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 #include <linux/module.h>
  23 #include <linux/moduleparam.h>
  24 #include <linux/sched.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/timer.h>
  27 #include <linux/errno.h>
  28 #include <linux/spinlock.h>
  29 #include <linux/slab.h>
  30 #include <linux/delay.h>
  31 #include <linux/list.h>
  32 #include <linux/notifier.h>
  33 #include <linux/mutex.h>
  34 #include <linux/kthread.h>
  35 #include <asm/irq.h>
  36 #include <linux/interrupt.h>
  37 #include <linux/rcupdate.h>
  38 #include <linux/ipmi.h>
  39 #include <linux/ipmi_smi.h>
  40 #include "ipmi_si.h"
  41 #include <linux/string.h>
  42 #include <linux/ctype.h>
  43
  44 #define PFX "ipmi_si: "
  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         /* Is the driver in maintenance mode? */
 225         bool in_maintenance_mode;
 226
 227         /*
 228          * Did we get an attention that we did not handle?
 229          */
 230         bool got_attn;
 231
 232         /* From the get device id response... */
 233         struct ipmi_device_id device_id;
 234
 235         /* Default driver model device. */
 236         struct platform_device *pdev;
 237
 238         /* Have we added the device group to the device? */
 239         bool dev_group_added;
 240
 241         /* Have we added the platform device? */
 242         bool pdev_registered;
 243
 244         /* Counters and things for the proc filesystem. */
 245         atomic_t stats[SI_NUM_STATS];
 246
 247         struct task_struct *thread;
 248
 249         struct list_head link;
 250 };
 251
 252 #define smi_inc_stat(smi, stat) \
 253         atomic_inc(&(smi)->stats[SI_STAT_ ## stat])
 254 #define smi_get_stat(smi, stat) \
 255         ((unsigned int) atomic_read(&(smi)->stats[SI_STAT_ ## stat]))
 256
 257 #define IPMI_MAX_INTFS 4
 258 static int force_kipmid[IPMI_MAX_INTFS];
 259 static int num_force_kipmid;
 260
 261 static unsigned int kipmid_max_busy_us[IPMI_MAX_INTFS];
 262 static int num_max_busy_us;
 263
 264 static bool unload_when_empty = true;
 265
 266 static int try_smi_init(struct smi_info *smi);
 267 static void cleanup_one_si(struct smi_info *smi_info);
 268 static void cleanup_ipmi_si(void);
 269
 270 #ifdef DEBUG_TIMING
 271 void debug_timestamp(char *msg)
 272 {
 273         struct timespec64 t;
 274
 275         getnstimeofday64(&t);
 276         pr_debug("**%s: %lld.%9.9ld\n", msg, (long long) t.tv_sec, t.tv_nsec);
 277 }
 278 #else
 279 #define debug_timestamp(x)
 280 #endif
 281
 282 static ATOMIC_NOTIFIER_HEAD(xaction_notifier_list);
 283 static int register_xaction_notifier(struct notifier_block *nb)
 284 {
 285         return atomic_notifier_chain_register(&xaction_notifier_list, nb);
 286 }
 287
 288 static void deliver_recv_msg(struct smi_info *smi_info,
 289                              struct ipmi_smi_msg *msg)
 290 {
 291         /* Deliver the message to the upper layer. */
 292         ipmi_smi_msg_received(smi_info->intf, msg);
 293 }
 294
 295 static void return_hosed_msg(struct smi_info *smi_info, int cCode)
 296 {
 297         struct ipmi_smi_msg *msg = smi_info->curr_msg;
 298
 299         if (cCode < 0 || cCode > IPMI_ERR_UNSPECIFIED)
 300                 cCode = IPMI_ERR_UNSPECIFIED;
 301         /* else use it as is */
 302
 303         /* Make it a response */
 304         msg->rsp[0] = msg->data[0] | 4;
 305         msg->rsp[1] = msg->data[1];
 306         msg->rsp[2] = cCode;
 307         msg->rsp_size = 3;
 308
 309         smi_info->curr_msg = NULL;
 310         deliver_recv_msg(smi_info, msg);
 311 }
 312
 313 static enum si_sm_result start_next_msg(struct smi_info *smi_info)
 314 {
 315         int              rv;
 316
 317         if (!smi_info->waiting_msg) {
 318                 smi_info->curr_msg = NULL;
 319                 rv = SI_SM_IDLE;
 320         } else {
 321                 int err;
 322
 323                 smi_info->curr_msg = smi_info->waiting_msg;
 324                 smi_info->waiting_msg = NULL;
 325                 debug_timestamp("Start2");
 326                 err = atomic_notifier_call_chain(&xaction_notifier_list,
 327                                 0, smi_info);
 328                 if (err & NOTIFY_STOP_MASK) {
 329                         rv = SI_SM_CALL_WITHOUT_DELAY;
 330                         goto out;
 331                 }
 332                 err = smi_info->handlers->start_transaction(
 333                         smi_info->si_sm,
 334                         smi_info->curr_msg->data,
 335                         smi_info->curr_msg->data_size);
 336                 if (err)
 337                         return_hosed_msg(smi_info, err);
 338
 339                 rv = SI_SM_CALL_WITHOUT_DELAY;
 340         }
 341 out:
 342         return rv;
 343 }
 344
 345 static void smi_mod_timer(struct smi_info *smi_info, unsigned long new_val)
 346 {
 347         if (!smi_info->timer_can_start)
 348                 return;
 349         smi_info->last_timeout_jiffies = jiffies;
 350         mod_timer(&smi_info->si_timer, new_val);
 351         smi_info->timer_running = true;
 352 }
 353
 354 /*
 355  * Start a new message and (re)start the timer and thread.
 356  */
 357 static void start_new_msg(struct smi_info *smi_info, unsigned char *msg,
 358                           unsigned int size)
 359 {
 360         smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
 361
 362         if (smi_info->thread)
 363                 wake_up_process(smi_info->thread);
 364
 365         smi_info->handlers->start_transaction(smi_info->si_sm, msg, size);
 366 }
 367
 368 static void start_check_enables(struct smi_info *smi_info)
 369 {
 370         unsigned char msg[2];
 371
 372         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 373         msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
 374
 375         start_new_msg(smi_info, msg, 2);
 376         smi_info->si_state = SI_CHECKING_ENABLES;
 377 }
 378
 379 static void start_clear_flags(struct smi_info *smi_info)
 380 {
 381         unsigned char msg[3];
 382
 383         /* Make sure the watchdog pre-timeout flag is not set at startup. */
 384         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 385         msg[1] = IPMI_CLEAR_MSG_FLAGS_CMD;
 386         msg[2] = WDT_PRE_TIMEOUT_INT;
 387
 388         start_new_msg(smi_info, msg, 3);
 389         smi_info->si_state = SI_CLEARING_FLAGS;
 390 }
 391
 392 static void start_getting_msg_queue(struct smi_info *smi_info)
 393 {
 394         smi_info->curr_msg->data[0] = (IPMI_NETFN_APP_REQUEST << 2);
 395         smi_info->curr_msg->data[1] = IPMI_GET_MSG_CMD;
 396         smi_info->curr_msg->data_size = 2;
 397
 398         start_new_msg(smi_info, smi_info->curr_msg->data,
 399                       smi_info->curr_msg->data_size);
 400         smi_info->si_state = SI_GETTING_MESSAGES;
 401 }
 402
 403 static void start_getting_events(struct smi_info *smi_info)
 404 {
 405         smi_info->curr_msg->data[0] = (IPMI_NETFN_APP_REQUEST << 2);
 406         smi_info->curr_msg->data[1] = IPMI_READ_EVENT_MSG_BUFFER_CMD;
 407         smi_info->curr_msg->data_size = 2;
 408
 409         start_new_msg(smi_info, smi_info->curr_msg->data,
 410                       smi_info->curr_msg->data_size);
 411         smi_info->si_state = SI_GETTING_EVENTS;
 412 }
 413
 414 /*
 415  * When we have a situtaion where we run out of memory and cannot
 416  * allocate messages, we just leave them in the BMC and run the system
 417  * polled until we can allocate some memory.  Once we have some
 418  * memory, we will re-enable the interrupt.
 419  *
 420  * Note that we cannot just use disable_irq(), since the interrupt may
 421  * be shared.
 422  */
 423 static inline bool disable_si_irq(struct smi_info *smi_info)
 424 {
 425         if ((smi_info->io.irq) && (!smi_info->interrupt_disabled)) {
 426                 smi_info->interrupt_disabled = true;
 427                 start_check_enables(smi_info);
 428                 return true;
 429         }
 430         return false;
 431 }
 432
 433 static inline bool enable_si_irq(struct smi_info *smi_info)
 434 {
 435         if ((smi_info->io.irq) && (smi_info->interrupt_disabled)) {
 436                 smi_info->interrupt_disabled = false;
 437                 start_check_enables(smi_info);
 438                 return true;
 439         }
 440         return false;
 441 }
 442
 443 /*
 444  * Allocate a message.  If unable to allocate, start the interrupt
 445  * disable process and return NULL.  If able to allocate but
 446  * interrupts are disabled, free the message and return NULL after
 447  * starting the interrupt enable process.
