drivers/misc/sgi-gru/grutables.h

   1 /* SPDX-License-Identifier: GPL-2.0-or-later */
   2 /*
   3  * SN Platform GRU Driver
   4  *
   5  *            GRU DRIVER TABLES, MACROS, externs, etc
   6  *
   7  *  Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
   8  */
   9
  10 #ifndef __GRUTABLES_H__
  11 #define __GRUTABLES_H__
  12
  13 /*
  14  * GRU Chiplet:
  15  *   The GRU is a user addressible memory accelerator. It provides
  16  *   several forms of load, store, memset, bcopy instructions. In addition, it
  17  *   contains special instructions for AMOs, sending messages to message
  18  *   queues, etc.
  19  *
  20  *   The GRU is an integral part of the node controller. It connects
  21  *   directly to the cpu socket. In its current implementation, there are 2
  22  *   GRU chiplets in the node controller on each blade (~node).
  23  *
  24  *   The entire GRU memory space is fully coherent and cacheable by the cpus.
  25  *
  26  *   Each GRU chiplet has a physical memory map that looks like the following:
  27  *
  28  *      +-----------------+
  29  *      |/////////////////|
  30  *      |/////////////////|
  31  *      |/////////////////|
  32  *      |/////////////////|
  33  *      |/////////////////|
  34  *      |/////////////////|
  35  *      |/////////////////|
  36  *      |/////////////////|
  37  *      +-----------------+
  38  *      |  system control |
  39  *      +-----------------+        _______ +-------------+
  40  *      |/////////////////|       /        |             |
  41  *      |/////////////////|      /         |             |
  42  *      |/////////////////|     /          | instructions|
  43  *      |/////////////////|    /           |             |
  44  *      |/////////////////|   /            |             |
  45  *      |/////////////////|  /             |-------------|
  46  *      |/////////////////| /              |             |
  47  *      +-----------------+                |             |
  48  *      |   context 15    |                |  data       |
  49  *      +-----------------+                |             |
  50  *      |    ......       | \              |             |
  51  *      +-----------------+  \____________ +-------------+
  52  *      |   context 1     |
  53  *      +-----------------+
  54  *      |   context 0     |
  55  *      +-----------------+
  56  *
  57  *   Each of the "contexts" is a chunk of memory that can be mmaped into user
  58  *   space. The context consists of 2 parts:
  59  *
  60  *      - an instruction space that can be directly accessed by the user
  61  *        to issue GRU instructions and to check instruction status.
  62  *
  63  *      - a data area that acts as normal RAM.
  64  *
  65  *   User instructions contain virtual addresses of data to be accessed by the
  66  *   GRU. The GRU contains a TLB that is used to convert these user virtual
  67  *   addresses to physical addresses.
  68  *
  69  *   The "system control" area of the GRU chiplet is used by the kernel driver
  70  *   to manage user contexts and to perform functions such as TLB dropin and
  71  *   purging.
  72  *
  73  *   One context may be reserved for the kernel and used for cross-partition
  74  *   communication. The GRU will also be used to asynchronously zero out
  75  *   large blocks of memory (not currently implemented).
  76  *
  77  *
  78  * Tables:
  79  *
  80  *      VDATA-VMA Data          - Holds a few parameters. Head of linked list of
  81  *                                GTS tables for threads using the GSEG
  82  *      GTS - Gru Thread State  - contains info for managing a GSEG context. A
  83  *                                GTS is allocated for each thread accessing a
  84  *                                GSEG.
  85  *      GTD - GRU Thread Data   - contains shadow copy of GRU data when GSEG is
  86  *                                not loaded into a GRU
  87  *      GMS - GRU Memory Struct - Used to manage TLB shootdowns. Tracks GRUs
  88  *                                where a GSEG has been loaded. Similar to
  89  *                                an mm_struct but for GRU.
  90  *
  91  *      GS  - GRU State         - Used to manage the state of a GRU chiplet
  92  *      BS  - Blade State       - Used to manage state of all GRU chiplets
  93  *                                on a blade
  94  *
  95  *
  96  *  Normal task tables for task using GRU.
