.
 495  */
 496 #define GLOBAL_ENABLES_MASK (IPMI_BMC_EVT_MSG_BUFF | IPMI_BMC_RCV_MSG_INTR | \
 497                              IPMI_BMC_EVT_MSG_INTR)
 498
 499 static u8 current_global_enables(struct smi_info *smi_info, u8 base,
 500                                  bool *irq_on)
 501 {
 502         u8 enables = 0;
 503
 504         if (smi_info->supports_event_msg_buff)
 505                 enables |= IPMI_BMC_EVT_MSG_BUFF;
 506
 507         if (((smi_info->io.irq && !smi_info->interrupt_disabled) ||
 508              smi_info->cannot_disable_irq) &&
 509             !smi_info->irq_enable_broken)
 510                 enables |= IPMI_BMC_RCV_MSG_INTR;
 511
 512         if (smi_info->supports_event_msg_buff &&
 513             smi_info->io.irq && !smi_info->interrupt_disabled &&
 514             !smi_info->irq_enable_broken)
 515                 enables |= IPMI_BMC_EVT_MSG_INTR;
 516
 517         *irq_on = enables & (IPMI_BMC_EVT_MSG_INTR | IPMI_BMC_RCV_MSG_INTR);
 518
 519         return enables;
 520 }
 521
 522 static void check_bt_irq(struct smi_info *smi_info, bool irq_on)
 523 {
 524         u8 irqstate = smi_info->io.inputb(&smi_info->io, IPMI_BT_INTMASK_REG);
 525
 526         irqstate &= IPMI_BT_INTMASK_ENABLE_IRQ_BIT;
 527
 528         if ((bool)irqstate == irq_on)
 529                 return;
 530
 531         if (irq_on)
 532                 smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG,
 533                                      IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
 534         else
 535                 smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG, 0);
 536 }
 537
 538 static void handle_transaction_done(struct smi_info *smi_info)
 539 {
 540         struct ipmi_smi_msg *msg;
 541
 542         debug_timestamp("Done");
 543         switch (smi_info->si_state) {
 544         case SI_NORMAL:
 545                 if (!smi_info->curr_msg)
 546                         break;
 547
 548                 smi_info->curr_msg->rsp_size
 549                         = smi_info->handlers->get_result(
 550                                 smi_info->si_sm,
 551                                 smi_info->curr_msg->rsp,
 552                                 IPMI_MAX_MSG_LENGTH);
 553
 554                 /*
 555                  * Do this here becase deliver_recv_msg() releases the
 556                  * lock, and a new message can be put in during the
 557                  * time the lock is released.
 558                  */
 559                 msg = smi_info->curr_msg;
 560                 smi_info->curr_msg = NULL;
 561                 deliver_recv_msg(smi_info, msg);
 562                 break;
 563
 564         case SI_GETTING_FLAGS:
 565         {
 566                 unsigned char msg[4];
 567                 unsigned int  len;
 568
 569                 /* We got the flags from the SMI, now handle them. */
 570                 len = smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
 571                 if (msg[2] != 0) {
 572                         /* Error fetching flags, just give up for now. */
 573                         smi_info->si_state = SI_NORMAL;
 574                 } else if (len < 4) {
 575                         /*
 576                          * Hmm, no flags.  That's technically illegal, but
 577                          * don't use uninitialized data.
 578                          */
 579                         smi_info->si_state = SI_NORMAL;
 580                 } else {
 581                         smi_info->msg_flags = msg[3];
 582                         handle_flags(smi_info);
 583                 }
 584                 break;
 585         }
 586
 587         case SI_CLEARING_FLAGS:
 588         {
 589                 unsigned char msg[3];
 590
 591                 /* We cleared the flags. */
 592                 smi_info->handlers->get_result(smi_info->si_sm, msg, 3);
 593                 if (msg[2] != 0) {
 594                         /* Error clearing flags */
 595                         dev_warn(smi_info->io.dev,
 596                                  "Error clearing flags: %2.2x\n", msg[2]);
 597                 }
 598                 smi_info->si_state = SI_NORMAL;
 599                 break;
 600         }
 601
 602         case SI_GETTING_EVENTS:
 603         {
 604                 smi_info->curr_msg->rsp_size
 605                         = smi_info->handlers->get_result(
 606                                 smi_info->si_sm,
 607                                 smi_info->curr_msg->rsp,
 608                                 IPMI_MAX_MSG_LENGTH);
 609
 610                 /*
 611                  * Do this here becase deliver_recv_msg() releases the
 612                  * lock, and a new message can be put in during the
 613                  * time the lock is released.
 614                  */
 615                 msg = smi_info->curr_msg;
 616                 smi_info->curr_msg = NULL;
 617                 if (msg->rsp[2] != 0) {
 618                         /* Error getting event, probably done. */
 619                         msg->done(msg);
 620
 621                         /* Take off the event flag. */
 622                         smi_info->msg_flags &= ~EVENT_MSG_BUFFER_FULL;
 623                         handle_flags(smi_info);
 624                 } else {
 625                         smi_inc_stat(smi_info, events);
 626
 627                         /*
 628                          * Do this before we deliver the message
 629                          * because delivering the message releases the
 630                          * lock and something else can mess with the
 631                          * state.
 632                          */
 633                         handle_flags(smi_info);
 634
 635                         deliver_recv_msg(smi_info, msg);
 636                 }
 637                 break;
 638         }
 639
 640         case SI_GETTING_MESSAGES:
 641         {
 642                 smi_info->curr_msg->rsp_size
 643                         = smi_info->handlers->get_result(
 644                                 smi_info->si_sm,
 645                                 smi_info->curr_msg->rsp,
 646                                 IPMI_MAX_MSG_LENGTH);
 647
 648                 /*
 649                  * Do this here becase deliver_recv_msg() releases the
 650                  * lock, and a new message can be put in during the
 651                  * time the lock is released.
 652                  */
 653                 msg = smi_info->curr_msg;
 654                 smi_info->curr_msg = NULL;
 655                 if (msg->rsp[2] != 0) {
 656                         /* Error getting event, probably done. */
 657                         msg->done(msg);
 658
 659                         /* Take off the msg flag. */
 660                         smi_info->msg_flags &= ~RECEIVE_MSG_AVAIL;
 661                         handle_flags(smi_info);
 662                 } else {
 663                         smi_inc_stat(smi_info, incoming_messages);
 664
 665                         /*
 666                          * Do this before we deliver the message
 667                          * because delivering the message releases the
 668                          * lock and something else can mess with the
 669                          * state.
 670                          */
 671                         handle_flags(smi_info);
 672
 673                         deliver_recv_msg(smi_info, msg);
 674                 }
 675                 break;
 676         }
 677
 678         case SI_CHECKING_ENABLES:
 679         {
 680                 unsigned char msg[4];
 681                 u8 enables;
 682                 bool irq_on;
 683
 684                 /* We got the flags from the SMI, now handle them. */
 685                 smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
 686                 if (msg[2] != 0) {
 687                         dev_warn(smi_info->io.dev,
 688                                  "Couldn't get irq info: %x.\n", msg[2]);
 689                         dev_warn(smi_info->io.dev,
 690                                  "Maybe ok, but ipmi might run very slowly.\n");
 691                         smi_info->si_state = SI_NORMAL;
 692                         break;
 693                 }
 694                 enables = current_global_enables(smi_info, 0, &irq_on);
 695                 if (smi_info->io.si_type == SI_BT)
 696                         /* BT has its own interrupt enable bit. */
 697                         check_bt_irq(smi_info, irq_on);
 698                 if (enables != (msg[3] & GLOBAL_ENABLES_MASK)) {
 699                         /* Enables are not correct, fix them. */
 700                         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 701                         msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
 702                         msg[2] = enables | (msg[3] & ~GLOBAL_ENABLES_MASK);
 703                         smi_info->handlers->start_transaction(
 704                                 smi_info->si_sm, msg, 3);
 705                         smi_info->si_state = SI_SETTING_ENABLES;
 706                 } else if (smi_info->supports_event_msg_buff) {
 707                         smi_info->curr_msg = ipmi_alloc_smi_msg();
 708                         if (!smi_info->curr_msg) {
 709                                 smi_info->si_state = SI_NORMAL;
 710                                 break;
 711                         }
 712                         start_getting_events(smi_info);
 713                 } else {
 714                         smi_info->si_state = SI_NORMAL;
 715                 }
 716                 break;
 717         }
 718
 719         case SI_SETTING_ENABLES:
 720         {
 721                 unsigned char msg[4];
 722
 723                 smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
 724                 if (msg[2] != 0)
 725                         dev_warn(smi_info->io.dev,
 726                                  "Could not set the global enables: 0x%x.\n",
 727                                  msg[2]);
 728
 729                 if (smi_info->supports_event_msg_buff) {
 730                         smi_info->curr_msg = ipmi_alloc_smi_msg();
 731                         if (!smi_info->curr_msg) {
 732                                 smi_info->si_state = SI_NORMAL;
 733                                 break;
 734                         }
 735                         start_getting_events(smi_info);
 736                 } else {
 737                         smi_info->si_state = SI_NORMAL;
 738                 }
 739                 break;
 740         }
 741         }
 742 }
 743
 744 /*
 745  * Called on timeouts and events.  Timeouts should pass the elapsed
 746  * time, interrupts should pass in zero.  Must be called with
 747  * si_lock held and interrupts disabled.
