drivers/misc/sgi-gru/grukservices.c

/* [<][>][^][v][top][bottom][index][help] */
This source file includes following definitions.
gru_load_kernel_context
gru_free_kernel_contexts
gru_lock_kernel_context
gru_unlock_kernel_context
gru_get_cpu_resources
gru_free_cpu_resources
gru_reserve_async_resources
gru_release_async_resources
gru_wait_async_cbr
gru_lock_async_resource
gru_unlock_async_resource
gru_get_cb_exception_detail
gru_get_cb_exception_detail_str
gru_wait_idle_or_exception
gru_retry_exception
gru_check_status_proc
gru_wait_proc
gru_abort
gru_wait_abort_proc
get_present2
restore_present2
gru_create_message_queue
send_noop_message
send_message_queue_full
send_message_put_nacked
send_message_failure
gru_send_message_gpa
gru_free_message
gru_get_next_message
gru_read_gpa
gru_copy_gpa
quicktest0
quicktest1
quicktest2
quicktest3
gru_ktest
gru_kservices_init
gru_kservices_exit
   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * SN Platform GRU Driver
   4  *
   5  *              KERNEL SERVICES THAT USE THE GRU
   6  *
   7  *  Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
   8  */
   9 
  10 #include <linux/kernel.h>
  11 #include <linux/errno.h>
  12 #include <linux/slab.h>
  13 #include <linux/mm.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/device.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/proc_fs.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/uaccess.h>
  20 #include <linux/delay.h>
  21 #include <linux/export.h>
  22 #include <asm/io_apic.h>
  23 #include "gru.h"
  24 #include "grulib.h"
  25 #include "grutables.h"
  26 #include "grukservices.h"
  27 #include "gru_instructions.h"
  28 #include <asm/uv/uv_hub.h>
  29 
  30 /*
  31  * Kernel GRU Usage
  32  *
  33  * The following is an interim algorithm for management of kernel GRU
  34  * resources. This will likely be replaced when we better understand the
  35  * kernel/user requirements.
  36  *
  37  * Blade percpu resources reserved for kernel use. These resources are
  38  * reserved whenever the the kernel context for the blade is loaded. Note
  39  * that the kernel context is not guaranteed to be always available. It is
  40  * loaded on demand & can be stolen by a user if the user demand exceeds the
  41  * kernel demand. The kernel can always reload the kernel context but
  42  * a SLEEP may be required!!!.
  43  *
  44  * Async Overview:
  45  *
  46  *      Each blade has one "kernel context" that owns GRU kernel resources
  47  *      located on the blade. Kernel drivers use GRU resources in this context
  48  *      for sending messages, zeroing memory, etc.
  49  *
  50  *      The kernel context is dynamically loaded on demand. If it is not in
  51  *      use by the kernel, the kernel context can be unloaded & given to a user.
  52  *      The kernel context will be reloaded when needed. This may require that
  53  *      a context be stolen from a user.
  54  *              NOTE: frequent unloading/reloading of the kernel context is
  55  *              expensive. We are depending on batch schedulers, cpusets, sane
  56  *              drivers or some other mechanism to prevent the need for frequent
  57  *              stealing/reloading.
  58  *
  59  *      The kernel context consists of two parts:
  60  *              - 1 CB & a few DSRs that are reserved for each cpu on the blade.
  61  *                Each cpu has it's own private resources & does not share them
  62  *                with other cpus. These resources are used serially, ie,
  63  *                locked, used & unlocked  on each call to a function in
  64  *                grukservices.
  65  *                      (Now that we have dynamic loading of kernel contexts, I
  66  *                       may rethink this & allow sharing between cpus....)
  67  *
  68  *              - Additional resources can be reserved long term & used directly
  69  *                by UV drivers located in the kernel. Drivers using these GRU
  70  *                resources can use asynchronous GRU instructions that send
  71  *                interrupts on completion.
  72  *                      - these resources must be explicitly locked/unlocked
  73  *                      - locked resources prevent (obviously) the kernel
  74  *                        context from being unloaded.
  75  *                      - drivers using these resource directly issue their own
  76  *                        GRU instruction and must wait/check completion.
  77  *
  78  *                When these resources are reserved, the caller can optionally
  79  *                associate a wait_queue with the resources and use asynchronous
  80  *                GRU instructions. When an async GRU instruction completes, the
  81  *                driver will do a wakeup on the event.
  82  *
  83  */
  84 
  85 
  86 #define ASYNC_HAN_TO_BID(h)     ((h) - 1)
  87 #define ASYNC_BID_TO_HAN(b)     ((b) + 1)
  88 #define ASYNC_HAN_TO_BS(h)      gru_base[ASYNC_HAN_TO_BID(h)]
  89 
  90 #define GRU_NUM_KERNEL_CBR      1
  91 #define GRU_NUM_KERNEL_DSR_BYTES 256
  92 #define GRU_NUM_KERNEL_DSR_CL   (GRU_NUM_KERNEL_DSR_BYTES /             \
  93                                         GRU_CACHE_LINE_BYTES)
  94 
  95 /* GRU instruction attributes for all instructions */
  96 #define IMA                     IMA_CB_DELAY
  97 
  98 /* GRU cacheline size is always 64 bytes - even on arches with 128 byte lines */
  99 #define __gru_cacheline_aligned__                               \
 100         __attribute__((__aligned__(GRU_CACHE_LINE_BYTES)))
 101 
 102 #define MAGIC   0x1234567887654321UL
 103 
 104 /* Default retry count for GRU errors on kernel instructions */
 105 #define EXCEPTION_RETRY_LIMIT   3
 106 
 107 /* Status of message queue sections */
 108 #define MQS_EMPTY               0
 109 #define MQS_FULL                1
 110 #define MQS_NOOP                2
 111 
 112 /*----------------- RESOURCE MANAGEMENT -------------------------------------*/
 113 /* optimized for x86_64 */
 114 struct message_queue {
 115         union gru_mesqhead      head __gru_cacheline_aligned__; /* CL 0 */
 116         int                     qlines;                         /* DW 1 */
 117         long                    hstatus[2];
 118         void                    *next __gru_cacheline_aligned__;/* CL 1 */
 119         void                    *limit;
 120         void                    *start;
 121         void                    *start2;
 122         char                    data ____cacheline_aligned;     /* CL 2 */
 123 };
 124 
 125 /* First word in every message - used by mesq interface */
 126 struct message_header {
 127         char    present;
 128         char    present2;
 129         char    lines;
 130         char    fill;
 131 };
 132 
 133 #define HSTATUS(mq, h)  ((mq) + offsetof(struct message_queue, hstatus[h]))
 134 
 135 /*
 136  * Reload the blade's kernel context into a GRU chiplet. Called holding
 137  * the bs_kgts_sema for READ. Will steal user contexts if necessary.
