drivers/s390/block/dasd_eer.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Character device driver for extended error reporting.
   4  *
   5  *  Copyright IBM Corp. 2005
   6  *  extended error reporting for DASD ECKD devices
   7  *  Author(s): Stefan Weinhuber <wein@de.ibm.com>
   8  */
   9
  10 #include <linux/init.h>
  11 #include <linux/fs.h>
  12 #include <linux/kernel.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/device.h>
  17 #include <linux/poll.h>
  18 #include <linux/mutex.h>
  19 #include <linux/err.h>
  20 #include <linux/slab.h>
  21
  22 #include <linux/uaccess.h>
  23 #include <linux/atomic.h>
  24 #include <asm/ebcdic.h>
  25
  26 #include "dasd_int.h"
  27 #include "dasd_eckd.h"
  28
  29 /*
  30  * SECTION: the internal buffer
  31  */
  32
  33 /*
  34  * The internal buffer is meant to store obaque blobs of data, so it does
  35  * not know of higher level concepts like triggers.
  36  * It consists of a number of pages that are used as a ringbuffer. Each data
  37  * blob is stored in a simple record that consists of an integer, which
  38  * contains the size of the following data, and the data bytes themselfes.
  39  *
  40  * To allow for multiple independent readers we create one internal buffer
  41  * each time the device is opened and destroy the buffer when the file is
  42  * closed again. The number of pages used for this buffer is determined by
  43  * the module parmeter eer_pages.
  44  *
  45  * One record can be written to a buffer by using the functions
  46  * - dasd_eer_start_record (one time per record to write the size to the
  47  *                          buffer and reserve the space for the data)
  48  * - dasd_eer_write_buffer (one or more times per record to write the data)
  49  * The data can be written in several steps but you will have to compute
  50  * the total size up front for the invocation of dasd_eer_start_record.
  51  * If the ringbuffer is full, dasd_eer_start_record will remove the required
  52  * number of old records.
  53  *
  54  * A record is typically read in two steps, first read the integer that
  55  * specifies the size of the following data, then read the data.
  56  * Both can be done by
  57  * - dasd_eer_read_buffer
  58  *
  59  * For all mentioned functions you need to get the bufferlock first and keep
  60  * it until a complete record is written or read.
  61  *
  62  * All information necessary to keep track of an internal buffer is kept in
  63  * a struct eerbuffer. The buffer specific to a file pointer is strored in
  64  * the private_data field of that file. To be able to write data to all
  65  * existing buffers, each buffer is also added to the bufferlist.
  66  * If the user does not want to read a complete record in one go, we have to
  67  * keep track of the rest of the record. residual stores the number of bytes
  68  * that are still to deliver. If the rest of the record is invalidated between
  69  * two reads then residual will be set to -1 so that the next read will fail.
  70  * All entries in the eerbuffer structure are protected with the bufferlock.
  71  * To avoid races between writing to a buffer on the one side and creating
  72  * and destroying buffers on the other side, the bufferlock must also be used
  73  * to protect the bufferlist.
  74  */
  75
  76 static int eer_pages = 5;
  77 module_param(eer_pages, int, S_IRUGO|S_IWUSR);
  78
  79 struct eerbuffer {
  80         struct list_head list;
  81         char **buffer;
  82         int buffersize;
  83         int buffer_page_count;
  84         int head;
  85         int tail;
  86         int residual;
  87 };
  88
  89 static LIST_HEAD(bufferlist);
  90 static DEFINE_SPINLOCK(bufferlock);
  91 static DECLARE_WAIT_QUEUE_HEAD(dasd_eer_read_wait_queue);
  92
  93 /*
  94  * How many free bytes are available on the buffer.
  95  * Needs to be called with bufferlock held.
  96  */
  97 static int dasd_eer_get_free_bytes(struct eerbuffer *eerb)
  98 {
  99         if (eerb->head < eerb->tail)
 100                 return eerb->tail - eerb->head - 1;
 101         return eerb->buffersize - eerb->head + eerb->tail -1;
 102 }
 103
 104 /*
 105  * How many bytes of buffer space are used.
 106  * Needs to be called with bufferlock held.
