drivers/s390/cio/cmf.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Linux on zSeries Channel Measurement Facility support
   4  *
   5  * Copyright IBM Corp. 2000, 2006
   6  *
   7  * Authors: Arnd Bergmann <arndb@de.ibm.com>
   8  *          Cornelia Huck <cornelia.huck@de.ibm.com>
   9  *
  10  * original idea from Natarajan Krishnaswami <nkrishna@us.ibm.com>
  11  */
  12
  13 #define KMSG_COMPONENT "cio"
  14 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  15
  16 #include <linux/bootmem.h>
  17 #include <linux/device.h>
  18 #include <linux/init.h>
  19 #include <linux/list.h>
  20 #include <linux/export.h>
  21 #include <linux/moduleparam.h>
  22 #include <linux/slab.h>
  23 #include <linux/timex.h>        /* get_tod_clock() */
  24
  25 #include <asm/ccwdev.h>
  26 #include <asm/cio.h>
  27 #include <asm/cmb.h>
  28 #include <asm/div64.h>
  29
  30 #include "cio.h"
  31 #include "css.h"
  32 #include "device.h"
  33 #include "ioasm.h"
  34 #include "chsc.h"
  35
  36 /*
  37  * parameter to enable cmf during boot, possible uses are:
  38  *  "s390cmf" -- enable cmf and allocate 2 MB of ram so measuring can be
  39  *               used on any subchannel
  40  *  "s390cmf=<num>" -- enable cmf and allocate enough memory to measure
  41  *                     <num> subchannel, where <num> is an integer
  42  *                     between 1 and 65535, default is 1024
  43  */
  44 #define ARGSTRING "s390cmf"
  45
  46 /* indices for READCMB */
  47 enum cmb_index {
  48         avg_utilization = -1,
  49  /* basic and exended format: */
  50         cmb_ssch_rsch_count = 0,
  51         cmb_sample_count,
  52         cmb_device_connect_time,
  53         cmb_function_pending_time,
  54         cmb_device_disconnect_time,
  55         cmb_control_unit_queuing_time,
  56         cmb_device_active_only_time,
  57  /* extended format only: */
  58         cmb_device_busy_time,
  59         cmb_initial_command_response_time,
  60 };
  61
  62 /**
  63  * enum cmb_format - types of supported measurement block formats
  64  *
  65  * @CMF_BASIC:      traditional channel measurement blocks supported
  66  *                  by all machines that we run on
  67  * @CMF_EXTENDED:   improved format that was introduced with the z990
  68  *                  machine
  69  * @CMF_AUTODETECT: default: use extended format when running on a machine
  70  *                  supporting extended format, otherwise fall back to
  71  *                  basic format
  72  */
  73 enum cmb_format {
  74         CMF_BASIC,
  75         CMF_EXTENDED,
  76         CMF_AUTODETECT = -1,
  77 };
  78
  79 /*
  80  * format - actual format for all measurement blocks
  81  *
  82  * The format module parameter can be set to a value of 0 (zero)
  83  * or 1, indicating basic or extended format as described for
  84  * enum cmb_format.
  85  */
  86 static int format = CMF_AUTODETECT;
  87 module_param(format, bint, 0444);
  88
  89 /**
  90  * struct cmb_operations - functions to use depending on cmb_format
  91  *
  92  * Most of these functions operate on a struct ccw_device. There is only
  93  * one instance of struct cmb_operations because the format of the measurement
  94  * data is guaranteed to be the same for every ccw_device.
  95  *
  96  * @alloc:      allocate memory for a channel measurement block,
  97  *              either with the help of a special pool or with kmalloc
  98  * @free:       free memory allocated with @alloc
  99  * @set:        enable or disable measurement
 100  * @read:       read a measurement entry at an index
 101  * @readall:    read a measurement block in a common format
 102  * @reset:      clear the data in the associated measurement block and
 103  *              reset its time stamp
 104  */
 105 struct cmb_operations {
 106         int  (*alloc)  (struct ccw_device *);
 107         void (*free)   (struct ccw_device *);
 108         int  (*set)    (struct ccw_device *, u32);
 109         u64  (*read)   (struct ccw_device *, int);
 110         int  (*readall)(struct ccw_device *, struct cmbdata *);
 111         void (*reset)  (struct ccw_device *);
 112 /* private: */
 113         struct attribute_group *attr_group;
 114 };
 115 static struct cmb_operations *cmbops;
 116
 117 struct cmb_data {
 118         void *hw_block;   /* Pointer to block updated by hardware */
 119         void *last_block; /* Last changed block copied from hardware block */
 120         int size;         /* Size of hw_block and last_block */
 121         unsigned long long last_update;  /* when last_block was updated */
 122 };
 123
 124 /*
 125  * Our user interface is designed in terms of nanoseconds,
 126  * while the hardware measures total times in its own
 127  * unit.
 128  */
 129 static inline u64 time_to_nsec(u32 value)
 130 {
 131         return ((u64)value) * 128000ull;
 132 }
 133
 134 /*
 135  * Users are usually interested in average times,
 136  * not accumulated time.
 137  * This also helps us with atomicity problems
 138  * when reading sinlge values.
 139  */
 140 static inline u64 time_to_avg_nsec(u32 value, u32 count)
 141 {
 142         u64 ret;
 143
 144         /* no samples yet, avoid division by 0 */
 145         if (count == 0)
 146                 return 0;
 147
 148         /* value comes in units of 128 µsec */
 149         ret = time_to_nsec(value);
 150         do_div(ret, count);
 151
 152         return ret;
 153 }
 154
 155 #define CMF_OFF 0
 156 #define CMF_ON  2
 157
 158 /*
 159  * Activate or deactivate the channel monitor. When area is NULL,
 160  * the monitor is deactivated. The channel monitor needs to
 161  * be active in order to measure subchannels, which also need
 162  * to be enabled.
