arch/ia64/kernel/fsys.S

   1 /*
   2  * This file contains the light-weight system call handlers (fsyscall-handlers).
   3  *
   4  * Copyright (C) 2003 Hewlett-Packard Co
   5  *      David Mosberger-Tang <davidm@hpl.hp.com>
   6  *
   7  * 25-Sep-03 davidm     Implement fsys_rt_sigprocmask().
   8  * 18-Feb-03 louisk     Implement fsys_gettimeofday().
   9  * 28-Feb-03 davidm     Fixed several bugs in fsys_gettimeofday().  Tuned it some more,
  10  *                      probably broke it along the way... ;-)
  11  * 13-Jul-04 clameter   Implement fsys_clock_gettime and revise fsys_gettimeofday to make
  12  *                      it capable of using memory based clocks without falling back to C code.
  13  * 08-Feb-07 Fenghua Yu Implement fsys_getcpu.
  14  *
  15  */
  16
  17 #include <asm/asmmacro.h>
  18 #include <asm/errno.h>
  19 #include <asm/asm-offsets.h>
  20 #include <asm/percpu.h>
  21 #include <asm/thread_info.h>
  22 #include <asm/sal.h>
  23 #include <asm/signal.h>
  24 #include <asm/system.h>
  25 #include <asm/unistd.h>
  26
  27 #include "entry.h"
  28
  29 /*
  30  * See Documentation/ia64/fsys.txt for details on fsyscalls.
  31  *
  32  * On entry to an fsyscall handler:
  33  *   r10        = 0 (i.e., defaults to "successful syscall return")
  34  *   r11        = saved ar.pfs (a user-level value)
  35  *   r15        = system call number
  36  *   r16        = "current" task pointer (in normal kernel-mode, this is in r13)
  37  *   r32-r39    = system call arguments
  38  *   b6         = return address (a user-level value)
  39  *   ar.pfs     = previous frame-state (a user-level value)
  40  *   PSR.be     = cleared to zero (i.e., little-endian byte order is in effect)
  41  *   all other registers may contain values passed in from user-mode
  42  *
  43  * On return from an fsyscall handler:
  44  *   r11        = saved ar.pfs (as passed into the fsyscall handler)
  45  *   r15        = system call number (as passed into the fsyscall handler)
  46  *   r32-r39    = system call arguments (as passed into the fsyscall handler)
  47  *   b6         = return address (as passed into the fsyscall handler)
  48  *   ar.pfs     = previous frame-state (as passed into the fsyscall handler)
  49  */
  50
  51 ENTRY(fsys_ni_syscall)
  52         .prologue
  53         .altrp b6
  54         .body
  55         mov r8=ENOSYS
  56         mov r10=-1
  57         FSYS_RETURN
  58 END(fsys_ni_syscall)
  59
  60 ENTRY(fsys_getpid)
  61         .prologue
  62         .altrp b6
  63         .body
  64         add r17=IA64_TASK_GROUP_LEADER_OFFSET,r16
  65         ;;
  66         ld8 r17=[r17]                           // r17 = current->group_leader
  67         add r9=TI_FLAGS+IA64_TASK_SIZE,r16
  68         ;;
  69         ld4 r9=[r9]
  70         add r17=IA64_TASK_TGIDLINK_OFFSET,r17
  71         ;;
  72         and r9=TIF_ALLWORK_MASK,r9
  73         ld8 r17=[r17]                           // r17 = current->group_leader->pids[PIDTYPE_PID].pid
  74         ;;
  75         add r8=IA64_PID_LEVEL_OFFSET,r17
  76         ;;
  77         ld4 r8=[r8]                             // r8 = pid->level
  78         add r17=IA64_PID_UPID_OFFSET,r17        // r17 = &pid->numbers[0]
  79         ;;
  80         shl r8=r8,IA64_UPID_SHIFT
  81         ;;
  82         add r17=r17,r8                          // r17 = &pid->numbers[pid->level]
  83         ;;
  84         ld4 r8=[r17]                            // r8 = pid->numbers[pid->level].nr
  85         ;;
  86         mov r17=0
  87         ;;
  88         cmp.ne p8,p0=0,r9
  89 (p8)    br.spnt.many fsys_fallback_syscall
  90         FSYS_RETURN
  91 END(fsys_getpid)
  92
  93 ENTRY(fsys_getppid)
  94         .prologue
  95         .altrp b6
  96         .body
  97         add r17=IA64_TASK_GROUP_LEADER_OFFSET,r16
  98         ;;
  99         ld8 r17=[r17]                           // r17 = current->group_leader
 100         add r9=TI_FLAGS+IA64_TASK_SIZE,r16
 101         ;;
 102
 103         ld4 r9=[r9]
 104         add r17=IA64_TASK_REAL_PARENT_OFFSET,r17 // r17 = &current->group_leader->real_parent
 105         ;;
 106         and r9=TIF_ALLWORK_MASK,r9
 107
 108 1:      ld8 r18=[r17]                           // r18 = current->group_leader->real_parent
 109         ;;
 110         cmp.ne p8,p0=0,r9
 111         add r8=IA64_TASK_TGID_OFFSET,r18        // r8 = &current->group_leader->real_parent->tgid
 112         ;;
 113
 114         /*
 115          * The .acq is needed to ensure that the read of tgid has returned its data before
 116          * we re-check "real_parent".
 117          */
 118         ld4.acq r8=[r8]                         // r8 = current->group_leader->real_parent->tgid
 119 #ifdef CONFIG_SMP
 120         /*
 121          * Re-read current->group_leader->real_parent.
 122          */
 123         ld8 r19=[r17]                           // r19 = current->group_leader->real_parent
 124 (p8)    br.spnt.many fsys_fallback_syscall
 125         ;;
 126         cmp.ne p6,p0=r18,r19                    // did real_parent change?
 127         mov r19=0                       // i must not leak kernel bits...
 128 (p6)    br.cond.spnt.few 1b                     // yes -> redo the read of tgid and the check
 129         ;;
 130         mov r17=0                       // i must not leak kernel bits...