 107  */
 108 static int dasd_eer_get_filled_bytes(struct eerbuffer *eerb)
 109 {
 110
 111         if (eerb->head >= eerb->tail)
 112                 return eerb->head - eerb->tail;
 113         return eerb->buffersize - eerb->tail + eerb->head;
 114 }
 115
 116 /*
 117  * The dasd_eer_write_buffer function just copies count bytes of data
 118  * to the buffer. Make sure to call dasd_eer_start_record first, to
 119  * make sure that enough free space is available.
 120  * Needs to be called with bufferlock held.
 121  */
 122 static void dasd_eer_write_buffer(struct eerbuffer *eerb,
 123                                   char *data, int count)
 124 {
 125
 126         unsigned long headindex,localhead;
 127         unsigned long rest, len;
 128         char *nextdata;
 129
 130         nextdata = data;
 131         rest = count;
 132         while (rest > 0) {
 133                 headindex = eerb->head / PAGE_SIZE;
 134                 localhead = eerb->head % PAGE_SIZE;
 135                 len = min(rest, PAGE_SIZE - localhead);
 136                 memcpy(eerb->buffer[headindex]+localhead, nextdata, len);
 137                 nextdata += len;
 138                 rest -= len;
 139                 eerb->head += len;
 140                 if (eerb->head == eerb->buffersize)
 141                         eerb->head = 0; /* wrap around */
 142                 BUG_ON(eerb->head > eerb->buffersize);
 143         }
 144 }
 145
 146 /*
 147  * Needs to be called with bufferlock held.
 148  */
 149 static int dasd_eer_read_buffer(struct eerbuffer *eerb, char *data, int count)
 150 {
 151
 152         unsigned long tailindex,localtail;
 153         unsigned long rest, len, finalcount;
 154         char *nextdata;
 155
 156         finalcount = min(count, dasd_eer_get_filled_bytes(eerb));
 157         nextdata = data;
 158         rest = finalcount;
 159         while (rest > 0) {
 160                 tailindex = eerb->tail / PAGE_SIZE;
 161                 localtail = eerb->tail % PAGE_SIZE;
 162                 len = min(rest, PAGE_SIZE - localtail);
 163                 memcpy(nextdata, eerb->buffer[tailindex] + localtail, len);
 164                 nextdata += len;
 165                 rest -= len;
 166                 eerb->tail += len;
 167                 if (eerb->tail == eerb->buffersize)
 168                         eerb->tail = 0; /* wrap around */
 169                 BUG_ON(eerb->tail > eerb->buffersize);
 170         }
 171         return finalcount;
 172 }
 173
 174 /*
 175  * Whenever you want to write a blob of data to the internal buffer you
 176  * have to start by using this function first. It will write the number
 177  * of bytes that will be written to the buffer. If necessary it will remove
 178  * old records to make room for the new one.
 179  * Needs to be called with bufferlock held.
 180  */
 181 static int dasd_eer_start_record(struct eerbuffer *eerb, int count)
 182 {
 183         int tailcount;
 184
 185         if (count + sizeof(count) > eerb->buffersize)
 186                 return -ENOMEM;
 187         while (dasd_eer_get_free_bytes(eerb) < count + sizeof(count)) {
 188                 if (eerb->residual > 0) {
 189                         eerb->tail += eerb->residual;
 190                         if (eerb->tail >= eerb->buffersize)
 191                                 eerb->tail -= eerb->buffersize;
 192                         eerb->residual = -1;
 193                 }
 194                 dasd_eer_read_buffer(eerb, (char *) &tailcount,
 195                                      sizeof(tailcount));
 196                 eerb->tail += tailcount;
 197                 if (eerb->tail >= eerb->buffersize)
 198                         eerb->tail -= eerb->buffersize;
 199         }
 200         dasd_eer_write_buffer(eerb, (char*) &count, sizeof(count));
 201
 202         return 0;
 203 };
 204
 205 /*
 206  * Release pages that are not used anymore.
 207  */
 208 static void dasd_eer_free_buffer_pages(char **buf, int no_pages)
 209 {
 210         int i;
 211
 212         for (i = 0; i < no_pages; i++)
 213                 free_page((unsigned long) buf[i]);
 214 }
 215
 216 /*
 217  * Allocate a new set of memory pages.