  97  *              - 2 threads in process
  98  *              - 2 GSEGs open in process
  99  *              - GSEG1 is being used by both threads
 100  *              - GSEG2 is used only by thread 2
 101  *
 102  *       task -->|
 103  *       task ---+---> mm ->------ (notifier) -------+-> gms
 104  *                     |                             |
 105  *                     |--> vma -> vdata ---> gts--->|          GSEG1 (thread1)
 106  *                     |                  |          |
 107  *                     |                  +-> gts--->|          GSEG1 (thread2)
 108  *                     |                             |
 109  *                     |--> vma -> vdata ---> gts--->|          GSEG2 (thread2)
 110  *                     .
 111  *                     .
 112  *
 113  *  GSEGs are marked DONTCOPY on fork
 114  *
 115  * At open
 116  *      file.private_data -> NULL
 117  *
 118  * At mmap,
 119  *      vma -> vdata
 120  *
 121  * After gseg reference
 122  *      vma -> vdata ->gts
 123  *
 124  * After fork
 125  *   parent
 126  *      vma -> vdata -> gts
 127  *   child
 128  *      (vma is not copied)
 129  *
 130  */
 131
 132 #include <linux/rmap.h>
 133 #include <linux/interrupt.h>
 134 #include <linux/mutex.h>
 135 #include <linux/wait.h>
 136 #include <linux/mmu_notifier.h>
 137 #include <linux/mm_types.h>
 138 #include "gru.h"
 139 #include "grulib.h"
 140 #include "gruhandles.h"
 141
 142 extern struct gru_stats_s gru_stats;
 143 extern struct gru_blade_state *gru_base[];
 144 extern unsigned long gru_start_paddr, gru_end_paddr;
 145 extern void *gru_start_vaddr;
 146 extern unsigned int gru_max_gids;
 147
 148 #define GRU_MAX_BLADES          MAX_NUMNODES
 149 #define GRU_MAX_GRUS            (GRU_MAX_BLADES * GRU_CHIPLETS_PER_BLADE)
 150
 151 #define GRU_DRIVER_ID_STR       "SGI GRU Device Driver"
 152 #define GRU_DRIVER_VERSION_STR  "0.85"
 153
 154 /*
 155  * GRU statistics.
 156  */
 157 struct gru_stats_s {
 158         atomic_long_t vdata_alloc;
 159         atomic_long_t vdata_free;
 160         atomic_long_t gts_alloc;
 161         atomic_long_t gts_free;
 162         atomic_long_t gms_alloc;
 163         atomic_long_t gms_free;
 164         atomic_long_t gts_double_allocate;
 165         atomic_long_t assign_context;
 166         atomic_long_t assign_context_failed;
 167         atomic_long_t free_context;
 168         atomic_long_t load_user_context;
 169         atomic_long_t load_kernel_context;
 170         atomic_long_t lock_kernel_context;
 171         atomic_long_t unlock_kernel_context;
 172         atomic_long_t steal_user_context;
 173         atomic_long_t steal_kernel_context;
 174         atomic_long_t steal_context_failed;
 175         atomic_long_t nopfn;
 176         atomic_long_t asid_new;
 177         atomic_long_t asid_next;
 178         atomic_long_t asid_wrap;
 179         atomic_long_t asid_reuse;
 180         atomic_long_t intr;
 181         atomic_long_t intr_cbr;
 182         atomic_long_t intr_tfh;
 183         atomic_long_t intr_spurious;
 184         atomic_long_t intr_mm_lock_failed;
 185         atomic_long_t call_os;
 186         atomic_long_t call_os_wait_queue;
 187         atomic_long_t user_flush_tlb;
 188         atomic_long_t user_unload_context;
 189         atomic_long_t user_exception;
 190         atomic_long_t set_context_option;
 191         atomic_long_t check_context_retarget_intr;
 192         atomic_long_t check_context_unload;
 193         atomic_long_t tlb_dropin;
