drivers/misc/sgi-gru/grufile.c

   1 /*
   2  * SN Platform GRU Driver
   3  *
   4  *              FILE OPERATIONS & DRIVER INITIALIZATION
   5  *
   6  * This file supports the user system call for file open, close, mmap, etc.
   7  * This also incudes the driver initialization code.
   8  *
   9  *  Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
  10  *
  11  *  This program is free software; you can redistribute it and/or modify
  12  *  it under the terms of the GNU General Public License as published by
  13  *  the Free Software Foundation; either version 2 of the License, or
  14  *  (at your option) any later version.
  15  *
  16  *  This program is distributed in the hope that it will be useful,
  17  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *  GNU General Public License for more details.
  20  *
  21  *  You should have received a copy of the GNU General Public License
  22  *  along with this program; if not, write to the Free Software
  23  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  24  */
  25
  26 #include <linux/module.h>
  27 #include <linux/kernel.h>
  28 #include <linux/errno.h>
  29 #include <linux/slab.h>
  30 #include <linux/mm.h>
  31 #include <linux/io.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/spinlock.h>
  34 #include <linux/device.h>
  35 #include <linux/miscdevice.h>
  36 #include <linux/interrupt.h>
  37 #include <linux/proc_fs.h>
  38 #include <linux/uaccess.h>
  39 #include <asm/uv/uv.h>
  40 #include "gru.h"
  41 #include "grulib.h"
  42 #include "grutables.h"
  43
  44 #include <asm/uv/uv_hub.h>
  45 #include <asm/uv/uv_mmrs.h>
  46
  47 struct gru_blade_state *gru_base[GRU_MAX_BLADES] __read_mostly;
  48 unsigned long gru_start_paddr __read_mostly;
  49 unsigned long gru_end_paddr __read_mostly;
  50 struct gru_stats_s gru_stats;
  51
  52 /* Guaranteed user available resources on each node */
  53 static int max_user_cbrs, max_user_dsr_bytes;
  54
  55 static struct file_operations gru_fops;
  56 static struct miscdevice gru_miscdev;
  57
  58
  59 /*
  60  * gru_vma_close
  61  *
  62  * Called when unmapping a device mapping. Frees all gru resources
  63  * and tables belonging to the vma.
  64  */
  65 static void gru_vma_close(struct vm_area_struct *vma)
  66 {
  67         struct gru_vma_data *vdata;
  68         struct gru_thread_state *gts;
  69         struct list_head *entry, *next;
  70
  71         if (!vma->vm_private_data)
  72                 return;
  73
  74         vdata = vma->vm_private_data;
  75         vma->vm_private_data = NULL;
  76         gru_dbg(grudev, "vma %p, file %p, vdata %p\n", vma, vma->vm_file,
  77                                 vdata);
  78         list_for_each_safe(entry, next, &vdata->vd_head) {
  79                 gts =
  80                     list_entry(entry, struct gru_thread_state, ts_next);
  81                 list_del(&gts->ts_next);
  82                 mutex_lock(&gts->ts_ctxlock);
  83                 if (gts->ts_gru)
  84                         gru_unload_context(gts, 0);
  85                 mutex_unlock(&gts->ts_ctxlock);
  86                 gts_drop(gts);
  87         }
  88         kfree(vdata);
  89         STAT(vdata_free);
  90 }
  91
  92 /*
  93  * gru_file_mmap
  94  *
  95  * Called when mmaping the device.  Initializes the vma with a fault handler
  96  * and private data structure necessary to allocate, track, and free the
  97  * underlying pages.