 448  */
 449 static struct ipmi_smi_msg *alloc_msg_handle_irq(struct smi_info *smi_info)
 450 {
 451         struct ipmi_smi_msg *msg;
 452
 453         msg = ipmi_alloc_smi_msg();
 454         if (!msg) {
 455                 if (!disable_si_irq(smi_info))
 456                         smi_info->si_state = SI_NORMAL;
 457         } else if (enable_si_irq(smi_info)) {
 458                 ipmi_free_smi_msg(msg);
 459                 msg = NULL;
 460         }
 461         return msg;
 462 }
 463
 464 static void handle_flags(struct smi_info *smi_info)
 465 {
 466 retry:
 467         if (smi_info->msg_flags & WDT_PRE_TIMEOUT_INT) {
 468                 /* Watchdog pre-timeout */
 469                 smi_inc_stat(smi_info, watchdog_pretimeouts);
 470
 471                 start_clear_flags(smi_info);
 472                 smi_info->msg_flags &= ~WDT_PRE_TIMEOUT_INT;
 473                 ipmi_smi_watchdog_pretimeout(smi_info->intf);
 474         } else if (smi_info->msg_flags & RECEIVE_MSG_AVAIL) {
 475                 /* Messages available. */
 476                 smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
 477                 if (!smi_info->curr_msg)
 478                         return;
 479
 480                 start_getting_msg_queue(smi_info);
 481         } else if (smi_info->msg_flags & EVENT_MSG_BUFFER_FULL) {
 482                 /* Events available. */
 483                 smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
 484                 if (!smi_info->curr_msg)
 485                         return;
 486
 487                 start_getting_events(smi_info);
 488         } else if (smi_info->msg_flags & OEM_DATA_AVAIL &&
 489                    smi_info->oem_data_avail_handler) {
 490                 if (smi_info->oem_data_avail_handler(smi_info))
 491                         goto retry;
 492         } else
 493                 smi_info->si_state = SI_NORMAL;
 494 }
 495
 496 /*
 497  * Global enables we care about.
 498  */
 499 #define GLOBAL_ENABLES_MASK (IPMI_BMC_EVT_MSG_BUFF | IPMI_BMC_RCV_MSG_INTR | \
 500                              IPMI_BMC_EVT_MSG_INTR)
 501
 502 static u8 current_global_enables(struct smi_info *smi_info, u8 base,
 503                                  bool *irq_on)
 504 {
 505         u8 enables = 0;
 506
 507         if (smi_info->supports_event_msg_buff)
 508                 enables |= IPMI_BMC_EVT_MSG_BUFF;
 509
 510         if (((smi_info->io.irq && !smi_info->interrupt_disabled) ||
 511              smi_info->cannot_disable_irq) &&
 512             !smi_info->irq_enable_broken)
 513                 enables |= IPMI_BMC_RCV_MSG_INTR;
 514
 515         if (smi_info->supports_event_msg_buff &&
 516             smi_info->io.irq && !smi_info->interrupt_disabled &&
 517             !smi_info->irq_enable_broken)
 518                 enables |= IPMI_BMC_EVT_MSG_INTR;
 519
 520         *irq_on = enables & (IPMI_BMC_EVT_MSG_INTR | IPMI_BMC_RCV_MSG_INTR);
 521
 522         return enables;
 523 }
 524
 525 static void check_bt_irq(struct smi_info *smi_info, bool irq_on)
 526 {
 527         u8 irqstate = smi_info->io.inputb(&smi_info->io, IPMI_BT_INTMASK_REG);
 528
 529         irqstate &= IPMI_BT_INTMASK_ENABLE_IRQ_BIT;
 530
 531         if ((bool)irqstate == irq_on)
 532                 return;
 533
 534         if (irq_on)
 535                 smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG,
 536                                      IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
 537         else
 538                 smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG, 0);
 539 }
 540
 541 static void handle_transaction_done(struct smi_info *smi_info)
 542 {
 543         struct ipmi_smi_msg *msg;
 544
 545         debug_timestamp("Done");
 546         switch (smi_info->si_state) {
 547         case SI_NORMAL:
 548                 if (!smi_info->curr_msg)
 549                         break;
 550
 551                 smi_info->curr_msg->rsp_size
 552                         = smi_info->handlers->get_result(
 553                                 smi_info->si_sm,
 554                                 smi_info->curr_msg->rsp,
 555                                 IPMI_MAX_MSG_LENGTH);
 556
 557                 /*
 558                  * Do this here becase deliver_recv_msg() releases the
 559                  * lock, and a new message can be put in during the
 560                  * time the lock is released.
 561                  */
 562                 msg = smi_info->curr_msg;
 563                 smi_info->curr_msg = NULL;
 564                 deliver_recv_msg(smi_info, msg);
 565                 break;
 566
 567         case SI_GETTING_FLAGS:
 568         {
 569                 unsigned char msg[4];
 570                 unsigned int  len;
 571
 572                 /* We got the flags from the SMI, now handle them. */
 573                 len = smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
 574                 if (msg[2] != 0) {
 575                         /* Error fetching flags, just give up for now. */
 576                         smi_info->si_state = SI_NORMAL;
 577                 } else if (len < 4) {
 578                         /*
 579                          * Hmm, no flags.  That's technically illegal, but
 580                          * don't use uninitialized data.
 581                          */
 582                         smi_info->si_state = SI_NORMAL;
 583                 } else {
 584                         smi_info->msg_flags = msg[3];
 585                         handle_flags(smi_info);
 586                 }
 587                 break;
 588         }
 589
 590         case SI_CLEARING_FLAGS:
 591         {
 592                 unsigned char msg[3];
 593
 594                 /* We cleared the flags. */
 595                 smi_info->handlers->get_result(smi_info->si_sm, msg, 3);
 596                 if (msg[2] != 0) {
 597                         /* Error clearing flags */
 598                         dev_warn(smi_info->io.dev,
 599                                  "Error clearing flags: %2.2x\n", msg[2]);
 600                 }
 601                 smi_info->si_state = SI_NORMAL;
 602                 break;
 603         }
 604
 605         case SI_GETTING_EVENTS:
 606         {
 607                 smi_info->curr_msg->rsp_size
 608                         = smi_info->handlers->get_result(
 609                                 smi_info->si_sm,
 610                                 smi_info->curr_msg->rsp,
 611                                 IPMI_MAX_MSG_LENGTH);
 612
 613                 /*
 614                  * Do this here becase deliver_recv_msg() releases the
 615                  * lock, and a new message can be put in during the
 616                  * time the lock is released.
 617                  */
 618                 msg = smi_info->curr_msg;
 619                 smi_info->curr_msg = NULL;
 620                 if (msg->rsp[2] != 0) {
 621                         /* Error getting event, probably done. */
 622                         msg->done(msg);
 623
 624                         /* Take off the event flag. */
 625                         smi_info->msg_flags &= ~EVENT_MSG_BUFFER_FULL;
 626                         handle_flags(smi_info);
 627                 } else {
 628                         smi_inc_stat(smi_info, events);
 629
 630                         /*
 631                          * Do this before we deliver the message
 632                          * because delivering the message releases the
 633                          * lock and something else can mess with the
 634                          * state.
 635                          */
 636                         handle_flags(smi_info);
 637
 638                         deliver_recv_msg(smi_info, msg);
 639                 }
 640                 break;
 641         }
 642
 643         case SI_GETTING_MESSAGES:
 644         {
 645                 smi_info->curr_msg->rsp_size
 646                         = smi_info->handlers->get_result(
 647                                 smi_info->si_sm,
 648                                 smi_info->curr_msg->rsp,
 649                                 IPMI_MAX_MSG_LENGTH);
 650
 651                 /*
 652                  * Do this here becase deliver_recv_msg() releases the
 653                  * lock, and a new message can be put in during the
 654                  * time the lock is released.
 655                  */
 656                 msg = smi_info->curr_msg;
 657                 smi_info->curr_msg = NULL;
 658                 if (msg->rsp[2] != 0) {
 659                         /* Error getting event, probably done. */
 660                         msg->done(msg);
 661
 662                         /* Take off the msg flag. */
 663                         smi_info->msg_flags &= ~RECEIVE_MSG_AVAIL;
 664                         handle_flags(smi_info);
 665                 } else {
 666                         smi_inc_stat(smi_info, incoming_messages);
 667
 668                         /*
 669                          * Do this before we deliver the message
 670                          * because delivering the message releases the
 671                          * lock and something else can mess with the
 672                          * state.
 673                          */
 674                         handle_flags(smi_info);
 675
 676                         deliver_recv_msg(smi_info, msg);
 677                 }
 678                 break;
 679         }
 680
 681         case SI_CHECKING_ENABLES:
 682         {
 683                 unsigned char msg[4];
 684                 u8 enables;
 685                 bool irq_on;
 686
 687                 /* We got the flags from the SMI, now handle them. */
 688                 smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
 689                 if (msg[2] != 0) {
 690                         dev_warn(smi_info->io.dev,
 691                                  "Couldn't get irq info: %x.\n", msg[2]);
 692                         dev_warn(smi_info->io.dev,
 693                                  "Maybe ok, but ipmi might run very slowly.\n");
 694                         smi_info->si_state = SI_NORMAL;
 695                         break;
 696                 }
 697                 enables = current_global_enables(smi_info, 0, &irq_on);
 698                 if (smi_info->io.si_type == SI_BT)
 699                         /* BT has its own interrupt enable bit. */
 700                         check_bt_irq(smi_info, irq_on);
 701                 if (enables != (msg[3] & GLOBAL_ENABLES_MASK)) {
 702                         /* Enables are not correct, fix them. */
 703                         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 704                         msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
 705                         msg[2] = enables | (msg[3] & ~GLOBAL_ENABLES_MASK);
 706                         smi_info->handlers->start_transaction(
 707                                 smi_info->si_sm, msg, 3);
 708                         smi_info->si_state = SI_SETTING_ENABLES;
 709                 } else if (smi_info->supports_event_msg_buff) {
 710                         smi_info->curr_msg = ipmi_alloc_smi_msg();
 711                         if (!smi_info->curr_msg) {
 712                                 smi_info->si_state = SI_NORMAL;
 713                                 break;
 714                         }
 715                         start_getting_events(smi_info);
 716                 } else {
 717                         smi_info->si_state = SI_NORMAL;
 718                 }
 719                 break;
 720         }
 721
 722         case SI_SETTING_ENABLES:
 723         {
 724                 unsigned char msg[4];
 725
 726                 smi_info->handlers->get_result(smi_info->si_sm, msg, 4);
 727                 if (msg[2] != 0)
 728                         dev_warn(smi_info->io.dev,
 729                                  "Could not set the global enables: 0x%x.\n",
 730                                  msg[2]);
 731
 732                 if (smi_info->supports_event_msg_buff) {
 733                         smi_info->curr_msg = ipmi_alloc_smi_msg();
 734                         if (!smi_info->curr_msg) {
 735                                 smi_info->si_state = SI_NORMAL;
 736                                 break;
 737                         }
 738                         start_getting_events(smi_info);
 739                 } else {
 740                         smi_info->si_state = SI_NORMAL;
 741                 }
 742                 break;
 743         }
 744         }
 745 }
 746
 747 /*
 748  * Called on timeouts and events.  Timeouts should pass the elapsed
 749  * time, interrupts should pass in zero.  Must be called with
 750  * si_lock held and interrupts disabled.