 194         atomic_long_t tlb_preload_page;
 195         atomic_long_t tlb_dropin_fail_no_asid;
 196         atomic_long_t tlb_dropin_fail_upm;
 197         atomic_long_t tlb_dropin_fail_invalid;
 198         atomic_long_t tlb_dropin_fail_range_active;
 199         atomic_long_t tlb_dropin_fail_idle;
 200         atomic_long_t tlb_dropin_fail_fmm;
 201         atomic_long_t tlb_dropin_fail_no_exception;
 202         atomic_long_t tfh_stale_on_fault;
 203         atomic_long_t mmu_invalidate_range;
 204         atomic_long_t mmu_invalidate_page;
 205         atomic_long_t flush_tlb;
 206         atomic_long_t flush_tlb_gru;
 207         atomic_long_t flush_tlb_gru_tgh;
 208         atomic_long_t flush_tlb_gru_zero_asid;
 209
 210         atomic_long_t copy_gpa;
 211         atomic_long_t read_gpa;
 212
 213         atomic_long_t mesq_receive;
 214         atomic_long_t mesq_receive_none;
 215         atomic_long_t mesq_send;
 216         atomic_long_t mesq_send_failed;
 217         atomic_long_t mesq_noop;
 218         atomic_long_t mesq_send_unexpected_error;
 219         atomic_long_t mesq_send_lb_overflow;
 220         atomic_long_t mesq_send_qlimit_reached;
 221         atomic_long_t mesq_send_amo_nacked;
 222         atomic_long_t mesq_send_put_nacked;
 223         atomic_long_t mesq_page_overflow;
 224         atomic_long_t mesq_qf_locked;
 225         atomic_long_t mesq_qf_noop_not_full;
 226         atomic_long_t mesq_qf_switch_head_failed;
 227         atomic_long_t mesq_qf_unexpected_error;
 228         atomic_long_t mesq_noop_unexpected_error;
 229         atomic_long_t mesq_noop_lb_overflow;
 230         atomic_long_t mesq_noop_qlimit_reached;
 231         atomic_long_t mesq_noop_amo_nacked;
 232         atomic_long_t mesq_noop_put_nacked;
 233         atomic_long_t mesq_noop_page_overflow;
 234
 235 };
 236
 237 enum mcs_op {cchop_allocate, cchop_start, cchop_interrupt, cchop_interrupt_sync,
 238         cchop_deallocate, tfhop_write_only, tfhop_write_restart,
 239         tghop_invalidate, mcsop_last};
 240
 241 struct mcs_op_statistic {
 242         atomic_long_t   count;
 243         atomic_long_t   total;
 244         unsigned long   max;
 245 };
 246
 247 extern struct mcs_op_statistic mcs_op_statistics[mcsop_last];
 248
 249 #define OPT_DPRINT              1
 250 #define OPT_STATS               2
 251
 252
 253 #define IRQ_GRU                 110     /* Starting IRQ number for interrupts */
 254
 255 /* Delay in jiffies between attempts to assign a GRU context */
 256 #define GRU_ASSIGN_DELAY        ((HZ * 20) / 1000)
 257
 258 /*
 259  * If a process has it's context stolen, min delay in jiffies before trying to
 260  * steal a context from another process.
 261  */
 262 #define GRU_STEAL_DELAY         ((HZ * 200) / 1000)
 263
 264 #define STAT(id)        do {                                            \
 265                                 if (gru_options & OPT_STATS)            \
 266                                         atomic_long_inc(&gru_stats.id); \
 267                         } while (0)
 268
 269 #ifdef CONFIG_SGI_GRU_DEBUG
 270 #define gru_dbg(dev, fmt, x...)                                         \
 271         do {                                                            \
 272                 if (gru_options & OPT_DPRINT)                           \
 273                         printk(KERN_DEBUG "GRU:%d %s: " fmt, smp_processor_id(), __func__, x);\
 274         } while (0)
 275 #else
 276 #define gru_dbg(x...)