  98  */
  99 static int gru_file_mmap(struct file *file, struct vm_area_struct *vma)
 100 {
 101         if ((vma->vm_flags & (VM_SHARED | VM_WRITE)) != (VM_SHARED | VM_WRITE))
 102                 return -EPERM;
 103
 104         if (vma->vm_start & (GRU_GSEG_PAGESIZE - 1) ||
 105                                 vma->vm_end & (GRU_GSEG_PAGESIZE - 1))
 106                 return -EINVAL;
 107
 108         vma->vm_flags |=
 109             (VM_IO | VM_DONTCOPY | VM_LOCKED | VM_DONTEXPAND | VM_PFNMAP |
 110                         VM_RESERVED);
 111         vma->vm_page_prot = PAGE_SHARED;
 112         vma->vm_ops = &gru_vm_ops;
 113
 114         vma->vm_private_data = gru_alloc_vma_data(vma, 0);
 115         if (!vma->vm_private_data)
 116                 return -ENOMEM;
 117
 118         gru_dbg(grudev, "file %p, vaddr 0x%lx, vma %p, vdata %p\n",
 119                 file, vma->vm_start, vma, vma->vm_private_data);
 120         return 0;
 121 }
 122
 123 /*
 124  * Create a new GRU context
 125  */
 126 static int gru_create_new_context(unsigned long arg)
 127 {
 128         struct gru_create_context_req req;
 129         struct vm_area_struct *vma;
 130         struct gru_vma_data *vdata;
 131         int ret = -EINVAL;
 132
 133
 134         if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
 135                 return -EFAULT;
 136
 137         if (req.data_segment_bytes == 0 ||
 138                                 req.data_segment_bytes > max_user_dsr_bytes)
 139                 return -EINVAL;
 140         if (!req.control_blocks || !req.maximum_thread_count ||
 141                                 req.control_blocks > max_user_cbrs)
 142                 return -EINVAL;
 143
 144         if (!(req.options & GRU_OPT_MISS_MASK))
 145                 req.options |= GRU_OPT_MISS_FMM_INTR;
 146
 147         down_write(&current->mm->mmap_sem);
 148         vma = gru_find_vma(req.gseg);
 149         if (vma) {
 150                 vdata = vma->vm_private_data;
 151                 vdata->vd_user_options = req.options;
 152                 vdata->vd_dsr_au_count =
 153                     GRU_DS_BYTES_TO_AU(req.data_segment_bytes);
 154                 vdata->vd_cbr_au_count = GRU_CB_COUNT_TO_AU(req.control_blocks);
 155                 ret = 0;
 156         }
 157         up_write(&current->mm->mmap_sem);
 158
 159         return ret;
 160 }
 161
 162 /*
 163  * Get GRU configuration info (temp - for emulator testing)
 164  */
 165 static long gru_get_config_info(unsigned long arg)
 166 {
 167         struct gru_config_info info;
 168         int nodesperblade;
 169
 170         if (num_online_nodes() > 1 &&
 171                         (uv_node_to_blade_id(1) == uv_node_to_blade_id(0)))
 172                 nodesperblade = 2;
 173         else
 174                 nodesperblade = 1;
 175         info.cpus = num_online_cpus();
 176         info.nodes = num_online_nodes();
 177         info.blades = info.nodes / nodesperblade;
 178         info.chiplets = GRU_CHIPLETS_PER_BLADE * info.blades;
 179
 180         if (copy_to_user((void __user *)arg, &info, sizeof(info)))
 181                 return -EFAULT;
 182         return 0;
 183 }
 184
 185 /*
 186  * Get GRU chiplet status
 187  */
 188 static long gru_get_chiplet_status(unsigned long arg)
 189 {
 190         struct gru_state *gru;
 191         struct gru_chiplet_info info;
 192
 193         if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
 194                 return -EFAULT;
 195
 196         if (info.node == -1)
 197                 info.node = numa_node_id();
 198         if (info.node >= num_possible_nodes() ||
 199                         info.chiplet >= GRU_CHIPLETS_PER_HUB ||
 200                         info.node < 0 || info.chiplet < 0)
 201                 return -EINVAL;
 202
 203         info.blade = uv_node_to_blade_id(info.node);
 204         gru = get_gru(info.blade, info.chiplet);
 205
 206         info.total_dsr_bytes = GRU_NUM_DSR_BYTES;
 207         info.total_cbr = GRU_NUM_CB;
 208         info.total_user_dsr_bytes = GRU_NUM_DSR_BYTES -
 209                 gru->gs_reserved_dsr_bytes;
 210         info.total_user_cbr = GRU_NUM_CB - gru->gs_reserved_cbrs;
 211         info.free_user_dsr_bytes = hweight64(gru->gs_dsr_map) *
 212                         GRU_DSR_AU_BYTES;
 213         info.free_user_cbr = hweight64(gru->gs_cbr_map) * GRU_CBR_AU_SIZE;
 214
 215         if (copy_to_user((void __user *)arg, &info, sizeof(info)))
 216                 return -EFAULT;
 217         return 0;
 218 }
 219
 220 /*
 221  * gru_file_unlocked_ioctl
 222  *
 223  * Called to update file attributes via IOCTL calls.