 138  */
 139 static void gru_load_kernel_context(struct gru_blade_state *bs, int blade_id)
 140 {
 141         struct gru_state *gru;
 142         struct gru_thread_state *kgts;
 143         void *vaddr;
 144         int ctxnum, ncpus;
 145 
 146         up_read(&bs->bs_kgts_sema);
 147         down_write(&bs->bs_kgts_sema);
 148 
 149         if (!bs->bs_kgts) {
 150                 do {
 151                         bs->bs_kgts = gru_alloc_gts(NULL, 0, 0, 0, 0, 0);
 152                         if (!IS_ERR(bs->bs_kgts))
 153                                 break;
 154                         msleep(1);
 155                 } while (true);
 156                 bs->bs_kgts->ts_user_blade_id = blade_id;
 157         }
 158         kgts = bs->bs_kgts;
 159 
 160         if (!kgts->ts_gru) {
 161                 STAT(load_kernel_context);
 162                 ncpus = uv_blade_nr_possible_cpus(blade_id);
 163                 kgts->ts_cbr_au_count = GRU_CB_COUNT_TO_AU(
 164                         GRU_NUM_KERNEL_CBR * ncpus + bs->bs_async_cbrs);
 165                 kgts->ts_dsr_au_count = GRU_DS_BYTES_TO_AU(
 166                         GRU_NUM_KERNEL_DSR_BYTES * ncpus +
 167                                 bs->bs_async_dsr_bytes);
 168                 while (!gru_assign_gru_context(kgts)) {
 169                         msleep(1);
 170                         gru_steal_context(kgts);
 171                 }
 172                 gru_load_context(kgts);
 173                 gru = bs->bs_kgts->ts_gru;
 174                 vaddr = gru->gs_gru_base_vaddr;
 175                 ctxnum = kgts->ts_ctxnum;
 176                 bs->kernel_cb = get_gseg_base_address_cb(vaddr, ctxnum, 0);
 177                 bs->kernel_dsr = get_gseg_base_address_ds(vaddr, ctxnum, 0);
 178         }
 179         downgrade_write(&bs->bs_kgts_sema);
 180 }
 181 
 182 /*
 183  * Free all kernel contexts that are not currently in use.
 184  *   Returns 0 if all freed, else number of inuse context.
 185  */
 186 static int gru_free_kernel_contexts(void)
 187 {
 188         struct gru_blade_state *bs;
 189         struct gru_thread_state *kgts;
 190         int bid, ret = 0;
 191 
 192         for (bid = 0; bid < GRU_MAX_BLADES; bid++) {
 193                 bs = gru_base[bid];
 194                 if (!bs)
 195                         continue;
 196 
 197                 /* Ignore busy contexts. Don't want to block here.  */
 198                 if (down_write_trylock(&bs->bs_kgts_sema)) {
 199                         kgts = bs->bs_kgts;
 200                         if (kgts && kgts->ts_gru)
 201                                 gru_unload_context(kgts, 0);
 202                         bs->bs_kgts = NULL;
 203                         up_write(&bs->bs_kgts_sema);
 204                         kfree(kgts);
 205                 } else {
 206                         ret++;
 207                 }
 208         }
 209         return ret;
 210 }
 211 
 212 /*
 213  * Lock & load the kernel context for the specified blade.
 214  */
 215 static struct gru_blade_state *gru_lock_kernel_context(int blade_id)
 216 {
 217         struct gru_blade_state *bs;
 218         int bid;
 219 
 220         STAT(lock_kernel_context);
 221 again:
 222         bid = blade_id < 0 ? uv_numa_blade_id() : blade_id;
 223         bs = gru_base[bid];
 224 
 225         /* Handle the case where migration occurred while waiting for the sema */
 226         down_read(&bs->bs_kgts_sema);
 227         if (blade_id < 0 && bid != uv_numa_blade_id()) {
 228                 up_read(&bs->bs_kgts_sema);
 229                 goto again;
 230         }
 231         if (!bs->bs_kgts || !bs->bs_kgts->ts_gru)
 232                 gru_load_kernel_context(bs, bid);
 233         return bs;
 234 
 235 }
 236 
 237 /*
 238  * Unlock the kernel context for the specified blade. Context is not
 239  * unloaded but may be stolen before next use.
 240  */
 241 static void gru_unlock_kernel_context(int blade_id)
 242 {
 243         struct gru_blade_state *bs;
 244 
 245         bs = gru_base[blade_id];
 246         up_read(&bs->bs_kgts_sema);
 247         STAT(unlock_kernel_context);
 248 }
 249 
 250 /*
 251  * Reserve & get pointers to the DSR/CBRs reserved for the current cpu.