 131         mov r18=0                       // i must not leak kernel bits...
 132 #else
 133         mov r17=0                       // i must not leak kernel bits...
 134         mov r18=0                       // i must not leak kernel bits...
 135         mov r19=0                       // i must not leak kernel bits...
 136 #endif
 137         FSYS_RETURN
 138 END(fsys_getppid)
 139
 140 ENTRY(fsys_set_tid_address)
 141         .prologue
 142         .altrp b6
 143         .body
 144         add r9=TI_FLAGS+IA64_TASK_SIZE,r16
 145         add r17=IA64_TASK_TGIDLINK_OFFSET,r16
 146         ;;
 147         ld4 r9=[r9]
 148         tnat.z p6,p7=r32                // check argument register for being NaT
 149         ld8 r17=[r17]                           // r17 = current->pids[PIDTYPE_PID].pid
 150         ;;
 151         and r9=TIF_ALLWORK_MASK,r9
 152         add r8=IA64_PID_LEVEL_OFFSET,r17
 153         add r18=IA64_TASK_CLEAR_CHILD_TID_OFFSET,r16
 154         ;;
 155         ld4 r8=[r8]                             // r8 = pid->level
 156         add r17=IA64_PID_UPID_OFFSET,r17        // r17 = &pid->numbers[0]
 157         ;;
 158         shl r8=r8,IA64_UPID_SHIFT
 159         ;;
 160         add r17=r17,r8                          // r17 = &pid->numbers[pid->level]
 161         ;;
 162         ld4 r8=[r17]                            // r8 = pid->numbers[pid->level].nr
 163         ;;
 164         cmp.ne p8,p0=0,r9
 165         mov r17=-1
 166         ;;
 167 (p6)    st8 [r18]=r32
 168 (p7)    st8 [r18]=r17
 169 (p8)    br.spnt.many fsys_fallback_syscall
 170         ;;
 171         mov r17=0                       // i must not leak kernel bits...
 172         mov r18=0                       // i must not leak kernel bits...
 173         FSYS_RETURN
 174 END(fsys_set_tid_address)
 175
 176 #if IA64_GTOD_LOCK_OFFSET !=0
 177 #error fsys_gettimeofday incompatible with changes to struct fsyscall_gtod_data_t
 178 #endif
 179 #if IA64_ITC_JITTER_OFFSET !=0
 180 #error fsys_gettimeofday incompatible with changes to struct itc_jitter_data_t
 181 #endif
 182 #define CLOCK_REALTIME 0
 183 #define CLOCK_MONOTONIC 1
 184 #define CLOCK_DIVIDE_BY_1000 0x4000
 185 #define CLOCK_ADD_MONOTONIC 0x8000
 186
 187 ENTRY(fsys_gettimeofday)
 188         .prologue
 189         .altrp b6
 190         .body
 191         mov r31 = r32
 192         tnat.nz p6,p0 = r33             // guard against NaT argument
 193 (p6)    br.cond.spnt.few .fail_einval
 194         mov r30 = CLOCK_DIVIDE_BY_1000
 195         ;;
 196 .gettime:
 197         // Register map
 198         // Incoming r31 = pointer to address where to place result
 199         //          r30 = flags determining how time is processed
 200         // r2,r3 = temp r4-r7 preserved
 201         // r8 = result nanoseconds
 202         // r9 = result seconds
 203         // r10 = temporary storage for clock difference
 204         // r11 = preserved: saved ar.pfs
 205         // r12 = preserved: memory stack
 206         // r13 = preserved: thread pointer
 207         // r14 = address of mask / mask value
 208         // r15 = preserved: system call number
 209         // r16 = preserved: current task pointer
 210         // r17 = (not used)
 211         // r18 = (not used)
 212         // r19 = address of itc_lastcycle
 213         // r20 = struct fsyscall_gtod_data (= address of gtod_lock.sequence)
 214         // r21 = address of mmio_ptr
 215         // r22 = address of wall_time or monotonic_time
 216         // r23 = address of shift / value
 217         // r24 = address mult factor / cycle_last value
 218         // r25 = itc_lastcycle value
 219         // r26 = address clocksource cycle_last
 220         // r27 = (not used)
 221         // r28 = sequence number at the beginning of critcal section
 222         // r29 = address of itc_jitter
 223         // r30 = time processing flags / memory address
 224         // r31 = pointer to result
 225         // Predicates
 226         // p6,p7 short term use
 227         // p8 = timesource ar.itc
 228         // p9 = timesource mmio64
 229         // p10 = timesource mmio32 - not used
 230         // p11 = timesource not to be handled by asm code
 231         // p12 = memory time source ( = p9 | p10) - not used
 232         // p13 = do cmpxchg with itc_lastcycle
 233         // p14 = Divide by 1000
 234         // p15 = Add monotonic
 235         //
 236         // Note that instructions are optimized for McKinley. McKinley can
 237         // process two bundles simultaneously and therefore we continuously
 238         // try to feed the CPU two bundles and then a stop.
 239         //
 240         // Additional note that code has changed a lot. Optimization is TBD.
 241         // Comments begin with "?" are maybe outdated.