 751  */
 752 static enum si_sm_result smi_event_handler(struct smi_info *smi_info,
 753                                            int time)
 754 {
 755         enum si_sm_result si_sm_result;
 756
 757 restart:
 758         /*
 759          * There used to be a loop here that waited a little while
 760          * (around 25us) before giving up.  That turned out to be
 761          * pointless, the minimum delays I was seeing were in the 300us
 762          * range, which is far too long to wait in an interrupt.  So
 763          * we just run until the state machine tells us something
 764          * happened or it needs a delay.
 765          */
 766         si_sm_result = smi_info->handlers->event(smi_info->si_sm, time);
 767         time = 0;
 768         while (si_sm_result == SI_SM_CALL_WITHOUT_DELAY)
 769                 si_sm_result = smi_info->handlers->event(smi_info->si_sm, 0);
 770
 771         if (si_sm_result == SI_SM_TRANSACTION_COMPLETE) {
 772                 smi_inc_stat(smi_info, complete_transactions);
 773
 774                 handle_transaction_done(smi_info);
 775                 goto restart;
 776         } else if (si_sm_result == SI_SM_HOSED) {
 777                 smi_inc_stat(smi_info, hosed_count);
 778
 779                 /*
 780                  * Do the before return_hosed_msg, because that
 781                  * releases the lock.
 782                  */
 783                 smi_info->si_state = SI_NORMAL;
 784                 if (smi_info->curr_msg != NULL) {
 785                         /*
 786                          * If we were handling a user message, format
 787                          * a response to send to the upper layer to
 788                          * tell it about the error.
 789                          */
 790                         return_hosed_msg(smi_info, IPMI_ERR_UNSPECIFIED);
 791                 }
 792                 goto restart;
 793         }
 794
 795         /*
 796          * We prefer handling attn over new messages.  But don't do
 797          * this if there is not yet an upper layer to handle anything.
 798          */
 799         if (si_sm_result == SI_SM_ATTN || smi_info->got_attn) {
 800                 unsigned char msg[2];
 801
 802                 if (smi_info->si_state != SI_NORMAL) {
 803                         /*
 804                          * We got an ATTN, but we are doing something else.
 805                          * Handle the ATTN later.
 806                          */
 807                         smi_info->got_attn = true;
 808                 } else {
 809                         smi_info->got_attn = false;
 810                         smi_inc_stat(smi_info, attentions);
 811
 812                         /*
 813                          * Got a attn, send down a get message flags to see
 814                          * what's causing it.  It would be better to handle
 815                          * this in the upper layer, but due to the way
 816                          * interrupts work with the SMI, that's not really
 817                          * possible.
 818                          */
 819                         msg[0] = (IPMI_NETFN_APP_REQUEST << 2);
 820                         msg[1] = IPMI_GET_MSG_FLAGS_CMD;
 821
 822                         start_new_msg(smi_info, msg, 2);
 823                         smi_info->si_state = SI_GETTING_FLAGS;
 824                         goto restart;
 825                 }
 826         }
 827
 828         /* If we are currently idle, try to start the next message. */
 829         if (si_sm_result == SI_SM_IDLE) {
 830                 smi_inc_stat(smi_info, idles);
 831
 832                 si_sm_result = start_next_msg(smi_info);
 833                 if (si_sm_result != SI_SM_IDLE)
 834                         goto restart;
 835         }
 836
 837         if ((si_sm_result == SI_SM_IDLE)
 838             && (atomic_read(&smi_info->req_events))) {
 839                 /*
 840                  * We are idle and the upper layer requested that I fetch
 841                  * events, so do so.
 842                  */
 843                 atomic_set(&smi_info->req_events, 0);
 844
 845                 /*
 846                  * Take this opportunity to check the interrupt and
 847                  * message enable state for the BMC.  The BMC can be
 848                  * asynchronously reset, and may thus get interrupts
 849                  * disable and messages disabled.
 850                  */
 851                 if (smi_info->supports_event_msg_buff || smi_info->io.irq) {
 852                         start_check_enables(smi_info);
 853                 } else {
 854                         smi_info->curr_msg = alloc_msg_handle_irq(smi_info);
 855                         if (!smi_info->curr_msg)
 856                                 goto out;
 857
 858                         start_getting_events(smi_info);
 859                 }
 860                 goto restart;
 861         }
 862
 863         if (si_sm_result == SI_SM_IDLE && smi_info->timer_running) {
 864                 /* Ok it if fails, the timer will just go off. */
 865                 if (del_timer(&smi_info->si_timer))
 866                         smi_info->timer_running = false;
 867         }
 868
 869 out:
 870         return si_sm_result;
 871 }
 872
 873 static void check_start_timer_thread(struct smi_info *smi_info)
 874 {
 875         if (smi_info->si_state == SI_NORMAL && smi_info->curr_msg == NULL) {
 876                 smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
 877
 878                 if (smi_info->thread)
 879                         wake_up_process(smi_info->thread);
 880
 881                 start_next_msg(smi_info);
 882                 smi_event_handler(smi_info, 0);
 883         }
 884 }
 885
 886 static void flush_messages(void *send_info)
 887 {
 888         struct smi_info *smi_info = send_info;
 889         enum si_sm_result result;
 890
 891         /*
 892          * Currently, this function is called only in run-to-completion
 893          * mode.  This means we are single-threaded, no need for locks.
 894          */
 895         result = smi_event_handler(smi_info, 0);
 896         while (result != SI_SM_IDLE) {
 897                 udelay(SI_SHORT_TIMEOUT_USEC);
 898                 result = smi_event_handler(smi_info, SI_SHORT_TIMEOUT_USEC);
 899         }
 900 }
 901
 902 static void sender(void                *send_info,
 903                    struct ipmi_smi_msg *msg)
 904 {
 905         struct smi_info   *smi_info = send_info;
 906         unsigned long     flags;
 907
 908         debug_timestamp("Enqueue");
 909
 910         if (smi_info->run_to_completion) {
 911                 /*
 912                  * If we are running to completion, start it.  Upper
 913                  * layer will call flush_messages to clear it out.
 914                  */
 915                 smi_info->waiting_msg = msg;
 916                 return;
 917         }
 918
 919         spin_lock_irqsave(&smi_info->si_lock, flags);
 920         /*
 921          * The following two lines don't need to be under the lock for
 922          * the lock's sake, but they do need SMP memory barriers to
 923          * avoid getting things out of order.  We are already claiming
 924          * the lock, anyway, so just do it under the lock to avoid the
 925          * ordering problem.
 926          */
 927         BUG_ON(smi_info->waiting_msg);
 928         smi_info->waiting_msg = msg;
 929         check_start_timer_thread(smi_info);
 930         spin_unlock_irqrestore(&smi_info->si_lock, flags);
 931 }
 932
 933 static void set_run_to_completion(void *send_info, bool i_run_to_completion)
 934 {
 935         struct smi_info   *smi_info = send_info;
 936
 937         smi_info->run_to_completion = i_run_to_completion;
 938         if (i_run_to_completion)
 939                 flush_messages(smi_info);
 940 }
 941
 942 /*
 943  * Use -1 in the nsec value of the busy waiting timespec to tell that
 944  * we are spinning in kipmid looking for something and not delaying
 945  * between checks
 946  */
 947 static inline void ipmi_si_set_not_busy(struct timespec64 *ts)
 948 {
 949         ts->tv_nsec = -1;
 950 }
 951 static inline int ipmi_si_is_busy(struct timespec64 *ts)
 952 {
 953         return ts->tv_nsec != -1;
 954 }
 955
 956 static inline int ipmi_thread_busy_wait(enum si_sm_result smi_result,
 957                                         const struct smi_info *smi_info,
 958                                         struct timespec64 *busy_until)
 959 {
 960         unsigned int max_busy_us = 0;
 961
 962         if (smi_info->si_num < num_max_busy_us)
 963                 max_busy_us = kipmid_max_busy_us[smi_info->si_num];
 964         if (max_busy_us == 0 || smi_result != SI_SM_CALL_WITH_DELAY)
 965                 ipmi_si_set_not_busy(busy_until);
 966         else if (!ipmi_si_is_busy(busy_until)) {
 967                 getnstimeofday64(busy_until);
 968                 timespec64_add_ns(busy_until, max_busy_us*NSEC_PER_USEC);
 969         } else {
 970                 struct timespec64 now;
 971
 972                 getnstimeofday64(&now);
 973                 if (unlikely(timespec64_compare(&now, busy_until) > 0)) {
 974                         ipmi_si_set_not_busy(busy_until);
 975                         return 0;
 976                 }
 977         }
 978         return 1;
 979 }
 980
 981
 982 /*
 983  * A busy-waiting loop for speeding up IPMI operation.