 252  *      - returns with preemption disabled
 253  */
 254 static int gru_get_cpu_resources(int dsr_bytes, void **cb, void **dsr)
 255 {
 256         struct gru_blade_state *bs;
 257         int lcpu;
 258 
 259         BUG_ON(dsr_bytes > GRU_NUM_KERNEL_DSR_BYTES);
 260         preempt_disable();
 261         bs = gru_lock_kernel_context(-1);
 262         lcpu = uv_blade_processor_id();
 263         *cb = bs->kernel_cb + lcpu * GRU_HANDLE_STRIDE;
 264         *dsr = bs->kernel_dsr + lcpu * GRU_NUM_KERNEL_DSR_BYTES;
 265         return 0;
 266 }
 267 
 268 /*
 269  * Free the current cpus reserved DSR/CBR resources.
 270  */
 271 static void gru_free_cpu_resources(void *cb, void *dsr)
 272 {
 273         gru_unlock_kernel_context(uv_numa_blade_id());
 274         preempt_enable();
 275 }
 276 
 277 /*
 278  * Reserve GRU resources to be used asynchronously.
 279  *   Note: currently supports only 1 reservation per blade.
 280  *
 281  *      input:
 282  *              blade_id  - blade on which resources should be reserved
 283  *              cbrs      - number of CBRs
 284  *              dsr_bytes - number of DSR bytes needed
 285  *      output:
 286  *              handle to identify resource
 287  *              (0 = async resources already reserved)
 288  */
 289 unsigned long gru_reserve_async_resources(int blade_id, int cbrs, int dsr_bytes,
 290                         struct completion *cmp)
 291 {
 292         struct gru_blade_state *bs;
 293         struct gru_thread_state *kgts;
 294         int ret = 0;
 295 
 296         bs = gru_base[blade_id];
 297 
 298         down_write(&bs->bs_kgts_sema);
 299 
 300         /* Verify no resources already reserved */
 301         if (bs->bs_async_dsr_bytes + bs->bs_async_cbrs)
 302                 goto done;
 303         bs->bs_async_dsr_bytes = dsr_bytes;
 304         bs->bs_async_cbrs = cbrs;
 305         bs->bs_async_wq = cmp;
 306         kgts = bs->bs_kgts;
 307 
 308         /* Resources changed. Unload context if already loaded */
 309         if (kgts && kgts->ts_gru)
 310                 gru_unload_context(kgts, 0);
 311         ret = ASYNC_BID_TO_HAN(blade_id);
 312 
 313 done:
 314         up_write(&bs->bs_kgts_sema);
 315         return ret;
 316 }
 317 
 318 /*
 319  * Release async resources previously reserved.
 320  *
 321  *      input:
 322  *              han - handle to identify resources
 323  */
 324 void gru_release_async_resources(unsigned long han)
 325 {
 326         struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
 327 
 328         down_write(&bs->bs_kgts_sema);
 329         bs->bs_async_dsr_bytes = 0;
 330         bs->bs_async_cbrs = 0;
 331         bs->bs_async_wq = NULL;
 332         up_write(&bs->bs_kgts_sema);
 333 }
 334 
 335 /*
 336  * Wait for async GRU instructions to complete.
 337  *
 338  *      input:
 339  *              han - handle to identify resources
 340  */
 341 void gru_wait_async_cbr(unsigned long han)
 342 {
 343         struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
 344 
 345         wait_for_completion(bs->bs_async_wq);
 346         mb();
 347 }
 348 
 349 /*
 350  * Lock previous reserved async GRU resources
 351  *
 352  *      input:
 353  *              han - handle to identify resources
 354  *      output:
 355  *              cb  - pointer to first CBR
 356  *              dsr - pointer to first DSR
 357  */
 358 void gru_lock_async_resource(unsigned long han,  void **cb, void **dsr)
 359 {
 360         struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
 361         int blade_id = ASYNC_HAN_TO_BID(han);
 362         int ncpus;
 363 
 364         gru_lock_kernel_context(blade_id);
 365         ncpus = uv_blade_nr_possible_cpus(blade_id);
 366         if (cb)
 367                 *cb = bs->kernel_cb + ncpus * GRU_HANDLE_STRIDE;
 368         if (dsr)
 369                 *dsr = bs->kernel_dsr + ncpus * GRU_NUM_KERNEL_DSR_BYTES;
 370 }
 371 
 372 /*
 373  * Unlock previous reserved async GRU resources
 374  *
 375  *      input:
 376  *              han - handle to identify resources
 377  */
 378 void gru_unlock_async_resource(unsigned long han)
 379 {
 380         int blade_id = ASYNC_HAN_TO_BID(han);
 381 
 382         gru_unlock_kernel_context(blade_id);
 383 }
 384 
 385 /*----------------------------------------------------------------------*/
 386 int gru_get_cb_exception_detail(void *cb,
 387                 struct control_block_extended_exc_detail *excdet)
 388 {
 389         struct gru_control_block_extended *cbe;
 390         struct gru_thread_state *kgts = NULL;
 391         unsigned long off;
 392         int cbrnum, bid;
 393 
 394         /*
 395          * Locate kgts for cb. This algorithm is SLOW but
 396          * this function is rarely called (ie., almost never).
 397          * Performance does not matter.