 218  */
 219 static int dasd_eer_allocate_buffer_pages(char **buf, int no_pages)
 220 {
 221         int i;
 222
 223         for (i = 0; i < no_pages; i++) {
 224                 buf[i] = (char *) get_zeroed_page(GFP_KERNEL);
 225                 if (!buf[i]) {
 226                         dasd_eer_free_buffer_pages(buf, i);
 227                         return -ENOMEM;
 228                 }
 229         }
 230         return 0;
 231 }
 232
 233 /*
 234  * SECTION: The extended error reporting functionality
 235  */
 236
 237 /*
 238  * When a DASD device driver wants to report an error, it calls the
 239  * function dasd_eer_write and gives the respective trigger ID as
 240  * parameter. Currently there are four kinds of triggers:
 241  *
 242  * DASD_EER_FATALERROR:  all kinds of unrecoverable I/O problems
 243  * DASD_EER_PPRCSUSPEND: PPRC was suspended
 244  * DASD_EER_NOPATH:      There is no path to the device left.
 245  * DASD_EER_STATECHANGE: The state of the device has changed.
 246  *
 247  * For the first three triggers all required information can be supplied by
 248  * the caller. For these triggers a record is written by the function
 249  * dasd_eer_write_standard_trigger.
 250  *
 251  * The DASD_EER_STATECHANGE trigger is special since a sense subsystem
 252  * status ccw need to be executed to gather the necessary sense data first.
 253  * The dasd_eer_snss function will queue the SNSS request and the request
 254  * callback will then call dasd_eer_write with the DASD_EER_STATCHANGE
 255  * trigger.
 256  *
 257  * To avoid memory allocations at runtime, the necessary memory is allocated
 258  * when the extended error reporting is enabled for a device (by
 259  * dasd_eer_probe). There is one sense subsystem status request for each
 260  * eer enabled DASD device. The presence of the cqr in device->eer_cqr
 261  * indicates that eer is enable for the device. The use of the snss request
 262  * is protected by the DASD_FLAG_EER_IN_USE bit. When this flag indicates
 263  * that the cqr is currently in use, dasd_eer_snss cannot start a second
 264  * request but sets the DASD_FLAG_EER_SNSS flag instead. The callback of
 265  * the SNSS request will check the bit and call dasd_eer_snss again.
 266  */
 267
 268 #define SNSS_DATA_SIZE 44
 269
 270 #define DASD_EER_BUSID_SIZE 10
 271 struct dasd_eer_header {
 272         __u32 total_size;
 273         __u32 trigger;
 274         __u64 tv_sec;
 275         __u64 tv_usec;
 276         char busid[DASD_EER_BUSID_SIZE];
 277 } __attribute__ ((packed));
 278
 279 /*
 280  * The following function can be used for those triggers that have
 281  * all necessary data available when the function is called.
 282  * If the parameter cqr is not NULL, the chain of requests will be searched
 283  * for valid sense data, and all valid sense data sets will be added to
 284  * the triggers data.
 285  */
 286 static void dasd_eer_write_standard_trigger(struct dasd_device *device,
 287                                             struct dasd_ccw_req *cqr,
 288                                             int trigger)
 289 {
 290         struct dasd_ccw_req *temp_cqr;
 291         int data_size;
 292         struct timespec64 ts;
 293         struct dasd_eer_header header;
 294         unsigned long flags;
 295         struct eerbuffer *eerb;
 296         char *sense;
 297
 298         /* go through cqr chain and count the valid sense data sets */
 299         data_size = 0;
 300         for (temp_cqr = cqr; temp_cqr; temp_cqr = temp_cqr->refers)
 301                 if (dasd_get_sense(&temp_cqr->irb))
 302                         data_size += 32;
 303
 304         header.total_size = sizeof(header) + data_size + 4; /* "EOR" */
 305         header.trigger = trigger;
 306         ktime_get_real_ts64(&ts);
 307         header.tv_sec = ts.tv_sec;
 308         header.tv_usec = ts.tv_nsec / NSEC_PER_USEC;
 309         strscpy(header.busid, dev_name(&device->cdev->dev),
 310                 DASD_EER_BUSID_SIZE);
 311
 312         spin_lock_irqsave(&bufferlock, flags);
 313         list_for_each_entry(eerb, &bufferlist, list) {
 314                 dasd_eer_start_record(eerb, header.total_size);
 315                 dasd_eer_write_buffer(eerb, (char *) &header, sizeof(header));
 316                 for (temp_cqr = cqr; temp_cqr; temp_cqr = temp_cqr->refers) {
 317                         sense = dasd_get_sense(&temp_cqr->irb);
 318                         if (sense)
 319                                 dasd_eer_write_buffer(eerb, sense, 32);
 320                 }
 321                 dasd_eer_write_buffer(eerb, "EOR", 4);
 322         }
 323         spin_unlock_irqrestore(&bufferlock, flags);
 324         wake_up_interruptible(&dasd_eer_read_wait_queue);
 325 }
 326
 327 /*
 328  * This function writes a DASD_EER_STATECHANGE trigger.