 984  *
 985  * Lousy hardware makes this hard.  This is only enabled for systems
 986  * that are not BT and do not have interrupts.  It starts spinning
 987  * when an operation is complete or until max_busy tells it to stop
 988  * (if that is enabled).  See the paragraph on kimid_max_busy_us in
 989  * Documentation/IPMI.txt for details.
 990  */
 991 static int ipmi_thread(void *data)
 992 {
 993         struct smi_info *smi_info = data;
 994         unsigned long flags;
 995         enum si_sm_result smi_result;
 996         struct timespec64 busy_until;
 997
 998         ipmi_si_set_not_busy(&busy_until);
 999         set_user_nice(current, MAX_NICE);
1000         while (!kthread_should_stop()) {
1001                 int busy_wait;
1002
1003                 spin_lock_irqsave(&(smi_info->si_lock), flags);
1004                 smi_result = smi_event_handler(smi_info, 0);
1005
1006                 /*
1007                  * If the driver is doing something, there is a possible
1008                  * race with the timer.  If the timer handler see idle,
1009                  * and the thread here sees something else, the timer
1010                  * handler won't restart the timer even though it is
1011                  * required.  So start it here if necessary.
1012                  */
1013                 if (smi_result != SI_SM_IDLE && !smi_info->timer_running)
1014                         smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_JIFFIES);
1015
1016                 spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1017                 busy_wait = ipmi_thread_busy_wait(smi_result, smi_info,
1018                                                   &busy_until);
1019                 if (smi_result == SI_SM_CALL_WITHOUT_DELAY) {
1020                         ; /* do nothing */
1021                 } else if (smi_result == SI_SM_CALL_WITH_DELAY && busy_wait) {
1022                         /*
1023                          * In maintenance mode we run as fast as
1024                          * possible to allow firmware updates to
1025                          * complete as fast as possible, but normally
1026                          * don't bang on the scheduler.
1027                          */
1028                         if (smi_info->in_maintenance_mode)
1029                                 schedule();
1030                         else
1031                                 usleep_range(100, 200);
1032                 } else if (smi_result == SI_SM_IDLE) {
1033                         if (atomic_read(&smi_info->need_watch)) {
1034                                 schedule_timeout_interruptible(100);
1035                         } else {
1036                                 /* Wait to be woken up when we are needed. */
1037                                 __set_current_state(TASK_INTERRUPTIBLE);
1038                                 schedule();
1039                         }
1040                 } else {
1041                         schedule_timeout_interruptible(1);
1042                 }
1043         }
1044         return 0;
1045 }
1046
1047
1048 static void poll(void *send_info)
1049 {
1050         struct smi_info *smi_info = send_info;
1051         unsigned long flags = 0;
1052         bool run_to_completion = smi_info->run_to_completion;
1053
1054         /*
1055          * Make sure there is some delay in the poll loop so we can
1056          * drive time forward and timeout things.
1057          */
1058         udelay(10);
1059         if (!run_to_completion)
1060                 spin_lock_irqsave(&smi_info->si_lock, flags);
1061         smi_event_handler(smi_info, 10);
1062         if (!run_to_completion)
1063                 spin_unlock_irqrestore(&smi_info->si_lock, flags);
1064 }
1065
1066 static void request_events(void *send_info)
1067 {
1068         struct smi_info *smi_info = send_info;
1069
1070         if (!smi_info->has_event_buffer)
1071                 return;
1072
1073         atomic_set(&smi_info->req_events, 1);
1074 }
1075
1076 static void set_need_watch(void *send_info, bool enable)
1077 {
1078         struct smi_info *smi_info = send_info;
1079         unsigned long flags;
1080
1081         atomic_set(&smi_info->need_watch, enable);
1082         spin_lock_irqsave(&smi_info->si_lock, flags);
1083         check_start_timer_thread(smi_info);
1084         spin_unlock_irqrestore(&smi_info->si_lock, flags);
1085 }
1086
1087 static void smi_timeout(struct timer_list *t)
1088 {
1089         struct smi_info   *smi_info = from_timer(smi_info, t, si_timer);
1090         enum si_sm_result smi_result;
1091         unsigned long     flags;
1092         unsigned long     jiffies_now;
1093         long              time_diff;
1094         long              timeout;
1095
1096         spin_lock_irqsave(&(smi_info->si_lock), flags);
1097         debug_timestamp("Timer");
1098
1099         jiffies_now = jiffies;
1100         time_diff = (((long)jiffies_now - (long)smi_info->last_timeout_jiffies)
1101                      * SI_USEC_PER_JIFFY);
1102         smi_result = smi_event_handler(smi_info, time_diff);
1103
1104         if ((smi_info->io.irq) && (!smi_info->interrupt_disabled)) {
1105                 /* Running with interrupts, only do long timeouts. */
1106                 timeout = jiffies + SI_TIMEOUT_JIFFIES;
1107                 smi_inc_stat(smi_info, long_timeouts);
1108                 goto do_mod_timer;
1109         }
1110
1111         /*
1112          * If the state machine asks for a short delay, then shorten
1113          * the timer timeout.
1114          */
1115         if (smi_result == SI_SM_CALL_WITH_DELAY) {
1116                 smi_inc_stat(smi_info, short_timeouts);
1117                 timeout = jiffies + 1;
1118         } else {
1119                 smi_inc_stat(smi_info, long_timeouts);
1120                 timeout = jiffies + SI_TIMEOUT_JIFFIES;
1121         }
1122
1123 do_mod_timer:
1124         if (smi_result != SI_SM_IDLE)
1125                 smi_mod_timer(smi_info, timeout);
1126         else
1127                 smi_info->timer_running = false;
1128         spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1129 }
1130
1131 irqreturn_t ipmi_si_irq_handler(int irq, void *data)
1132 {
1133         struct smi_info *smi_info = data;
1134         unsigned long   flags;
1135
1136         if (smi_info->io.si_type == SI_BT)
1137                 /* We need to clear the IRQ flag for the BT interface. */
1138                 smi_info->io.outputb(&smi_info->io, IPMI_BT_INTMASK_REG,
1139                                      IPMI_BT_INTMASK_CLEAR_IRQ_BIT
1140                                      | IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
1141
1142         spin_lock_irqsave(&(smi_info->si_lock), flags);
1143
1144         smi_inc_stat(smi_info, interrupts);
1145
1146         debug_timestamp("Interrupt");
1147
1148         smi_event_handler(smi_info, 0);
1149         spin_unlock_irqrestore(&(smi_info->si_lock), flags);
1150         return IRQ_HANDLED;
1151 }
1152
1153 static int smi_start_processing(void            *send_info,
1154                                 struct ipmi_smi *intf)
1155 {
1156         struct smi_info *new_smi = send_info;
1157         int             enable = 0;
1158
1159         new_smi->intf = intf;
1160
1161         /* Set up the timer that drives the interface. */
1162         timer_setup(&new_smi->si_timer, smi_timeout, 0);
1163         new_smi->timer_can_start = true;
1164         smi_mod_timer(new_smi, jiffies + SI_TIMEOUT_JIFFIES);
1165
1166         /* Try to claim any interrupts. */
1167         if (new_smi->io.irq_setup) {
1168                 new_smi->io.irq_handler_data = new_smi;
1169                 new_smi->io.irq_setup(&new_smi->io);
1170         }
1171
1172         /*
1173          * Check if the user forcefully enabled the daemon.
1174          */
1175         if (new_smi->si_num < num_force_kipmid)
1176                 enable = force_kipmid[new_smi->si_num];
1177         /*
1178          * The BT interface is efficient enough to not need a thread,
1179          * and there is no need for a thread if we have interrupts.