 242         tnat.nz p6,p0 = r31     // ? branch deferred to fit later bundle
 243         mov pr = r30,0xc000     // Set predicates according to function
 244         add r2 = TI_FLAGS+IA64_TASK_SIZE,r16
 245         movl r20 = fsyscall_gtod_data // load fsyscall gettimeofday data address
 246         ;;
 247         movl r29 = itc_jitter_data      // itc_jitter
 248         add r22 = IA64_GTOD_WALL_TIME_OFFSET,r20        // wall_time
 249         ld4 r2 = [r2]           // process work pending flags
 250         ;;
 251 (p15)   add r22 = IA64_GTOD_MONO_TIME_OFFSET,r20        // monotonic_time
 252         add r21 = IA64_CLKSRC_MMIO_OFFSET,r20
 253         add r19 = IA64_ITC_LASTCYCLE_OFFSET,r29
 254         and r2 = TIF_ALLWORK_MASK,r2
 255 (p6)    br.cond.spnt.few .fail_einval   // ? deferred branch
 256         ;;
 257         add r26 = IA64_CLKSRC_CYCLE_LAST_OFFSET,r20 // clksrc_cycle_last
 258         cmp.ne p6, p0 = 0, r2   // Fallback if work is scheduled
 259 (p6)    br.cond.spnt.many fsys_fallback_syscall
 260         ;;
 261         // Begin critical section
 262 .time_redo:
 263         ld4.acq r28 = [r20]     // gtod_lock.sequence, Must take first
 264         ;;
 265         and r28 = ~1,r28        // And make sequence even to force retry if odd
 266         ;;
 267         ld8 r30 = [r21]         // clocksource->mmio_ptr
 268         add r24 = IA64_CLKSRC_MULT_OFFSET,r20
 269         ld4 r2 = [r29]          // itc_jitter value
 270         add r23 = IA64_CLKSRC_SHIFT_OFFSET,r20
 271         add r14 = IA64_CLKSRC_MASK_OFFSET,r20
 272         ;;
 273         ld4 r3 = [r24]          // clocksource mult value
 274         ld8 r14 = [r14]         // clocksource mask value
 275         cmp.eq p8,p9 = 0,r30    // use cpu timer if no mmio_ptr
 276         ;;
 277         setf.sig f7 = r3        // Setup for mult scaling of counter
 278 (p8)    cmp.ne p13,p0 = r2,r0   // need itc_jitter compensation, set p13
 279         ld4 r23 = [r23]         // clocksource shift value
 280         ld8 r24 = [r26]         // get clksrc_cycle_last value
 281 (p9)    cmp.eq p13,p0 = 0,r30   // if mmio_ptr, clear p13 jitter control
 282         ;;
 283         .pred.rel.mutex p8,p9
 284 (p8)    mov r2 = ar.itc         // CPU_TIMER. 36 clocks latency!!!
 285 (p9)    ld8 r2 = [r30]          // MMIO_TIMER. Could also have latency issues..
 286 (p13)   ld8 r25 = [r19]         // get itc_lastcycle value
 287         ;;              // ? could be removed by moving the last add upward
 288         ld8 r9 = [r22],IA64_TIMESPEC_TV_NSEC_OFFSET     // tv_sec
 289         ;;
 290         ld8 r8 = [r22],-IA64_TIMESPEC_TV_NSEC_OFFSET    // tv_nsec
 291 (p13)   sub r3 = r25,r2         // Diff needed before comparison (thanks davidm)
 292         ;;
 293 (p13)   cmp.gt.unc p6,p7 = r3,r0 // check if it is less than last. p6,p7 cleared
 294         sub r10 = r2,r24        // current_cycle - last_cycle
 295         ;;
 296 (p6)    sub r10 = r25,r24       // time we got was less than last_cycle
 297 (p7)    mov ar.ccv = r25        // more than last_cycle. Prep for cmpxchg
 298         ;;
 299 (p7)    cmpxchg8.rel r3 = [r19],r2,ar.ccv
 300         ;;
 301 (p7)    cmp.ne p7,p0 = r25,r3   // if cmpxchg not successful
 302         ;;
 303 (p7)    sub r10 = r3,r24        // then use new last_cycle instead
 304         ;;
 305         and r10 = r10,r14       // Apply mask
 306         ;;
 307         setf.sig f8 = r10
 308         nop.i 123
 309         ;;
 310         // fault check takes 5 cycles and we have spare time
 311 EX(.fail_efault, probe.w.fault r31, 3)
 312         xmpy.l f8 = f8,f7       // nsec_per_cyc*(counter-last_counter)
 313         ;;
 314         // ? simulate tbit.nz.or p7,p0 = r28,0
 315         getf.sig r2 = f8
 316         mf
 317         ;;
 318         ld4 r10 = [r20]         // gtod_lock.sequence
 319         shr.u r2 = r2,r23       // shift by factor
 320         ;;              // ? overloaded 3 bundles!
 321         add r8 = r8,r2          // Add xtime.nsecs
 322         cmp4.ne p7,p0 = r28,r10
 323 (p7)    br.cond.dpnt.few .time_redo     // sequence number changed, redo
 324         // End critical section.
 325         // Now r8=tv->tv_nsec and r9=tv->tv_sec
 326         mov r10 = r0
 327         movl r2 = 1000000000
 328         add r23 = IA64_TIMESPEC_TV_NSEC_OFFSET, r31
 329 (p14)   movl r3 = 2361183241434822607   // Prep for / 1000 hack
 330         ;;
 331 .time_normalize:
 332         mov r21 = r8
 333         cmp.ge p6,p0 = r8,r2
 334 (p14)   shr.u r20 = r8, 3 // We can repeat this if necessary just wasting time
 335         ;;
 336 (p14)   setf.sig f8 = r20
 337 (p6)    sub r8 = r8,r2
 338 (p6)    add r9 = 1,r9           // two nops before the branch.
 339 (p14)   setf.sig f7 = r3        // Chances for repeats are 1 in 10000 for gettod
 340 (p6)    br.cond.dpnt.few .time_normalize
 341         ;;
 342         // Divided by 8 though shift. Now divide by 125
 343         // The compiler was able to do that with a multiply
 344         // and a shift and we do the same
 345 EX(.fail_efault, probe.w.fault r23, 3)  // This also costs 5 cycles
 346 (p14)   xmpy.hu f8 = f8, f7             // xmpy has 5 cycles latency so use it
 347         ;;
 348         mov r8 = r0
 349 (p14)   getf.sig r2 = f8
 350         ;;
 351 (p14)   shr.u r21 = r2, 4
 352         ;;
 353 EX(.fail_efault, st8 [r31] = r9)
 354 EX(.fail_efault, st8 [r23] = r21)
 355         FSYS_RETURN
 356 .fail_einval:
 357         mov r8 = EINVAL
 358         mov r10 = -1
 359         FSYS_RETURN
 360 .fail_efault:
 361         mov r8 = EFAULT
 362         mov r10 = -1
 363         FSYS_RETURN
 364 END(fsys_gettimeofday)
 365
 366 ENTRY(fsys_clock_gettime)
 367         .prologue
 368         .altrp b6
 369         .body
 370         cmp4.ltu p6, p0 = CLOCK_MONOTONIC, r32
 371         // Fallback if this is not CLOCK_REALTIME or CLOCK_MONOTONIC
 372 (p6)    br.spnt.few fsys_fallback_syscall
 373         mov r31 = r33
 374         shl r30 = r32,15
 375         br.many .gettime
 376 END(fsys_clock_gettime)
 377
 378 /*
 379  * long fsys_rt_sigprocmask (int how, sigset_t *set, sigset_t *oset, size_t sigsetsize).