 163  */
 164 static inline void cmf_activate(void *area, unsigned int onoff)
 165 {
 166         register void * __gpr2 asm("2");
 167         register long __gpr1 asm("1");
 168
 169         __gpr2 = area;
 170         __gpr1 = onoff;
 171         /* activate channel measurement */
 172         asm("schm" : : "d" (__gpr2), "d" (__gpr1) );
 173 }
 174
 175 static int set_schib(struct ccw_device *cdev, u32 mme, int mbfc,
 176                      unsigned long address)
 177 {
 178         struct subchannel *sch = to_subchannel(cdev->dev.parent);
 179         int ret;
 180
 181         sch->config.mme = mme;
 182         sch->config.mbfc = mbfc;
 183         /* address can be either a block address or a block index */
 184         if (mbfc)
 185                 sch->config.mba = address;
 186         else
 187                 sch->config.mbi = address;
 188
 189         ret = cio_commit_config(sch);
 190         if (!mme && ret == -ENODEV) {
 191                 /*
 192                  * The task was to disable measurement block updates but
 193                  * the subchannel is already gone. Report success.
 194                  */
 195                 ret = 0;
 196         }
 197         return ret;
 198 }
 199
 200 struct set_schib_struct {
 201         u32 mme;
 202         int mbfc;
 203         unsigned long address;
 204         wait_queue_head_t wait;
 205         int ret;
 206 };
 207
 208 #define CMF_PENDING 1
 209 #define SET_SCHIB_TIMEOUT (10 * HZ)
 210
 211 static int set_schib_wait(struct ccw_device *cdev, u32 mme,
 212                           int mbfc, unsigned long address)
 213 {
 214         struct set_schib_struct set_data;
 215         int ret = -ENODEV;
 216
 217         spin_lock_irq(cdev->ccwlock);
 218         if (!cdev->private->cmb)
 219                 goto out;
 220
 221         ret = set_schib(cdev, mme, mbfc, address);
 222         if (ret != -EBUSY)
 223                 goto out;
 224
 225         /* if the device is not online, don't even try again */
 226         if (cdev->private->state != DEV_STATE_ONLINE)
 227                 goto out;
 228
 229         init_waitqueue_head(&set_data.wait);
 230         set_data.mme = mme;
 231         set_data.mbfc = mbfc;
 232         set_data.address = address;
 233         set_data.ret = CMF_PENDING;
 234
 235         cdev->private->state = DEV_STATE_CMFCHANGE;
 236         cdev->private->cmb_wait = &set_data;
 237         spin_unlock_irq(cdev->ccwlock);
 238
 239         ret = wait_event_interruptible_timeout(set_data.wait,
 240                                                set_data.ret != CMF_PENDING,
 241                                                SET_SCHIB_TIMEOUT);
 242         spin_lock_irq(cdev->ccwlock);
 243         if (ret <= 0) {
 244                 if (set_data.ret == CMF_PENDING) {
 245                         set_data.ret = (ret == 0) ? -ETIME : ret;
 246                         if (cdev->private->state == DEV_STATE_CMFCHANGE)
 247                                 cdev->private->state = DEV_STATE_ONLINE;
 248                 }
 249         }
 250         cdev->private->cmb_wait = NULL;
 251         ret = set_data.ret;
 252 out:
 253         spin_unlock_irq(cdev->ccwlock);
 254         return ret;
 255 }
 256
 257 void retry_set_schib(struct ccw_device *cdev)
 258 {
 259         struct set_schib_struct *set_data = cdev->private->cmb_wait;
 260
 261         if (!set_data)
 262                 return;
 263
 264         set_data->ret = set_schib(cdev, set_data->mme, set_data->mbfc,
 265                                   set_data->address);
 266         wake_up(&set_data->wait);
 267 }
 268
 269 static int cmf_copy_block(struct ccw_device *cdev)
 270 {
 271         struct subchannel *sch = to_subchannel(cdev->dev.parent);
 272         struct cmb_data *cmb_data;
 273         void *hw_block;
 274
 275         if (cio_update_schib(sch))
 276                 return -ENODEV;
 277
 278         if (scsw_fctl(&sch->schib.scsw) & SCSW_FCTL_START_FUNC) {
 279                 /* Don't copy if a start function is in progress. */
 280                 if ((!(scsw_actl(&sch->schib.scsw) & SCSW_ACTL_SUSPENDED)) &&
 281                     (scsw_actl(&sch->schib.scsw) &
 282                      (SCSW_ACTL_DEVACT | SCSW_ACTL_SCHACT)) &&
 283                     (!(scsw_stctl(&sch->schib.scsw) & SCSW_STCTL_SEC_STATUS)))
 284                         return -EBUSY;
 285         }
 286         cmb_data = cdev->private->cmb;
 287         hw_block = cmb_data->hw_block;
 288         memcpy(cmb_data->last_block, hw_block, cmb_data->size);
 289         cmb_data->last_update = get_tod_clock();
 290         return 0;
 291 }
 292
 293 struct copy_block_struct {
 294         wait_queue_head_t wait;
 295         int ret;
 296 };
 297
 298 static int cmf_cmb_copy_wait(struct ccw_device *cdev)
 299 {
 300         struct copy_block_struct copy_block;
 301         int ret = -ENODEV;
 302
 303         spin_lock_irq(cdev->ccwlock);
 304         if (!cdev->private->cmb)
 305                 goto out;
 306
 307         ret = cmf_copy_block(cdev);
 308         if (ret != -EBUSY)
 309                 goto out;
 310
 311         if (cdev->private->state != DEV_STATE_ONLINE)
 312                 goto out;
 313
 314         init_waitqueue_head(&copy_block.wait);
 315         copy_block.ret = CMF_PENDING;
 316
 317         cdev->private->state = DEV_STATE_CMFUPDATE;
 318         cdev->private->cmb_wait = &copy_block;
 319         spin_unlock_irq(cdev->ccwlock);
 320
 321         ret = wait_event_interruptible(copy_block.wait,
 322                                        copy_block.ret != CMF_PENDING);
 323         spin_lock_irq(cdev->ccwlock);
 324         if (ret) {
 325                 if (copy_block.ret == CMF_PENDING) {
 326                         copy_block.ret = -ERESTARTSYS;
 327                         if (cdev->private->state == DEV_STATE_CMFUPDATE)
 328                                 cdev->private->state = DEV_STATE_ONLINE;
 329                 }
 330         }
 331         cdev->private->cmb_wait = NULL;
 332         ret = copy_block.ret;
 333 out:
 334         spin_unlock_irq(cdev->ccwlock);
 335         return ret;
 336 }
 337
 338 void cmf_retry_copy_block(struct ccw_device *cdev)
 339 {
 340         struct copy_block_struct *copy_block = cdev->private->cmb_wait;
 341
 342         if (!copy_block)
 343                 return;
 344
 345         copy_block->ret = cmf_copy_block(cdev);
 346         wake_up(&copy_block->wait);
 347 }
 348
 349 static void cmf_generic_reset(struct ccw_device *cdev)
 350 {
 351         struct cmb_data *cmb_data;
 352
 353         spin_lock_irq(cdev->ccwlock);
 354         cmb_data = cdev->private->cmb;
 355         if (cmb_data) {
 356                 memset(cmb_data->last_block, 0, cmb_data->size);
 357                 /*
 358                  * Need to reset hw block as well to make the hardware start
 359                  * from 0 again.