 329  */
 330 static void dasd_eer_write_snss_trigger(struct dasd_device *device,
 331                                         struct dasd_ccw_req *cqr,
 332                                         int trigger)
 333 {
 334         int data_size;
 335         int snss_rc;
 336         struct timespec64 ts;
 337         struct dasd_eer_header header;
 338         unsigned long flags;
 339         struct eerbuffer *eerb;
 340
 341         snss_rc = (cqr->status == DASD_CQR_DONE) ? 0 : -EIO;
 342         if (snss_rc)
 343                 data_size = 0;
 344         else
 345                 data_size = SNSS_DATA_SIZE;
 346
 347         header.total_size = sizeof(header) + data_size + 4; /* "EOR" */
 348         header.trigger = DASD_EER_STATECHANGE;
 349         ktime_get_real_ts64(&ts);
 350         header.tv_sec = ts.tv_sec;
 351         header.tv_usec = ts.tv_nsec / NSEC_PER_USEC;
 352         strscpy(header.busid, dev_name(&device->cdev->dev),
 353                 DASD_EER_BUSID_SIZE);
 354
 355         spin_lock_irqsave(&bufferlock, flags);
 356         list_for_each_entry(eerb, &bufferlist, list) {
 357                 dasd_eer_start_record(eerb, header.total_size);
 358                 dasd_eer_write_buffer(eerb, (char *) &header , sizeof(header));
 359                 if (!snss_rc)
 360                         dasd_eer_write_buffer(eerb, cqr->data, SNSS_DATA_SIZE);
 361                 dasd_eer_write_buffer(eerb, "EOR", 4);
 362         }
 363         spin_unlock_irqrestore(&bufferlock, flags);
 364         wake_up_interruptible(&dasd_eer_read_wait_queue);
 365 }
 366
 367 /*
 368  * This function is called for all triggers. It calls the appropriate
 369  * function that writes the actual trigger records.
 370  */
 371 void dasd_eer_write(struct dasd_device *device, struct dasd_ccw_req *cqr,
 372                     unsigned int id)
 373 {
 374         if (!device->eer_cqr)
 375                 return;
 376         switch (id) {
 377         case DASD_EER_FATALERROR:
 378         case DASD_EER_PPRCSUSPEND:
 379                 dasd_eer_write_standard_trigger(device, cqr, id);
 380                 break;
 381         case DASD_EER_NOPATH:
 382         case DASD_EER_NOSPC:
 383         case DASD_EER_AUTOQUIESCE:
 384                 dasd_eer_write_standard_trigger(device, NULL, id);
 385                 break;
 386         case DASD_EER_STATECHANGE:
 387                 dasd_eer_write_snss_trigger(device, cqr, id);
 388                 break;
 389         default: /* unknown trigger, so we write it without any sense data */
 390                 dasd_eer_write_standard_trigger(device, NULL, id);
 391                 break;
 392         }
 393 }
 394 EXPORT_SYMBOL(dasd_eer_write);
 395
 396 /*
 397  * Start a sense subsystem status request.
 398  * Needs to be called with the device held.