 748  */
 749 static enum si_sm_result smi_event_handler(struct smi_info *smi_info,
 750                                            int time)
 751 {
 752         enum si_sm_result si_sm_result;
 753
 754 restart:
 755         /*
 756          * There used to be a loop here that waited a little while
 757          * (around 25us) before giving up.  That turned out to be
 758          * pointless, the minimum delays I was seeing were in the 300us
 759          * range, which is far too long to wait in an interrupt.  So
 760          * we just run until the state machine tells us something
 761          * happened or it needs a delay.
 762          */
 763         si_sm_result = smi_info->handlers->event(smi_info->si_sm, time);
 764         time = 0;
 765         while (si_sm_result == SI_SM_CALL_WITHOUT_DELAY)
 766                 si_sm_result = smi_info->handlers->event(smi_info->si_sm, 0);
 767
 768         if (si_sm_result == SI_SM_TRANSACTION_COMPLETE) {
 769                 smi_inc_stat(smi_info, complete_transactions);
 770
 771                 handle_transaction_done(smi_info);
 772                 goto restart;
 773         } else if (si_sm_result == SI_SM_HOSED) {
 774                 smi_inc_stat(smi_info, hosed_count);
 775
 776                 /*
 777                  * Do the before return_hosed_msg, because that
 778                  * releases the lock.
 779                  */
 780                 smi_info->si_state = SI_NORMAL;
 781                 if (smi_info->curr_msg != NULL) {
 782                         /*
 783                          * If we were handling a user message, format
 784                          * a response to send to the upper layer to
 785                          * tell it about the error.
 786                          */
 787                         return_hosed_msg(smi_info, IPMI_ERR_UNSPECIFIED);
 788                 }
 789                 goto restart;
 790         }
 791
 792         /*
 793          * We prefer handling attn over new messages.  But don't do
 794          * this if there is not yet an upper layer to handle anything.
 795          */
 796         if (si_sm_result == SI_SM_ATTN || smi_info->got_attn) {
 797                 unsigned char msg[2];
 798
 799                 if (smi_info->si_state != SI_NORMAL) {
 800                         /*
 801                          * We got an ATTN, but we are doing something else.
 802                          * Handle the ATTN later.
 803                          */
 804                         smi_info->got_attn = true;
 805                 } else {
 806                         smi_info->got_attn = false;
 807                         smi_inc_stat(smi_info, attentions);
 808
 809                         /*
 810                          * Got a attn, send down a get message flags to see
 811                          * what's causing it.  It would be better to handle
 812                          * this in the upper layer, but due to the way
 813                          * interrupts work with the SMI, that's not really
 814                          * possible.
 815                          */
 816                         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 817                         msg[1] = IPMI_GET_MSG_FLAGS_CMD;
 818
 819                         start_new_msg(smi_info, msg, 2);
 820                         smi_info->si_state = SI_GETTING_FLAGS;
 821                         goto restart;
 822                 }
 823         }
 824
 825         /* If we are currently idle, try to start the next message. */
 826         if (si_sm_result == SI_SM_IDLE) {
 827                 smi_inc_stat(smi_info, idles);
 828
 829                 si_sm_result = start_next_msg(smi_info);
 830                 if (si_sm_result != SI_SM_IDLE)
 831                         goto restart;
 832         }
 833
 834         if ((si_sm_result == SI_SM_IDLE)
 835             && (atomic_read(&smi_info->req_events))) {
 836                 /*
 837                  * We are idle and the upper layer requested that I fetch
 838                  * events, so do so.
 839                  */
 840                 atomic_set(&smi_info->req_events, 0);
 841
 842                 /*
 843                  * Take this opportunity to check the interrupt and
 844                  * message enable state for the BMC.  The BMC can be
 845                  * asynchronously reset, and may thus get interrupts
 846                  * disable and messages disabled.
 847                  */
 848                 if (smi_info->supports_event_msg_buff || smi_info->io.irq) {
 849                         start_check_enables(smi_info);
 850                 } else {
 851                         smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
 852                         if (!smi_info->curr_msg)
 853                                 goto out;
 854
 855                         start_getting_events(smi_info);
 856                 }
 857                 goto restart;
 858         }
 859
 860         if (si_sm_result == SI_SM_IDLE && smi_info->timer_running) {
 861                 /* Ok it if fails, the timer will just go off. */
 862                 if (del_timer(&smi_info->si_timer))
 863                         smi_info->timer_running = false;
 864         }
 865
 866 out:
 867         return si_sm_result;
 868 }
 869
 870 static void check_start_timer_thread(struct smi_info *smi_info)
 871 {
 872         if (smi_info->si_state == SI_NORMAL && smi_info->curr_msg == NULL) {
 873                 smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
 874
 875                 if (smi_info->thread)
 876                         wake_up_process(smi_info->thread);
 877
 878                 start_next_msg(smi_info);
 879                 smi_event_handler(smi_info, 0);
 880         }
 881 }
 882
 883 static void flush_messages(void *send_info)
 884 {
 885         struct smi_info *smi_info = send_info;
 886         enum si_sm_result result;
 887
 888         /*
 889          * Currently, this function is called only in run-to-completion
 890          * mode.  This means we are single-threaded, no need for locks.
 891          */
 892         result = smi_event_handler(smi_info, 0);
 893         while (result != SI_SM_IDLE) {
 894                 udelay(SI_SHORT_TIMEOUT_USEC);
 895                 result = smi_event_handler(smi_info, SI_SHORT_TIMEOUT_USEC);
 896         }
 897 }
 898
 899 static void sender(void                *send_info,
 900                    struct ipmi_smi_msg *msg)
 901 {
 902         struct smi_info   *smi_info = send_info;
 903         unsigned long     flags;
 904
 905         debug_timestamp("Enqueue");
 906
 907         if (smi_info->run_to_completion) {
 908                 /*
 909                  * If we are running to completion, start it.  Upper
 910                  * layer will call flush_messages to clear it out.
 911                  */
 912                 smi_info->waiting_msg = msg;
 913                 return;
 914         }
 915
 916         spin_lock_irqsave(&smi_info->si_lock, flags);
 917         /*
 918          * The following two lines don't need to be under the lock for
 919          * the lock's sake, but they do need SMP memory barriers to
 920          * avoid getting things out of order.  We are already claiming
 921          * the lock, anyway, so just do it under the lock to avoid the
 922          * ordering problem.
 923          */
 924         BUG_ON(smi_info->waiting_msg);
 925         smi_info->waiting_msg = msg;
 926         check_start_timer_thread(smi_info);
 927         spin_unlock_irqrestore(&smi_info->si_lock, flags);
 928 }
 929
 930 static void set_run_to_completion(void *send_info, bool i_run_to_completion)
 931 {
 932         struct smi_info   *smi_info = send_info;
 933
 934         smi_info->run_to_completion = i_run_to_completion;
 935         if (i_run_to_completion)
 936                 flush_messages(smi_info);
 937 }
 938
 939 /*
 940  * Use -1 in the nsec value of the busy waiting timespec to tell that
 941  * we are spinning in kipmid looking for something and not delaying
 942  * between checks
 943  */
 944 static inline void ipmi_si_set_not_busy(struct timespec64 *ts)
 945 {
 946         ts->tv_nsec = -1;
 947 }
 948 static inline int ipmi_si_is_busy(struct timespec64 *ts)
 949 {
 950         return ts->tv_nsec != -1;
 951 }
 952
 953 static inline int ipmi_thread_busy_wait(enum si_sm_result smi_result,
 954                                         const struct smi_info *smi_info,
 955                                         struct timespec64 *busy_until)
 956 {
 957         unsigned int max_busy_us = 0;
 958
 959         if (smi_info->si_num < num_max_busy_us)
 960                 max_busy_us = kipmid_max_busy_us[smi_info->si_num];
 961         if (max_busy_us == 0 || smi_result != SI_SM_CALL_WITH_DELAY)
 962                 ipmi_si_set_not_busy(busy_until);
 963         else if (!ipmi_si_is_busy(busy_until)) {
 964                 ktime_get_ts64(busy_until);
 965                 timespec64_add_ns(busy_until, max_busy_us*NSEC_PER_USEC);
 966         } else {
 967                 struct timespec64 now;
 968
 969                 ktime_get_ts64(&now);
 970                 if (unlikely(timespec64_compare(&now, busy_until) > 0)) {
 971                         ipmi_si_set_not_busy(busy_until);
 972                         return 0;
 973                 }
 974         }
 975         return 1;
 976 }
 977
 978
 979 /*
 980  * A busy-waiting loop for speeding up IPMI operation.