 360                  */
 361                 memset(cmb_data->hw_block, 0, cmb_data->size);
 362                 cmb_data->last_update = 0;
 363         }
 364         cdev->private->cmb_start_time = get_tod_clock();
 365         spin_unlock_irq(cdev->ccwlock);
 366 }
 367
 368 /**
 369  * struct cmb_area - container for global cmb data
 370  *
 371  * @mem:        pointer to CMBs (only in basic measurement mode)
 372  * @list:       contains a linked list of all subchannels
 373  * @num_channels: number of channels to be measured
 374  * @lock:       protect concurrent access to @mem and @list
 375  */
 376 struct cmb_area {
 377         struct cmb *mem;
 378         struct list_head list;
 379         int num_channels;
 380         spinlock_t lock;
 381 };
 382
 383 static struct cmb_area cmb_area = {
 384         .lock = __SPIN_LOCK_UNLOCKED(cmb_area.lock),
 385         .list = LIST_HEAD_INIT(cmb_area.list),
 386         .num_channels  = 1024,
 387 };
 388
 389 /* ****** old style CMB handling ********/
 390
 391 /*
 392  * Basic channel measurement blocks are allocated in one contiguous
 393  * block of memory, which can not be moved as long as any channel
 394  * is active. Therefore, a maximum number of subchannels needs to
 395  * be defined somewhere. This is a module parameter, defaulting to
 396  * a reasonable value of 1024, or 32 kb of memory.
 397  * Current kernels don't allow kmalloc with more than 128kb, so the
 398  * maximum is 4096.
 399  */
 400
 401 module_param_named(maxchannels, cmb_area.num_channels, uint, 0444);
 402
 403 /**
 404  * struct cmb - basic channel measurement block
 405  * @ssch_rsch_count: number of ssch and rsch
 406  * @sample_count: number of samples
 407  * @device_connect_time: time of device connect
 408  * @function_pending_time: time of function pending
 409  * @device_disconnect_time: time of device disconnect
 410  * @control_unit_queuing_time: time of control unit queuing
 411  * @device_active_only_time: time of device active only
 412  * @reserved: unused in basic measurement mode
 413  *
 414  * The measurement block as used by the hardware. The fields are described
 415  * further in z/Architecture Principles of Operation, chapter 17.
 416  *
 417  * The cmb area made up from these blocks must be a contiguous array and may
 418  * not be reallocated or freed.
 419  * Only one cmb area can be present in the system.
 420  */
 421 struct cmb {
 422         u16 ssch_rsch_count;
 423         u16 sample_count;
 424         u32 device_connect_time;
 425         u32 function_pending_time;
 426         u32 device_disconnect_time;
 427         u32 control_unit_queuing_time;
 428         u32 device_active_only_time;
 429         u32 reserved[2];
 430 };
 431
 432 /*
 433  * Insert a single device into the cmb_area list.
 434  * Called with cmb_area.lock held from alloc_cmb.
 435  */
 436 static int alloc_cmb_single(struct ccw_device *cdev,
 437                             struct cmb_data *cmb_data)
 438 {
 439         struct cmb *cmb;
 440         struct ccw_device_private *node;
 441         int ret;
 442
 443         spin_lock_irq(cdev->ccwlock);
 444         if (!list_empty(&cdev->private->cmb_list)) {
 445                 ret = -EBUSY;
 446                 goto out;
 447         }
 448
 449         /*
 450          * Find first unused cmb in cmb_area.mem.
 451          * This is a little tricky: cmb_area.list
 452          * remains sorted by ->cmb->hw_data pointers.