 398          */
 399         for_each_possible_blade(bid) {
 400                 if (!gru_base[bid])
 401                         break;
 402                 kgts = gru_base[bid]->bs_kgts;
 403                 if (!kgts || !kgts->ts_gru)
 404                         continue;
 405                 off = cb - kgts->ts_gru->gs_gru_base_vaddr;
 406                 if (off < GRU_SIZE)
 407                         break;
 408                 kgts = NULL;
 409         }
 410         BUG_ON(!kgts);
 411         cbrnum = thread_cbr_number(kgts, get_cb_number(cb));
 412         cbe = get_cbe(GRUBASE(cb), cbrnum);
 413         gru_flush_cache(cbe);   /* CBE not coherent */
 414         sync_core();
 415         excdet->opc = cbe->opccpy;
 416         excdet->exopc = cbe->exopccpy;
 417         excdet->ecause = cbe->ecause;
 418         excdet->exceptdet0 = cbe->idef1upd;
 419         excdet->exceptdet1 = cbe->idef3upd;
 420         gru_flush_cache(cbe);
 421         return 0;
 422 }
 423 
 424 static char *gru_get_cb_exception_detail_str(int ret, void *cb,
 425                                              char *buf, int size)
 426 {
 427         struct gru_control_block_status *gen = (void *)cb;
 428         struct control_block_extended_exc_detail excdet;
 429 
 430         if (ret > 0 && gen->istatus == CBS_EXCEPTION) {
 431                 gru_get_cb_exception_detail(cb, &excdet);
 432                 snprintf(buf, size,
 433                         "GRU:%d exception: cb %p, opc %d, exopc %d, ecause 0x%x,"
 434                         "excdet0 0x%lx, excdet1 0x%x", smp_processor_id(),
 435                         gen, excdet.opc, excdet.exopc, excdet.ecause,
 436                         excdet.exceptdet0, excdet.exceptdet1);
 437         } else {
 438                 snprintf(buf, size, "No exception");
 439         }
 440         return buf;
 441 }
 442 
 443 static int gru_wait_idle_or_exception(struct gru_control_block_status *gen)
 444 {
 445         while (gen->istatus >= CBS_ACTIVE) {
 446                 cpu_relax();
 447                 barrier();
 448         }
 449         return gen->istatus;
 450 }
 451 
 452 static int gru_retry_exception(void *cb)
 453 {
 454         struct gru_control_block_status *gen = (void *)cb;
 455         struct control_block_extended_exc_detail excdet;
 456         int retry = EXCEPTION_RETRY_LIMIT;
 457 
 458         while (1)  {
 459                 if (gru_wait_idle_or_exception(gen) == CBS_IDLE)
 460                         return CBS_IDLE;
 461                 if (gru_get_cb_message_queue_substatus(cb))
 462                         return CBS_EXCEPTION;
 463                 gru_get_cb_exception_detail(cb, &excdet);
 464                 if ((excdet.ecause & ~EXCEPTION_RETRY_BITS) ||
 465                                 (excdet.cbrexecstatus & CBR_EXS_ABORT_OCC))
 466                         break;
 467                 if (retry-- == 0)
 468                         break;
 469                 gen->icmd = 1;
 470                 gru_flush_cache(gen);
 471         }
 472         return CBS_EXCEPTION;
 473 }
 474 
 475 int gru_check_status_proc(void *cb)
 476 {
 477         struct gru_control_block_status *gen = (void *)cb;
 478         int ret;
 479 
 480         ret = gen->istatus;
 481         if (ret == CBS_EXCEPTION)
 482                 ret = gru_retry_exception(cb);
 483         rmb();
 484         return ret;
 485 
 486 }
 487 
 488 int gru_wait_proc(void *cb)
 489 {
 490         struct gru_control_block_status *gen = (void *)cb;
 491         int ret;
 492 
 493         ret = gru_wait_idle_or_exception(gen);
 494         if (ret == CBS_EXCEPTION)
 495                 ret = gru_retry_exception(cb);
 496         rmb();
 497         return ret;
 498 }
 499 
 500 static void gru_abort(int ret, void *cb, char *str)
 501 {
 502         char buf[GRU_EXC_STR_SIZE];
 503 
 504         panic("GRU FATAL ERROR: %s - %s\n", str,
 505               gru_get_cb_exception_detail_str(ret, cb, buf, sizeof(buf)));
 506 }
 507 
 508 void gru_wait_abort_proc(void *cb)
 509 {
 510         int ret;
 511 
 512         ret = gru_wait_proc(cb);
 513         if (ret)
 514                 gru_abort(ret, cb, "gru_wait_abort");
 515 }
 516 
 517 
 518 /*------------------------------ MESSAGE QUEUES -----------------------------*/
 519 
 520 /* Internal status . These are NOT returned to the user. */
 521 #define MQIE_AGAIN              -1      /* try again */
 522 
 523 
 524 /*
 525  * Save/restore the "present" flag that is in the second line of 2-line
 526  * messages
 527  */
 528 static inline int get_present2(void *p)
 529 {
 530         struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
 531         return mhdr->present;
 532 }
 533 
 534 static inline void restore_present2(void *p, int val)
 535 {
 536         struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
 537         mhdr->present = val;
 538 }
 539 
 540 /*
 541  * Create a message queue.
 542  *      qlines - message queue size in cache lines. Includes 2-line header.
 543  */
 544 int gru_create_message_queue(struct gru_message_queue_desc *mqd,
 545                 void *p, unsigned int bytes, int nasid, int vector, int apicid)
 546 {
 547         struct message_queue *mq = p;
 548         unsigned int qlines;
 549 
 550         qlines = bytes / GRU_CACHE_LINE_BYTES - 2;
 551         memset(mq, 0, bytes);
 552         mq->start = &mq->data;
 553         mq->start2 = &mq->data + (qlines / 2 - 1) * GRU_CACHE_LINE_BYTES;
 554         mq->next = &mq->data;
 555         mq->limit = &mq->data + (qlines - 2) * GRU_CACHE_LINE_BYTES;
 556         mq->qlines = qlines;
 557         mq->hstatus[0] = 0;
 558         mq->hstatus[1] = 1;
 559         mq->head = gru_mesq_head(2, qlines / 2 + 1);
 560         mqd->mq = mq;
 561         mqd->mq_gpa = uv_gpa(mq);
 562         mqd->qlines = qlines;
 563         mqd->interrupt_pnode = nasid >> 1;
 564         mqd->interrupt_vector = vector;
 565         mqd->interrupt_apicid = apicid;
 566         return 0;
 567 }
 568 EXPORT_SYMBOL_GPL(gru_create_message_queue);
 569 
 570 /*
 571  * Send a NOOP message to a message queue
 572  *      Returns:
 573  *               0 - if queue is full after the send. This is the normal case
 574  *                   but various races can change this.