 981  *
 982  * Lousy hardware makes this hard.  This is only enabled for systems
 983  * that are not BT and do not have interrupts.  It starts spinning
 984  * when an operation is complete or until max_busy tells it to stop
 985  * (if that is enabled).  See the paragraph on kimid_max_busy_us in
 986  * Documentation/IPMI.txt for details.
 987  */
 988 static int ipmi_thread(void *data)
 989 {
 990         struct smi_info *smi_info = data;
 991         unsigned long flags;
 992         enum si_sm_result smi_result;
 993         struct timespec64 busy_until;
 994
 995         ipmi_si_set_not_busy(&busy_until);
 996         set_user_nice(current, MAX_NICE);
 997         while (!kthread_should_stop()) {
 998                 int busy_wait;
 999
1000                 spin_lock_irqsave(&(smi_info->si_lock), flags);
1001                 smi_result = smi_event_handler(smi_info, 0);
1002
1003                 /*
1004                  * If the driver is doing something, there is a possible
1005                  * race with the timer.  If the timer handler see idle,
1006                  * and the thread here sees something else, the timer
1007                  * handler won't restart the timer even though it is
1008                  * required.  So start it here if necessary.
1009                  */
1010                 if (smi_result != SI_SM_IDLE && !smi_info->timer_running)
1011                         smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
1012
1013                 spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1014                 busy_wait = ipmi_thread_busy_wait(smi_result, smi_info,
1015                                                   &busy_until);
1016                 if (smi_result == SI_SM_CALL_WITHOUT_DELAY)
1017                         ; /* do nothing */
1018                 else if (smi_result == SI_SM_CALL_WITH_DELAY && busy_wait)
1019                         schedule();
1020                 else if (smi_result == SI_SM_IDLE) {
1021                         if (atomic_read(&smi_info->need_watch)) {
1022                                 schedule_timeout_interruptible(100);
1023                         } else {
1024                                 /* Wait to be woken up when we are needed. */
1025                                 __set_current_state(TASK_INTERRUPTIBLE);
1026                                 schedule();
1027                         }
1028                 } else
1029                         schedule_timeout_interruptible(1);
1030         }
1031         return 0;
1032 }
1033
1034
1035 static void poll(void *send_info)
1036 {
1037         struct smi_info *smi_info = send_info;
1038         unsigned long flags = 0;
1039         bool run_to_completion = smi_info->run_to_completion;
1040
1041         /*
1042          * Make sure there is some delay in the poll loop so we can
1043          * drive time forward and timeout things.
1044          */
1045         udelay(10);
1046         if (!run_to_completion)
1047                 spin_lock_irqsave(&smi_info->si_lock, flags);
1048         smi_event_handler(smi_info, 10);
1049         if (!run_to_completion)
1050                 spin_unlock_irqrestore(&smi_info->si_lock, flags);
1051 }
1052
1053 static void request_events(void *send_info)
1054 {
1055         struct smi_info *smi_info = send_info;
1056
1057         if (!smi_info->has_event_buffer)
1058                 return;
1059
1060         atomic_set(&smi_info->req_events, 1);
1061 }
1062
1063 static void set_need_watch(void *send_info, unsigned int watch_mask)
1064 {
1065         struct smi_info *smi_info = send_info;
1066         unsigned long flags;
1067         int enable;
1068
1069         enable = !!watch_mask;
1070
1071         atomic_set(&smi_info->need_watch, enable);
1072         spin_lock_irqsave(&smi_info->si_lock, flags);
1073         check_start_timer_thread(smi_info);
1074         spin_unlock_irqrestore(&smi_info->si_lock, flags);
1075 }
1076
1077 static void smi_timeout(struct timer_list *t)
1078 {
1079         struct smi_info   *smi_info = from_timer(smi_info, t, si_timer);
1080         enum si_sm_result smi_result;
1081         unsigned long     flags;
1082         unsigned long     jiffies_now;
1083         long              time_diff;
1084         long              timeout;
1085
1086         spin_lock_irqsave(&(smi_info->si_lock), flags);
1087         debug_timestamp("Timer");
1088
1089         jiffies_now = jiffies;
1090         time_diff = (((long)jiffies_now - (long)smi_info->last_timeout_jiffies)
1091                      * SI_USEC_PER_JIFFY);
1092         smi_result = smi_event_handler(smi_info, time_diff);
1093
1094         if ((smi_info->io.irq) && (!smi_info->interrupt_disabled)) {
1095                 /* Running with interrupts, only do long timeouts. */
1096                 timeout = jiffies + SI_TIMEOUT_JIFFIES;
1097                 smi_inc_stat(smi_info, long_timeouts);
1098                 goto do_mod_timer;
1099         }
1100
1101         /*
1102          * If the state machine asks for a short delay, then shorten
1103          * the timer timeout.
1104          */
1105         if (smi_result == SI_SM_CALL_WITH_DELAY) {
1106                 smi_inc_stat(smi_info, short_timeouts);
1107                 timeout = jiffies + 1;
1108         } else {
1109                 smi_inc_stat(smi_info, long_timeouts);
1110                 timeout = jiffies + SI_TIMEOUT_JIFFIES;
1111         }
1112
1113 do_mod_timer:
1114         if (smi_result != SI_SM_IDLE)
1115                 smi_mod_timer(smi_info, timeout);
1116         else
1117                 smi_info->timer_running = false;
1118         spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1119 }
1120
1121 irqreturn_t ipmi_si_irq_handler(int irq, void *data)
1122 {
1123         struct smi_info *smi_info = data;
1124         unsigned long   flags;
1125
1126         if (smi_info->io.si_type == SI_BT)
1127                 /* We need to clear the IRQ flag for the BT interface. */
1128                 smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG,
1129                                      IPMI_BT_INTMASK_CLEAR_IRQ_BIT
1130                                      | IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
1131
1132         spin_lock_irqsave(&(smi_info->si_lock), flags);
1133
1134         smi_inc_stat(smi_info, interrupts);
1135
1136         debug_timestamp("Interrupt");
1137
1138         smi_event_handler(smi_info, 0);
1139         spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1140         return IRQ_HANDLED;
1141 }
1142
1143 static int smi_start_processing(void            *send_info,
1144                                 struct ipmi_smi *intf)
1145 {
1146         struct smi_info *new_smi = send_info;
1147         int             enable = 0;
1148
1149         new_smi->intf = intf;
1150
1151         /* Set up the timer that drives the interface. */
1152         timer_setup(&new_smi->si_timer, smi_timeout, 0);
1153         new_smi->timer_can_start = true;
1154         smi_mod_timer(new_smi, jiffies + SI_TIMEOUT_JIFFIES);
1155
1156         /* Try to claim any interrupts. */
1157         if (new_smi->io.irq_setup) {
1158                 new_smi->io.irq_handler_data = new_smi;
1159                 new_smi->io.irq_setup(&new_smi->io);
1160         }
1161
1162         /*
1163          * Check if the user forcefully enabled the daemon.
1164          */
1165         if (new_smi->si_num < num_force_kipmid)
1166                 enable = force_kipmid[new_smi->si_num];
1167         /*
1168          * The BT interface is efficient enough to not need a thread,
1169          * and there is no need for a thread if we have interrupts.