 380  */
 381 #if _NSIG_WORDS != 1
 382 # error Sorry, fsys_rt_sigprocmask() needs to be updated for _NSIG_WORDS != 1.
 383 #endif
 384 ENTRY(fsys_rt_sigprocmask)
 385         .prologue
 386         .altrp b6
 387         .body
 388
 389         add r2=IA64_TASK_BLOCKED_OFFSET,r16
 390         add r9=TI_FLAGS+IA64_TASK_SIZE,r16
 391         cmp4.ltu p6,p0=SIG_SETMASK,r32
 392
 393         cmp.ne p15,p0=r0,r34                    // oset != NULL?
 394         tnat.nz p8,p0=r34
 395         add r31=IA64_TASK_SIGHAND_OFFSET,r16
 396         ;;
 397         ld8 r3=[r2]                             // read/prefetch current->blocked
 398         ld4 r9=[r9]
 399         tnat.nz.or p6,p0=r35
 400
 401         cmp.ne.or p6,p0=_NSIG_WORDS*8,r35
 402         tnat.nz.or p6,p0=r32
 403 (p6)    br.spnt.few .fail_einval                // fail with EINVAL
 404         ;;
 405 #ifdef CONFIG_SMP
 406         ld8 r31=[r31]                           // r31 <- current->sighand
 407 #endif
 408         and r9=TIF_ALLWORK_MASK,r9
 409         tnat.nz.or p8,p0=r33
 410         ;;
 411         cmp.ne p7,p0=0,r9
 412         cmp.eq p6,p0=r0,r33                     // set == NULL?
 413         add r31=IA64_SIGHAND_SIGLOCK_OFFSET,r31 // r31 <- current->sighand->siglock
 414 (p8)    br.spnt.few .fail_efault                // fail with EFAULT
 415 (p7)    br.spnt.many fsys_fallback_syscall      // got pending kernel work...
 416 (p6)    br.dpnt.many .store_mask                // -> short-circuit to just reading the signal mask
 417
 418         /* Argh, we actually have to do some work and _update_ the signal mask: */
 419
 420 EX(.fail_efault, probe.r.fault r33, 3)          // verify user has read-access to *set
 421 EX(.fail_efault, ld8 r14=[r33])                 // r14 <- *set
 422         mov r17=(1 << (SIGKILL - 1)) | (1 << (SIGSTOP - 1))
 423         ;;
 424
 425         rsm psr.i                               // mask interrupt delivery
 426         mov ar.ccv=0
 427         andcm r14=r14,r17                       // filter out SIGKILL & SIGSTOP
 428
 429 #ifdef CONFIG_SMP
 430         mov r17=1
 431         ;;
 432         cmpxchg4.acq r18=[r31],r17,ar.ccv       // try to acquire the lock
 433         mov r8=EINVAL                   // default to EINVAL
 434         ;;
 435         ld8 r3=[r2]                     // re-read current->blocked now that we hold the lock
 436         cmp4.ne p6,p0=r18,r0
 437 (p6)    br.cond.spnt.many .lock_contention
 438         ;;
 439 #else
 440         ld8 r3=[r2]                     // re-read current->blocked now that we hold the lock
 441         mov r8=EINVAL                   // default to EINVAL
 442 #endif
 443         add r18=IA64_TASK_PENDING_OFFSET+IA64_SIGPENDING_SIGNAL_OFFSET,r16
 444         add r19=IA64_TASK_SIGNAL_OFFSET,r16
 445         cmp4.eq p6,p0=SIG_BLOCK,r32
 446         ;;
 447         ld8 r19=[r19]                   // r19 <- current->signal
 448         cmp4.eq p7,p0=SIG_UNBLOCK,r32
 449         cmp4.eq p8,p0=SIG_SETMASK,r32
 450         ;;
 451         ld8 r18=[r18]                   // r18 <- current->pending.signal
 452         .pred.rel.mutex p6,p7,p8
 453 (p6)    or r14=r3,r14                   // SIG_BLOCK
 454 (p7)    andcm r14=r3,r14                // SIG_UNBLOCK
 455
 456 (p8)    mov r14=r14                     // SIG_SETMASK
 457 (p6)    mov r8=0                        // clear error code
 458         // recalc_sigpending()
 459         add r17=IA64_SIGNAL_GROUP_STOP_COUNT_OFFSET,r19
 460
 461         add r19=IA64_SIGNAL_SHARED_PENDING_OFFSET+IA64_SIGPENDING_SIGNAL_OFFSET,r19
 462         ;;
 463         ld4 r17=[r17]           // r17 <- current->signal->group_stop_count
 464 (p7)    mov r8=0                // clear error code
 465
 466         ld8 r19=[r19]           // r19 <- current->signal->shared_pending
 467         ;;
 468         cmp4.gt p6,p7=r17,r0    // p6/p7 <- (current->signal->group_stop_count > 0)?