 399  */
 400 void dasd_eer_snss(struct dasd_device *device)
 401 {
 402         struct dasd_ccw_req *cqr;
 403
 404         cqr = device->eer_cqr;
 405         if (!cqr)       /* Device not eer enabled. */
 406                 return;
 407         if (test_and_set_bit(DASD_FLAG_EER_IN_USE, &device->flags)) {
 408                 /* Sense subsystem status request in use. */
 409                 set_bit(DASD_FLAG_EER_SNSS, &device->flags);
 410                 return;
 411         }
 412         /* cdev is already locked, can't use dasd_add_request_head */
 413         clear_bit(DASD_FLAG_EER_SNSS, &device->flags);
 414         cqr->status = DASD_CQR_QUEUED;
 415         list_add(&cqr->devlist, &device->ccw_queue);
 416         dasd_schedule_device_bh(device);
 417 }
 418
 419 /*
 420  * Callback function for use with sense subsystem status request.
 421  */
 422 static void dasd_eer_snss_cb(struct dasd_ccw_req *cqr, void *data)
 423 {
 424         struct dasd_device *device = cqr->startdev;
 425         unsigned long flags;
 426
 427         dasd_eer_write(device, cqr, DASD_EER_STATECHANGE);
 428         spin_lock_irqsave(get_ccwdev_lock(device->cdev), flags);
 429         if (device->eer_cqr == cqr) {
 430                 clear_bit(DASD_FLAG_EER_IN_USE, &device->flags);
 431                 if (test_bit(DASD_FLAG_EER_SNSS, &device->flags))
 432                         /* Another SNSS has been requested in the meantime. */
 433                         dasd_eer_snss(device);
 434                 cqr = NULL;
 435         }
 436         spin_unlock_irqrestore(get_ccwdev_lock(device->cdev), flags);
 437         if (cqr)
 438                 /*
 439                  * Extended error recovery has been switched off while
 440                  * the SNSS request was running. It could even have
 441                  * been switched off and on again in which case there
 442                  * is a new ccw in device->eer_cqr. Free the "old"
 443                  * snss request now.
 444                  */
 445                 dasd_sfree_request(cqr, device);
 446 }
 447
 448 /*
 449  * Enable error reporting on a given device.
 450  */
 451 int dasd_eer_enable(struct dasd_device *device)
 452 {
 453         struct dasd_ccw_req *cqr = NULL;
 454         unsigned long flags;
 455         struct ccw1 *ccw;
 456         int rc = 0;
 457
 458         spin_lock_irqsave(get_ccwdev_lock(device->cdev), flags);
 459         if (device->eer_cqr)
 460                 goto out;
 461         else if (!device->discipline ||
 462                  strcmp(device->discipline->name, "ECKD"))
 463                 rc = -EMEDIUMTYPE;
 464         else if (test_bit(DASD_FLAG_OFFLINE, &device->flags))
 465                 rc = -EBUSY;
 466
 467         if (rc)
 468                 goto out;
 469
 470         cqr = dasd_smalloc_request(DASD_ECKD_MAGIC, 1 /* SNSS */,
 471                                    SNSS_DATA_SIZE, device, NULL);
 472         if (IS_ERR(cqr)) {
 473                 rc = -ENOMEM;
 474                 cqr = NULL;
 475                 goto out;
 476         }
 477
 478         cqr->startdev = device;
 479         cqr->retries = 255;
 480         cqr->expires = 10 * HZ;
 481         clear_bit(DASD_CQR_FLAGS_USE_ERP, &cqr->flags);
 482         set_bit(DASD_CQR_ALLOW_SLOCK, &cqr->flags);
 483
 484         ccw = cqr->cpaddr;
 485         ccw->cmd_code = DASD_ECKD_CCW_SNSS;
 486         ccw->count = SNSS_DATA_SIZE;
 487         ccw->flags = 0;
 488         ccw->cda = virt_to_dma32(cqr->data);
 489
 490         cqr->buildclk = get_tod_clock();
 491         cqr->status = DASD_CQR_FILLED;
 492         cqr->callback = dasd_eer_snss_cb;
 493
 494         if (!device->eer_cqr) {
 495                 device->eer_cqr = cqr;
 496                 cqr = NULL;
 497         }
 498
 499 out:
 500         spin_unlock_irqrestore(get_ccwdev_lock(device->cdev), flags);
 501
 502         if (cqr)
 503                 dasd_sfree_request(cqr, device);
 504
 505         return rc;
 506 }
 507
 508 /*
 509  * Disable error reporting on a given device.