1170          */
1171         else if ((new_smi->io.si_type != SI_BT) && (!new_smi->io.irq))
1172                 enable = 1;
1173
1174         if (enable) {
1175                 new_smi->thread = kthread_run(ipmi_thread, new_smi,
1176                                               "kipmi%d", new_smi->si_num);
1177                 if (IS_ERR(new_smi->thread)) {
1178                         dev_notice(new_smi->io.dev, "Could not start"
1179                                    " kernel thread due to error %ld, only using"
1180                                    " timers to drive the interface\n",
1181                                    PTR_ERR(new_smi->thread));
1182                         new_smi->thread = NULL;
1183                 }
1184         }
1185
1186         return 0;
1187 }
1188
1189 static int get_smi_info(void *send_info, struct ipmi_smi_info *data)
1190 {
1191         struct smi_info *smi = send_info;
1192
1193         data->addr_src = smi->io.addr_source;
1194         data->dev = smi->io.dev;
1195         data->addr_info = smi->io.addr_info;
1196         get_device(smi->io.dev);
1197
1198         return 0;
1199 }
1200
1201 static void set_maintenance_mode(void *send_info, bool enable)
1202 {
1203         struct smi_info   *smi_info = send_info;
1204
1205         if (!enable)
1206                 atomic_set(&smi_info->req_events, 0);
1207 }
1208
1209 static void shutdown_smi(void *send_info);
1210 static const struct ipmi_smi_handlers handlers = {
1211         .owner                  = THIS_MODULE,
1212         .start_processing       = smi_start_processing,
1213         .shutdown               = shutdown_smi,
1214         .get_smi_info           = get_smi_info,
1215         .sender                 = sender,
1216         .request_events         = request_events,
1217         .set_need_watch         = set_need_watch,
1218         .set_maintenance_mode   = set_maintenance_mode,
1219         .set_run_to_completion  = set_run_to_completion,
1220         .flush_messages         = flush_messages,
1221         .poll                   = poll,
1222 };
1223
1224 static LIST_HEAD(smi_infos);
1225 static DEFINE_MUTEX(smi_infos_lock);
1226 static int smi_num; /* Used to sequence the SMIs */
1227
1228 static const char * const addr_space_to_str[] = { "i/o", "mem" };
1229
1230 module_param_array(force_kipmid, int, &num_force_kipmid, 0);
1231 MODULE_PARM_DESC(force_kipmid, "Force the kipmi daemon to be enabled (1) or"
1232                  " disabled(0).  Normally the IPMI driver auto-detects"
1233                  " this, but the value may be overridden by this parm.");
1234 module_param(unload_when_empty, bool, 0);
1235 MODULE_PARM_DESC(unload_when_empty, "Unload the module if no interfaces are"
1236                  " specified or found, default is 1.  Setting to 0"
1237                  " is useful for hot add of devices using hotmod.");
1238 module_param_array(kipmid_max_busy_us, uint, &num_max_busy_us, 0644);
1239 MODULE_PARM_DESC(kipmid_max_busy_us,
1240                  "Max time (in microseconds) to busy-wait for IPMI data before"
1241                  " sleeping. 0 (default) means to wait forever. Set to 100-500"
1242                  " if kipmid is using up a lot of CPU time.");
1243
1244 void ipmi_irq_finish_setup(struct si_sm_io *io)
1245 {
1246         if (io->si_type == SI_BT)
1247                 /* Enable the interrupt in the BT interface. */
1248                 io->outputb(io, IPMI_BT_INTMASK_REG,
1249                             IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
1250 }
1251
1252 void ipmi_irq_start_cleanup(struct si_sm_io *io)
1253 {
1254         if (io->si_type == SI_BT)
1255                 /* Disable the interrupt in the BT interface. */
1256                 io->outputb(io, IPMI_BT_INTMASK_REG, 0);
1257 }
1258
1259 static void std_irq_cleanup(struct si_sm_io *io)
1260 {
1261         ipmi_irq_start_cleanup(io);
1262         free_irq(io->irq, io->irq_handler_data);
1263 }
1264
1265 int ipmi_std_irq_setup(struct si_sm_io *io)
1266 {
1267         int rv;
1268
1269         if (!io->irq)
1270                 return 0;
1271
1272         rv = request_irq(io->irq,
1273                          ipmi_si_irq_handler,
1274                          IRQF_SHARED,
1275                          DEVICE_NAME,
1276                          io->irq_handler_data);
1277         if (rv) {
1278                 dev_warn(io->dev, "%s unable to claim interrupt %d,"
1279                          " running polled\n",
1280                          DEVICE_NAME, io->irq);
1281                 io->irq = 0;
1282         } else {
1283                 io->irq_cleanup = std_irq_cleanup;
1284                 ipmi_irq_finish_setup(io);
1285                 dev_info(io->dev, "Using irq %d\n", io->irq);
1286         }
1287
1288         return rv;
1289 }
1290
1291 static int wait_for_msg_done(struct smi_info *smi_info)
1292 {
1293         enum si_sm_result     smi_result;
1294
1295         smi_result = smi_info->handlers->event(smi_info->si_sm, 0);
1296         for (;;) {
1297                 if (smi_result == SI_SM_CALL_WITH_DELAY ||
1298                     smi_result == SI_SM_CALL_WITH_TICK_DELAY) {
1299                         schedule_timeout_uninterruptible(1);
1300                         smi_result = smi_info->handlers->event(
1301                                 smi_info->si_sm, jiffies_to_usecs(1));
1302                 } else if (smi_result == SI_SM_CALL_WITHOUT_DELAY) {
1303                         smi_result = smi_info->handlers->event(
1304                                 smi_info->si_sm, 0);
1305                 } else
1306                         break;
1307         }
1308         if (smi_result == SI_SM_HOSED)
1309                 /*
1310                  * We couldn't get the state machine to run, so whatever's at
1311                  * the port is probably not an IPMI SMI interface.
1312                  */
1313                 return -ENODEV;
1314
1315         return 0;
1316 }
1317
1318 static int try_get_dev_id(struct smi_info *smi_info)
1319 {
1320         unsigned char         msg[2];
1321         unsigned char         *resp;
1322         unsigned long         resp_len;
1323         int                   rv = 0;
1324
1325         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1326         if (!resp)
1327                 return -ENOMEM;
1328
1329         /*
1330          * Do a Get Device ID command, since it comes back with some
1331          * useful info.
1332          */
1333         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1334         msg[1] = IPMI_GET_DEVICE_ID_CMD;
1335         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1336
1337         rv = wait_for_msg_done(smi_info);
1338         if (rv)
1339                 goto out;
1340
1341         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1342                                                   resp, IPMI_MAX_MSG_LENGTH);
1343
1344         /* Check and record info from the get device id, in case we need it. */
1345         rv = ipmi_demangle_device_id(resp[0] >> 2, resp[1],
1346                         resp + 2, resp_len - 2, &smi_info->device_id);
1347
1348 out:
1349         kfree(resp);
1350         return rv;
1351 }
1352
1353 static int get_global_enables(struct smi_info *smi_info, u8 *enables)
1354 {
1355         unsigned char         msg[3];
1356         unsigned char         *resp;
1357         unsigned long         resp_len;
1358         int                   rv;
1359
1360         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1361         if (!resp)
1362                 return -ENOMEM;
1363
1364         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1365         msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
1366         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1367
1368         rv = wait_for_msg_done(smi_info);
1369         if (rv) {
1370                 dev_warn(smi_info->io.dev,
1371                          "Error getting response from get global enables command: %d\n",
1372                          rv);
1373                 goto out;
1374         }
1375
1376         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1377                                                   resp, IPMI_MAX_MSG_LENGTH);
1378
1379         if (resp_len < 4 ||
1380                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1381                         resp[1] != IPMI_GET_BMC_GLOBAL_ENABLES_CMD   ||
1382                         resp[2] != 0) {
1383                 dev_warn(smi_info->io.dev,
1384                          "Invalid return from get global enables command: %ld %x %x %x\n",
1385                          resp_len, resp[0], resp[1], resp[2]);
1386                 rv = -EINVAL;
1387                 goto out;
1388         } else {
1389                 *enables = resp[3];
1390         }
1391
1392 out:
1393         kfree(resp);
1394         return rv;
1395 }
1396
1397 /*
1398  * Returns 1 if it gets an error from the command.
1399  */
1400 static int set_global_enables(struct smi_info *smi_info, u8 enables)
1401 {
1402         unsigned char         msg[3];
1403         unsigned char         *resp;
1404         unsigned long         resp_len;
1405         int                   rv;
1406
1407         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1408         if (!resp)
1409                 return -ENOMEM;
1410
1411         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1412         msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
1413         msg[2] = enables;
1414         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 3);
1415
1416         rv = wait_for_msg_done(smi_info);
1417         if (rv) {
1418                 dev_warn(smi_info->io.dev,
1419                          "Error getting response from set global enables command: %d\n",
1420                          rv);
1421                 goto out;
1422         }
1423
1424         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1425                                                   resp, IPMI_MAX_MSG_LENGTH);
1426
1427         if (resp_len < 3 ||
1428                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1429                         resp[1] != IPMI_SET_BMC_GLOBAL_ENABLES_CMD) {
1430                 dev_warn(smi_info->io.dev,
1431                          "Invalid return from set global enables command: %ld %x %x\n",
1432                          resp_len, resp[0], resp[1]);
1433                 rv = -EINVAL;
1434                 goto out;
1435         }
1436
1437         if (resp[2] != 0)
1438                 rv = 1;
1439
1440 out:
1441         kfree(resp);
1442         return rv;
1443 }
1444
1445 /*
1446  * Some BMCs do not support clearing the receive irq bit in the global
1447  * enables (even if they don't support interrupts on the BMC).  Check
1448  * for this and handle it properly.