 453          */
 454         cmb = cmb_area.mem;
 455         list_for_each_entry(node, &cmb_area.list, cmb_list) {
 456                 struct cmb_data *data;
 457                 data = node->cmb;
 458                 if ((struct cmb*)data->hw_block > cmb)
 459                         break;
 460                 cmb++;
 461         }
 462         if (cmb - cmb_area.mem >= cmb_area.num_channels) {
 463                 ret = -ENOMEM;
 464                 goto out;
 465         }
 466
 467         /* insert new cmb */
 468         list_add_tail(&cdev->private->cmb_list, &node->cmb_list);
 469         cmb_data->hw_block = cmb;
 470         cdev->private->cmb = cmb_data;
 471         ret = 0;
 472 out:
 473         spin_unlock_irq(cdev->ccwlock);
 474         return ret;
 475 }
 476
 477 static int alloc_cmb(struct ccw_device *cdev)
 478 {
 479         int ret;
 480         struct cmb *mem;
 481         ssize_t size;
 482         struct cmb_data *cmb_data;
 483
 484         /* Allocate private cmb_data. */
 485         cmb_data = kzalloc(sizeof(struct cmb_data), GFP_KERNEL);
 486         if (!cmb_data)
 487                 return -ENOMEM;
 488
 489         cmb_data->last_block = kzalloc(sizeof(struct cmb), GFP_KERNEL);
 490         if (!cmb_data->last_block) {
 491                 kfree(cmb_data);
 492                 return -ENOMEM;
 493         }
 494         cmb_data->size = sizeof(struct cmb);
 495         spin_lock(&cmb_area.lock);
 496
 497         if (!cmb_area.mem) {
 498                 /* there is no user yet, so we need a new area */
 499                 size = sizeof(struct cmb) * cmb_area.num_channels;
 500                 WARN_ON(!list_empty(&cmb_area.list));
 501
 502                 spin_unlock(&cmb_area.lock);
 503                 mem = (void*)__get_free_pages(GFP_KERNEL | GFP_DMA,
 504                                  get_order(size));
 505                 spin_lock(&cmb_area.lock);
 506
 507                 if (cmb_area.mem) {
 508                         /* ok, another thread was faster */
 509                         free_pages((unsigned long)mem, get_order(size));
 510                 } else if (!mem) {
 511                         /* no luck */
 512                         ret = -ENOMEM;
 513                         goto out;
 514                 } else {
 515                         /* everything ok */
 516                         memset(mem, 0, size);
 517                         cmb_area.mem = mem;
 518                         cmf_activate(cmb_area.mem, CMF_ON);
 519                 }
 520         }
 521
 522         /* do the actual allocation */
 523         ret = alloc_cmb_single(cdev, cmb_data);
 524 out:
 525         spin_unlock(&cmb_area.lock);
 526         if (ret) {
 527                 kfree(cmb_data->last_block);
 528                 kfree(cmb_data);
 529         }
 530         return ret;
 531 }
 532
 533 static void free_cmb(struct ccw_device *cdev)
 534 {
 535         struct ccw_device_private *priv;
 536         struct cmb_data *cmb_data;
 537
 538         spin_lock(&cmb_area.lock);
 539         spin_lock_irq(cdev->ccwlock);
 540
 541         priv = cdev->private;
 542         cmb_data = priv->cmb;
 543         priv->cmb = NULL;
 544         if (cmb_data)
 545                 kfree(cmb_data->last_block);
 546         kfree(cmb_data);
 547         list_del_init(&priv->cmb_list);
 548
 549         if (list_empty(&cmb_area.list)) {
 550                 ssize_t size;
 551                 size = sizeof(struct cmb) * cmb_area.num_channels;
 552                 cmf_activate(NULL, CMF_OFF);
 553                 free_pages((unsigned long)cmb_area.mem, get_order(size));
 554                 cmb_area.mem = NULL;
 555         }
 556         spin_unlock_irq(cdev->ccwlock);
 557         spin_unlock(&cmb_area.lock);
 558 }
 559
 560 static int set_cmb(struct ccw_device *cdev, u32 mme)
 561 {
 562         u16 offset;
 563         struct cmb_data *cmb_data;
 564         unsigned long flags;
 565
 566         spin_lock_irqsave(cdev->ccwlock, flags);
 567         if (!cdev->private->cmb) {
 568                 spin_unlock_irqrestore(cdev->ccwlock, flags);
 569                 return -EINVAL;
 570         }
 571         cmb_data = cdev->private->cmb;
 572         offset = mme ? (struct cmb *)cmb_data->hw_block - cmb_area.mem : 0;
 573         spin_unlock_irqrestore(cdev->ccwlock, flags);
 574
 575         return set_schib_wait(cdev, mme, 0, offset);
 576 }
 577
 578 /* calculate utilization in 0.1 percent units */
 579 static u64 __cmb_utilization(u64 device_connect_time, u64 function_pending_time,
 580                              u64 device_disconnect_time, u64 start_time)
 581 {
 582         u64 utilization, elapsed_time;
 583
 584         utilization = time_to_nsec(device_connect_time +
 585                                    function_pending_time +
 586                                    device_disconnect_time);
 587
 588         elapsed_time = get_tod_clock() - start_time;
 589         elapsed_time = tod_to_ns(elapsed_time);
 590         elapsed_time /= 1000;
 591
 592         return elapsed_time ? (utilization / elapsed_time) : 0;
 593 }
 594
 595 static u64 read_cmb(struct ccw_device *cdev, int index)
 596 {
 597         struct cmb_data *cmb_data;
 598         unsigned long flags;
 599         struct cmb *cmb;
 600         u64 ret = 0;
 601         u32 val;
 602
 603         spin_lock_irqsave(cdev->ccwlock, flags);
 604         cmb_data = cdev->private->cmb;
 605         if (!cmb_data)
 606                 goto out;
 607
 608         cmb = cmb_data->hw_block;
 609         switch (index) {
 610         case avg_utilization:
 611                 ret = __cmb_utilization(cmb->device_connect_time,
 612                                         cmb->function_pending_time,
 613                                         cmb->device_disconnect_time,
 614                                         cdev->private->cmb_start_time);
 615                 goto out;
 616         case cmb_ssch_rsch_count:
 617                 ret = cmb->ssch_rsch_count;
 618                 goto out;