 277 #endif
 278
 279 /*-----------------------------------------------------------------------------
 280  * ASID management
 281  */
 282 #define MAX_ASID        0xfffff0
 283 #define MIN_ASID        8
 284 #define ASID_INC        8       /* number of regions */
 285
 286 /* Generate a GRU asid value from a GRU base asid & a virtual address. */
 287 #define VADDR_HI_BIT            64
 288 #define GRUREGION(addr)         ((addr) >> (VADDR_HI_BIT - 3) & 3)
 289 #define GRUASID(asid, addr)     ((asid) + GRUREGION(addr))
 290
 291 /*------------------------------------------------------------------------------
 292  *  File & VMS Tables
 293  */
 294
 295 struct gru_state;
 296
 297 /*
 298  * This structure is pointed to from the mmstruct via the notifier pointer.
 299  * There is one of these per address space.
 300  */
 301 struct gru_mm_tracker {                         /* pack to reduce size */
 302         unsigned int            mt_asid_gen:24; /* ASID wrap count */
 303         unsigned int            mt_asid:24;     /* current base ASID for gru */
 304         unsigned short          mt_ctxbitmap:16;/* bitmap of contexts using
 305                                                    asid */
 306 } __attribute__ ((packed));
 307
 308 struct gru_mm_struct {
 309         struct mmu_notifier     ms_notifier;
 310         spinlock_t              ms_asid_lock;   /* protects ASID assignment */
 311         atomic_t                ms_range_active;/* num range_invals active */
 312         wait_queue_head_t       ms_wait_queue;
 313         DECLARE_BITMAP(ms_asidmap, GRU_MAX_GRUS);
 314         struct gru_mm_tracker   ms_asids[GRU_MAX_GRUS];
 315 };
 316
 317 /*
 318  * One of these structures is allocated when a GSEG is mmaped. The
 319  * structure is pointed to by the vma->vm_private_data field in the vma struct.
 320  */
 321 struct gru_vma_data {
 322         spinlock_t              vd_lock;        /* Serialize access to vma */
 323         struct list_head        vd_head;        /* head of linked list of gts */
 324         long                    vd_user_options;/* misc user option flags */
 325         int                     vd_cbr_au_count;
 326         int                     vd_dsr_au_count;
 327         unsigned char           vd_tlb_preload_count;
 328 };
 329
 330 /*
 331  * One of these is allocated for each thread accessing a mmaped GRU. A linked
 332  * list of these structure is hung off the struct gru_vma_data in the mm_struct.
 333  */
 334 struct gru_thread_state {
 335         struct list_head        ts_next;        /* list - head at vma-private */
 336         struct mutex            ts_ctxlock;     /* load/unload CTX lock */
 337         struct mm_struct        *ts_mm;         /* mm currently mapped to
 338                                                    context */
 339         struct vm_area_struct   *ts_vma;        /* vma of GRU context */
 340         struct gru_state        *ts_gru;        /* GRU where the context is
 341                                                    loaded */
 342         struct gru_mm_struct    *ts_gms;        /* asid & ioproc struct */
 343         unsigned char           ts_tlb_preload_count; /* TLB preload pages */
 344         unsigned long           ts_cbr_map;     /* map of allocated CBRs */
 345         unsigned long           ts_dsr_map;     /* map of allocated DATA
 346                                                    resources */
 347         unsigned long           ts_steal_jiffies;/* jiffies when context last
 348                                                     stolen */
 349         long                    ts_user_options;/* misc user option flags */
 350         pid_t                   ts_tgid_owner;  /* task that is using the
 351                                                    context - for migration */
 352         short                   ts_user_blade_id;/* user selected blade */
 353         char                    ts_user_chiplet_id;/* user selected chiplet */
 354         unsigned short          ts_sizeavail;   /* Pagesizes in use */
 355         int                     ts_tsid;        /* thread that owns the
 356                                                    structure */
 357         int                     ts_tlb_int_select;/* target cpu if interrupts
 358                                                      enabled */
 359         int                     ts_ctxnum;      /* context number where the
 360                                                    context is loaded */
 361         atomic_t                ts_refcnt;      /* reference count GTS */
 362         unsigned char           ts_dsr_au_count;/* Number of DSR resources
 363                                                    required for contest */
 364         unsigned char           ts_cbr_au_count;/* Number of CBR resources
 365                                                    required for contest */