1449  */
1450 static void check_clr_rcv_irq(struct smi_info *smi_info)
1451 {
1452         u8 enables = 0;
1453         int rv;
1454
1455         rv = get_global_enables(smi_info, &enables);
1456         if (!rv) {
1457                 if ((enables & IPMI_BMC_RCV_MSG_INTR) == 0)
1458                         /* Already clear, should work ok. */
1459                         return;
1460
1461                 enables &= ~IPMI_BMC_RCV_MSG_INTR;
1462                 rv = set_global_enables(smi_info, enables);
1463         }
1464
1465         if (rv < 0) {
1466                 dev_err(smi_info->io.dev,
1467                         "Cannot check clearing the rcv irq: %d\n", rv);
1468                 return;
1469         }
1470
1471         if (rv) {
1472                 /*
1473                  * An error when setting the event buffer bit means
1474                  * clearing the bit is not supported.
1475                  */
1476                 dev_warn(smi_info->io.dev,
1477                          "The BMC does not support clearing the recv irq bit, compensating, but the BMC needs to be fixed.\n");
1478                 smi_info->cannot_disable_irq = true;
1479         }
1480 }
1481
1482 /*
1483  * Some BMCs do not support setting the interrupt bits in the global
1484  * enables even if they support interrupts.  Clearly bad, but we can
1485  * compensate.
1486  */
1487 static void check_set_rcv_irq(struct smi_info *smi_info)
1488 {
1489         u8 enables = 0;
1490         int rv;
1491
1492         if (!smi_info->io.irq)
1493                 return;
1494
1495         rv = get_global_enables(smi_info, &enables);
1496         if (!rv) {
1497                 enables |= IPMI_BMC_RCV_MSG_INTR;
1498                 rv = set_global_enables(smi_info, enables);
1499         }
1500
1501         if (rv < 0) {
1502                 dev_err(smi_info->io.dev,
1503                         "Cannot check setting the rcv irq: %d\n", rv);
1504                 return;
1505         }
1506
1507         if (rv) {
1508                 /*
1509                  * An error when setting the event buffer bit means
1510                  * setting the bit is not supported.
1511                  */
1512                 dev_warn(smi_info->io.dev,
1513                          "The BMC does not support setting the recv irq bit, compensating, but the BMC needs to be fixed.\n");
1514                 smi_info->cannot_disable_irq = true;
1515                 smi_info->irq_enable_broken = true;
1516         }
1517 }
1518
1519 static int try_enable_event_buffer(struct smi_info *smi_info)
1520 {
1521         unsigned char         msg[3];
1522         unsigned char         *resp;
1523         unsigned long         resp_len;
1524         int                   rv = 0;
1525
1526         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1527         if (!resp)
1528                 return -ENOMEM;
1529
1530         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1531         msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
1532         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1533
1534         rv = wait_for_msg_done(smi_info);
1535         if (rv) {
1536                 pr_warn("Error getting response from get global enables command, the event buffer is not enabled\n");
1537                 goto out;
1538         }
1539
1540         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1541                                                   resp, IPMI_MAX_MSG_LENGTH);
1542
1543         if (resp_len < 4 ||
1544                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1545                         resp[1] != IPMI_GET_BMC_GLOBAL_ENABLES_CMD   ||
1546                         resp[2] != 0) {
1547                 pr_warn("Invalid return from get global enables command, cannot enable the event buffer\n");
1548                 rv = -EINVAL;
1549                 goto out;
1550         }
1551
1552         if (resp[3] & IPMI_BMC_EVT_MSG_BUFF) {
1553                 /* buffer is already enabled, nothing to do. */
1554                 smi_info->supports_event_msg_buff = true;
1555                 goto out;
1556         }
1557
1558         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1559         msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
1560         msg[2] = resp[3] | IPMI_BMC_EVT_MSG_BUFF;
1561         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 3);
1562
1563         rv = wait_for_msg_done(smi_info);
1564         if (rv) {
1565                 pr_warn("Error getting response from set global, enables command, the event buffer is not enabled\n");
1566                 goto out;
1567         }
1568
1569         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1570                                                   resp, IPMI_MAX_MSG_LENGTH);
1571
1572         if (resp_len < 3 ||
1573                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1574                         resp[1] != IPMI_SET_BMC_GLOBAL_ENABLES_CMD) {
1575                 pr_warn("Invalid return from get global, enables command, not enable the event buffer\n");
1576                 rv = -EINVAL;
1577                 goto out;
1578         }
1579
1580         if (resp[2] != 0)
1581                 /*
1582                  * An error when setting the event buffer bit means
1583                  * that the event buffer is not supported.
1584                  */
1585                 rv = -ENOENT;
1586         else
1587                 smi_info->supports_event_msg_buff = true;
1588
1589 out:
1590         kfree(resp);
1591         return rv;
1592 }
1593
1594 #define IPMI_SI_ATTR(name) \
1595 static ssize_t ipmi_##name##_show(struct device *dev,                   \
1596                                   struct device_attribute *attr,        \
1597                                   char *buf)                            \
1598 {                                                                       \
1599         struct smi_info *smi_info = dev_get_drvdata(dev);               \
1600                                                                         \
1601         return snprintf(buf, 10, "%u\n", smi_get_stat(smi_info, name)); \
1602 }                                                                       \
1603 static DEVICE_ATTR(name, S_IRUGO, ipmi_##name##_show, NULL)
1604
1605 static ssize_t ipmi_type_show(struct device *dev,
1606                               struct device_attribute *attr,
1607                               char *buf)
1608 {
1609         struct smi_info *smi_info = dev_get_drvdata(dev);
1610
1611         return snprintf(buf, 10, "%s\n", si_to_str[smi_info->io.si_type]);
1612 }
1613 static DEVICE_ATTR(type, S_IRUGO, ipmi_type_show, NULL);
1614
1615 static ssize_t ipmi_interrupts_enabled_show(struct device *dev,
1616                                             struct device_attribute *attr,
1617                                             char *buf)
1618 {
1619         struct smi_info *smi_info = dev_get_drvdata(dev);
1620         int enabled = smi_info->io.irq && !smi_info->interrupt_disabled;
1621
1622         return snprintf(buf, 10, "%d\n", enabled);
1623 }
1624 static DEVICE_ATTR(interrupts_enabled, S_IRUGO,
1625                    ipmi_interrupts_enabled_show, NULL);
1626
1627 IPMI_SI_ATTR(short_timeouts);
1628 IPMI_SI_ATTR(long_timeouts);
1629 IPMI_SI_ATTR(idles);
1630 IPMI_SI_ATTR(interrupts);
1631 IPMI_SI_ATTR(attentions);
1632 IPMI_SI_ATTR(flag_fetches);
1633 IPMI_SI_ATTR(hosed_count);
1634 IPMI_SI_ATTR(complete_transactions);
1635 IPMI_SI_ATTR(events);
1636 IPMI_SI_ATTR(watchdog_pretimeouts);
1637 IPMI_SI_ATTR(incoming_messages);
1638
1639 static ssize_t ipmi_params_show(struct device *dev,
1640                                 struct device_attribute *attr,
1641                                 char *buf)
1642 {
1643         struct smi_info *smi_info = dev_get_drvdata(dev);
1644
1645         return snprintf(buf, 200,
1646                         "%s,%s,0x%lx,rsp=%d,rsi=%d,rsh=%d,irq=%d,ipmb=%d\n",
1647                         si_to_str[smi_info->io.si_type],
1648                         addr_space_to_str[smi_info->io.addr_space],
1649                         smi_info->io.addr_data,
1650                         smi_info->io.regspacing,
1651                         smi_info->io.regsize,
1652                         smi_info->io.regshift,
1653                         smi_info->io.irq,
1654                         smi_info->io.slave_addr);
1655 }
1656 static DEVICE_ATTR(params, S_IRUGO, ipmi_params_show, NULL);
1657
1658 static struct attribute *ipmi_si_dev_attrs[] = {
1659         &dev_attr_type.attr,
1660         &dev_attr_interrupts_enabled.attr,
1661         &dev_attr_short_timeouts.attr,
1662         &dev_attr_long_timeouts.attr,
1663         &dev_attr_idles.attr,
1664         &dev_attr_interrupts.attr,
1665         &dev_attr_attentions.attr,
1666         &dev_attr_flag_fetches.attr,
1667         &dev_attr_hosed_count.attr,
1668         &dev_attr_complete_transactions.attr,
1669         &dev_attr_events.attr,
1670         &dev_attr_watchdog_pretimeouts.attr,
1671         &dev_attr_incoming_messages.attr,
1672         &dev_attr_params.attr,
1673         NULL
1674 };
1675
1676 static const struct attribute_group ipmi_si_dev_attr_group = {
1677         .attrs          = ipmi_si_dev_attrs,
1678 };
1679
1680 /*
1681  * oem_data_avail_to_receive_msg_avail
1682  * @info - smi_info structure with msg_flags set
1683  *
1684  * Converts flags from OEM_DATA_AVAIL to RECEIVE_MSG_AVAIL
1685  * Returns 1 indicating need to re-run handle_flags().