 619         case cmb_sample_count:
 620                 ret = cmb->sample_count;
 621                 goto out;
 622         case cmb_device_connect_time:
 623                 val = cmb->device_connect_time;
 624                 break;
 625         case cmb_function_pending_time:
 626                 val = cmb->function_pending_time;
 627                 break;
 628         case cmb_device_disconnect_time:
 629                 val = cmb->device_disconnect_time;
 630                 break;
 631         case cmb_control_unit_queuing_time:
 632                 val = cmb->control_unit_queuing_time;
 633                 break;
 634         case cmb_device_active_only_time:
 635                 val = cmb->device_active_only_time;
 636                 break;
 637         default:
 638                 goto out;
 639         }
 640         ret = time_to_avg_nsec(val, cmb->sample_count);
 641 out:
 642         spin_unlock_irqrestore(cdev->ccwlock, flags);
 643         return ret;
 644 }
 645
 646 static int readall_cmb(struct ccw_device *cdev, struct cmbdata *data)
 647 {
 648         struct cmb *cmb;
 649         struct cmb_data *cmb_data;
 650         u64 time;
 651         unsigned long flags;
 652         int ret;
 653
 654         ret = cmf_cmb_copy_wait(cdev);
 655         if (ret < 0)
 656                 return ret;
 657         spin_lock_irqsave(cdev->ccwlock, flags);
 658         cmb_data = cdev->private->cmb;
 659         if (!cmb_data) {
 660                 ret = -ENODEV;
 661                 goto out;
 662         }
 663         if (cmb_data->last_update == 0) {
 664                 ret = -EAGAIN;
 665                 goto out;
 666         }
 667         cmb = cmb_data->last_block;
 668         time = cmb_data->last_update - cdev->private->cmb_start_time;
 669
 670         memset(data, 0, sizeof(struct cmbdata));
 671
 672         /* we only know values before device_busy_time */
 673         data->size = offsetof(struct cmbdata, device_busy_time);
 674
 675         data->elapsed_time = tod_to_ns(time);
 676
 677         /* copy data to new structure */
 678         data->ssch_rsch_count = cmb->ssch_rsch_count;
 679         data->sample_count = cmb->sample_count;
 680
 681         /* time fields are converted to nanoseconds while copying */
 682         data->device_connect_time = time_to_nsec(cmb->device_connect_time);
 683         data->function_pending_time = time_to_nsec(cmb->function_pending_time);
 684         data->device_disconnect_time =
 685                 time_to_nsec(cmb->device_disconnect_time);
 686         data->control_unit_queuing_time
 687                 = time_to_nsec(cmb->control_unit_queuing_time);
 688         data->device_active_only_time
 689                 = time_to_nsec(cmb->device_active_only_time);
 690         ret = 0;
 691 out:
 692         spin_unlock_irqrestore(cdev->ccwlock, flags);
 693         return ret;
 694 }
 695
 696 static void reset_cmb(struct ccw_device *cdev)
 697 {
 698         cmf_generic_reset(cdev);
 699 }
 700
 701 static int cmf_enabled(struct ccw_device *cdev)
 702 {
 703         int enabled;
 704
 705         spin_lock_irq(cdev->ccwlock);
 706         enabled = !!cdev->private->cmb;
 707         spin_unlock_irq(cdev->ccwlock);
 708
 709         return enabled;
 710 }
 711
 712 static struct attribute_group cmf_attr_group;
 713
 714 static struct cmb_operations cmbops_basic = {
 715         .alloc  = alloc_cmb,
 716         .free   = free_cmb,
 717         .set    = set_cmb,
 718         .read   = read_cmb,
 719         .readall    = readall_cmb,
 720         .reset      = reset_cmb,
 721         .attr_group = &cmf_attr_group,
 722 };
 723
 724 /* ******** extended cmb handling ********/
 725
 726 /**
 727  * struct cmbe - extended channel measurement block
 728  * @ssch_rsch_count: number of ssch and rsch
 729  * @sample_count: number of samples
 730  * @device_connect_time: time of device connect
 731  * @function_pending_time: time of function pending
 732  * @device_disconnect_time: time of device disconnect
 733  * @control_unit_queuing_time: time of control unit queuing
 734  * @device_active_only_time: time of device active only
 735  * @device_busy_time: time of device busy
 736  * @initial_command_response_time: initial command response time
 737  * @reserved: unused
 738  *
 739  * The measurement block as used by the hardware. May be in any 64 bit physical
 740  * location.
 741  * The fields are described further in z/Architecture Principles of Operation,
 742  * third edition, chapter 17.
 743  */
 744 struct cmbe {
 745         u32 ssch_rsch_count;
 746         u32 sample_count;
 747         u32 device_connect_time;
 748         u32 function_pending_time;
 749         u32 device_disconnect_time;
 750         u32 control_unit_queuing_time;
 751         u32 device_active_only_time;
 752         u32 device_busy_time;
 753         u32 initial_command_response_time;
 754         u32 reserved[7];
 755 } __packed __aligned(64);
 756
 757 static struct kmem_cache *cmbe_cache;
 758
 759 static int alloc_cmbe(struct ccw_device *cdev)
 760 {
 761         struct cmb_data *cmb_data;
 762         struct cmbe *cmbe;
 763         int ret = -ENOMEM;
 764
 765         cmbe = kmem_cache_zalloc(cmbe_cache, GFP_KERNEL);
 766         if (!cmbe)
 767                 return ret;
 768
 769         cmb_data = kzalloc(sizeof(*cmb_data), GFP_KERNEL);
 770         if (!cmb_data)
 771                 goto out_free;
 772
 773         cmb_data->last_block = kzalloc(sizeof(struct cmbe), GFP_KERNEL);
 774         if (!cmb_data->last_block)
 775                 goto out_free;
 776
 777         cmb_data->size = sizeof(*cmbe);
 778         cmb_data->hw_block = cmbe;
 779
 780         spin_lock(&cmb_area.lock);
 781         spin_lock_irq(cdev->ccwlock);
 782         if (cdev->private->cmb)
 783                 goto out_unlock;
 784
 785         cdev->private->cmb = cmb_data;
 786
 787         /* activate global measurement if this is the first channel */
 788         if (list_empty(&cmb_area.list))
 789                 cmf_activate(NULL, CMF_ON);