 366         char                    ts_cch_req_slice;/* CCH packet slice */
 367         char                    ts_blade;       /* If >= 0, migrate context if
 368                                                    ref from different blade */
 369         char                    ts_force_cch_reload;
 370         char                    ts_cbr_idx[GRU_CBR_AU];/* CBR numbers of each
 371                                                           allocated CB */
 372         int                     ts_data_valid;  /* Indicates if ts_gdata has
 373                                                    valid data */
 374         struct gru_gseg_statistics ustats;      /* User statistics */
 375         unsigned long           ts_gdata[0];    /* save area for GRU data (CB,
 376                                                    DS, CBE) */
 377 };
 378
 379 /*
 380  * Threaded programs actually allocate an array of GSEGs when a context is
 381  * created. Each thread uses a separate GSEG. TSID is the index into the GSEG
 382  * array.
 383  */
 384 #define TSID(a, v)              (((a) - (v)->vm_start) / GRU_GSEG_PAGESIZE)
 385 #define UGRUADDR(gts)           ((gts)->ts_vma->vm_start +              \
 386                                         (gts)->ts_tsid * GRU_GSEG_PAGESIZE)
 387
 388 #define NULLCTX                 (-1)    /* if context not loaded into GRU */
 389
 390 /*-----------------------------------------------------------------------------
 391  *  GRU State Tables
 392  */
 393
 394 /*
 395  * One of these exists for each GRU chiplet.
 396  */
 397 struct gru_state {
 398         struct gru_blade_state  *gs_blade;              /* GRU state for entire
 399                                                            blade */
 400         unsigned long           gs_gru_base_paddr;      /* Physical address of
 401                                                            gru segments (64) */
 402         void                    *gs_gru_base_vaddr;     /* Virtual address of
 403                                                            gru segments (64) */
 404         unsigned short          gs_gid;                 /* unique GRU number */
 405         unsigned short          gs_blade_id;            /* blade of GRU */
 406         unsigned char           gs_chiplet_id;          /* blade chiplet of GRU */
 407         unsigned char           gs_tgh_local_shift;     /* used to pick TGH for
 408                                                            local flush */
 409         unsigned char           gs_tgh_first_remote;    /* starting TGH# for
 410                                                            remote flush */
 411         spinlock_t              gs_asid_lock;           /* lock used for
 412                                                            assigning asids */
 413         spinlock_t              gs_lock;                /* lock used for
 414                                                            assigning contexts */
 415
 416         /* -- the following are protected by the gs_asid_lock spinlock ---- */
 417         unsigned int            gs_asid;                /* Next availe ASID */
 418         unsigned int            gs_asid_limit;          /* Limit of available
 419                                                            ASIDs */
 420         unsigned int            gs_asid_gen;            /* asid generation.
 421                                                            Inc on wrap */
 422
 423         /* --- the following fields are protected by the gs_lock spinlock --- */
 424         unsigned long           gs_context_map;         /* bitmap to manage
 425                                                            contexts in use */
 426         unsigned long           gs_cbr_map;             /* bitmap to manage CB
 427                                                            resources */
 428         unsigned long           gs_dsr_map;             /* bitmap used to manage
 429                                                            DATA resources */
 430         unsigned int            gs_reserved_cbrs;       /* Number of kernel-
 431                                                            reserved cbrs */
 432         unsigned int            gs_reserved_dsr_bytes;  /* Bytes of kernel-
 433                                                            reserved dsrs */
 434         unsigned short          gs_active_contexts;     /* number of contexts
 435                                                            in use */
 436         struct gru_thread_state *gs_gts[GRU_NUM_CCH];   /* GTS currently using
 437                                                            the context */
 438         int                     gs_irq[GRU_NUM_TFM];    /* Interrupt irqs */
 439 };
 440
 441 /*
 442  * This structure contains the GRU state for all the GRUs on a blade.