 575  *              -1 - if mesq sent successfully but queue not full
 576  *              >0 - unexpected error. MQE_xxx returned
 577  */
 578 static int send_noop_message(void *cb, struct gru_message_queue_desc *mqd,
 579                                 void *mesg)
 580 {
 581         const struct message_header noop_header = {
 582                                         .present = MQS_NOOP, .lines = 1};
 583         unsigned long m;
 584         int substatus, ret;
 585         struct message_header save_mhdr, *mhdr = mesg;
 586 
 587         STAT(mesq_noop);
 588         save_mhdr = *mhdr;
 589         *mhdr = noop_header;
 590         gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), 1, IMA);
 591         ret = gru_wait(cb);
 592 
 593         if (ret) {
 594                 substatus = gru_get_cb_message_queue_substatus(cb);
 595                 switch (substatus) {
 596                 case CBSS_NO_ERROR:
 597                         STAT(mesq_noop_unexpected_error);
 598                         ret = MQE_UNEXPECTED_CB_ERR;
 599                         break;
 600                 case CBSS_LB_OVERFLOWED:
 601                         STAT(mesq_noop_lb_overflow);
 602                         ret = MQE_CONGESTION;
 603                         break;
 604                 case CBSS_QLIMIT_REACHED:
 605                         STAT(mesq_noop_qlimit_reached);
 606                         ret = 0;
 607                         break;
 608                 case CBSS_AMO_NACKED:
 609                         STAT(mesq_noop_amo_nacked);
 610                         ret = MQE_CONGESTION;
 611                         break;
 612                 case CBSS_PUT_NACKED:
 613                         STAT(mesq_noop_put_nacked);
 614                         m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
 615                         gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, 1, 1,
 616                                                 IMA);
 617                         if (gru_wait(cb) == CBS_IDLE)
 618                                 ret = MQIE_AGAIN;
 619                         else
 620                                 ret = MQE_UNEXPECTED_CB_ERR;
 621                         break;
 622                 case CBSS_PAGE_OVERFLOW:
 623                         STAT(mesq_noop_page_overflow);
 624                         /* fall through */
 625                 default:
 626                         BUG();
 627                 }
 628         }
 629         *mhdr = save_mhdr;
 630         return ret;
 631 }
 632 
 633 /*
 634  * Handle a gru_mesq full.
 635  */
 636 static int send_message_queue_full(void *cb, struct gru_message_queue_desc *mqd,
 637                                 void *mesg, int lines)
 638 {
 639         union gru_mesqhead mqh;
 640         unsigned int limit, head;
 641         unsigned long avalue;
 642         int half, qlines;
 643 
 644         /* Determine if switching to first/second half of q */
 645         avalue = gru_get_amo_value(cb);
 646         head = gru_get_amo_value_head(cb);
 647         limit = gru_get_amo_value_limit(cb);
 648 
 649         qlines = mqd->qlines;
 650         half = (limit != qlines);
 651 
 652         if (half)
 653                 mqh = gru_mesq_head(qlines / 2 + 1, qlines);
 654         else
 655                 mqh = gru_mesq_head(2, qlines / 2 + 1);
 656 
 657         /* Try to get lock for switching head pointer */
 658         gru_gamir(cb, EOP_IR_CLR, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, IMA);
 659         if (gru_wait(cb) != CBS_IDLE)
 660                 goto cberr;
 661         if (!gru_get_amo_value(cb)) {
 662                 STAT(mesq_qf_locked);
 663                 return MQE_QUEUE_FULL;
 664         }
 665 
 666         /* Got the lock. Send optional NOP if queue not full, */
 667         if (head != limit) {
 668                 if (send_noop_message(cb, mqd, mesg)) {
 669                         gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half),
 670                                         XTYPE_DW, IMA);
 671                         if (gru_wait(cb) != CBS_IDLE)
 672                                 goto cberr;
 673                         STAT(mesq_qf_noop_not_full);
 674                         return MQIE_AGAIN;
 675                 }
 676                 avalue++;
 677         }
 678 
 679         /* Then flip queuehead to other half of queue. */
 680         gru_gamer(cb, EOP_ERR_CSWAP, mqd->mq_gpa, XTYPE_DW, mqh.val, avalue,
 681                                                         IMA);
 682         if (gru_wait(cb) != CBS_IDLE)
 683                 goto cberr;
 684 
 685         /* If not successfully in swapping queue head, clear the hstatus lock */
 686         if (gru_get_amo_value(cb) != avalue) {
 687                 STAT(mesq_qf_switch_head_failed);
 688                 gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), XTYPE_DW,
 689                                                         IMA);
 690                 if (gru_wait(cb) != CBS_IDLE)
 691                         goto cberr;
 692         }
 693         return MQIE_AGAIN;
 694 cberr:
 695         STAT(mesq_qf_unexpected_error);
 696         return MQE_UNEXPECTED_CB_ERR;
 697 }
 698 
 699 /*
 700  * Handle a PUT failure. Note: if message was a 2-line message, one of the
 701  * lines might have successfully have been written. Before sending the
 702  * message, "present" must be cleared in BOTH lines to prevent the receiver
 703  * from prematurely seeing the full message.
