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/unistd.h>
  25
  26 #include "entry.h"
  27 #include "paravirt_inst.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_SEQ_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         add r2 = TI_FLAGS+IA64_TASK_SIZE,r16
 241         tnat.nz p6,p0 = r31             // guard against Nat argument
 242 (p6)    br.cond.spnt.few .fail_einval
 243         movl r20 = fsyscall_gtod_data // load fsyscall gettimeofday data address
 244         ;;
 245         ld4 r2 = [r2]                   // process work pending flags
 246         movl r29 = itc_jitter_data      // itc_jitter
 247         add r22 = IA64_GTOD_WALL_TIME_OFFSET,r20        // wall_time
 248         add r21 = IA64_CLKSRC_MMIO_OFFSET,r20
 249         mov pr = r30,0xc000     // Set predicates according to function
 250         ;;
 251         and r2 = TIF_ALLWORK_MASK,r2
 252         add r19 = IA64_ITC_LASTCYCLE_OFFSET,r29
 253 (p15)   add r22 = IA64_GTOD_MONO_TIME_OFFSET,r20        // monotonic_time
 254         ;;
 255         add r26 = IA64_CLKSRC_CYCLE_LAST_OFFSET,r20     // clksrc_cycle_last
 256         cmp.ne p6, p0 = 0, r2   // Fallback if work is scheduled
 257 (p6)    br.cond.spnt.many fsys_fallback_syscall
 258         ;;
 259         // Begin critical section
 260 .time_redo:
 261         ld4.acq r28 = [r20]     // gtod_lock.sequence, Must take first
 262         ;;
 263         and r28 = ~1,r28        // And make sequence even to force retry if odd
 264         ;;
 265         ld8 r30 = [r21]         // clocksource->mmio_ptr
 266         add r24 = IA64_CLKSRC_MULT_OFFSET,r20
 267         ld4 r2 = [r29]          // itc_jitter value
 268         add r23 = IA64_CLKSRC_SHIFT_OFFSET,r20
 269         add r14 = IA64_CLKSRC_MASK_OFFSET,r20
 270         ;;
 271         ld4 r3 = [r24]          // clocksource mult value
 272         ld8 r14 = [r14]         // clocksource mask value
 273         cmp.eq p8,p9 = 0,r30    // use cpu timer if no mmio_ptr
 274         ;;
 275         setf.sig f7 = r3        // Setup for mult scaling of counter
 276 (p8)    cmp.ne p13,p0 = r2,r0   // need itc_jitter compensation, set p13
 277         ld4 r23 = [r23]         // clocksource shift value
 278         ld8 r24 = [r26]         // get clksrc_cycle_last value
 279 (p9)    cmp.eq p13,p0 = 0,r30   // if mmio_ptr, clear p13 jitter control
 280         ;;
 281         .pred.rel.mutex p8,p9
 282         MOV_FROM_ITC(p8, p6, r2, r10)   // CPU_TIMER. 36 clocks latency!!!
 283 (p9)    ld8 r2 = [r30]          // MMIO_TIMER. Could also have latency issues..