 469 (p8)    mov r8=0                // clear error code
 470
 471         or r18=r18,r19          // r18 <- current->pending | current->signal->shared_pending
 472         ;;
 473         // r18 <- (current->pending | current->signal->shared_pending) & ~current->blocked:
 474         andcm r18=r18,r14
 475         add r9=TI_FLAGS+IA64_TASK_SIZE,r16
 476         ;;
 477
 478 (p7)    cmp.ne.or.andcm p6,p7=r18,r0            // p6/p7 <- signal pending
 479         mov r19=0                                       // i must not leak kernel bits...
 480 (p6)    br.cond.dpnt.many .sig_pending
 481         ;;
 482
 483 1:      ld4 r17=[r9]                            // r17 <- current->thread_info->flags
 484         ;;
 485         mov ar.ccv=r17
 486         and r18=~_TIF_SIGPENDING,r17            // r18 <- r17 & ~(1 << TIF_SIGPENDING)
 487         ;;
 488
 489         st8 [r2]=r14                            // update current->blocked with new mask
 490         cmpxchg4.acq r8=[r9],r18,ar.ccv         // current->thread_info->flags <- r18
 491         ;;
 492         cmp.ne p6,p0=r17,r8                     // update failed?
 493 (p6)    br.cond.spnt.few 1b                     // yes -> retry
 494
 495 #ifdef CONFIG_SMP
 496         st4.rel [r31]=r0                        // release the lock
 497 #endif
 498         ssm psr.i
 499         ;;
 500
 501         srlz.d                                  // ensure psr.i is set again
 502         mov r18=0                                       // i must not leak kernel bits...
 503
 504 .store_mask:
 505 EX(.fail_efault, (p15) probe.w.fault r34, 3)    // verify user has write-access to *oset
 506 EX(.fail_efault, (p15) st8 [r34]=r3)
 507         mov r2=0                                        // i must not leak kernel bits...
 508         mov r3=0                                        // i must not leak kernel bits...
 509         mov r8=0                                // return 0
 510         mov r9=0                                        // i must not leak kernel bits...
 511         mov r14=0                                       // i must not leak kernel bits...
 512         mov r17=0                                       // i must not leak kernel bits...
 513         mov r31=0                                       // i must not leak kernel bits...
 514         FSYS_RETURN
 515
 516 .sig_pending:
 517 #ifdef CONFIG_SMP
 518         st4.rel [r31]=r0                        // release the lock
 519 #endif
 520         ssm psr.i
 521         ;;
 522         srlz.d
 523         br.sptk.many fsys_fallback_syscall      // with signal pending, do the heavy-weight syscall
 524
 525 #ifdef CONFIG_SMP
 526 .lock_contention:
 527         /* Rather than spinning here, fall back on doing a heavy-weight syscall.  */
 528         ssm psr.i
 529         ;;
 530         srlz.d
 531         br.sptk.many fsys_fallback_syscall
 532 #endif
 533 END(fsys_rt_sigprocmask)
 534
 535 /*
 536  * fsys_getcpu doesn't use the third parameter in this implementation. It reads
 537  * current_thread_info()->cpu and corresponding node in cpu_to_node_map.
 538  */
 539 ENTRY(fsys_getcpu)
 540         .prologue
 541         .altrp b6
 542         .body
 543         ;;
 544         add r2=TI_FLAGS+IA64_TASK_SIZE,r16
 545         tnat.nz p6,p0 = r32                     // guard against NaT argument
 546         add r3=TI_CPU+IA64_TASK_SIZE,r16
 547         ;;
 548         ld4 r3=[r3]                             // M r3 = thread_info->cpu
 549         ld4 r2=[r2]                             // M r2 = thread_info->flags
 550 (p6)    br.cond.spnt.few .fail_einval           // B
 551         ;;
 552         tnat.nz p7,p0 = r33                     // I guard against NaT argument
 553 (p7)    br.cond.spnt.few .fail_einval           // B
 554 #ifdef CONFIG_NUMA
 555         movl r17=cpu_to_node_map
 556         ;;
 557 EX(.fail_efault, probe.w.fault r32, 3)          // M This takes 5 cycles
 558 EX(.fail_efault, probe.w.fault r33, 3)          // M This takes 5 cycles
 559         shladd r18=r3,1,r17
 560         ;;
 561         ld2 r20=[r18]                           // r20 = cpu_to_node_map[cpu]
 562         and r2 = TIF_ALLWORK_MASK,r2
 563         ;;
 564         cmp.ne p8,p0=0,r2
 565 (p8)    br.spnt.many fsys_fallback_syscall
 566         ;;
 567         ;;
 568 EX(.fail_efault, st4 [r32] = r3)
 569 EX(.fail_efault, st2 [r33] = r20)
 570         mov r8=0
 571         ;;
 572 #else
 573 EX(.fail_efault, probe.w.fault r32, 3)          // M This takes 5 cycles
 574 EX(.fail_efault, probe.w.fault r33, 3)          // M This takes 5 cycles
 575         and r2 = TIF_ALLWORK_MASK,r2
 576         ;;
 577         cmp.ne p8,p0=0,r2
 578 (p8)    br.spnt.many fsys_fallback_syscall
 579         ;;
 580 EX(.fail_efault, st4 [r32] = r3)
 581 EX(.fail_efault, st2 [r33] = r0)
 582         mov r8=0
 583         ;;
 584 #endif
 585         FSYS_RETURN
 586 END(fsys_getcpu)
 587
 588 ENTRY(fsys_fallback_syscall)
 589         .prologue
 590         .altrp b6
 591         .body
 592         /*
 593          * We only get here from light-weight syscall handlers.  Thus, we already
 594          * know that r15 contains a valid syscall number.  No need to re-check.