1686  */
1687 static int oem_data_avail_to_receive_msg_avail(struct smi_info *smi_info)
1688 {
1689         smi_info->msg_flags = ((smi_info->msg_flags & ~OEM_DATA_AVAIL) |
1690                                RECEIVE_MSG_AVAIL);
1691         return 1;
1692 }
1693
1694 /*
1695  * setup_dell_poweredge_oem_data_handler
1696  * @info - smi_info.device_id must be populated
1697  *
1698  * Systems that match, but have firmware version < 1.40 may assert
1699  * OEM0_DATA_AVAIL on their own, without being told via Set Flags that
1700  * it's safe to do so.  Such systems will de-assert OEM1_DATA_AVAIL
1701  * upon receipt of IPMI_GET_MSG_CMD, so we should treat these flags
1702  * as RECEIVE_MSG_AVAIL instead.
1703  *
1704  * As Dell has no plans to release IPMI 1.5 firmware that *ever*
1705  * assert the OEM[012] bits, and if it did, the driver would have to
1706  * change to handle that properly, we don't actually check for the
1707  * firmware version.
1708  * Device ID = 0x20                BMC on PowerEdge 8G servers
1709  * Device Revision = 0x80
1710  * Firmware Revision1 = 0x01       BMC version 1.40
1711  * Firmware Revision2 = 0x40       BCD encoded
1712  * IPMI Version = 0x51             IPMI 1.5
1713  * Manufacturer ID = A2 02 00      Dell IANA
1714  *
1715  * Additionally, PowerEdge systems with IPMI < 1.5 may also assert
1716  * OEM0_DATA_AVAIL and needs to be treated as RECEIVE_MSG_AVAIL.
1717  *
1718  */
1719 #define DELL_POWEREDGE_8G_BMC_DEVICE_ID  0x20
1720 #define DELL_POWEREDGE_8G_BMC_DEVICE_REV 0x80
1721 #define DELL_POWEREDGE_8G_BMC_IPMI_VERSION 0x51
1722 #define DELL_IANA_MFR_ID 0x0002a2
1723 static void setup_dell_poweredge_oem_data_handler(struct smi_info *smi_info)
1724 {
1725         struct ipmi_device_id *id = &smi_info->device_id;
1726         if (id->manufacturer_id == DELL_IANA_MFR_ID) {
1727                 if (id->device_id       == DELL_POWEREDGE_8G_BMC_DEVICE_ID  &&
1728                     id->device_revision == DELL_POWEREDGE_8G_BMC_DEVICE_REV &&
1729                     id->ipmi_version   == DELL_POWEREDGE_8G_BMC_IPMI_VERSION) {
1730                         smi_info->oem_data_avail_handler =
1731                                 oem_data_avail_to_receive_msg_avail;
1732                 } else if (ipmi_version_major(id) < 1 ||
1733                            (ipmi_version_major(id) == 1 &&
1734                             ipmi_version_minor(id) < 5)) {
1735                         smi_info->oem_data_avail_handler =
1736                                 oem_data_avail_to_receive_msg_avail;
1737                 }
1738         }
1739 }
1740
1741 #define CANNOT_RETURN_REQUESTED_LENGTH 0xCA
1742 static void return_hosed_msg_badsize(struct smi_info *smi_info)
1743 {
1744         struct ipmi_smi_msg *msg = smi_info->curr_msg;
1745
1746         /* Make it a response */
1747         msg->rsp[0] = msg->data[0] | 4;
1748         msg->rsp[1] = msg->data[1];
1749         msg->rsp[2] = CANNOT_RETURN_REQUESTED_LENGTH;
1750         msg->rsp_size = 3;
1751         smi_info->curr_msg = NULL;
1752         deliver_recv_msg(smi_info, msg);
1753 }
1754
1755 /*
1756  * dell_poweredge_bt_xaction_handler
1757  * @info - smi_info.device_id must be populated
1758  *
1759  * Dell PowerEdge servers with the BT interface (x6xx and 1750) will
1760  * not respond to a Get SDR command if the length of the data
1761  * requested is exactly 0x3A, which leads to command timeouts and no
1762  * data returned.  This intercepts such commands, and causes userspace
1763  * callers to try again with a different-sized buffer, which succeeds.
1764  */
1765
1766 #define STORAGE_NETFN 0x0A
1767 #define STORAGE_CMD_GET_SDR 0x23
1768 static int dell_poweredge_bt_xaction_handler(struct notifier_block *self,
1769                                              unsigned long unused,
1770                                              void *in)
1771 {
1772         struct smi_info *smi_info = in;
1773         unsigned char *data = smi_info->curr_msg->data;
1774         unsigned int size   = smi_info->curr_msg->data_size;
1775         if (size >= 8 &&
1776             (data[0]>>2) == STORAGE_NETFN &&
1777             data[1] == STORAGE_CMD_GET_SDR &&
1778             data[7] == 0x3A) {
1779                 return_hosed_msg_badsize(smi_info);
1780                 return NOTIFY_STOP;
1781         }
1782         return NOTIFY_DONE;
1783 }
1784
1785 static struct notifier_block dell_poweredge_bt_xaction_notifier = {
1786         .notifier_call  = dell_poweredge_bt_xaction_handler,
1787 };
1788
1789 /*
1790  * setup_dell_poweredge_bt_xaction_handler
1791  * @info - smi_info.device_id must be filled in already
1792  *
1793  * Fills in smi_info.device_id.start_transaction_pre_hook
1794  * when we know what function to use there.
1795  */
1796 static void
1797 setup_dell_poweredge_bt_xaction_handler(struct smi_info *smi_info)
1798 {
1799         struct ipmi_device_id *id = &smi_info->device_id;
1800         if (id->manufacturer_id == DELL_IANA_MFR_ID &&
1801             smi_info->io.si_type == SI_BT)
1802                 register_xaction_notifier(&dell_poweredge_bt_xaction_notifier);
1803 }
1804
1805 /*
1806  * setup_oem_data_handler
1807  * @info - smi_info.device_id must be filled in already
1808  *
1809  * Fills in smi_info.device_id.oem_data_available_handler
1810  * when we know what function to use there.
1811  */
1812
1813 static void setup_oem_data_handler(struct smi_info *smi_info)
1814 {
1815         setup_dell_poweredge_oem_data_handler(smi_info);
1816 }
1817
1818 static void setup_xaction_handlers(struct smi_info *smi_info)
1819 {
1820         setup_dell_poweredge_bt_xaction_handler(smi_info);
1821 }
1822
1823 static void check_for_broken_irqs(struct smi_info *smi_info)
1824 {
1825         check_clr_rcv_irq(smi_info);
1826         check_set_rcv_irq(smi_info);
1827 }
1828
1829 static inline void stop_timer_and_thread(struct smi_info *smi_info)
1830 {
1831         if (smi_info->thread != NULL) {
1832                 kthread_stop(smi_info->thread);
1833                 smi_info->thread = NULL;
1834         }
1835
1836         smi_info->timer_can_start = false;
1837         if (smi_info->timer_running)
1838                 del_timer_sync(&smi_info->si_timer);
1839 }
1840
1841 static struct smi_info *find_dup_si(struct smi_info *info)
1842 {
1843         struct smi_info *e;
1844
1845         list_for_each_entry(e, &smi_infos, link) {
1846                 if (e->io.addr_space != info->io.addr_space)
1847                         continue;
1848                 if (e->io.addr_data == info->io.addr_data) {
1849                         /*
1850                          * This is a cheap hack, ACPI doesn't have a defined
1851                          * slave address but SMBIOS does.  Pick it up from
1852                          * any source that has it available.