 284 (p13)   ld8 r25 = [r19]         // get itc_lastcycle value
 285         ld8 r9 = [r22],IA64_TIMESPEC_TV_NSEC_OFFSET     // tv_sec
 286         ;;
 287         ld8 r8 = [r22],-IA64_TIMESPEC_TV_NSEC_OFFSET    // tv_nsec
 288 (p13)   sub r3 = r25,r2         // Diff needed before comparison (thanks davidm)
 289         ;;
 290 (p13)   cmp.gt.unc p6,p7 = r3,r0 // check if it is less than last. p6,p7 cleared
 291         sub r10 = r2,r24        // current_cycle - last_cycle
 292         ;;
 293 (p6)    sub r10 = r25,r24       // time we got was less than last_cycle
 294 (p7)    mov ar.ccv = r25        // more than last_cycle. Prep for cmpxchg
 295         ;;
 296 (p7)    cmpxchg8.rel r3 = [r19],r2,ar.ccv
 297         ;;
 298 (p7)    cmp.ne p7,p0 = r25,r3   // if cmpxchg not successful
 299         ;;
 300 (p7)    sub r10 = r3,r24        // then use new last_cycle instead
 301         ;;
 302         and r10 = r10,r14       // Apply mask
 303         ;;
 304         setf.sig f8 = r10
 305         nop.i 123
 306         ;;
 307         // fault check takes 5 cycles and we have spare time
 308 EX(.fail_efault, probe.w.fault r31, 3)
 309         xmpy.l f8 = f8,f7       // nsec_per_cyc*(counter-last_counter)
 310         ;;
 311         getf.sig r2 = f8
 312         mf
 313         ;;
 314         ld4 r10 = [r20]         // gtod_lock.sequence
 315         shr.u r2 = r2,r23       // shift by factor
 316         ;;
 317         add r8 = r8,r2          // Add xtime.nsecs
 318         cmp4.ne p7,p0 = r28,r10
 319 (p7)    br.cond.dpnt.few .time_redo     // sequence number changed, redo
 320         // End critical section.
 321         // Now r8=tv->tv_nsec and r9=tv->tv_sec
 322         mov r10 = r0
 323         movl r2 = 1000000000
 324         add r23 = IA64_TIMESPEC_TV_NSEC_OFFSET, r31
 325 (p14)   movl r3 = 2361183241434822607   // Prep for / 1000 hack
 326         ;;
 327 .time_normalize:
 328         mov r21 = r8
 329         cmp.ge p6,p0 = r8,r2
 330 (p14)   shr.u r20 = r8, 3 // We can repeat this if necessary just wasting time
 331         ;;
 332 (p14)   setf.sig f8 = r20
 333 (p6)    sub r8 = r8,r2
 334 (p6)    add r9 = 1,r9           // two nops before the branch.
 335 (p14)   setf.sig f7 = r3        // Chances for repeats are 1 in 10000 for gettod
 336 (p6)    br.cond.dpnt.few .time_normalize
 337         ;;
 338         // Divided by 8 though shift. Now divide by 125
 339         // The compiler was able to do that with a multiply
 340         // and a shift and we do the same
 341 EX(.fail_efault, probe.w.fault r23, 3)  // This also costs 5 cycles
 342 (p14)   xmpy.hu f8 = f8, f7             // xmpy has 5 cycles latency so use it
 343         ;;
 344 (p14)   getf.sig r2 = f8
 345         ;;
 346         mov r8 = r0
 347 (p14)   shr.u r21 = r2, 4
 348         ;;
 349 EX(.fail_efault, st8 [r31] = r9)
 350 EX(.fail_efault, st8 [r23] = r21)
 351         FSYS_RETURN
 352 .fail_einval:
 353         mov r8 = EINVAL
 354         mov r10 = -1
 355         FSYS_RETURN
 356 .fail_efault:
 357         mov r8 = EFAULT
 358         mov r10 = -1
 359         FSYS_RETURN
 360 END(fsys_gettimeofday)
 361
 362 ENTRY(fsys_clock_gettime)
 363         .prologue
 364         .altrp b6
 365         .body
 366         cmp4.ltu p6, p0 = CLOCK_MONOTONIC, r32
 367         // Fallback if this is not CLOCK_REALTIME or CLOCK_MONOTONIC
 368 (p6)    br.spnt.few fsys_fallback_syscall
 369         mov r31 = r33
 370         shl r30 = r32,15
 371         br.many .gettime
 372 END(fsys_clock_gettime)
 373
 374 /*
 375  * fsys_getcpu doesn't use the third parameter in this implementation. It reads
 376  * current_thread_info()->cpu and corresponding node in cpu_to_node_map.