 595          */
 596         adds r17=-1024,r15
 597         movl r14=sys_call_table
 598         ;;
 599         rsm psr.i
 600         shladd r18=r17,3,r14
 601         ;;
 602         ld8 r18=[r18]                           // load normal (heavy-weight) syscall entry-point
 603         mov r29=psr                             // read psr (12 cyc load latency)
 604         mov r27=ar.rsc
 605         mov r21=ar.fpsr
 606         mov r26=ar.pfs
 607 END(fsys_fallback_syscall)
 608         /* FALL THROUGH */
 609 GLOBAL_ENTRY(fsys_bubble_down)
 610         .prologue
 611         .altrp b6
 612         .body
 613         /*
 614          * We get here for syscalls that don't have a lightweight
 615          * handler.  For those, we need to bubble down into the kernel
 616          * and that requires setting up a minimal pt_regs structure,
 617          * and initializing the CPU state more or less as if an
 618          * interruption had occurred.  To make syscall-restarts work,
 619          * we setup pt_regs such that cr_iip points to the second
 620          * instruction in syscall_via_break.  Decrementing the IP
 621          * hence will restart the syscall via break and not
 622          * decrementing IP will return us to the caller, as usual.
 623          * Note that we preserve the value of psr.pp rather than
 624          * initializing it from dcr.pp.  This makes it possible to
 625          * distinguish fsyscall execution from other privileged
 626          * execution.
 627          *
 628          * On entry:
 629          *      - normal fsyscall handler register usage, except
 630          *        that we also have:
 631          *      - r18: address of syscall entry point
 632          *      - r21: ar.fpsr
 633          *      - r26: ar.pfs
 634          *      - r27: ar.rsc
 635          *      - r29: psr
 636          *
 637          * We used to clear some PSR bits here but that requires slow
 638          * serialization.  Fortuntely, that isn't really necessary.
 639          * The rationale is as follows: we used to clear bits
 640          * ~PSR_PRESERVED_BITS in PSR.L.  Since
 641          * PSR_PRESERVED_BITS==PSR.{UP,MFL,MFH,PK,DT,PP,SP,RT,IC}, we
 642          * ended up clearing PSR.{BE,AC,I,DFL,DFH,DI,DB,SI,TB}.
 643          * However,
 644          *
 645          * PSR.BE : already is turned off in __kernel_syscall_via_epc()
 646          * PSR.AC : don't care (kernel normally turns PSR.AC on)
 647          * PSR.I  : already turned off by the time fsys_bubble_down gets
 648          *          invoked
 649          * PSR.DFL: always 0 (kernel never turns it on)
 650          * PSR.DFH: don't care --- kernel never touches f32-f127 on its own
 651          *          initiative
 652          * PSR.DI : always 0 (kernel never turns it on)
 653          * PSR.SI : always 0 (kernel never turns it on)
 654          * PSR.DB : don't care --- kernel never enables kernel-level
 655          *          breakpoints
 656          * PSR.TB : must be 0 already; if it wasn't zero on entry to
 657          *          __kernel_syscall_via_epc, the branch to fsys_bubble_down
 658          *          will trigger a taken branch; the taken-trap-handler then
 659          *          converts the syscall into a break-based system-call.
 660          */
 661         /*
 662          * Reading psr.l gives us only bits 0-31, psr.it, and psr.mc.
 663          * The rest we have to synthesize.
 664          */
 665 #       define PSR_ONE_BITS             ((3 << IA64_PSR_CPL0_BIT)       \
 666                                          | (0x1 << IA64_PSR_RI_BIT)     \
 667                                          | IA64_PSR_BN | IA64_PSR_I)
 668
 669         invala                                  // M0|1
 670         movl r14=ia64_ret_from_syscall          // X
 671
 672         nop.m 0
 673         movl r28=__kernel_syscall_via_break     // X    create cr.iip
 674         ;;
 675
 676         mov r2=r16                              // A    get task addr to addl-addressable register
 677         adds r16=IA64_TASK_THREAD_ON_USTACK_OFFSET,r16 // A
 678         mov r31=pr                              // I0   save pr (2 cyc)
 679         ;;
 680         st1 [r16]=r0                            // M2|3 clear current->thread.on_ustack flag
 681         addl r22=IA64_RBS_OFFSET,r2             // A    compute base of RBS
 682         add r3=TI_FLAGS+IA64_TASK_SIZE,r2       // A
 683         ;;
 684         ld4 r3=[r3]                             // M0|1 r3 = current_thread_info()->flags
 685         lfetch.fault.excl.nt1 [r22]             // M0|1 prefetch register backing-store
 686         nop.i 0
 687         ;;
 688         mov ar.rsc=0                            // M2   set enforced lazy mode, pl 0, LE, loadrs=0
 689         nop.m 0
 690         nop.i 0
 691         ;;
 692         mov r23=ar.bspstore                     // M2 (12 cyc) save ar.bspstore
 693         mov.m r24=ar.rnat                       // M2 (5 cyc) read ar.rnat (dual-issues!)