1180          */
1181         else if ((new_smi->io.si_type != SI_BT) && (!new_smi->io.irq))
1182                 enable = 1;
1183
1184         if (enable) {
1185                 new_smi->thread = kthread_run(ipmi_thread, new_smi,
1186                                               "kipmi%d", new_smi->si_num);
1187                 if (IS_ERR(new_smi->thread)) {
1188                         dev_notice(new_smi->io.dev, "Could not start"
1189                                    " kernel thread due to error %ld, only using"
1190                                    " timers to drive the interface\n",
1191                                    PTR_ERR(new_smi->thread));
1192                         new_smi->thread = NULL;
1193                 }
1194         }
1195
1196         return 0;
1197 }
1198
1199 static int get_smi_info(void *send_info, struct ipmi_smi_info *data)
1200 {
1201         struct smi_info *smi = send_info;
1202
1203         data->addr_src = smi->io.addr_source;
1204         data->dev = smi->io.dev;
1205         data->addr_info = smi->io.addr_info;
1206         get_device(smi->io.dev);
1207
1208         return 0;
1209 }
1210
1211 static void set_maintenance_mode(void *send_info, bool enable)
1212 {
1213         struct smi_info   *smi_info = send_info;
1214
1215         if (!enable)
1216                 atomic_set(&smi_info->req_events, 0);
1217         smi_info->in_maintenance_mode = enable;
1218 }
1219
1220 static void shutdown_smi(void *send_info);
1221 static const struct ipmi_smi_handlers handlers = {
1222         .owner                  = THIS_MODULE,
1223         .start_processing       = smi_start_processing,
1224         .shutdown               = shutdown_smi,
1225         .get_smi_info           = get_smi_info,
1226         .sender                 = sender,
1227         .request_events         = request_events,
1228         .set_need_watch         = set_need_watch,
1229         .set_maintenance_mode   = set_maintenance_mode,
1230         .set_run_to_completion  = set_run_to_completion,
1231         .flush_messages         = flush_messages,
1232         .poll                   = poll,
1233 };
1234
1235 static LIST_HEAD(smi_infos);
1236 static DEFINE_MUTEX(smi_infos_lock);
1237 static int smi_num; /* Used to sequence the SMIs */
1238
1239 static const char * const addr_space_to_str[] = { "i/o", "mem" };
1240
1241 module_param_array(force_kipmid, int, &num_force_kipmid, 0);
1242 MODULE_PARM_DESC(force_kipmid, "Force the kipmi daemon to be enabled (1) or"
1243                  " disabled(0).  Normally the IPMI driver auto-detects"
1244                  " this, but the value may be overridden by this parm.");
1245 module_param(unload_when_empty, bool, 0);
1246 MODULE_PARM_DESC(unload_when_empty, "Unload the module if no interfaces are"
1247                  " specified or found, default is 1.  Setting to 0"
1248                  " is useful for hot add of devices using hotmod.");
1249 module_param_array(kipmid_max_busy_us, uint, &num_max_busy_us, 0644);
1250 MODULE_PARM_DESC(kipmid_max_busy_us,
1251                  "Max time (in microseconds) to busy-wait for IPMI data before"
1252                  " sleeping. 0 (default) means to wait forever. Set to 100-500"
1253                  " if kipmid is using up a lot of CPU time.");
1254
1255 void ipmi_irq_finish_setup(struct si_sm_io *io)
1256 {
1257         if (io->si_type == SI_BT)
1258                 /* Enable the interrupt in the BT interface. */
1259                 io->outputb(io, IPMI_BT_INTMASK_REG,
1260                             IPMI_BT_INTMASK_ENABLE_IRQ_BIT);
1261 }
1262
1263 void ipmi_irq_start_cleanup(struct si_sm_io *io)
1264 {
1265         if (io->si_type == SI_BT)
1266                 /* Disable the interrupt in the BT interface. */
1267                 io->outputb(io, IPMI_BT_INTMASK_REG, 0);
1268 }
1269
1270 static void std_irq_cleanup(struct si_sm_io *io)
1271 {
1272         ipmi_irq_start_cleanup(io);
1273         free_irq(io->irq, io->irq_handler_data);
1274 }
1275
1276 int ipmi_std_irq_setup(struct si_sm_io *io)
1277 {
1278         int rv;
1279
1280         if (!io->irq)
1281                 return 0;
1282
1283         rv = request_irq(io->irq,
1284                          ipmi_si_irq_handler,
1285                          IRQF_SHARED,
1286                          DEVICE_NAME,
1287                          io->irq_handler_data);
1288         if (rv) {
1289                 dev_warn(io->dev, "%s unable to claim interrupt %d,"
1290                          " running polled\n",
1291                          DEVICE_NAME, io->irq);
1292                 io->irq = 0;
1293         } else {
1294                 io->irq_cleanup = std_irq_cleanup;
1295                 ipmi_irq_finish_setup(io);
1296                 dev_info(io->dev, "Using irq %d\n", io->irq);
1297         }
1298
1299         return rv;
1300 }
1301
1302 static int wait_for_msg_done(struct smi_info *smi_info)
1303 {
1304         enum si_sm_result     smi_result;
1305
1306         smi_result = smi_info->handlers->event(smi_info->si_sm, 0);
1307         for (;;) {
1308                 if (smi_result == SI_SM_CALL_WITH_DELAY ||
1309                     smi_result == SI_SM_CALL_WITH_TICK_DELAY) {
1310                         schedule_timeout_uninterruptible(1);
1311                         smi_result = smi_info->handlers->event(
1312                                 smi_info->si_sm, jiffies_to_usecs(1));
1313                 } else if (smi_result == SI_SM_CALL_WITHOUT_DELAY) {
1314                         smi_result = smi_info->handlers->event(
1315                                 smi_info->si_sm, 0);
1316                 } else
1317                         break;
1318         }
1319         if (smi_result == SI_SM_HOSED)
1320                 /*
1321                  * We couldn't get the state machine to run, so whatever's at
1322                  * the port is probably not an IPMI SMI interface.
1323                  */
1324                 return -ENODEV;
1325
1326         return 0;
1327 }
1328
1329 static int try_get_dev_id(struct smi_info *smi_info)
1330 {
1331         unsigned char         msg[2];
1332         unsigned char         *resp;
1333         unsigned long         resp_len;
1334         int                   rv = 0;
1335
1336         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1337         if (!resp)
1338                 return -ENOMEM;
1339
1340         /*
1341          * Do a Get Device ID command, since it comes back with some
1342          * useful info.
1343          */
1344         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1345         msg[1] = IPMI_GET_DEVICE_ID_CMD;
1346         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1347
1348         rv = wait_for_msg_done(smi_info);
1349         if (rv)
1350                 goto out;
1351
1352         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1353                                                   resp, IPMI_MAX_MSG_LENGTH);
1354
1355         /* Check and record info from the get device id, in case we need it. */
1356         rv = ipmi_demangle_device_id(resp[0] >> 2, resp[1],
1357                         resp + 2, resp_len - 2, &smi_info->device_id);
1358
1359 out:
1360         kfree(resp);
1361         return rv;
1362 }
1363
1364 static int get_global_enables(struct smi_info *smi_info, u8 *enables)
1365 {
1366         unsigned char         msg[3];
1367         unsigned char         *resp;
1368         unsigned long         resp_len;
1369         int                   rv;
1370
1371         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1372         if (!resp)
1373                 return -ENOMEM;
1374
1375         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1376         msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
1377         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1378
1379         rv = wait_for_msg_done(smi_info);
1380         if (rv) {
1381                 dev_warn(smi_info->io.dev,
1382                          "Error getting response from get global enables command: %d\n",
1383                          rv);
1384                 goto out;
1385         }
1386
1387         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1388                                                   resp, IPMI_MAX_MSG_LENGTH);
1389
1390         if (resp_len < 4 ||
1391                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1392                         resp[1] != IPMI_GET_BMC_GLOBAL_ENABLES_CMD   ||
1393                         resp[2] != 0) {
1394                 dev_warn(smi_info->io.dev,
1395                          "Invalid return from get global enables command: %ld %x %x %x\n",
1396                          resp_len, resp[0], resp[1], resp[2]);
1397                 rv = -EINVAL;
1398                 goto out;
1399         } else {
1400                 *enables = resp[3];
1401         }
1402
1403 out:
1404         kfree(resp);
1405         return rv;
1406 }
1407
1408 /*
1409  * Returns 1 if it gets an error from the command.
1410  */
1411 static int set_global_enables(struct smi_info *smi_info, u8 enables)
1412 {
1413         unsigned char         msg[3];
1414         unsigned char         *resp;
1415         unsigned long         resp_len;
1416         int                   rv;
1417
1418         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1419         if (!resp)
1420                 return -ENOMEM;
1421
1422         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1423         msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
1424         msg[2] = enables;
1425         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 3);
1426
1427         rv = wait_for_msg_done(smi_info);
1428         if (rv) {
1429                 dev_warn(smi_info->io.dev,
1430                          "Error getting response from set global enables command: %d\n",
1431                          rv);
1432                 goto out;
1433         }
1434
1435         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1436                                                   resp, IPMI_MAX_MSG_LENGTH);
1437
1438         if (resp_len < 3 ||
1439                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1440                         resp[1] != IPMI_SET_BMC_GLOBAL_ENABLES_CMD) {
1441                 dev_warn(smi_info->io.dev,
1442                          "Invalid return from set global enables command: %ld %x %x\n",
1443                          resp_len, resp[0], resp[1]);
1444                 rv = -EINVAL;
1445                 goto out;
1446         }
1447
1448         if (resp[2] != 0)
1449                 rv = 1;
1450
1451 out:
1452         kfree(resp);
1453         return rv;
1454 }
1455
1456 /*
1457  * Some BMCs do not support clearing the receive irq bit in the global
1458  * enables (even if they don't support interrupts on the BMC).  Check
1459  * for this and handle it properly.
1460  */
1461 static void check_clr_rcv_irq(struct smi_info *smi_info)
1462 {
1463         u8 enables = 0;
1464         int rv;
1465
1466         rv = get_global_enables(smi_info, &enables);
1467         if (!rv) {
1468                 if ((enables & IPMI_BMC_RCV_MSG_INTR) == 0)
1469                         /* Already clear, should work ok. */
1470                         return;
1471
1472                 enables &= ~IPMI_BMC_RCV_MSG_INTR;
1473                 rv = set_global_enables(smi_info, enables);
1474         }
1475
1476         if (rv < 0) {
1477                 dev_err(smi_info->io.dev,
1478                         "Cannot check clearing the rcv irq: %d\n", rv);
1479                 return;
1480         }
1481
1482         if (rv) {
1483                 /*
1484                  * An error when setting the event buffer bit means
1485                  * clearing the bit is not supported.
1486                  */
1487                 dev_warn(smi_info->io.dev,
1488                          "The BMC does not support clearing the recv irq bit, compensating, but the BMC needs to be fixed.\n");
1489                 smi_info->cannot_disable_irq = true;
1490         }
1491 }
1492
1493 /*
1494  * Some BMCs do not support setting the interrupt bits in the global
1495  * enables even if they support interrupts.  Clearly bad, but we can
1496  * compensate.
1497  */
1498 static void check_set_rcv_irq(struct smi_info *smi_info)
1499 {
1500         u8 enables = 0;
1501         int rv;
1502
1503         if (!smi_info->io.irq)
1504                 return;
1505
1506         rv = get_global_enables(smi_info, &enables);
1507         if (!rv) {
1508                 enables |= IPMI_BMC_RCV_MSG_INTR;
1509                 rv = set_global_enables(smi_info, enables);
1510         }
1511
1512         if (rv < 0) {
1513                 dev_err(smi_info->io.dev,
1514                         "Cannot check setting the rcv irq: %d\n", rv);
1515                 return;
1516         }
1517
1518         if (rv) {
1519                 /*
1520                  * An error when setting the event buffer bit means
1521                  * setting the bit is not supported.