 790         list_add_tail(&cdev->private->cmb_list, &cmb_area.list);
 791
 792         spin_unlock_irq(cdev->ccwlock);
 793         spin_unlock(&cmb_area.lock);
 794         return 0;
 795
 796 out_unlock:
 797         spin_unlock_irq(cdev->ccwlock);
 798         spin_unlock(&cmb_area.lock);
 799         ret = -EBUSY;
 800 out_free:
 801         if (cmb_data)
 802                 kfree(cmb_data->last_block);
 803         kfree(cmb_data);
 804         kmem_cache_free(cmbe_cache, cmbe);
 805
 806         return ret;
 807 }
 808
 809 static void free_cmbe(struct ccw_device *cdev)
 810 {
 811         struct cmb_data *cmb_data;
 812
 813         spin_lock(&cmb_area.lock);
 814         spin_lock_irq(cdev->ccwlock);
 815         cmb_data = cdev->private->cmb;
 816         cdev->private->cmb = NULL;
 817         if (cmb_data) {
 818                 kfree(cmb_data->last_block);
 819                 kmem_cache_free(cmbe_cache, cmb_data->hw_block);
 820         }
 821         kfree(cmb_data);
 822
 823         /* deactivate global measurement if this is the last channel */
 824         list_del_init(&cdev->private->cmb_list);
 825         if (list_empty(&cmb_area.list))
 826                 cmf_activate(NULL, CMF_OFF);
 827         spin_unlock_irq(cdev->ccwlock);
 828         spin_unlock(&cmb_area.lock);
 829 }
 830
 831 static int set_cmbe(struct ccw_device *cdev, u32 mme)
 832 {
 833         unsigned long mba;
 834         struct cmb_data *cmb_data;
 835         unsigned long flags;
 836
 837         spin_lock_irqsave(cdev->ccwlock, flags);
 838         if (!cdev->private->cmb) {
 839                 spin_unlock_irqrestore(cdev->ccwlock, flags);
 840                 return -EINVAL;
 841         }
 842         cmb_data = cdev->private->cmb;
 843         mba = mme ? (unsigned long) cmb_data->hw_block : 0;
 844         spin_unlock_irqrestore(cdev->ccwlock, flags);
 845
 846         return set_schib_wait(cdev, mme, 1, mba);
 847 }
 848
 849 static u64 read_cmbe(struct ccw_device *cdev, int index)
 850 {
 851         struct cmb_data *cmb_data;
 852         unsigned long flags;
 853         struct cmbe *cmb;
 854         u64 ret = 0;
 855         u32 val;
 856
 857         spin_lock_irqsave(cdev->ccwlock, flags);
 858         cmb_data = cdev->private->cmb;
 859         if (!cmb_data)
 860                 goto out;
 861
 862         cmb = cmb_data->hw_block;
 863         switch (index) {
 864         case avg_utilization:
 865                 ret = __cmb_utilization(cmb->device_connect_time,
 866                                         cmb->function_pending_time,
 867                                         cmb->device_disconnect_time,
 868                                         cdev->private->cmb_start_time);
 869                 goto out;
 870         case cmb_ssch_rsch_count:
 871                 ret = cmb->ssch_rsch_count;
 872                 goto out;
 873         case cmb_sample_count:
 874                 ret = cmb->sample_count;
 875                 goto out;
 876         case cmb_device_connect_time:
 877                 val = cmb->device_connect_time;
 878                 break;
 879         case cmb_function_pending_time:
 880                 val = cmb->function_pending_time;
 881                 break;
 882         case cmb_device_disconnect_time:
 883                 val = cmb->device_disconnect_time;
 884                 break;
 885         case cmb_control_unit_queuing_time:
 886                 val = cmb->control_unit_queuing_time;
 887                 break;
 888         case cmb_device_active_only_time:
 889                 val = cmb->device_active_only_time;
 890                 break;
 891         case cmb_device_busy_time:
 892                 val = cmb->device_busy_time;
 893                 break;
 894         case cmb_initial_command_response_time:
 895                 val = cmb->initial_command_response_time;
 896                 break;
 897         default:
 898                 goto out;
 899         }
 900         ret = time_to_avg_nsec(val, cmb->sample_count);
 901 out:
 902         spin_unlock_irqrestore(cdev->ccwlock, flags);
 903         return ret;
 904 }
 905
 906 static int readall_cmbe(struct ccw_device *cdev, struct cmbdata *data)
 907 {
 908         struct cmbe *cmb;
 909         struct cmb_data *cmb_data;
 910         u64 time;
 911         unsigned long flags;
 912         int ret;
 913
 914         ret = cmf_cmb_copy_wait(cdev);
 915         if (ret < 0)
 916                 return ret;
 917         spin_lock_irqsave(cdev->ccwlock, flags);
 918         cmb_data = cdev->private->cmb;
 919         if (!cmb_data) {
 920                 ret = -ENODEV;
 921                 goto out;
 922         }
 923         if (cmb_data->last_update == 0) {
 924                 ret = -EAGAIN;
 925                 goto out;
 926         }
 927         time = cmb_data->last_update - cdev->private->cmb_start_time;
 928
 929         memset (data, 0, sizeof(struct cmbdata));
 930
 931         /* we only know values before device_busy_time */
 932         data->size = offsetof(struct cmbdata, device_busy_time);
 933
 934         data->elapsed_time = tod_to_ns(time);
 935
 936         cmb = cmb_data->last_block;
 937         /* copy data to new structure */
 938         data->ssch_rsch_count = cmb->ssch_rsch_count;
 939         data->sample_count = cmb->sample_count;
 940
 941         /* time fields are converted to nanoseconds while copying */
 942         data->device_connect_time = time_to_nsec(cmb->device_connect_time);
 943         data->function_pending_time = time_to_nsec(cmb->function_pending_time);
 944         data->device_disconnect_time =
 945                 time_to_nsec(cmb->device_disconnect_time);
 946         data->control_unit_queuing_time
 947                 = time_to_nsec(cmb->control_unit_queuing_time);
 948         data->device_active_only_time
 949                 = time_to_nsec(cmb->device_active_only_time);
 950         data->device_busy_time = time_to_nsec(cmb->device_busy_time);
 951         data->initial_command_response_time
 952                 = time_to_nsec(cmb->initial_command_response_time);