 443  */
 444 struct gru_blade_state {
 445         void                    *kernel_cb;             /* First kernel
 446                                                            reserved cb */
 447         void                    *kernel_dsr;            /* First kernel
 448                                                            reserved DSR */
 449         struct rw_semaphore     bs_kgts_sema;           /* lock for kgts */
 450         struct gru_thread_state *bs_kgts;               /* GTS for kernel use */
 451
 452         /* ---- the following are used for managing kernel async GRU CBRs --- */
 453         int                     bs_async_dsr_bytes;     /* DSRs for async */
 454         int                     bs_async_cbrs;          /* CBRs AU for async */
 455         struct completion       *bs_async_wq;
 456
 457         /* ---- the following are protected by the bs_lock spinlock ---- */
 458         spinlock_t              bs_lock;                /* lock used for
 459                                                            stealing contexts */
 460         int                     bs_lru_ctxnum;          /* STEAL - last context
 461                                                            stolen */
 462         struct gru_state        *bs_lru_gru;            /* STEAL - last gru
 463                                                            stolen */
 464
 465         struct gru_state        bs_grus[GRU_CHIPLETS_PER_BLADE];
 466 };
 467
 468 /*-----------------------------------------------------------------------------
 469  * Address Primitives
 470  */
 471 #define get_tfm_for_cpu(g, c)                                           \
 472         ((struct gru_tlb_fault_map *)get_tfm((g)->gs_gru_base_vaddr, (c)))
 473 #define get_tfh_by_index(g, i)                                          \
 474         ((struct gru_tlb_fault_handle *)get_tfh((g)->gs_gru_base_vaddr, (i)))
 475 #define get_tgh_by_index(g, i)                                          \
 476         ((struct gru_tlb_global_handle *)get_tgh((g)->gs_gru_base_vaddr, (i)))
 477 #define get_cbe_by_index(g, i)                                          \
 478         ((struct gru_control_block_extended *)get_cbe((g)->gs_gru_base_vaddr,\
 479                         (i)))
 480
 481 /*-----------------------------------------------------------------------------
 482  * Useful Macros
 483  */
 484
 485 /* Given a blade# & chiplet#, get a pointer to the GRU */
 486 #define get_gru(b, c)           (&gru_base[b]->bs_grus[c])
 487
 488 /* Number of bytes to save/restore when unloading/loading GRU contexts */
 489 #define DSR_BYTES(dsr)          ((dsr) * GRU_DSR_AU_BYTES)
 490 #define CBR_BYTES(cbr)          ((cbr) * GRU_HANDLE_BYTES * GRU_CBR_AU_SIZE * 2)
 491
 492 /* Convert a user CB number to the actual CBRNUM */
 493 #define thread_cbr_number(gts, n) ((gts)->ts_cbr_idx[(n) / GRU_CBR_AU_SIZE] \
 494                                   * GRU_CBR_AU_SIZE + (n) % GRU_CBR_AU_SIZE)
 495
 496 /* Convert a gid to a pointer to the GRU */
 497 #define GID_TO_GRU(gid)                                                 \
 498         (gru_base[(gid) / GRU_CHIPLETS_PER_BLADE] ?                     \
 499                 (&gru_base[(gid) / GRU_CHIPLETS_PER_BLADE]->            \
 500                         bs_grus[(gid) % GRU_CHIPLETS_PER_BLADE]) :      \
 501          NULL)
 502
 503 /* Scan all active GRUs in a GRU bitmap */
 504 #define for_each_gru_in_bitmap(gid, map)                                \
 505         for_each_set_bit((gid), (map), GRU_MAX_GRUS)
 506
 507 /* Scan all active GRUs on a specific blade */