 510  */
 511 void dasd_eer_disable(struct dasd_device *device)
 512 {
 513         struct dasd_ccw_req *cqr;
 514         unsigned long flags;
 515         int in_use;
 516
 517         if (!device->eer_cqr)
 518                 return;
 519         spin_lock_irqsave(get_ccwdev_lock(device->cdev), flags);
 520         cqr = device->eer_cqr;
 521         device->eer_cqr = NULL;
 522         clear_bit(DASD_FLAG_EER_SNSS, &device->flags);
 523         in_use = test_and_clear_bit(DASD_FLAG_EER_IN_USE, &device->flags);
 524         spin_unlock_irqrestore(get_ccwdev_lock(device->cdev), flags);
 525         if (cqr && !in_use)
 526                 dasd_sfree_request(cqr, device);
 527 }
 528
 529 /*
 530  * SECTION: the device operations
 531  */
 532
 533 /*
 534  * On the one side we need a lock to access our internal buffer, on the
 535  * other side a copy_to_user can sleep. So we need to copy the data we have
 536  * to transfer in a readbuffer, which is protected by the readbuffer_mutex.
 537  */
 538 static char readbuffer[PAGE_SIZE];
 539 static DEFINE_MUTEX(readbuffer_mutex);
 540
 541 static int dasd_eer_open(struct inode *inp, struct file *filp)
 542 {
 543         struct eerbuffer *eerb;
 544         unsigned long flags;
 545
 546         eerb = kzalloc(sizeof(struct eerbuffer), GFP_KERNEL);
 547         if (!eerb)
 548                 return -ENOMEM;
 549         eerb->buffer_page_count = eer_pages;
 550         if (eerb->buffer_page_count < 1 ||
 551             eerb->buffer_page_count > INT_MAX / PAGE_SIZE) {
 552                 kfree(eerb);
 553                 DBF_EVENT(DBF_WARNING, "can't open device since module "
 554                         "parameter eer_pages is smaller than 1 or"
 555                         " bigger than %d", (int)(INT_MAX / PAGE_SIZE));
 556                 return -EINVAL;
 557         }
 558         eerb->buffersize = eerb->buffer_page_count * PAGE_SIZE;
 559         eerb->buffer = kmalloc_array(eerb->buffer_page_count, sizeof(char *),
 560                                      GFP_KERNEL);
 561         if (!eerb->buffer) {
 562                 kfree(eerb);
 563                 return -ENOMEM;
 564         }
 565         if (dasd_eer_allocate_buffer_pages(eerb->buffer,
 566                                            eerb->buffer_page_count)) {
 567                 kfree(eerb->buffer);
 568                 kfree(eerb);
 569                 return -ENOMEM;
 570         }
 571         filp->private_data = eerb;
 572         spin_lock_irqsave(&bufferlock, flags);
 573         list_add(&eerb->list, &bufferlist);
 574         spin_unlock_irqrestore(&bufferlock, flags);
 575
 576         return nonseekable_open(inp,filp);
 577 }
 578
 579 static int dasd_eer_close(struct inode *inp, struct file *filp)
 580 {
 581         struct eerbuffer *eerb;
 582         unsigned long flags;
 583
 584         eerb = (struct eerbuffer *) filp->private_data;
 585         spin_lock_irqsave(&bufferlock, flags);
 586         list_del(&eerb->list);
 587         spin_unlock_irqrestore(&bufferlock, flags);
 588         dasd_eer_free_buffer_pages(eerb->buffer, eerb->buffer_page_count);
 589         kfree(eerb->buffer);
 590         kfree(eerb);
 591
 592         return 0;
 593 }
 594
 595 static ssize_t dasd_eer_read(struct file *filp, char __user *buf,
 596                              size_t count, loff_t *ppos)
 597 {
 598         int tc,rc;
 599         int tailcount,effective_count;
 600         unsigned long flags;
 601         struct eerbuffer *eerb;
 602
 603         eerb = (struct eerbuffer *) filp->private_data;
 604         if (mutex_lock_interruptible(&readbuffer_mutex))
 605                 return -ERESTARTSYS;
 606
 607         spin_lock_irqsave(&bufferlock, flags);
 608
 609         if (eerb->residual < 0) { /* the remainder of this record */
 610                                   /* has been deleted             */
 611                 eerb->residual = 0;
 612                 spin_unlock_irqrestore(&bufferlock, flags);
 613                 mutex_unlock(&readbuffer_mutex);
 614                 return -EIO;