1522                  */
1523                 dev_warn(smi_info->io.dev,
1524                          "The BMC does not support setting the recv irq bit, compensating, but the BMC needs to be fixed.\n");
1525                 smi_info->cannot_disable_irq = true;
1526                 smi_info->irq_enable_broken = true;
1527         }
1528 }
1529
1530 static int try_enable_event_buffer(struct smi_info *smi_info)
1531 {
1532         unsigned char         msg[3];
1533         unsigned char         *resp;
1534         unsigned long         resp_len;
1535         int                   rv = 0;
1536
1537         resp = kmalloc(IPMI_MAX_MSG_LENGTH, GFP_KERNEL);
1538         if (!resp)
1539                 return -ENOMEM;
1540
1541         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1542         msg[1] = IPMI_GET_BMC_GLOBAL_ENABLES_CMD;
1543         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 2);
1544
1545         rv = wait_for_msg_done(smi_info);
1546         if (rv) {
1547                 pr_warn(PFX "Error getting response from get global enables command, the event buffer is not enabled.\n");
1548                 goto out;
1549         }
1550
1551         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1552                                                   resp, IPMI_MAX_MSG_LENGTH);
1553
1554         if (resp_len < 4 ||
1555                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1556                         resp[1] != IPMI_GET_BMC_GLOBAL_ENABLES_CMD   ||
1557                         resp[2] != 0) {
1558                 pr_warn(PFX "Invalid return from get global enables command, cannot enable the event buffer.\n");
1559                 rv = -EINVAL;
1560                 goto out;
1561         }
1562
1563         if (resp[3] & IPMI_BMC_EVT_MSG_BUFF) {
1564                 /* buffer is already enabled, nothing to do. */
1565                 smi_info->supports_event_msg_buff = true;
1566                 goto out;
1567         }
1568
1569         msg[0] = IPMI_NETFN_APP_REQUEST << 2;
1570         msg[1] = IPMI_SET_BMC_GLOBAL_ENABLES_CMD;
1571         msg[2] = resp[3] | IPMI_BMC_EVT_MSG_BUFF;
1572         smi_info->handlers->start_transaction(smi_info->si_sm, msg, 3);
1573
1574         rv = wait_for_msg_done(smi_info);
1575         if (rv) {
1576                 pr_warn(PFX "Error getting response from set global, enables command, the event buffer is not enabled.\n");
1577                 goto out;
1578         }
1579
1580         resp_len = smi_info->handlers->get_result(smi_info->si_sm,
1581                                                   resp, IPMI_MAX_MSG_LENGTH);
1582
1583         if (resp_len < 3 ||
1584                         resp[0] != (IPMI_NETFN_APP_REQUEST | 1) << 2 ||
1585                         resp[1] != IPMI_SET_BMC_GLOBAL_ENABLES_CMD) {
1586                 pr_warn(PFX "Invalid return from get global, enables command, not enable the event buffer.\n");
1587                 rv = -EINVAL;
1588                 goto out;
1589         }
1590
1591         if (resp[2] != 0)
1592                 /*
1593                  * An error when setting the event buffer bit means
1594                  * that the event buffer is not supported.
1595                  */
1596                 rv = -ENOENT;
1597         else
1598                 smi_info->supports_event_msg_buff = true;
1599
1600 out:
1601         kfree(resp);
1602         return rv;
1603 }
1604
1605 #define IPMI_SI_ATTR(name) \
1606 static ssize_t ipmi_##name##_show(struct device *dev,                   \
1607                                   struct device_attribute *attr,        \
1608                                   char *buf)                            \
1609 {                                                                       \
1610         struct smi_info *smi_info = dev_get_drvdata(dev);               \
1611                                                                         \
1612         return snprintf(buf, 10, "%u\n", smi_get_stat(smi_info, name)); \
1613 }                                                                       \
1614 static DEVICE_ATTR(name, S_IRUGO, ipmi_##name##_show, NULL)
1615
1616 static ssize_t ipmi_type_show(struct device *dev,
1617                               struct device_attribute *attr,
1618                               char *buf)
1619 {
1620         struct smi_info *smi_info = dev_get_drvdata(dev);
1621
1622         return snprintf(buf, 10, "%s\n", si_to_str[smi_info->io.si_type]);
1623 }
1624 static DEVICE_ATTR(type, S_IRUGO, ipmi_type_show, NULL);
1625
1626 static ssize_t ipmi_interrupts_enabled_show(struct device *dev,
1627                                             struct device_attribute *attr,
1628                                             char *buf)
1629 {
1630         struct smi_info *smi_info = dev_get_drvdata(dev);
1631         int enabled = smi_info->io.irq && !smi_info->interrupt_disabled;
1632
1633         return snprintf(buf, 10, "%d\n", enabled);
1634 }
1635 static DEVICE_ATTR(interrupts_enabled, S_IRUGO,
1636                    ipmi_interrupts_enabled_show, NULL);
1637
1638 IPMI_SI_ATTR(short_timeouts);
1639 IPMI_SI_ATTR(long_timeouts);
1640 IPMI_SI_ATTR(idles);
1641 IPMI_SI_ATTR(interrupts);
1642 IPMI_SI_ATTR(attentions);
1643 IPMI_SI_ATTR(flag_fetches);
1644 IPMI_SI_ATTR(hosed_count);
1645 IPMI_SI_ATTR(complete_transactions);
1646 IPMI_SI_ATTR(events);
1647 IPMI_SI_ATTR(watchdog_pretimeouts);
1648 IPMI_SI_ATTR(incoming_messages);
1649
1650 static ssize_t ipmi_params_show(struct device *dev,
1651                                 struct device_attribute *attr,
1652                                 char *buf)
1653 {
1654         struct smi_info *smi_info = dev_get_drvdata(dev);
1655
1656         return snprintf(buf, 200,
1657                         "%s,%s,0x%lx,rsp=%d,rsi=%d,rsh=%d,irq=%d,ipmb=%d\n",
1658                         si_to_str[smi_info->io.si_type],
1659                         addr_space_to_str[smi_info->io.addr_type],
1660                         smi_info->io.addr_data,
1661                         smi_info->io.regspacing,
1662                         smi_info->io.regsize,
1663                         smi_info->io.regshift,
1664                         smi_info->io.irq,
1665                         smi_info->io.slave_addr);
1666 }
1667 static DEVICE_ATTR(params, S_IRUGO, ipmi_params_show, NULL);
1668
1669 static struct attribute *ipmi_si_dev_attrs[] = {
1670         &dev_attr_type.attr,
1671         &dev_attr_interrupts_enabled.attr,
1672         &dev_attr_short_timeouts.attr,
1673         &dev_attr_long_timeouts.attr,
1674         &dev_attr_idles.attr,
1675         &dev_attr_interrupts.attr,
1676         &dev_attr_attentions.attr,
1677         &dev_attr_flag_fetches.attr,
1678         &dev_attr_hosed_count.attr,
1679         &dev_attr_complete_transactions.attr,
1680         &dev_attr_events.attr,
1681         &dev_attr_watchdog_pretimeouts.attr,
1682         &dev_attr_incoming_messages.attr,
1683         &dev_attr_params.attr,
1684         NULL
1685 };
1686
1687 static const struct attribute_group ipmi_si_dev_attr_group = {
1688         .attrs          = ipmi_si_dev_attrs,
1689 };
1690
1691 /*
1692  * oem_data_avail_to_receive_msg_avail
1693  * @info - smi_info structure with msg_flags set
1694  *
1695  * Converts flags from OEM_DATA_AVAIL to RECEIVE_MSG_AVAIL
1696  * Returns 1 indicating need to re-run handle_flags().
1697  */
1698 static int oem_data_avail_to_receive_msg_avail(struct smi_info *smi_info)
1699 {
1700         smi_info->msg_flags = ((smi_info->msg_flags & ~OEM_DATA_AVAIL) |
1701                                RECEIVE_MSG_AVAIL);
1702         return 1;
1703 }
1704
1705 /*
1706  * setup_dell_poweredge_oem_data_handler
1707  * @info - smi_info.device_id must be populated
1708  *
1709  * Systems that match, but have firmware version < 1.40 may assert
1710  * OEM0_DATA_AVAIL on their own, without being told via Set Flags that
1711  * it's safe to do so.  Such systems will de-assert OEM1_DATA_AVAIL
1712  * upon receipt of IPMI_GET_MSG_CMD, so we should treat these flags
1713  * as RECEIVE_MSG_AVAIL instead.
1714  *
1715  * As Dell has no plans to release IPMI 1.5 firmware that *ever*
1716  * assert the OEM[012] bits, and if it did, the driver would have to
1717  * change to handle that properly, we don't actually check for the
1718  * firmware version.
1719  * Device ID = 0x20                BMC on PowerEdge 8G servers
1720  * Device Revision = 0x80
1721  * Firmware Revision1 = 0x01       BMC version 1.40
1722  * Firmware Revision2 = 0x40       BCD encoded
1723  * IPMI Version = 0x51             IPMI 1.5
1724  * Manufacturer ID = A2 02 00      Dell IANA
1725  *
1726  * Additionally, PowerEdge systems with IPMI < 1.5 may also assert
1727  * OEM0_DATA_AVAIL and needs to be treated as RECEIVE_MSG_AVAIL.