 953
 954         ret = 0;
 955 out:
 956         spin_unlock_irqrestore(cdev->ccwlock, flags);
 957         return ret;
 958 }
 959
 960 static void reset_cmbe(struct ccw_device *cdev)
 961 {
 962         cmf_generic_reset(cdev);
 963 }
 964
 965 static struct attribute_group cmf_attr_group_ext;
 966
 967 static struct cmb_operations cmbops_extended = {
 968         .alloc      = alloc_cmbe,
 969         .free       = free_cmbe,
 970         .set        = set_cmbe,
 971         .read       = read_cmbe,
 972         .readall    = readall_cmbe,
 973         .reset      = reset_cmbe,
 974         .attr_group = &cmf_attr_group_ext,
 975 };
 976
 977 static ssize_t cmb_show_attr(struct device *dev, char *buf, enum cmb_index idx)
 978 {
 979         return sprintf(buf, "%lld\n",
 980                 (unsigned long long) cmf_read(to_ccwdev(dev), idx));
 981 }
 982
 983 static ssize_t cmb_show_avg_sample_interval(struct device *dev,
 984                                             struct device_attribute *attr,
 985                                             char *buf)
 986 {
 987         struct ccw_device *cdev = to_ccwdev(dev);
 988         unsigned long count;
 989         long interval;
 990
 991         count = cmf_read(cdev, cmb_sample_count);
 992         spin_lock_irq(cdev->ccwlock);
 993         if (count) {
 994                 interval = get_tod_clock() - cdev->private->cmb_start_time;
 995                 interval = tod_to_ns(interval);
 996                 interval /= count;
 997         } else
 998                 interval = -1;
 999         spin_unlock_irq(cdev->ccwlock);
1000         return sprintf(buf, "%ld\n", interval);
1001 }
1002
1003 static ssize_t cmb_show_avg_utilization(struct device *dev,
1004                                         struct device_attribute *attr,
1005                                         char *buf)
1006 {
1007         unsigned long u = cmf_read(to_ccwdev(dev), avg_utilization);
1008
1009         return sprintf(buf, "%02lu.%01lu%%\n", u / 10, u % 10);
1010 }
1011
1012 #define cmf_attr(name) \
1013 static ssize_t show_##name(struct device *dev, \
1014                            struct device_attribute *attr, char *buf)    \
1015 { return cmb_show_attr((dev), buf, cmb_##name); } \
1016 static DEVICE_ATTR(name, 0444, show_##name, NULL);
1017
1018 #define cmf_attr_avg(name) \
1019 static ssize_t show_avg_##name(struct device *dev, \
1020                                struct device_attribute *attr, char *buf) \
1021 { return cmb_show_attr((dev), buf, cmb_##name); } \
1022 static DEVICE_ATTR(avg_##name, 0444, show_avg_##name, NULL);
1023
1024 cmf_attr(ssch_rsch_count);
1025 cmf_attr(sample_count);
1026 cmf_attr_avg(device_connect_time);
1027 cmf_attr_avg(function_pending_time);
1028 cmf_attr_avg(device_disconnect_time);
1029 cmf_attr_avg(control_unit_queuing_time);
1030 cmf_attr_avg(device_active_only_time);
1031 cmf_attr_avg(device_busy_time);
1032 cmf_attr_avg(initial_command_response_time);
1033
1034 static DEVICE_ATTR(avg_sample_interval, 0444, cmb_show_avg_sample_interval,
1035                    NULL);
1036 static DEVICE_ATTR(avg_utilization, 0444, cmb_show_avg_utilization, NULL);
1037
1038 static struct attribute *cmf_attributes[] = {
1039         &dev_attr_avg_sample_interval.attr,
1040         &dev_attr_avg_utilization.attr,
1041         &dev_attr_ssch_rsch_count.attr,
1042         &dev_attr_sample_count.attr,
1043         &dev_attr_avg_device_connect_time.attr,
1044         &dev_attr_avg_function_pending_time.attr,
1045         &dev_attr_avg_device_disconnect_time.attr,
1046         &dev_attr_avg_control_unit_queuing_time.attr,
1047         &dev_attr_avg_device_active_only_time.attr,
1048         NULL,
1049 };
1050
1051 static struct attribute_group cmf_attr_group = {
1052         .name  = "cmf",
1053         .attrs = cmf_attributes,
1054 };
1055
1056 static struct attribute *cmf_attributes_ext[] = {
1057         &dev_attr_avg_sample_interval.attr,
1058         &dev_attr_avg_utilization.attr,
1059         &dev_attr_ssch_rsch_count.attr,
1060         &dev_attr_sample_count.attr,
1061         &dev_attr_avg_device_connect_time.attr,
1062         &dev_attr_avg_function_pending_time.attr,
1063         &dev_attr_avg_device_disconnect_time.attr,
1064         &dev_attr_avg_control_unit_queuing_time.attr,
1065         &dev_attr_avg_device_active_only_time.attr,
1066         &dev_attr_avg_device_busy_time.attr,
1067         &dev_attr_avg_initial_command_response_time.attr,
1068         NULL,
1069 };
1070
1071 static struct attribute_group cmf_attr_group_ext = {
1072         .name  = "cmf",
1073         .attrs = cmf_attributes_ext,
1074 };
1075
1076 static ssize_t cmb_enable_show(struct device *dev,
1077                                struct device_attribute *attr,
1078                                char *buf)
1079 {
1080         struct ccw_device *cdev = to_ccwdev(dev);
1081
1082         return sprintf(buf, "%d\n", cmf_enabled(cdev));
1083 }
1084
1085 static ssize_t cmb_enable_store(struct device *dev,
1086                                 struct device_attribute *attr, const char *buf,
1087                                 size_t c)
1088 {
1089         struct ccw_device *cdev = to_ccwdev(dev);
1090         unsigned long val;
1091         int ret;
1092
1093         ret = kstrtoul(buf, 16, &val);
1094         if (ret)
1095                 return ret;
1096
1097         switch (val) {
1098         case 0:
1099                 ret = disable_cmf(cdev);
1100                 break;
1101         case 1:
1102                 ret = enable_cmf(cdev);
1103                 break;
1104         default:
1105                 ret = -EINVAL;
1106         }
1107
1108         return ret ? ret : c;
1109 }
1110 DEVICE_ATTR_RW(cmb_enable);
1111
1112 int ccw_set_cmf(struct ccw_device *cdev, int enable)
1113 {
1114         return cmbops->set(cdev, enable ? 2 : 0);
1115 }
1116
1117 /**
1118  * enable_cmf() - switch on the channel measurement for a specific device
1119  *  @cdev:      The ccw device to be enabled
1120  *
1121  *  Returns %0 for success or a negative error value.