 224  */
 225 static long gru_file_unlocked_ioctl(struct file *file, unsigned int req,
 226                                     unsigned long arg)
 227 {
 228         int err = -EBADRQC;
 229
 230         gru_dbg(grudev, "file %p\n", file);
 231
 232         switch (req) {
 233         case GRU_CREATE_CONTEXT:
 234                 err = gru_create_new_context(arg);
 235                 break;
 236         case GRU_SET_TASK_SLICE:
 237                 err = gru_set_task_slice(arg);
 238                 break;
 239         case GRU_USER_GET_EXCEPTION_DETAIL:
 240                 err = gru_get_exception_detail(arg);
 241                 break;
 242         case GRU_USER_UNLOAD_CONTEXT:
 243                 err = gru_user_unload_context(arg);
 244                 break;
 245         case GRU_GET_CHIPLET_STATUS:
 246                 err = gru_get_chiplet_status(arg);
 247                 break;
 248         case GRU_USER_FLUSH_TLB:
 249                 err = gru_user_flush_tlb(arg);
 250                 break;
 251         case GRU_USER_CALL_OS:
 252                 err = gru_handle_user_call_os(arg);
 253                 break;
 254         case GRU_GET_CONFIG_INFO:
 255                 err = gru_get_config_info(arg);
 256                 break;
 257         }
 258         return err;
 259 }
 260
 261 /*
 262  * Called at init time to build tables for all GRUs that are present in the
 263  * system.
 264  */
 265 static void gru_init_chiplet(struct gru_state *gru, unsigned long paddr,
 266                              void *vaddr, int nid, int bid, int grunum)
 267 {
 268         spin_lock_init(&gru->gs_lock);
 269         spin_lock_init(&gru->gs_asid_lock);
 270         gru->gs_gru_base_paddr = paddr;
 271         gru->gs_gru_base_vaddr = vaddr;
 272         gru->gs_gid = bid * GRU_CHIPLETS_PER_BLADE + grunum;
 273         gru->gs_blade = gru_base[bid];
 274         gru->gs_blade_id = bid;
 275         gru->gs_cbr_map = (GRU_CBR_AU == 64) ? ~0 : (1UL << GRU_CBR_AU) - 1;
 276         gru->gs_dsr_map = (1UL << GRU_DSR_AU) - 1;
 277         gru_tgh_flush_init(gru);
 278         gru_dbg(grudev, "bid %d, nid %d, gru %x, vaddr %p (0x%lx)\n",
 279                 bid, nid, gru->gs_gid, gru->gs_gru_base_vaddr,
 280                 gru->gs_gru_base_paddr);
 281         gru_kservices_init(gru);
 282 }
 283
 284 static int gru_init_tables(unsigned long gru_base_paddr, void *gru_base_vaddr)
 285 {
 286         int pnode, nid, bid, chip;
 287         int cbrs, dsrbytes, n;
 288         int order = get_order(sizeof(struct gru_blade_state));
 289         struct page *page;
 290         struct gru_state *gru;