1853                          */
1854                         if (info->io.slave_addr && !e->io.slave_addr)
1855                                 e->io.slave_addr = info->io.slave_addr;
1856                         return e;
1857                 }
1858         }
1859
1860         return NULL;
1861 }
1862
1863 int ipmi_si_add_smi(struct si_sm_io *io)
1864 {
1865         int rv = 0;
1866         struct smi_info *new_smi, *dup;
1867
1868         /*
1869          * If the user gave us a hard-coded device at the same
1870          * address, they presumably want us to use it and not what is
1871          * in the firmware.
1872          */
1873         if (io->addr_source != SI_HARDCODED &&
1874             ipmi_si_hardcode_match(io->addr_space, io->addr_data)) {
1875                 dev_info(io->dev,
1876                          "Hard-coded device at this address already exists");
1877                 return -ENODEV;
1878         }
1879
1880         if (!io->io_setup) {
1881                 if (io->addr_space == IPMI_IO_ADDR_SPACE) {
1882                         io->io_setup = ipmi_si_port_setup;
1883                 } else if (io->addr_space == IPMI_MEM_ADDR_SPACE) {
1884                         io->io_setup = ipmi_si_mem_setup;
1885                 } else {
1886                         return -EINVAL;
1887                 }
1888         }
1889
1890         new_smi = kzalloc(sizeof(*new_smi), GFP_KERNEL);
1891         if (!new_smi)
1892                 return -ENOMEM;
1893         spin_lock_init(&new_smi->si_lock);
1894
1895         new_smi->io = *io;
1896
1897         mutex_lock(&smi_infos_lock);
1898         dup = find_dup_si(new_smi);
1899         if (dup) {
1900                 if (new_smi->io.addr_source == SI_ACPI &&
1901                     dup->io.addr_source == SI_SMBIOS) {
1902                         /* We prefer ACPI over SMBIOS. */
1903                         dev_info(dup->io.dev,
1904                                  "Removing SMBIOS-specified %s state machine in favor of ACPI\n",
1905                                  si_to_str[new_smi->io.si_type]);
1906                         cleanup_one_si(dup);
1907                 } else {
1908                         dev_info(new_smi->io.dev,
1909                                  "%s-specified %s state machine: duplicate\n",
1910                                  ipmi_addr_src_to_str(new_smi->io.addr_source),
1911                                  si_to_str[new_smi->io.si_type]);
1912                         rv = -EBUSY;
1913                         kfree(new_smi);
1914                         goto out_err;
1915                 }
1916         }
1917
1918         pr_info("Adding %s-specified %s state machine\n",
1919                 ipmi_addr_src_to_str(new_smi->io.addr_source),
1920                 si_to_str[new_smi->io.si_type]);
1921
1922         list_add_tail(&new_smi->link, &smi_infos);
1923
1924         if (initialized)
1925                 rv = try_smi_init(new_smi);
1926 out_err:
1927         mutex_unlock(&smi_infos_lock);
1928         return rv;
1929 }
1930
1931 /*
1932  * Try to start up an interface.  Must be called with smi_infos_lock
1933  * held, primarily to keep smi_num consistent, we only one to do these
1934  * one at a time.
1935  */
1936 static int try_smi_init(struct smi_info *new_smi)
1937 {
1938         int rv = 0;
1939         int i;
1940         char *init_name = NULL;
1941
1942         pr_info("Trying %s-specified %s state machine at %s address 0x%lx, slave address 0x%x, irq %d\n",
1943                 ipmi_addr_src_to_str(new_smi->io.addr_source),
1944                 si_to_str[new_smi->io.si_type],
1945                 addr_space_to_str[new_smi->io.addr_space],
1946                 new_smi->io.addr_data,
1947                 new_smi->io.slave_addr, new_smi->io.irq);
1948
1949         switch (new_smi->io.si_type) {
1950         case SI_KCS:
1951                 new_smi->handlers = &kcs_smi_handlers;
1952                 break;
1953
1954         case SI_SMIC:
1955                 new_smi->handlers = &smic_smi_handlers;
1956                 break;
1957
1958         case SI_BT:
1959                 new_smi->handlers = &bt_smi_handlers;
1960                 break;
1961
1962         default:
1963                 /* No support for anything else yet. */
1964                 rv = -EIO;
1965                 goto out_err;
1966         }
1967
1968         new_smi->si_num = smi_num;
1969
1970         /* Do this early so it's available for logs. */
1971         if (!new_smi->io.dev) {
1972                 init_name = kasprintf(GFP_KERNEL, "ipmi_si.%d",
1973                                       new_smi->si_num);
1974
1975                 /*
1976                  * If we don't already have a device from something
1977                  * else (like PCI), then register a new one.
1978                  */
1979                 new_smi->pdev = platform_device_alloc("ipmi_si",
1980                                                       new_smi->si_num);
1981                 if (!new_smi->pdev) {
1982                         pr_err("Unable to allocate platform device\n");
1983                         rv = -ENOMEM;
1984                         goto out_err;
1985                 }
1986                 new_smi->io.dev = &new_smi->pdev->dev;
1987                 new_smi->io.dev->driver = &ipmi_platform_driver.driver;
1988                 /* Nulled by device_add() */
1989                 new_smi->io.dev->init_name = init_name;
1990         }
1991
1992         /* Allocate the state machine's data and initialize it. */
1993         new_smi->si_sm = kmalloc(new_smi->handlers->size(), GFP_KERNEL);
1994         if (!new_smi->si_sm) {
1995                 rv = -ENOMEM;
1996                 goto out_err;
1997         }
1998         new_smi->io.io_size = new_smi->handlers->init_data(new_smi->si_sm,
1999                                                            &new_smi->io);
2000
2001         /* Now that we know the I/O size, we can set up the I/O. */
2002         rv = new_smi->io.io_setup(&new_smi->io);
2003         if (rv) {
2004                 dev_err(new_smi->io.dev, "Could not set up I/O space\n");
2005                 goto out_err;
2006         }
2007
2008         /* Do low-level detection first. */
2009         if (new_smi->handlers->detect(new_smi->si_sm)) {
2010                 if (new_smi->io.addr_source)
2011                         dev_err(new_smi->io.dev,
2012                                 "Interface detection failed\n");
2013                 rv = -ENODEV;
2014                 goto out_err;
2015         }
2016
2017         /*
2018          * Attempt a get device id command.  If it fails, we probably
2019          * don't have a BMC here.
2020          */
2021         rv = try_get_dev_id(new_smi);
2022         if (rv) {
2023                 if (new_smi->io.addr_source)
2024                         dev_err(new_smi->io.dev,
2025                                "There appears to be no BMC at this location\n");
2026                 goto out_err;
2027         }
2028
2029         setup_oem_data_handler(new_smi);
2030         setup_xaction_handlers(new_smi);
2031         check_for_broken_irqs(new_smi);
2032
2033         new_smi->waiting_msg = NULL;
2034         new_smi->curr_msg = NULL;
2035         atomic_set(&new_smi->req_events, 0);
2036         new_smi->run_to_completion = false;
2037         for (i = 0; i < SI_NUM_STATS; i++)
2038                 atomic_set(&new_smi->stats[i], 0);
2039
2040         new_smi->interrupt_disabled = true;
2041         atomic_set(&new_smi->need_watch, 0);
2042
2043         rv = try_enable_event_buffer(new_smi);
2044         if (rv == 0)
2045                 new_smi->has_event_buffer = true;
2046
2047         /*
2048          * Start clearing the flags before we enable interrupts or the
2049          * timer to avoid racing with the timer.
2050          */
2051         start_clear_flags(new_smi);
2052
2053         /*
2054          * IRQ is defined to be set when non-zero.  req_events will
2055          * cause a global flags check that will enable interrupts.