 508 #define for_each_gru_on_blade(gru, nid, i)                              \
 509         for ((gru) = gru_base[nid]->bs_grus, (i) = 0;                   \
 510                         (i) < GRU_CHIPLETS_PER_BLADE;                   \
 511                         (i)++, (gru)++)
 512
 513 /* Scan all GRUs */
 514 #define foreach_gid(gid)                                                \
 515         for ((gid) = 0; (gid) < gru_max_gids; (gid)++)
 516
 517 /* Scan all active GTSs on a gru. Note: must hold ss_lock to use this macro. */
 518 #define for_each_gts_on_gru(gts, gru, ctxnum)                           \
 519         for ((ctxnum) = 0; (ctxnum) < GRU_NUM_CCH; (ctxnum)++)          \
 520                 if (((gts) = (gru)->gs_gts[ctxnum]))
 521
 522 /* Scan each CBR whose bit is set in a TFM (or copy of) */
 523 #define for_each_cbr_in_tfm(i, map)                                     \
 524         for_each_set_bit((i), (map), GRU_NUM_CBE)
 525
 526 /* Scan each CBR in a CBR bitmap. Note: multiple CBRs in an allocation unit */
 527 #define for_each_cbr_in_allocation_map(i, map, k)                       \
 528         for_each_set_bit((k), (map), GRU_CBR_AU)                        \
 529                 for ((i) = (k)*GRU_CBR_AU_SIZE;                         \
 530                                 (i) < ((k) + 1) * GRU_CBR_AU_SIZE; (i)++)
 531
 532 /* Scan each DSR in a DSR bitmap. Note: multiple DSRs in an allocation unit */
 533 #define for_each_dsr_in_allocation_map(i, map, k)                       \
 534         for_each_set_bit((k), (const unsigned long *)(map), GRU_DSR_AU) \
 535                 for ((i) = (k) * GRU_DSR_AU_CL;                         \
 536                                 (i) < ((k) + 1) * GRU_DSR_AU_CL; (i)++)
 537
 538 #define gseg_physical_address(gru, ctxnum)                              \
 539                 ((gru)->gs_gru_base_paddr + ctxnum * GRU_GSEG_STRIDE)
 540 #define gseg_virtual_address(gru, ctxnum)                               \
 541                 ((gru)->gs_gru_base_vaddr + ctxnum * GRU_GSEG_STRIDE)
 542
 543 /*-----------------------------------------------------------------------------
 544  * Lock / Unlock GRU handles
 545  *      Use the "delresp" bit in the handle as a "lock" bit.
 546  */
 547
 548 /* Lock hierarchy checking enabled only in emulator */
 549
 550 /* 0 = lock failed, 1 = locked */
 551 static inline int __trylock_handle(void *h)
 552 {
 553         return !test_and_set_bit(1, h);
 554 }
 555
 556 static inline void __lock_handle(void *h)
 557 {
 558         while (test_and_set_bit(1, h))
 559                 cpu_relax();
 560 }
 561
 562 static inline void __unlock_handle(void *h)
 563 {
 564         clear_bit(1, h);
 565 }
 566
 567 static inline int trylock_cch_handle(struct gru_context_configuration_handle *cch)
 568 {
 569         return __trylock_handle(cch);
 570 }
 571
 572 static inline void lock_cch_handle(struct gru_context_configuration_handle *cch)
 573 {
 574         __lock_handle(cch);
 575 }
 576
 577 static inline void unlock_cch_handle(struct gru_context_configuration_handle
 578                                      *cch)
 579 {
 580         __unlock_handle(cch);
 581 }
 582
 583 static inline void lock_tgh_handle(struct gru_tlb_global_handle *tgh)
 584 {
 585         __lock_handle(tgh);
 586 }
 587
 588 static inline void unlock_tgh_handle(struct gru_tlb_global_handle *tgh)
 589 {
 590         __unlock_handle(tgh);
 591 }
 592
 593 static inline int is_kernel_context(struct gru_thread_state *gts)
 594 {
 595         return !gts->ts_mm;
 596 }
 597
 598 /*
 599  * The following are for Nehelem-EX. A more general scheme is needed for
 600  * future processors.