1728  *
1729  */
1730 #define DELL_POWEREDGE_8G_BMC_DEVICE_ID  0x20
1731 #define DELL_POWEREDGE_8G_BMC_DEVICE_REV 0x80
1732 #define DELL_POWEREDGE_8G_BMC_IPMI_VERSION 0x51
1733 #define DELL_IANA_MFR_ID 0x0002a2
1734 static void setup_dell_poweredge_oem_data_handler(struct smi_info *smi_info)
1735 {
1736         struct ipmi_device_id *id = &smi_info->device_id;
1737         if (id->manufacturer_id == DELL_IANA_MFR_ID) {
1738                 if (id->device_id       == DELL_POWEREDGE_8G_BMC_DEVICE_ID  &&
1739                     id->device_revision == DELL_POWEREDGE_8G_BMC_DEVICE_REV &&
1740                     id->ipmi_version   == DELL_POWEREDGE_8G_BMC_IPMI_VERSION) {
1741                         smi_info->oem_data_avail_handler =
1742                                 oem_data_avail_to_receive_msg_avail;
1743                 } else if (ipmi_version_major(id) < 1 ||
1744                            (ipmi_version_major(id) == 1 &&
1745                             ipmi_version_minor(id) < 5)) {
1746                         smi_info->oem_data_avail_handler =
1747                                 oem_data_avail_to_receive_msg_avail;
1748                 }
1749         }
1750 }
1751
1752 #define CANNOT_RETURN_REQUESTED_LENGTH 0xCA
1753 static void return_hosed_msg_badsize(struct smi_info *smi_info)
1754 {
1755         struct ipmi_smi_msg *msg = smi_info->curr_msg;
1756
1757         /* Make it a response */
1758         msg->rsp[0] = msg->data[0] | 4;
1759         msg->rsp[1] = msg->data[1];
1760         msg->rsp[2] = CANNOT_RETURN_REQUESTED_LENGTH;
1761         msg->rsp_size = 3;
1762         smi_info->curr_msg = NULL;
1763         deliver_recv_msg(smi_info, msg);
1764 }
1765
1766 /*
1767  * dell_poweredge_bt_xaction_handler
1768  * @info - smi_info.device_id must be populated
1769  *
1770  * Dell PowerEdge servers with the BT interface (x6xx and 1750) will
1771  * not respond to a Get SDR command if the length of the data
1772  * requested is exactly 0x3A, which leads to command timeouts and no
1773  * data returned.  This intercepts such commands, and causes userspace
1774  * callers to try again with a different-sized buffer, which succeeds.
1775  */
1776
1777 #define STORAGE_NETFN 0x0A
1778 #define STORAGE_CMD_GET_SDR 0x23
1779 static int dell_poweredge_bt_xaction_handler(struct notifier_block *self,
1780                                              unsigned long unused,
1781                                              void *in)
1782 {
1783         struct smi_info *smi_info = in;
1784         unsigned char *data = smi_info->curr_msg->data;
1785         unsigned int size   = smi_info->curr_msg->data_size;
1786         if (size >= 8 &&
1787             (data[0]>>2) == STORAGE_NETFN &&
1788             data[1] == STORAGE_CMD_GET_SDR &&
1789             data[7] == 0x3A) {
1790                 return_hosed_msg_badsize(smi_info);
1791                 return NOTIFY_STOP;
1792         }
1793         return NOTIFY_DONE;
1794 }
1795
1796 static struct notifier_block dell_poweredge_bt_xaction_notifier = {
1797         .notifier_call  = dell_poweredge_bt_xaction_handler,
1798 };
1799
1800 /*
1801  * setup_dell_poweredge_bt_xaction_handler
1802  * @info - smi_info.device_id must be filled in already
1803  *
1804  * Fills in smi_info.device_id.start_transaction_pre_hook
1805  * when we know what function to use there.
1806  */
1807 static void
1808 setup_dell_poweredge_bt_xaction_handler(struct smi_info *smi_info)
1809 {
1810         struct ipmi_device_id *id = &smi_info->device_id;
1811         if (id->manufacturer_id == DELL_IANA_MFR_ID &&
1812             smi_info->io.si_type == SI_BT)
1813                 register_xaction_notifier(&dell_poweredge_bt_xaction_notifier);
1814 }
1815
1816 /*
1817  * setup_oem_data_handler
1818  * @info - smi_info.device_id must be filled in already
1819  *
1820  * Fills in smi_info.device_id.oem_data_available_handler
1821  * when we know what function to use there.
1822  */
1823
1824 static void setup_oem_data_handler(struct smi_info *smi_info)
1825 {
1826         setup_dell_poweredge_oem_data_handler(smi_info);
1827 }
1828
1829 static void setup_xaction_handlers(struct smi_info *smi_info)
1830 {
1831         setup_dell_poweredge_bt_xaction_handler(smi_info);
1832 }
1833
1834 static void check_for_broken_irqs(struct smi_info *smi_info)
1835 {
1836         check_clr_rcv_irq(smi_info);
1837         check_set_rcv_irq(smi_info);
1838 }
1839
1840 static inline void stop_timer_and_thread(struct smi_info *smi_info)
1841 {
1842         if (smi_info->thread != NULL) {
1843                 kthread_stop(smi_info->thread);
1844                 smi_info->thread = NULL;
1845         }
1846
1847         smi_info->timer_can_start = false;
1848         if (smi_info->timer_running)
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_type != info->io.addr_type)
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 &&
1885             ipmi_si_hardcode_match(io->addr_type, 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_type == IPMI_IO_ADDR_SPACE) {
1893                         io->io_setup = ipmi_si_port_setup;
1894                 } else if (io->addr_type == 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(PFX "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         char *init_name = NULL;
1952
1953         pr_info(PFX "Trying %s-specified %s state machine at %s address 0x%lx, slave address 0x%x, irq %d\n",
1954                 ipmi_addr_src_to_str(new_smi->io.addr_source),
1955                 si_to_str[new_smi->io.si_type],
1956                 addr_space_to_str[new_smi->io.addr_type],
1957                 new_smi->io.addr_data,
1958                 new_smi->io.slave_addr, new_smi->io.irq);
1959
1960         switch (new_smi->io.si_type) {
1961         case SI_KCS:
1962                 new_smi->handlers = &kcs_smi_handlers;
1963                 break;
1964
1965         case SI_SMIC:
1966                 new_smi->handlers = &smic_smi_handlers;
1967                 break;
1968
1969         case SI_BT:
1970                 new_smi->handlers = &bt_smi_handlers;
1971                 break;
1972
1973         default:
1974                 /* No support for anything else yet. */
1975                 rv = -EIO;
1976                 goto out_err;
1977         }
1978
1979         new_smi->si_num = smi_num;
1980
1981         /* Do this early so it's available for logs. */
1982         if (!new_smi->io.dev) {
1983                 init_name = kasprintf(GFP_KERNEL, "ipmi_si.%d",
1984                                       new_smi->si_num);
1985
1986                 /*
1987                  * If we don't already have a device from something
1988                  * else (like PCI), then register a new one.
1989                  */
1990                 new_smi->pdev = platform_device_alloc("ipmi_si",
1991                                                       new_smi->si_num);
1992                 if (!new_smi->pdev) {
1993                         pr_err(PFX "Unable to allocate platform device\n");
1994                         rv = -ENOMEM;
1995                         goto out_err;
1996                 }
1997                 new_smi->io.dev = &new_smi->pdev->dev;
1998                 new_smi->io.dev->driver = &ipmi_platform_driver.driver;
1999                 /* Nulled by device_add() */
2000                 new_smi->io.dev->init_name = init_name;
2001         }
2002
2003         /* Allocate the state machine's data and initialize it. */
2004         new_smi->si_sm = kmalloc(new_smi->handlers->size(), GFP_KERNEL);
2005         if (!new_smi->si_sm) {
2006                 rv = -ENOMEM;
2007                 goto out_err;
2008         }
2009         new_smi->io.io_size = new_smi->handlers->init_data(new_smi->si_sm,
2010                                                            &new_smi->io);
2011
2012         /* Now that we know the I/O size, we can set up the I/O. */
2013         rv = new_smi->io.io_setup(&new_smi->io);
2014         if (rv) {
2015                 dev_err(new_smi->io.dev, "Could not set up I/O space\n");
2016                 goto out_err;
2017         }
2018
2019         /* Do low-level detection first. */
2020         if (new_smi->handlers->detect(new_smi->si_sm)) {
2021                 if (new_smi->io.addr_source)
2022                         dev_err(new_smi->io.dev,
2023                                 "Interface detection failed\n");
2024                 rv = -ENODEV;
2025                 goto out_err;
2026         }
2027
2028         /*
2029          * Attempt a get device id command.  If it fails, we probably
2030          * don't have a BMC here.
2031          */
2032         rv = try_get_dev_id(new_smi);
2033         if (rv) {
2034                 if (new_smi->io.addr_source)
2035                         dev_err(new_smi->io.dev,
2036                                "There appears to be no BMC at this location\n");
2037                 goto out_err;
2038         }
2039
2040         setup_oem_data_handler(new_smi);
2041         setup_xaction_handlers(new_smi);
2042         check_for_broken_irqs(new_smi);
2043
2044         new_smi->waiting_msg = NULL;
2045         new_smi->curr_msg = NULL;
2046         atomic_set(&new_smi->req_events, 0);
2047         new_smi->run_to_completion = false;
2048         for (i = 0; i < SI_NUM_STATS; i++)
2049                 atomic_set(&new_smi->stats[i], 0);
2050
2051         new_smi->interrupt_disabled = true;
2052         atomic_set(&new_smi->need_watch, 0);
2053
2054         rv = try_enable_event_buffer(new_smi);
2055         if (rv == 0)
2056                 new_smi->has_event_buffer = true;
2057
2058         /*
2059          * Start clearing the flags before we enable interrupts or the
2060          * timer to avoid racing with the timer.