 694         nop.i 0
 695         ;;
 696         mov ar.bspstore=r22                     // M2 (6 cyc) switch to kernel RBS
 697         movl r8=PSR_ONE_BITS                    // X
 698         ;;
 699         mov r25=ar.unat                         // M2 (5 cyc) save ar.unat
 700         mov r19=b6                              // I0   save b6 (2 cyc)
 701         mov r20=r1                              // A    save caller's gp in r20
 702         ;;
 703         or r29=r8,r29                           // A    construct cr.ipsr value to save
 704         mov b6=r18                              // I0   copy syscall entry-point to b6 (7 cyc)
 705         addl r1=IA64_STK_OFFSET-IA64_PT_REGS_SIZE,r2 // A compute base of memory stack
 706
 707         mov r18=ar.bsp                          // M2   save (kernel) ar.bsp (12 cyc)
 708         cmp.ne pKStk,pUStk=r0,r0                // A    set pKStk <- 0, pUStk <- 1
 709         br.call.sptk.many b7=ia64_syscall_setup // B
 710         ;;
 711         mov ar.rsc=0x3                          // M2   set eager mode, pl 0, LE, loadrs=0
 712         mov rp=r14                              // I0   set the real return addr
 713         and r3=_TIF_SYSCALL_TRACEAUDIT,r3       // A
 714         ;;
 715         ssm psr.i                               // M2   we're on kernel stacks now, reenable irqs
 716         cmp.eq p8,p0=r3,r0                      // A
 717 (p10)   br.cond.spnt.many ia64_ret_from_syscall // B    return if bad call-frame or r15 is a NaT
 718
 719         nop.m 0
 720 (p8)    br.call.sptk.many b6=b6                 // B    (ignore return address)
 721         br.cond.spnt ia64_trace_syscall         // B
 722 END(fsys_bubble_down)
 723
 724         .rodata
 725         .align 8
 726         .globl fsyscall_table
 727
 728         data8 fsys_bubble_down
 729 fsyscall_table:
 730         data8 fsys_ni_syscall
 731         data8 0                         // exit                 // 1025
 732         data8 0                         // read
 733         data8 0                         // write
 734         data8 0                         // open
 735         data8 0                         // close
 736         data8 0                         // creat                // 1030
 737         data8 0                         // link
 738         data8 0                         // unlink
 739         data8 0                         // execve
 740         data8 0                         // chdir
 741         data8 0                         // fchdir               // 1035
 742         data8 0                         // utimes
 743         data8 0                         // mknod
 744         data8 0                         // chmod
 745         data8 0                         // chown
 746         data8 0                         // lseek                // 1040
 747         data8 fsys_getpid               // getpid
 748         data8 fsys_getppid              // getppid
 749         data8 0                         // mount
 750         data8 0                         // umount
 751         data8 0                         // setuid               // 1045
 752         data8 0                         // getuid
 753         data8 0                         // geteuid
 754         data8 0                         // ptrace
 755         data8 0                         // access
 756         data8 0                         // sync                 // 1050
 757         data8 0                         // fsync
 758         data8 0                         // fdatasync
 759         data8 0                         // kill
 760         data8 0                         // rename
 761         data8 0                         // mkdir                // 1055
 762         data8 0                         // rmdir
 763         data8 0                         // dup
 764         data8 0                         // pipe
 765         data8 0                         // times
 766         data8 0                         // brk                  // 1060
 767         data8 0                         // setgid
 768         data8 0                         // getgid
 769         data8 0                         // getegid
 770         data8 0                         // acct
 771         data8 0                         // ioctl                // 1065
 772         data8 0                         // fcntl
 773         data8 0                         // umask
 774         data8 0                         // chroot
 775         data8 0                         // ustat
 776         data8 0                         // dup2                 // 1070
 777         data8 0                         // setreuid
 778         data8 0                         // setregid
 779         data8 0                         // getresuid
 780         data8 0                         // setresuid
 781         data8 0                         // getresgid            // 1075
 782         data8 0                         // setresgid
 783         data8 0                         // getgroups
 784         data8 0                         // setgroups
 785         data8 0                         // getpgid
 786         data8 0                         // setpgid              // 1080
 787         data8 0                         // setsid
 788         data8 0                         // getsid
 789         data8 0                         // sethostname
 790         data8 0                         // setrlimit
 791         data8 0                         // getrlimit            // 1085
 792         data8 0                         // getrusage
 793         data8 fsys_gettimeofday         // gettimeofday
 794         data8 0                         // settimeofday
 795         data8 0                         // select
 796         data8 0                         // poll                 // 1090
 797         data8 0                         // symlink
 798         data8 0                         // readlink
 799         data8 0                         // uselib
 800         data8 0                         // swapon
 801         data8 0                         // swapoff              // 1095
 802         data8 0                         // reboot
 803         data8 0                         // truncate
 804         data8 0                         // ftruncate
 805         data8 0                         // fchmod
 806         data8 0                         // fchown               // 1100
 807         data8 0                         // getpriority
 808         data8 0                         // setpriority
 809         data8 0                         // statfs
 810         data8 0                         // fstatfs
 811         data8 0                         // gettid               // 1105
 812         data8 0                         // semget
 813         data8 0                         // semop
 814         data8 0                         // semctl
 815         data8 0                         // msgget
 816         data8 0                         // msgsnd               // 1110
 817         data8 0                         // msgrcv
 818         data8 0                         // msgctl
 819         data8 0                         // shmget
 820         data8 0                         // shmat
 821         data8 0                         // shmdt                // 1115
 822         data8 0                         // shmctl
 823         data8 0                         // syslog
 824         data8 0                         // setitimer
 825         data8 0                         // getitimer
 826         data8 0                                                 // 1120
 827         data8 0
 828         data8 0
 829         data8 0                         // vhangup
 830         data8 0                         // lchown
 831         data8 0                         // remap_file_pages     // 1125
 832         data8 0                         // wait4
 833         data8 0                         // sysinfo
 834         data8 0                         // clone
 835         data8 0                         // setdomainname
 836         data8 0                         // newuname             // 1130
 837         data8 0                         // adjtimex
 838         data8 0
 839         data8 0                         // init_module
 840         data8 0                         // delete_module
 841         data8 0                                                 // 1135
 842         data8 0
 843         data8 0                         // quotactl
 844         data8 0                         // bdflush
 845         data8 0                         // sysfs
 846         data8 0                         // personality          // 1140
 847         data8 0                         // afs_syscall
 848         data8 0                         // setfsuid
 849         data8 0                         // setfsgid
 850         data8 0                         // getdents
 851         data8 0                         // flock                // 1145
 852         data8 0                         // readv
 853         data8 0                         // writev
 854         data8 0                         // pread64
 855         data8 0                         // pwrite64
 856         data8 0                         // sysctl               // 1150
 857         data8 0                         // mmap
 858         data8 0                         // munmap
 859         data8 0                         // mlock