 704  */
 705 static int send_message_put_nacked(void *cb, struct gru_message_queue_desc *mqd,
 706                         void *mesg, int lines)
 707 {
 708         unsigned long m;
 709         int ret, loops = 200;   /* experimentally determined */
 710 
 711         m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
 712         if (lines == 2) {
 713                 gru_vset(cb, m, 0, XTYPE_CL, lines, 1, IMA);
 714                 if (gru_wait(cb) != CBS_IDLE)
 715                         return MQE_UNEXPECTED_CB_ERR;
 716         }
 717         gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, lines, 1, IMA);
 718         if (gru_wait(cb) != CBS_IDLE)
 719                 return MQE_UNEXPECTED_CB_ERR;
 720 
 721         if (!mqd->interrupt_vector)
 722                 return MQE_OK;
 723 
 724         /*
 725          * Send a noop message in order to deliver a cross-partition interrupt
 726          * to the SSI that contains the target message queue. Normally, the
 727          * interrupt is automatically delivered by hardware following mesq
 728          * operations, but some error conditions require explicit delivery.
 729          * The noop message will trigger delivery. Otherwise partition failures
 730          * could cause unrecovered errors.
 731          */
 732         do {
 733                 ret = send_noop_message(cb, mqd, mesg);
 734         } while ((ret == MQIE_AGAIN || ret == MQE_CONGESTION) && (loops-- > 0));
 735 
 736         if (ret == MQIE_AGAIN || ret == MQE_CONGESTION) {
 737                 /*
 738                  * Don't indicate to the app to resend the message, as it's
 739                  * already been successfully sent.  We simply send an OK
 740                  * (rather than fail the send with MQE_UNEXPECTED_CB_ERR),
 741                  * assuming that the other side is receiving enough
 742                  * interrupts to get this message processed anyway.
 743                  */
 744                 ret = MQE_OK;
 745         }
 746         return ret;
 747 }
 748 
 749 /*
 750  * Handle a gru_mesq failure. Some of these failures are software recoverable
 751  * or retryable.
 752  */
 753 static int send_message_failure(void *cb, struct gru_message_queue_desc *mqd,
 754                                 void *mesg, int lines)
 755 {
 756         int substatus, ret = 0;
 757 
 758         substatus = gru_get_cb_message_queue_substatus(cb);
 759         switch (substatus) {
 760         case CBSS_NO_ERROR:
 761                 STAT(mesq_send_unexpected_error);
 762                 ret = MQE_UNEXPECTED_CB_ERR;
 763                 break;
 764         case CBSS_LB_OVERFLOWED:
 765                 STAT(mesq_send_lb_overflow);
 766                 ret = MQE_CONGESTION;
 767                 break;
 768         case CBSS_QLIMIT_REACHED:
 769                 STAT(mesq_send_qlimit_reached);
 770                 ret = send_message_queue_full(cb, mqd, mesg, lines);
 771                 break;
 772         case CBSS_AMO_NACKED:
 773                 STAT(mesq_send_amo_nacked);
 774                 ret = MQE_CONGESTION;
 775                 break;
 776         case CBSS_PUT_NACKED:
 777                 STAT(mesq_send_put_nacked);
 778                 ret = send_message_put_nacked(cb, mqd, mesg, lines);
 779                 break;
 780         case CBSS_PAGE_OVERFLOW:
 781                 STAT(mesq_page_overflow);
 782                 /* fall through */
 783         default:
 784                 BUG();
 785         }
 786         return ret;
 787 }
 788 
 789 /*
 790  * Send a message to a message queue
 791  *      mqd     message queue descriptor
 792  *      mesg    message. ust be vaddr within a GSEG
 793  *      bytes   message size (<= 2 CL)
 794  */
 795 int gru_send_message_gpa(struct gru_message_queue_desc *mqd, void *mesg,
 796                                 unsigned int bytes)
 797 {
 798         struct message_header *mhdr;
 799         void *cb;
 800         void *dsr;
 801         int istatus, clines, ret;
 802 
 803         STAT(mesq_send);
 804         BUG_ON(bytes < sizeof(int) || bytes > 2 * GRU_CACHE_LINE_BYTES);
 805 
 806         clines = DIV_ROUND_UP(bytes, GRU_CACHE_LINE_BYTES);
 807         if (gru_get_cpu_resources(bytes, &cb, &dsr))
 808                 return MQE_BUG_NO_RESOURCES;
 809         memcpy(dsr, mesg, bytes);
 810         mhdr = dsr;
 811         mhdr->present = MQS_FULL;
 812         mhdr->lines = clines;
 813         if (clines == 2) {
 814                 mhdr->present2 = get_present2(mhdr);
 815                 restore_present2(mhdr, MQS_FULL);
 816         }
 817 
 818         do {
 819                 ret = MQE_OK;
 820                 gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), clines, IMA);
 821                 istatus = gru_wait(cb);
 822                 if (istatus != CBS_IDLE)
 823                         ret = send_message_failure(cb, mqd, dsr, clines);
 824         } while (ret == MQIE_AGAIN);
 825         gru_free_cpu_resources(cb, dsr);
 826 
 827         if (ret)
 828                 STAT(mesq_send_failed);
 829         return ret;
 830 }
 831 EXPORT_SYMBOL_GPL(gru_send_message_gpa);
 832 
 833 /*
 834  * Advance the receive pointer for the queue to the next message.
 835  */
 836 void gru_free_message(struct gru_message_queue_desc *mqd, void *mesg)
 837 {
 838         struct message_queue *mq = mqd->mq;
 839         struct message_header *mhdr = mq->next;
 840         void *next, *pnext;
 841         int half = -1;
 842         int lines = mhdr->lines;
 843 
 844         if (lines == 2)
 845                 restore_present2(mhdr, MQS_EMPTY);
 846         mhdr->present = MQS_EMPTY;
 847 
 848         pnext = mq->next;
 849         next = pnext + GRU_CACHE_LINE_BYTES * lines;
 850         if (next == mq->limit) {
 851                 next = mq->start;
 852                 half = 1;
 853         } else if (pnext < mq->start2 && next >= mq->start2) {
 854                 half = 0;
 855         }
 856 
 857         if (half >= 0)
 858                 mq->hstatus[half] = 1;
 859         mq->next = next;
 860 }
 861 EXPORT_SYMBOL_GPL(gru_free_message);
 862 
 863 /*
 864  * Get next message from message queue. Return NULL if no message
 865  * present. User must call next_message() to move to next message.