 291         unsigned long paddr;
 292         void *vaddr;
 293
 294         max_user_cbrs = GRU_NUM_CB;
 295         max_user_dsr_bytes = GRU_NUM_DSR_BYTES;
 296         for_each_online_node(nid) {
 297                 bid = uv_node_to_blade_id(nid);
 298                 pnode = uv_node_to_pnode(nid);
 299                 if (bid < 0 || gru_base[bid])
 300                         continue;
 301                 page = alloc_pages_node(nid, GFP_KERNEL, order);
 302                 if (!page)
 303                         goto fail;
 304                 gru_base[bid] = page_address(page);
 305                 memset(gru_base[bid], 0, sizeof(struct gru_blade_state));
 306                 gru_base[bid]->bs_lru_gru = &gru_base[bid]->bs_grus[0];
 307                 spin_lock_init(&gru_base[bid]->bs_lock);
 308
 309                 dsrbytes = 0;
 310                 cbrs = 0;
 311                 for (gru = gru_base[bid]->bs_grus, chip = 0;
 312                                 chip < GRU_CHIPLETS_PER_BLADE;
 313                                 chip++, gru++) {
 314                         paddr = gru_chiplet_paddr(gru_base_paddr, pnode, chip);
 315                         vaddr = gru_chiplet_vaddr(gru_base_vaddr, pnode, chip);
 316                         gru_init_chiplet(gru, paddr, vaddr, nid, bid, chip);
 317                         n = hweight64(gru->gs_cbr_map) * GRU_CBR_AU_SIZE;
 318                         cbrs = max(cbrs, n);
 319                         n = hweight64(gru->gs_dsr_map) * GRU_DSR_AU_BYTES;
 320                         dsrbytes = max(dsrbytes, n);
 321                 }
 322                 max_user_cbrs = min(max_user_cbrs, cbrs);
 323                 max_user_dsr_bytes = min(max_user_dsr_bytes, dsrbytes);
 324         }
 325
 326         return 0;
 327
 328 fail:
 329         for (nid--; nid >= 0; nid--)
 330                 free_pages((unsigned long)gru_base[nid], order);
 331         return -ENOMEM;
 332 }
 333
 334 #ifdef CONFIG_IA64
 335
 336 static int get_base_irq(void)
 337 {
 338         return IRQ_GRU;
 339 }
 340
 341 #elif defined CONFIG_X86_64
 342
 343 static void noop(unsigned int irq)
 344 {
 345 }
 346
 347 static struct irq_chip gru_chip = {
 348         .name           = "gru",
 349         .mask           = noop,
 350         .unmask         = noop,
 351         .ack            = noop,
 352 };
 353
 354 static int get_base_irq(void)
 355 {
 356         set_irq_chip(IRQ_GRU, &gru_chip);
 357         set_irq_chip(IRQ_GRU + 1, &gru_chip);
 358         return IRQ_GRU;
 359 }
 360 #endif
 361
 362 /*
 363  * gru_init
 364  *
 365  * Called at boot or module load time to initialize the GRUs.