 860         data8 0                         // mlockall
 861         data8 0                         // mprotect             // 1155
 862         data8 0                         // mremap
 863         data8 0                         // msync
 864         data8 0                         // munlock
 865         data8 0                         // munlockall
 866         data8 0                         // sched_getparam       // 1160
 867         data8 0                         // sched_setparam
 868         data8 0                         // sched_getscheduler
 869         data8 0                         // sched_setscheduler
 870         data8 0                         // sched_yield
 871         data8 0                         // sched_get_priority_max       // 1165
 872         data8 0                         // sched_get_priority_min
 873         data8 0                         // sched_rr_get_interval
 874         data8 0                         // nanosleep
 875         data8 0                         // nfsservctl
 876         data8 0                         // prctl                // 1170
 877         data8 0                         // getpagesize
 878         data8 0                         // mmap2
 879         data8 0                         // pciconfig_read
 880         data8 0                         // pciconfig_write
 881         data8 0                         // perfmonctl           // 1175
 882         data8 0                         // sigaltstack
 883         data8 0                         // rt_sigaction
 884         data8 0                         // rt_sigpending
 885         data8 fsys_rt_sigprocmask       // rt_sigprocmask
 886         data8 0                         // rt_sigqueueinfo      // 1180
 887         data8 0                         // rt_sigreturn
 888         data8 0                         // rt_sigsuspend
 889         data8 0                         // rt_sigtimedwait
 890         data8 0                         // getcwd
 891         data8 0                         // capget               // 1185
 892         data8 0                         // capset
 893         data8 0                         // sendfile
 894         data8 0
 895         data8 0
 896         data8 0                         // socket               // 1190
 897         data8 0                         // bind
 898         data8 0                         // connect
 899         data8 0                         // listen
 900         data8 0                         // accept
 901         data8 0                         // getsockname          // 1195
 902         data8 0                         // getpeername
 903         data8 0                         // socketpair
 904         data8 0                         // send
 905         data8 0                         // sendto
 906         data8 0                         // recv                 // 1200
 907         data8 0                         // recvfrom
 908         data8 0                         // shutdown
 909         data8 0                         // setsockopt
 910         data8 0                         // getsockopt
 911         data8 0                         // sendmsg              // 1205
 912         data8 0                         // recvmsg
 913         data8 0                         // pivot_root
 914         data8 0                         // mincore
 915         data8 0                         // madvise
 916         data8 0                         // newstat              // 1210
 917         data8 0                         // newlstat
 918         data8 0                         // newfstat
 919         data8 0                         // clone2
 920         data8 0                         // getdents64
 921         data8 0                         // getunwind            // 1215
 922         data8 0                         // readahead
 923         data8 0                         // setxattr
 924         data8 0                         // lsetxattr
 925         data8 0                         // fsetxattr
 926         data8 0                         // getxattr             // 1220
 927         data8 0                         // lgetxattr
 928         data8 0                         // fgetxattr
 929         data8 0                         // listxattr
 930         data8 0                         // llistxattr
 931         data8 0                         // flistxattr           // 1225
 932         data8 0                         // removexattr
 933         data8 0                         // lremovexattr
 934         data8 0                         // fremovexattr
 935         data8 0                         // tkill
 936         data8 0                         // futex                // 1230
 937         data8 0                         // sched_setaffinity
 938         data8 0                         // sched_getaffinity
 939         data8 fsys_set_tid_address      // set_tid_address
 940         data8 0                         // fadvise64_64
 941         data8 0                         // tgkill               // 1235
 942         data8 0                         // exit_group
 943         data8 0                         // lookup_dcookie
 944         data8 0                         // io_setup
 945         data8 0                         // io_destroy
 946         data8 0                         // io_getevents         // 1240
 947         data8 0                         // io_submit
 948         data8 0                         // io_cancel
 949         data8 0                         // epoll_create
 950         data8 0                         // epoll_ctl
 951         data8 0                         // epoll_wait           // 1245
 952         data8 0                         // restart_syscall
 953         data8 0                         // semtimedop
 954         data8 0                         // timer_create
 955         data8 0                         // timer_settime
 956         data8 0                         // timer_gettime        // 1250
 957         data8 0                         // timer_getoverrun
 958         data8 0                         // timer_delete
 959         data8 0                         // clock_settime
 960         data8 fsys_clock_gettime        // clock_gettime
 961         data8 0                         // clock_getres         // 1255
 962         data8 0                         // clock_nanosleep
 963         data8 0                         // fstatfs64
 964         data8 0                         // statfs64
 965         data8 0                         // mbind
 966         data8 0                         // get_mempolicy        // 1260
 967         data8 0                         // set_mempolicy
 968         data8 0                         // mq_open
 969         data8 0                         // mq_unlink
 970         data8 0                         // mq_timedsend
 971         data8 0                         // mq_timedreceive      // 1265
 972         data8 0                         // mq_notify
 973         data8 0                         // mq_getsetattr
 974         data8 0                         // kexec_load
 975         data8 0                         // vserver
 976         data8 0                         // waitid               // 1270
 977         data8 0                         // add_key
 978         data8 0                         // request_key
 979         data8 0                         // keyctl
 980         data8 0                         // ioprio_set
 981         data8 0                         // ioprio_get           // 1275
 982         data8 0                         // move_pages
 983         data8 0                         // inotify_init
 984         data8 0                         // inotify_add_watch
 985         data8 0                         // inotify_rm_watch
 986         data8 0                         // migrate_pages        // 1280
 987         data8 0                         // openat
 988         data8 0                         // mkdirat
 989         data8 0                         // mknodat
 990         data8 0                         // fchownat
 991         data8 0                         // futimesat            // 1285
 992         data8 0                         // newfstatat
 993         data8 0                         // unlinkat
 994         data8 0                         // renameat
 995         data8 0                         // linkat
 996         data8 0                         // symlinkat            // 1290
 997         data8 0                         // readlinkat
 998         data8 0                         // fchmodat
 999         data8 0                         // faccessat
1000         data8 0
1001         data8 0                                                 // 1295
1002         data8 0                         // unshare
1003         data8 0                         // splice
1004         data8 0                         // set_robust_list
1005         data8 0                         // get_robust_list
1006         data8 0                         // sync_file_range      // 1300
1007         data8 0                         // tee
1008         data8 0                         // vmsplice
1009         data8 0
1010         data8 fsys_getcpu               // getcpu               // 1304
1011
1012         // fill in zeros for the remaining entries
1013         .zero:
1014         .space fsyscall_table + 8*NR_syscalls - .zero, 0