 601  */
 602 #define UV_MAX_INT_CORES                8
 603 #define uv_cpu_socket_number(p)         ((cpu_physical_id(p) >> 5) & 1)
 604 #define uv_cpu_ht_number(p)             (cpu_physical_id(p) & 1)
 605 #define uv_cpu_core_number(p)           (((cpu_physical_id(p) >> 2) & 4) |      \
 606                                         ((cpu_physical_id(p) >> 1) & 3))
 607 /*-----------------------------------------------------------------------------
 608  * Function prototypes & externs
 609  */
 610 struct gru_unload_context_req;
 611
 612 extern const struct vm_operations_struct gru_vm_ops;
 613 extern struct device *grudev;
 614
 615 extern struct gru_vma_data *gru_alloc_vma_data(struct vm_area_struct *vma,
 616                                 int tsid);
 617 extern struct gru_thread_state *gru_find_thread_state(struct vm_area_struct
 618                                 *vma, int tsid);
 619 extern struct gru_thread_state *gru_alloc_thread_state(struct vm_area_struct
 620                                 *vma, int tsid);
 621 extern struct gru_state *gru_assign_gru_context(struct gru_thread_state *gts);
 622 extern void gru_load_context(struct gru_thread_state *gts);
 623 extern void gru_steal_context(struct gru_thread_state *gts);
 624 extern void gru_unload_context(struct gru_thread_state *gts, int savestate);
 625 extern int gru_update_cch(struct gru_thread_state *gts);
 626 extern void gts_drop(struct gru_thread_state *gts);
 627 extern void gru_tgh_flush_init(struct gru_state *gru);
 628 extern int gru_kservices_init(void);
 629 extern void gru_kservices_exit(void);
 630 extern irqreturn_t gru0_intr(int irq, void *dev_id);
 631 extern irqreturn_t gru1_intr(int irq, void *dev_id);
 632 extern irqreturn_t gru_intr_mblade(int irq, void *dev_id);
 633 extern int gru_dump_chiplet_request(unsigned long arg);
 634 extern long gru_get_gseg_statistics(unsigned long arg);
 635 extern int gru_handle_user_call_os(unsigned long address);
 636 extern int gru_user_flush_tlb(unsigned long arg);
 637 extern int gru_user_unload_context(unsigned long arg);
 638 extern int gru_get_exception_detail(unsigned long arg);
 639 extern int gru_set_context_option(unsigned long address);
 640 extern void gru_check_context_placement(struct gru_thread_state *gts);
 641 extern int gru_cpu_fault_map_id(void);
 642 extern struct vm_area_struct *gru_find_vma(unsigned long vaddr);
 643 extern void gru_flush_all_tlb(struct gru_state *gru);
 644 extern int gru_proc_init(void);
 645 extern void gru_proc_exit(void);
 646
 647 extern struct gru_thread_state *gru_alloc_gts(struct vm_area_struct *vma,
 648                 int cbr_au_count, int dsr_au_count,
 649                 unsigned char tlb_preload_count, int options, int tsid);
 650 extern unsigned long gru_reserve_cb_resources(struct gru_state *gru,
 651                 int cbr_au_count, char *cbmap);
 652 extern unsigned long gru_reserve_ds_resources(struct gru_state *gru,
 653                 int dsr_au_count, char *dsmap);
 654 extern vm_fault_t gru_fault(struct vm_fault *vmf);
 655 extern struct gru_mm_struct *gru_register_mmu_notifier(void);
 656 extern void gru_drop_mmu_notifier(struct gru_mm_struct *gms);
 657
 658 extern int gru_ktest(unsigned long arg);
 659 extern void gru_flush_tlb_range(struct gru_mm_struct *gms, unsigned long start,
 660                                         unsigned long len);
 661
 662 extern unsigned long gru_options;
 663
 664 #endif /* __GRUTABLES_H__ */