 615         } else if (eerb->residual > 0) {
 616                 /* OK we still have a second half of a record to deliver */
 617                 effective_count = min(eerb->residual, (int) count);
 618                 eerb->residual -= effective_count;
 619         } else {
 620                 tc = 0;
 621                 while (!tc) {
 622                         tc = dasd_eer_read_buffer(eerb, (char *) &tailcount,
 623                                                   sizeof(tailcount));
 624                         if (!tc) {
 625                                 /* no data available */
 626                                 spin_unlock_irqrestore(&bufferlock, flags);
 627                                 mutex_unlock(&readbuffer_mutex);
 628                                 if (filp->f_flags & O_NONBLOCK)
 629                                         return -EAGAIN;
 630                                 rc = wait_event_interruptible(
 631                                         dasd_eer_read_wait_queue,
 632                                         eerb->head != eerb->tail);
 633                                 if (rc)
 634                                         return rc;
 635                                 if (mutex_lock_interruptible(&readbuffer_mutex))
 636                                         return -ERESTARTSYS;
 637                                 spin_lock_irqsave(&bufferlock, flags);
 638                         }
 639                 }
 640                 WARN_ON(tc != sizeof(tailcount));
 641                 effective_count = min(tailcount,(int)count);
 642                 eerb->residual = tailcount - effective_count;
 643         }
 644
 645         tc = dasd_eer_read_buffer(eerb, readbuffer, effective_count);
 646         WARN_ON(tc != effective_count);
 647
 648         spin_unlock_irqrestore(&bufferlock, flags);
 649
 650         if (copy_to_user(buf, readbuffer, effective_count)) {
 651                 mutex_unlock(&readbuffer_mutex);
 652                 return -EFAULT;
 653         }
 654
 655         mutex_unlock(&readbuffer_mutex);
 656         return effective_count;
 657 }
 658
 659 static __poll_t dasd_eer_poll(struct file *filp, poll_table *ptable)
 660 {
 661         __poll_t mask;
 662         unsigned long flags;
 663         struct eerbuffer *eerb;
 664
 665         eerb = (struct eerbuffer *) filp->private_data;
 666         poll_wait(filp, &dasd_eer_read_wait_queue, ptable);
 667         spin_lock_irqsave(&bufferlock, flags);
 668         if (eerb->head != eerb->tail)
 669                 mask = EPOLLIN | EPOLLRDNORM ;
 670         else
 671                 mask = 0;
 672         spin_unlock_irqrestore(&bufferlock, flags);
 673         return mask;
 674 }
 675
 676 static const struct file_operations dasd_eer_fops = {
 677         .open           = &dasd_eer_open,
 678         .release        = &dasd_eer_close,
 679         .read           = &dasd_eer_read,
 680         .poll           = &dasd_eer_poll,
 681         .owner          = THIS_MODULE,
 682         .llseek         = noop_llseek,
 683 };
 684
 685 static struct miscdevice *dasd_eer_dev = NULL;
 686
 687 int __init dasd_eer_init(void)
 688 {
 689         int rc;
 690
 691         dasd_eer_dev = kzalloc(sizeof(*dasd_eer_dev), GFP_KERNEL);
 692         if (!dasd_eer_dev)
 693                 return -ENOMEM;
 694
 695         dasd_eer_dev->minor = MISC_DYNAMIC_MINOR;
 696         dasd_eer_dev->name  = "dasd_eer";
 697         dasd_eer_dev->fops  = &dasd_eer_fops;
 698
 699         rc = misc_register(dasd_eer_dev);
 700         if (rc) {
 701                 kfree(dasd_eer_dev);
 702                 dasd_eer_dev = NULL;
 703                 DBF_EVENT(DBF_ERR, "%s", "dasd_eer_init could not "
 704                        "register misc device");
 705                 return rc;
 706         }
 707
 708         return 0;
 709 }
 710
 711 void dasd_eer_exit(void)
 712 {
 713         if (dasd_eer_dev) {
 714                 misc_deregister(dasd_eer_dev);
 715                 kfree(dasd_eer_dev);
 716                 dasd_eer_dev = NULL;
 717         }
 718 }