1122  *  Note: If this is called on a device for which channel measurement is already
1123  *        enabled a reset of the measurement data is triggered.
1124  *  Context:
1125  *    non-atomic
1126  */
1127 int enable_cmf(struct ccw_device *cdev)
1128 {
1129         int ret = 0;
1130
1131         device_lock(&cdev->dev);
1132         if (cmf_enabled(cdev)) {
1133                 cmbops->reset(cdev);
1134                 goto out_unlock;
1135         }
1136         get_device(&cdev->dev);
1137         ret = cmbops->alloc(cdev);
1138         if (ret)
1139                 goto out;
1140         cmbops->reset(cdev);
1141         ret = sysfs_create_group(&cdev->dev.kobj, cmbops->attr_group);
1142         if (ret) {
1143                 cmbops->free(cdev);
1144                 goto out;
1145         }
1146         ret = cmbops->set(cdev, 2);
1147         if (ret) {
1148                 sysfs_remove_group(&cdev->dev.kobj, cmbops->attr_group);
1149                 cmbops->free(cdev);
1150         }
1151 out:
1152         if (ret)
1153                 put_device(&cdev->dev);
1154 out_unlock:
1155         device_unlock(&cdev->dev);
1156         return ret;
1157 }
1158
1159 /**
1160  * __disable_cmf() - switch off the channel measurement for a specific device
1161  *  @cdev:      The ccw device to be disabled
1162  *
1163  *  Returns %0 for success or a negative error value.
1164  *
1165  *  Context:
1166  *    non-atomic, device_lock() held.
1167  */
1168 int __disable_cmf(struct ccw_device *cdev)
1169 {
1170         int ret;
1171
1172         ret = cmbops->set(cdev, 0);
1173         if (ret)
1174                 return ret;
1175
1176         sysfs_remove_group(&cdev->dev.kobj, cmbops->attr_group);
1177         cmbops->free(cdev);
1178         put_device(&cdev->dev);
1179
1180         return ret;
1181 }
1182
1183 /**
1184  * disable_cmf() - switch off the channel measurement for a specific device
1185  *  @cdev:      The ccw device to be disabled
1186  *
1187  *  Returns %0 for success or a negative error value.
1188  *
1189  *  Context:
1190  *    non-atomic
1191  */
1192 int disable_cmf(struct ccw_device *cdev)
1193 {
1194         int ret;
1195
1196         device_lock(&cdev->dev);
1197         ret = __disable_cmf(cdev);
1198         device_unlock(&cdev->dev);
1199
1200         return ret;
1201 }
1202
1203 /**
1204  * cmf_read() - read one value from the current channel measurement block
1205  * @cdev:       the channel to be read
1206  * @index:      the index of the value to be read
1207  *
1208  * Returns the value read or %0 if the value cannot be read.
1209  *
1210  *  Context:
1211  *    any
1212  */
1213 u64 cmf_read(struct ccw_device *cdev, int index)
1214 {
1215         return cmbops->read(cdev, index);
1216 }
1217
1218 /**
1219  * cmf_readall() - read the current channel measurement block
1220  * @cdev:       the channel to be read
1221  * @data:       a pointer to a data block that will be filled
1222  *
1223  * Returns %0 on success, a negative error value otherwise.
1224  *
1225  *  Context:
1226  *    any
1227  */
1228 int cmf_readall(struct ccw_device *cdev, struct cmbdata *data)
1229 {
1230         return cmbops->readall(cdev, data);
1231 }
1232
1233 /* Reenable cmf when a disconnected device becomes available again. */
1234 int cmf_reenable(struct ccw_device *cdev)
1235 {
1236         cmbops->reset(cdev);
1237         return cmbops->set(cdev, 2);
1238 }
1239
1240 /**
1241  * cmf_reactivate() - reactivate measurement block updates
1242  *
1243  * Use this during resume from hibernate.
1244  */
1245 void cmf_reactivate(void)
1246 {
1247         spin_lock(&cmb_area.lock);
1248         if (!list_empty(&cmb_area.list))
1249                 cmf_activate(cmb_area.mem, CMF_ON);
1250         spin_unlock(&cmb_area.lock);
1251 }
1252
1253 static int __init init_cmbe(void)
1254 {
1255         cmbe_cache = kmem_cache_create("cmbe_cache", sizeof(struct cmbe),
1256                                        __alignof__(struct cmbe), 0, NULL);
1257
1258         return cmbe_cache ? 0 : -ENOMEM;
1259 }
1260
1261 static int __init init_cmf(void)
1262 {
1263         char *format_string;
1264         char *detect_string;
1265         int ret;
1266
1267         /*
1268          * If the user did not give a parameter, see if we are running on a
1269          * machine supporting extended measurement blocks, otherwise fall back
1270          * to basic mode.
1271          */
1272         if (format == CMF_AUTODETECT) {
1273                 if (!css_general_characteristics.ext_mb) {
1274                         format = CMF_BASIC;
1275                 } else {
1276                         format = CMF_EXTENDED;
1277                 }
1278                 detect_string = "autodetected";
1279         } else {
1280                 detect_string = "parameter";
1281         }
1282
1283         switch (format) {
1284         case CMF_BASIC:
1285                 format_string = "basic";
1286                 cmbops = &cmbops_basic;
1287                 break;
1288         case CMF_EXTENDED:
1289                 format_string = "extended";
1290                 cmbops = &cmbops_extended;
1291
1292                 ret = init_cmbe();
1293                 if (ret)
1294                         return ret;
1295                 break;
1296         default:
1297                 return -EINVAL;
1298         }
1299         pr_info("Channel measurement facility initialized using format "
1300                 "%s (mode %s)\n", format_string, detect_string);
1301         return 0;
1302 }
1303 device_initcall(init_cmf);
1304
1305 EXPORT_SYMBOL_GPL(enable_cmf);
1306 EXPORT_SYMBOL_GPL(disable_cmf);
1307 EXPORT_SYMBOL_GPL(cmf_read);
1308 EXPORT_SYMBOL_GPL(cmf_readall);