 866  *      rmq     message queue
 867  */
 868 void *gru_get_next_message(struct gru_message_queue_desc *mqd)
 869 {
 870         struct message_queue *mq = mqd->mq;
 871         struct message_header *mhdr = mq->next;
 872         int present = mhdr->present;
 873 
 874         /* skip NOOP messages */
 875         while (present == MQS_NOOP) {
 876                 gru_free_message(mqd, mhdr);
 877                 mhdr = mq->next;
 878                 present = mhdr->present;
 879         }
 880 
 881         /* Wait for both halves of 2 line messages */
 882         if (present == MQS_FULL && mhdr->lines == 2 &&
 883                                 get_present2(mhdr) == MQS_EMPTY)
 884                 present = MQS_EMPTY;
 885 
 886         if (!present) {
 887                 STAT(mesq_receive_none);
 888                 return NULL;
 889         }
 890 
 891         if (mhdr->lines == 2)
 892                 restore_present2(mhdr, mhdr->present2);
 893 
 894         STAT(mesq_receive);
 895         return mhdr;
 896 }
 897 EXPORT_SYMBOL_GPL(gru_get_next_message);
 898 
 899 /* ---------------------- GRU DATA COPY FUNCTIONS ---------------------------*/
 900 
 901 /*
 902  * Load a DW from a global GPA. The GPA can be a memory or MMR address.
 903  */
 904 int gru_read_gpa(unsigned long *value, unsigned long gpa)
 905 {
 906         void *cb;
 907         void *dsr;
 908         int ret, iaa;
 909 
 910         STAT(read_gpa);
 911         if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr))
 912                 return MQE_BUG_NO_RESOURCES;
 913         iaa = gpa >> 62;
 914         gru_vload_phys(cb, gpa, gru_get_tri(dsr), iaa, IMA);
 915         ret = gru_wait(cb);
 916         if (ret == CBS_IDLE)
 917                 *value = *(unsigned long *)dsr;
 918         gru_free_cpu_resources(cb, dsr);
 919         return ret;
 920 }
 921 EXPORT_SYMBOL_GPL(gru_read_gpa);
 922 
 923 
 924 /*
 925  * Copy a block of data using the GRU resources
 926  */
 927 int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa,
 928                                 unsigned int bytes)
 929 {
 930         void *cb;
 931         void *dsr;
 932         int ret;
 933 
 934         STAT(copy_gpa);
 935         if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr))
 936                 return MQE_BUG_NO_RESOURCES;
 937         gru_bcopy(cb, src_gpa, dest_gpa, gru_get_tri(dsr),
 938                   XTYPE_B, bytes, GRU_NUM_KERNEL_DSR_CL, IMA);
 939         ret = gru_wait(cb);
 940         gru_free_cpu_resources(cb, dsr);
 941         return ret;
 942 }
 943 EXPORT_SYMBOL_GPL(gru_copy_gpa);
 944 
 945 /* ------------------- KERNEL QUICKTESTS RUN AT STARTUP ----------------*/
 946 /*      Temp - will delete after we gain confidence in the GRU          */
 947 
 948 static int quicktest0(unsigned long arg)
 949 {
 950         unsigned long word0;
 951         unsigned long word1;
 952         void *cb;
 953         void *dsr;
 954         unsigned long *p;
 955         int ret = -EIO;
 956 
 957         if (gru_get_cpu_resources(GRU_CACHE_LINE_BYTES, &cb, &dsr))
 958                 return MQE_BUG_NO_RESOURCES;
 959         p = dsr;
 960         word0 = MAGIC;
 961         word1 = 0;
 962 
 963         gru_vload(cb, uv_gpa(&word0), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
 964         if (gru_wait(cb) != CBS_IDLE) {
 965                 printk(KERN_DEBUG "GRU:%d quicktest0: CBR failure 1\n", smp_processor_id());
 966                 goto done;
 967         }
 968 
 969         if (*p != MAGIC) {
 970                 printk(KERN_DEBUG "GRU:%d quicktest0 bad magic 0x%lx\n", smp_processor_id(), *p);
 971                 goto done;
 972         }
 973         gru_vstore(cb, uv_gpa(&word1), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
 974         if (gru_wait(cb) != CBS_IDLE) {
 975                 printk(KERN_DEBUG "GRU:%d quicktest0: CBR failure 2\n", smp_processor_id());
 976                 goto done;
 977         }
 978 
 979         if (word0 != word1 || word1 != MAGIC) {
 980                 printk(KERN_DEBUG
 981                        "GRU:%d quicktest0 err: found 0x%lx, expected 0x%lx\n",
 982                      smp_processor_id(), word1, MAGIC);
 983                 goto done;
 984         }
 985         ret = 0;
 986 
 987 done:
 988         gru_free_cpu_resources(cb, dsr);
 989         return ret;
 990 }
 991 
 992 #define ALIGNUP(p, q)   ((void *)(((unsigned long)(p) + (q) - 1) & ~(q - 1)))
 993 
 994 static int quicktest1(unsigned long arg)
 995 {
 996         struct gru_message_queue_desc mqd;
 997         void *p, *mq;
 998         int i, ret = -EIO;
 999         char mes[GRU_CACHE_LINE_BYTES], *m;
1000 
1001         /* Need  1K cacheline aligned that does not cross page boundary */
1002         p = kmalloc(4096, 0);
1003         if (p == NULL)
1004                 return -ENOMEM;
1005         mq = ALIGNUP(p, 1024);
1006         memset(mes, 0xee, sizeof(mes));
1007 
1008         gru_create_message_queue(&mqd, mq, 8 * GRU_CACHE_LINE_BYTES, 0, 0, 0);
1009         for (i = 0; i < 6; i++) {
1010                 mes[8] = i;