2056          */
2057         if (new_smi->io.irq) {
2058                 new_smi->interrupt_disabled = false;
2059                 atomic_set(&new_smi->req_events, 1);
2060         }
2061
2062         if (new_smi->pdev && !new_smi->pdev_registered) {
2063                 rv = platform_device_add(new_smi->pdev);
2064                 if (rv) {
2065                         dev_err(new_smi->io.dev,
2066                                 "Unable to register system interface device: %d\n",
2067                                 rv);
2068                         goto out_err;
2069                 }
2070                 new_smi->pdev_registered = true;
2071         }
2072
2073         dev_set_drvdata(new_smi->io.dev, new_smi);
2074         rv = device_add_group(new_smi->io.dev, &ipmi_si_dev_attr_group);
2075         if (rv) {
2076                 dev_err(new_smi->io.dev,
2077                         "Unable to add device attributes: error %d\n",
2078                         rv);
2079                 goto out_err;
2080         }
2081         new_smi->dev_group_added = true;
2082
2083         rv = ipmi_register_smi(&handlers,
2084                                new_smi,
2085                                new_smi->io.dev,
2086                                new_smi->io.slave_addr);
2087         if (rv) {
2088                 dev_err(new_smi->io.dev,
2089                         "Unable to register device: error %d\n",
2090                         rv);
2091                 goto out_err;
2092         }
2093
2094         /* Don't increment till we know we have succeeded. */
2095         smi_num++;
2096
2097         dev_info(new_smi->io.dev, "IPMI %s interface initialized\n",
2098                  si_to_str[new_smi->io.si_type]);
2099
2100         WARN_ON(new_smi->io.dev->init_name != NULL);
2101
2102  out_err:
2103         if (rv && new_smi->io.io_cleanup) {
2104                 new_smi->io.io_cleanup(&new_smi->io);
2105                 new_smi->io.io_cleanup = NULL;
2106         }
2107
2108         kfree(init_name);
2109         return rv;
2110 }
2111
2112 static int __init init_ipmi_si(void)
2113 {
2114         struct smi_info *e;
2115         enum ipmi_addr_src type = SI_INVALID;
2116
2117         if (initialized)
2118                 return 0;
2119
2120         ipmi_hardcode_init();
2121
2122         pr_info("IPMI System Interface driver\n");
2123
2124         ipmi_si_platform_init();
2125
2126         ipmi_si_pci_init();
2127
2128         ipmi_si_parisc_init();
2129
2130         /* We prefer devices with interrupts, but in the case of a machine
2131            with multiple BMCs we assume that there will be several instances
2132            of a given type so if we succeed in registering a type then also
2133            try to register everything else of the same type */
2134         mutex_lock(&smi_infos_lock);
2135         list_for_each_entry(e, &smi_infos, link) {
2136                 /* Try to register a device if it has an IRQ and we either
2137                    haven't successfully registered a device yet or this
2138                    device has the same type as one we successfully registered */
2139                 if (e->io.irq && (!type || e->io.addr_source == type)) {
2140                         if (!try_smi_init(e)) {
2141                                 type = e->io.addr_source;
2142                         }
2143                 }
2144         }
2145
2146         /* type will only have been set if we successfully registered an si */
2147         if (type)
2148                 goto skip_fallback_noirq;
2149
2150         /* Fall back to the preferred device */
2151
2152         list_for_each_entry(e, &smi_infos, link) {
2153                 if (!e->io.irq && (!type || e->io.addr_source == type)) {
2154                         if (!try_smi_init(e)) {
2155                                 type = e->io.addr_source;
2156                         }
2157                 }
2158         }
2159
2160 skip_fallback_noirq:
2161         initialized = 1;
2162         mutex_unlock(&smi_infos_lock);
2163
2164         if (type)
2165                 return 0;
2166
2167         mutex_lock(&smi_infos_lock);
2168         if (unload_when_empty && list_empty(&smi_infos)) {
2169                 mutex_unlock(&smi_infos_lock);
2170                 cleanup_ipmi_si();
2171                 pr_warn("Unable to find any System Interface(s)\n");
2172                 return -ENODEV;
2173         } else {
2174                 mutex_unlock(&smi_infos_lock);
2175                 return 0;
2176         }
2177 }
2178 module_init(init_ipmi_si);
2179
2180 static void shutdown_smi(void *send_info)
2181 {
2182         struct smi_info *smi_info = send_info;
2183
2184         if (smi_info->dev_group_added) {
2185                 device_remove_group(smi_info->io.dev, &ipmi_si_dev_attr_group);
2186                 smi_info->dev_group_added = false;
2187         }
2188         if (smi_info->io.dev)
2189                 dev_set_drvdata(smi_info->io.dev, NULL);
2190
2191         /*
2192          * Make sure that interrupts, the timer and the thread are
2193          * stopped and will not run again.
2194          */
2195         smi_info->interrupt_disabled = true;
2196         if (smi_info->io.irq_cleanup) {
2197                 smi_info->io.irq_cleanup(&smi_info->io);
2198                 smi_info->io.irq_cleanup = NULL;
2199         }
2200         stop_timer_and_thread(smi_info);
2201
2202         /*
2203          * Wait until we know that we are out of any interrupt
2204          * handlers might have been running before we freed the
2205          * interrupt.
2206          */
2207         synchronize_rcu();
2208
2209         /*
2210          * Timeouts are stopped, now make sure the interrupts are off
2211          * in the BMC.  Note that timers and CPU interrupts are off,
2212          * so no need for locks.
2213          */
2214         while (smi_info->curr_msg || (smi_info->si_state != SI_NORMAL)) {
2215                 poll(smi_info);
2216                 schedule_timeout_uninterruptible(1);
2217         }
2218         if (smi_info->handlers)
2219                 disable_si_irq(smi_info);
2220         while (smi_info->curr_msg || (smi_info->si_state != SI_NORMAL)) {
2221                 poll(smi_info);
2222                 schedule_timeout_uninterruptible(1);
2223         }
2224         if (smi_info->handlers)
2225                 smi_info->handlers->cleanup(smi_info->si_sm);
2226
2227         if (smi_info->io.addr_source_cleanup) {
2228                 smi_info->io.addr_source_cleanup(&smi_info->io);
2229                 smi_info->io.addr_source_cleanup = NULL;
2230         }
2231         if (smi_info->io.io_cleanup) {
2232                 smi_info->io.io_cleanup(&smi_info->io);
2233                 smi_info->io.io_cleanup = NULL;
2234         }
2235
2236         kfree(smi_info->si_sm);
2237         smi_info->si_sm = NULL;
2238
2239         smi_info->intf = NULL;
2240 }
2241
2242 /*
2243  * Must be called with smi_infos_lock held, to serialize the
2244  * smi_info->intf check.
2245  */
2246 static void cleanup_one_si(struct smi_info *smi_info)
2247 {
2248         if (!smi_info)
2249                 return;
2250
2251         list_del(&smi_info->link);
2252
2253         if (smi_info->intf)
2254                 ipmi_unregister_smi(smi_info->intf);
2255
2256         if (smi_info->pdev) {
2257                 if (smi_info->pdev_registered)
2258                         platform_device_unregister(smi_info->pdev);
2259                 else
2260                         platform_device_put(smi_info->pdev);
2261         }
2262
2263         kfree(smi_info);
2264 }
2265
2266 int ipmi_si_remove_by_dev(struct device *dev)
2267 {
2268         struct smi_info *e;
2269         int rv = -ENOENT;
2270
2271         mutex_lock(&smi_infos_lock);
2272         list_for_each_entry(e, &smi_infos, link) {
2273                 if (e->io.dev == dev) {
2274                         cleanup_one_si(e);
2275                         rv = 0;
2276                         break;
2277                 }
2278         }
2279         mutex_unlock(&smi_infos_lock);
2280
2281         return rv;
2282 }
2283
2284 void ipmi_si_remove_by_data(int addr_space, enum si_type si_type,
2285                             unsigned long addr)
2286 {
2287         /* remove */
2288         struct smi_info *e, *tmp_e;
2289
2290         mutex_lock(&smi_infos_lock);
2291         list_for_each_entry_safe(e, tmp_e, &smi_infos, link) {
2292                 if (e->io.addr_space != addr_space)
2293                         continue;
2294                 if (e->io.si_type != si_type)
2295                         continue;
2296                 if (e->io.addr_data == addr)
2297                         cleanup_one_si(e);
2298         }
2299         mutex_unlock(&smi_infos_lock);
2300 }
2301
2302 static void cleanup_ipmi_si(void)
2303 {
2304         struct smi_info *e, *tmp_e;
2305
2306         if (!initialized)
2307                 return;
2308
2309         ipmi_si_pci_shutdown();
2310
2311         ipmi_si_parisc_shutdown();
2312
2313         ipmi_si_platform_shutdown();
2314
2315         mutex_lock(&smi_infos_lock);
2316         list_for_each_entry_safe(e, tmp_e, &smi_infos, link)
2317                 cleanup_one_si(e);
2318         mutex_unlock(&smi_infos_lock);
2319
2320         ipmi_si_hardcode_exit();
2321 }
2322 module_exit(cleanup_ipmi_si);
2323
2324 MODULE_ALIAS("platform:dmi-ipmi-si");
2325 MODULE_LICENSE("GPL");
2326 MODULE_AUTHOR("Corey Minyard <minyard@mvista.com>");
2327 MODULE_DESCRIPTION("Interface to the IPMI driver for the KCS, SMIC, and BT"
2328                    " system interfaces.");