2061          */
2062         start_clear_flags(new_smi);
2063
2064         /*
2065          * IRQ is defined to be set when non-zero.  req_events will
2066          * cause a global flags check that will enable interrupts.
2067          */
2068         if (new_smi->io.irq) {
2069                 new_smi->interrupt_disabled = false;
2070                 atomic_set(&new_smi->req_events, 1);
2071         }
2072
2073         if (new_smi->pdev && !new_smi->pdev_registered) {
2074                 rv = platform_device_add(new_smi->pdev);
2075                 if (rv) {
2076                         dev_err(new_smi->io.dev,
2077                                 "Unable to register system interface device: %d\n",
2078                                 rv);
2079                         goto out_err;
2080                 }
2081                 new_smi->pdev_registered = true;
2082         }
2083
2084         dev_set_drvdata(new_smi->io.dev, new_smi);
2085         rv = device_add_group(new_smi->io.dev, &ipmi_si_dev_attr_group);
2086         if (rv) {
2087                 dev_err(new_smi->io.dev,
2088                         "Unable to add device attributes: error %d\n",
2089                         rv);
2090                 goto out_err;
2091         }
2092         new_smi->dev_group_added = true;
2093
2094         rv = ipmi_register_smi(&handlers,
2095                                new_smi,
2096                                new_smi->io.dev,
2097                                new_smi->io.slave_addr);
2098         if (rv) {
2099                 dev_err(new_smi->io.dev,
2100                         "Unable to register device: error %d\n",
2101                         rv);
2102                 goto out_err;
2103         }
2104
2105         /* Don't increment till we know we have succeeded. */
2106         smi_num++;
2107
2108         dev_info(new_smi->io.dev, "IPMI %s interface initialized\n",
2109                  si_to_str[new_smi->io.si_type]);
2110
2111         WARN_ON(new_smi->io.dev->init_name != NULL);
2112
2113  out_err:
2114         if (rv && new_smi->io.io_cleanup) {
2115                 new_smi->io.io_cleanup(&new_smi->io);
2116                 new_smi->io.io_cleanup = NULL;
2117         }
2118
2119         kfree(init_name);
2120         return rv;
2121 }
2122
2123 static int __init init_ipmi_si(void)
2124 {
2125         struct smi_info *e;
2126         enum ipmi_addr_src type = SI_INVALID;
2127
2128         if (initialized)
2129                 return 0;
2130
2131         ipmi_hardcode_init();
2132         pr_info("IPMI System Interface driver.\n");
2133
2134         ipmi_si_platform_init();
2135
2136         ipmi_si_pci_init();
2137
2138         ipmi_si_parisc_init();
2139
2140         /* We prefer devices with interrupts, but in the case of a machine
2141            with multiple BMCs we assume that there will be several instances
2142            of a given type so if we succeed in registering a type then also
2143            try to register everything else of the same type */
2144         mutex_lock(&smi_infos_lock);
2145         list_for_each_entry(e, &smi_infos, link) {
2146                 /* Try to register a device if it has an IRQ and we either
2147                    haven't successfully registered a device yet or this
2148                    device has the same type as one we successfully registered */
2149                 if (e->io.irq && (!type || e->io.addr_source == type)) {
2150                         if (!try_smi_init(e)) {
2151                                 type = e->io.addr_source;
2152                         }
2153                 }
2154         }
2155
2156         /* type will only have been set if we successfully registered an si */
2157         if (type)
2158                 goto skip_fallback_noirq;
2159
2160         /* Fall back to the preferred device */
2161
2162         list_for_each_entry(e, &smi_infos, link) {
2163                 if (!e->io.irq && (!type || e->io.addr_source == type)) {
2164                         if (!try_smi_init(e)) {
2165                                 type = e->io.addr_source;
2166                         }
2167                 }
2168         }
2169
2170 skip_fallback_noirq:
2171         initialized = 1;
2172         mutex_unlock(&smi_infos_lock);
2173
2174         if (type)
2175                 return 0;
2176
2177         mutex_lock(&smi_infos_lock);
2178         if (unload_when_empty && list_empty(&smi_infos)) {
2179                 mutex_unlock(&smi_infos_lock);
2180                 cleanup_ipmi_si();
2181                 pr_warn(PFX "Unable to find any System Interface(s)\n");
2182                 return -ENODEV;
2183         } else {
2184                 mutex_unlock(&smi_infos_lock);
2185                 return 0;
2186         }
2187 }
2188 module_init(init_ipmi_si);
2189
2190 static void shutdown_smi(void *send_info)
2191 {
2192         struct smi_info *smi_info = send_info;
2193
2194         if (smi_info->dev_group_added) {
2195                 device_remove_group(smi_info->io.dev, &ipmi_si_dev_attr_group);
2196                 smi_info->dev_group_added = false;
2197         }
2198         if (smi_info->io.dev)
2199                 dev_set_drvdata(smi_info->io.dev, NULL);
2200
2201         /*
2202          * Make sure that interrupts, the timer and the thread are
2203          * stopped and will not run again.
2204          */
2205         smi_info->interrupt_disabled = true;
2206         if (smi_info->io.irq_cleanup) {
2207                 smi_info->io.irq_cleanup(&smi_info->io);
2208                 smi_info->io.irq_cleanup = NULL;
2209         }
2210         stop_timer_and_thread(smi_info);
2211
2212         /*
2213          * Wait until we know that we are out of any interrupt
2214          * handlers might have been running before we freed the
2215          * interrupt.
2216          */
2217         synchronize_sched();
2218
2219         /*
2220          * Timeouts are stopped, now make sure the interrupts are off
2221          * in the BMC.  Note that timers and CPU interrupts are off,
2222          * so no need for locks.
2223          */
2224         while (smi_info->curr_msg || (smi_info->si_state != SI_NORMAL)) {
2225                 poll(smi_info);
2226                 schedule_timeout_uninterruptible(1);
2227         }
2228         if (smi_info->handlers)
2229                 disable_si_irq(smi_info);
2230         while (smi_info->curr_msg || (smi_info->si_state != SI_NORMAL)) {
2231                 poll(smi_info);
2232                 schedule_timeout_uninterruptible(1);
2233         }
2234         if (smi_info->handlers)
2235                 smi_info->handlers->cleanup(smi_info->si_sm);
2236
2237         if (smi_info->io.addr_source_cleanup) {
2238                 smi_info->io.addr_source_cleanup(&smi_info->io);
2239                 smi_info->io.addr_source_cleanup = NULL;
2240         }
2241         if (smi_info->io.io_cleanup) {
2242                 smi_info->io.io_cleanup(&smi_info->io);
2243                 smi_info->io.io_cleanup = NULL;
2244         }
2245
2246         kfree(smi_info->si_sm);
2247         smi_info->si_sm = NULL;
2248
2249         smi_info->intf = NULL;
2250 }
2251
2252 /*
2253  * Must be called with smi_infos_lock held, to serialize the
2254  * smi_info->intf check.
2255  */
2256 static void cleanup_one_si(struct smi_info *smi_info)
2257 {
2258         if (!smi_info)
2259                 return;
2260
2261         list_del(&smi_info->link);
2262
2263         if (smi_info->intf)
2264                 ipmi_unregister_smi(smi_info->intf);
2265
2266         if (smi_info->pdev) {
2267                 if (smi_info->pdev_registered)
2268                         platform_device_unregister(smi_info->pdev);
2269                 else
2270                         platform_device_put(smi_info->pdev);
2271         }
2272
2273         kfree(smi_info);
2274 }
2275
2276 int ipmi_si_remove_by_dev(struct device *dev)
2277 {
2278         struct smi_info *e;
2279         int rv = -ENOENT;
2280
2281         mutex_lock(&smi_infos_lock);
2282         list_for_each_entry(e, &smi_infos, link) {
2283                 if (e->io.dev == dev) {
2284                         cleanup_one_si(e);
2285                         rv = 0;
2286                         break;
2287                 }
2288         }
2289         mutex_unlock(&smi_infos_lock);
2290
2291         return rv;
2292 }
2293
2294 void ipmi_si_remove_by_data(int addr_space, enum si_type si_type,
2295                             unsigned long addr)
2296 {
2297         /* remove */
2298         struct smi_info *e, *tmp_e;
2299
2300         mutex_lock(&smi_infos_lock);
2301         list_for_each_entry_safe(e, tmp_e, &smi_infos, link) {
2302                 if (e->io.addr_type != addr_space)
2303                         continue;
2304                 if (e->io.si_type != si_type)
2305                         continue;
2306                 if (e->io.addr_data == addr)
2307                         cleanup_one_si(e);
2308         }
2309         mutex_unlock(&smi_infos_lock);
2310 }
2311
2312 static void cleanup_ipmi_si(void)
2313 {
2314         struct smi_info *e, *tmp_e;
2315
2316         if (!initialized)
2317                 return;
2318
2319         ipmi_si_pci_shutdown();
2320
2321         ipmi_si_parisc_shutdown();
2322
2323         ipmi_si_platform_shutdown();
2324
2325         mutex_lock(&smi_infos_lock);
2326         list_for_each_entry_safe(e, tmp_e, &smi_infos, link)
2327                 cleanup_one_si(e);
2328         mutex_unlock(&smi_infos_lock);
2329
2330         ipmi_si_hardcode_exit();
2331 }
2332 module_exit(cleanup_ipmi_si);
2333
2334 MODULE_ALIAS("platform:dmi-ipmi-si");
2335 MODULE_LICENSE("GPL");
2336 MODULE_AUTHOR("Corey Minyard <minyard@mvista.com>");
2337 MODULE_DESCRIPTION("Interface to the IPMI driver for the KCS, SMIC, and BT"
2338                    " system interfaces.");