1011                 do {
1012                         ret = gru_send_message_gpa(&mqd, mes, sizeof(mes));
1013                 } while (ret == MQE_CONGESTION);
1014                 if (ret)
1015                         break;
1016         }
1017         if (ret != MQE_QUEUE_FULL || i != 4) {
1018                 printk(KERN_DEBUG "GRU:%d quicktest1: unexpect status %d, i %d\n",
1019                        smp_processor_id(), ret, i);
1020                 goto done;
1021         }
1022 
1023         for (i = 0; i < 6; i++) {
1024                 m = gru_get_next_message(&mqd);
1025                 if (!m || m[8] != i)
1026                         break;
1027                 gru_free_message(&mqd, m);
1028         }
1029         if (i != 4) {
1030                 printk(KERN_DEBUG "GRU:%d quicktest2: bad message, i %d, m %p, m8 %d\n",
1031                         smp_processor_id(), i, m, m ? m[8] : -1);
1032                 goto done;
1033         }
1034         ret = 0;
1035 
1036 done:
1037         kfree(p);
1038         return ret;
1039 }
1040 
1041 static int quicktest2(unsigned long arg)
1042 {
1043         static DECLARE_COMPLETION(cmp);
1044         unsigned long han;
1045         int blade_id = 0;
1046         int numcb = 4;
1047         int ret = 0;
1048         unsigned long *buf;
1049         void *cb0, *cb;
1050         struct gru_control_block_status *gen;
1051         int i, k, istatus, bytes;
1052 
1053         bytes = numcb * 4 * 8;
1054         buf = kmalloc(bytes, GFP_KERNEL);
1055         if (!buf)
1056                 return -ENOMEM;
1057 
1058         ret = -EBUSY;
1059         han = gru_reserve_async_resources(blade_id, numcb, 0, &cmp);
1060         if (!han)
1061                 goto done;
1062 
1063         gru_lock_async_resource(han, &cb0, NULL);
1064         memset(buf, 0xee, bytes);
1065         for (i = 0; i < numcb; i++)
1066                 gru_vset(cb0 + i * GRU_HANDLE_STRIDE, uv_gpa(&buf[i * 4]), 0,
1067                                 XTYPE_DW, 4, 1, IMA_INTERRUPT);
1068 
1069         ret = 0;
1070         k = numcb;
1071         do {
1072                 gru_wait_async_cbr(han);
1073                 for (i = 0; i < numcb; i++) {
1074                         cb = cb0 + i * GRU_HANDLE_STRIDE;
1075                         istatus = gru_check_status(cb);
1076                         if (istatus != CBS_ACTIVE && istatus != CBS_CALL_OS)
1077                                 break;
1078                 }
1079                 if (i == numcb)
1080                         continue;
1081                 if (istatus != CBS_IDLE) {
1082                         printk(KERN_DEBUG "GRU:%d quicktest2: cb %d, exception\n", smp_processor_id(), i);
1083                         ret = -EFAULT;
1084                 } else if (buf[4 * i] || buf[4 * i + 1] || buf[4 * i + 2] ||
1085                                 buf[4 * i + 3]) {
1086                         printk(KERN_DEBUG "GRU:%d quicktest2:cb %d,  buf 0x%lx, 0x%lx, 0x%lx, 0x%lx\n",
1087                                smp_processor_id(), i, buf[4 * i], buf[4 * i + 1], buf[4 * i + 2], buf[4 * i + 3]);
1088                         ret = -EIO;
1089                 }
1090                 k--;
1091                 gen = cb;
1092                 gen->istatus = CBS_CALL_OS; /* don't handle this CBR again */
1093         } while (k);
1094         BUG_ON(cmp.done);
1095 
1096         gru_unlock_async_resource(han);
1097         gru_release_async_resources(han);
1098 done:
1099         kfree(buf);
1100         return ret;
1101 }
1102 
1103 #define BUFSIZE 200
1104 static int quicktest3(unsigned long arg)
1105 {
1106         char buf1[BUFSIZE], buf2[BUFSIZE];
1107         int ret = 0;
1108 
1109         memset(buf2, 0, sizeof(buf2));
1110         memset(buf1, get_cycles() & 255, sizeof(buf1));
1111         gru_copy_gpa(uv_gpa(buf2), uv_gpa(buf1), BUFSIZE);
1112         if (memcmp(buf1, buf2, BUFSIZE)) {
1113                 printk(KERN_DEBUG "GRU:%d quicktest3 error\n", smp_processor_id());
1114                 ret = -EIO;
1115         }
1116         return ret;
1117 }
1118 
1119 /*
1120  * Debugging only. User hook for various kernel tests
1121  * of driver & gru.
1122  */
1123 int gru_ktest(unsigned long arg)
1124 {
1125         int ret = -EINVAL;
1126 
1127         switch (arg & 0xff) {
1128         case 0:
1129                 ret = quicktest0(arg);
1130                 break;
1131         case 1:
1132                 ret = quicktest1(arg);
1133                 break;
1134         case 2:
1135                 ret = quicktest2(arg);
1136                 break;
1137         case 3:
1138                 ret = quicktest3(arg);
1139                 break;
1140         case 99:
1141                 ret = gru_free_kernel_contexts();
1142                 break;
1143         }
1144         return ret;
1145 
1146 }
1147 
1148 int gru_kservices_init(void)
1149 {
1150         return 0;
1151 }
1152 
1153 void gru_kservices_exit(void)
1154 {
1155         if (gru_free_kernel_contexts())
1156                 BUG();
1157 }
1158
/* [<][>][^][v][top][bottom][index][help] */
root/drivers/misc/sgi-gru/grukservices.c

DEFINITIONS