 366  */
 367 static int __init gru_init(void)
 368 {
 369         int ret, irq, chip;
 370         char id[10];
 371         void *gru_start_vaddr;
 372
 373         if (!is_uv_system())
 374                 return -ENODEV;
 375
 376 #if defined CONFIG_IA64
 377         gru_start_paddr = 0xd000000000UL; /* ZZZZZZZZZZZZZZZZZZZ fixme */
 378 #else
 379         gru_start_paddr = uv_read_local_mmr(UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR) &
 380                                 0x7fffffffffffUL;
 381 #endif
 382         gru_start_vaddr = __va(gru_start_paddr);
 383         gru_end_paddr = gru_start_paddr + GRU_MAX_BLADES * GRU_SIZE;
 384         printk(KERN_INFO "GRU space: 0x%lx - 0x%lx\n",
 385                gru_start_paddr, gru_end_paddr);
 386         irq = get_base_irq();
 387         for (chip = 0; chip < GRU_CHIPLETS_PER_BLADE; chip++) {
 388                 ret = request_irq(irq + chip, gru_intr, 0, id, NULL);
 389                 /* TODO: fix irq handling on x86. For now ignore failure because
 390                  * interrupts are not required & not yet fully supported */
 391                 if (ret) {
 392                         printk(KERN_WARNING
 393                                "!!!WARNING: GRU ignoring request failure!!!\n");
 394                         ret = 0;
 395                 }
 396                 if (ret) {
 397                         printk(KERN_ERR "%s: request_irq failed\n",
 398                                GRU_DRIVER_ID_STR);
 399                         goto exit1;
 400                 }
 401         }
 402
 403         ret = misc_register(&gru_miscdev);
 404         if (ret) {
 405                 printk(KERN_ERR "%s: misc_register failed\n",
 406                        GRU_DRIVER_ID_STR);
 407                 goto exit1;
 408         }
 409
 410         ret = gru_proc_init();
 411         if (ret) {
 412                 printk(KERN_ERR "%s: proc init failed\n", GRU_DRIVER_ID_STR);
 413                 goto exit2;
 414         }
 415
 416         ret = gru_init_tables(gru_start_paddr, gru_start_vaddr);
 417         if (ret) {
 418                 printk(KERN_ERR "%s: init tables failed\n", GRU_DRIVER_ID_STR);
 419                 goto exit3;
 420         }
 421
 422         printk(KERN_INFO "%s: v%s\n", GRU_DRIVER_ID_STR,
 423                GRU_DRIVER_VERSION_STR);
 424         return 0;
 425
 426 exit3:
 427         gru_proc_exit();
 428 exit2:
 429         misc_deregister(&gru_miscdev);
 430 exit1:
 431         for (--chip; chip >= 0; chip--)
 432                 free_irq(irq + chip, NULL);
 433         return ret;
 434
 435 }
 436
 437 static void __exit gru_exit(void)
 438 {
 439         int i, bid;
 440         int order = get_order(sizeof(struct gru_state) *
 441                               GRU_CHIPLETS_PER_BLADE);
 442
 443         if (!is_uv_system())
 444                 return;
 445
 446         for (i = 0; i < GRU_CHIPLETS_PER_BLADE; i++)
 447                 free_irq(IRQ_GRU + i, NULL);
 448
 449         for (bid = 0; bid < GRU_MAX_BLADES; bid++)
 450                 free_pages((unsigned long)gru_base[bid], order);
 451
 452         misc_deregister(&gru_miscdev);
 453         gru_proc_exit();
 454 }
 455
 456 static struct file_operations gru_fops = {
 457         .owner          = THIS_MODULE,
 458         .unlocked_ioctl = gru_file_unlocked_ioctl,
 459         .mmap           = gru_file_mmap,
 460 };
 461
 462 static struct miscdevice gru_miscdev = {
 463         .minor          = MISC_DYNAMIC_MINOR,
 464         .name           = "gru",
 465         .fops           = &gru_fops,
 466 };
 467
 468 struct vm_operations_struct gru_vm_ops = {
 469         .close          = gru_vma_close,
 470         .fault          = gru_fault,
 471 };
 472
 473 #ifndef MODULE
 474 fs_initcall(gru_init);
 475 #else
 476 module_init(gru_init);
 477 #endif
 478 module_exit(gru_exit);
 479
 480 module_param(gru_options, ulong, 0644);
 481 MODULE_PARM_DESC(gru_options, "Various debug options");
 482
 483 MODULE_AUTHOR("Silicon Graphics, Inc.");
 484 MODULE_LICENSE("GPL");
 485 MODULE_DESCRIPTION(GRU_DRIVER_ID_STR GRU_DRIVER_VERSION_STR);
 486 MODULE_VERSION(GRU_DRIVER_VERSION_STR);
 487