 377  */
 378 ENTRY(fsys_getcpu)
 379         .prologue
 380         .altrp b6
 381         .body
 382         ;;
 383         add r2=TI_FLAGS+IA64_TASK_SIZE,r16
 384         tnat.nz p6,p0 = r32                     // guard against NaT argument
 385         add r3=TI_CPU+IA64_TASK_SIZE,r16
 386         ;;
 387         ld4 r3=[r3]                             // M r3 = thread_info->cpu
 388         ld4 r2=[r2]                             // M r2 = thread_info->flags
 389 (p6)    br.cond.spnt.few .fail_einval           // B
 390         ;;
 391         tnat.nz p7,p0 = r33                     // I guard against NaT argument
 392 (p7)    br.cond.spnt.few .fail_einval           // B
 393         ;;
 394         cmp.ne p6,p0=r32,r0
 395         cmp.ne p7,p0=r33,r0
 396         ;;
 397 #ifdef CONFIG_NUMA
 398         movl r17=cpu_to_node_map
 399         ;;
 400 EX(.fail_efault, (p6) probe.w.fault r32, 3)             // M This takes 5 cycles
 401 EX(.fail_efault, (p7) probe.w.fault r33, 3)             // M This takes 5 cycles
 402         shladd r18=r3,1,r17
 403         ;;
 404         ld2 r20=[r18]                           // r20 = cpu_to_node_map[cpu]
 405         and r2 = TIF_ALLWORK_MASK,r2
 406         ;;
 407         cmp.ne p8,p0=0,r2
 408 (p8)    br.spnt.many fsys_fallback_syscall
 409         ;;
 410         ;;
 411 EX(.fail_efault, (p6) st4 [r32] = r3)
 412 EX(.fail_efault, (p7) st2 [r33] = r20)
 413         mov r8=0
 414         ;;
 415 #else
 416 EX(.fail_efault, (p6) probe.w.fault r32, 3)             // M This takes 5 cycles
 417 EX(.fail_efault, (p7) probe.w.fault r33, 3)             // M This takes 5 cycles
 418         and r2 = TIF_ALLWORK_MASK,r2
 419         ;;
 420         cmp.ne p8,p0=0,r2
 421 (p8)    br.spnt.many fsys_fallback_syscall
 422         ;;
 423 EX(.fail_efault, (p6) st4 [r32] = r3)
 424 EX(.fail_efault, (p7) st2 [r33] = r0)
 425         mov r8=0
 426         ;;
 427 #endif
 428         FSYS_RETURN
 429 END(fsys_getcpu)
 430
 431 ENTRY(fsys_fallback_syscall)
 432         .prologue
 433         .altrp b6
 434         .body
 435         /*
 436          * We only get here from light-weight syscall handlers.  Thus, we already
 437          * know that r15 contains a valid syscall number.  No need to re-check.
 438          */
 439         adds r17=-1024,r15
 440         movl r14=sys_call_table
 441         ;;
 442         RSM_PSR_I(p0, r26, r27)
 443         shladd r18=r17,3,r14
 444         ;;
 445         ld8 r18=[r18]                           // load normal (heavy-weight) syscall entry-point
 446         MOV_FROM_PSR(p0, r29, r26)              // read psr (12 cyc load latency)
 447         mov r27=ar.rsc
 448         mov r21=ar.fpsr
 449         mov r26=ar.pfs
 450 END(fsys_fallback_syscall)
 451         /* FALL THROUGH */
 452 GLOBAL_ENTRY(paravirt_fsys_bubble_down)
 453         .prologue
 454         .altrp b6
 455         .body
 456         /*
 457          * We get here for syscalls that don't have a lightweight
 458          * handler.  For those, we need to bubble down into the kernel
 459          * and that requires setting up a minimal pt_regs structure,
 460          * and initializing the CPU state more or less as if an
 461          * interruption had occurred.  To make syscall-restarts work,
 462          * we setup pt_regs such that cr_iip points to the second
 463          * instruction in syscall_via_break.  Decrementing the IP
 464          * hence will restart the syscall via break and not
 465          * decrementing IP will return us to the caller, as usual.
 466          * Note that we preserve the value of psr.pp rather than
 467          * initializing it from dcr.pp.  This makes it possible to
 468          * distinguish fsyscall execution from other privileged
 469          * execution.
 470          *
 471          * On entry:
 472          *      - normal fsyscall handler register usage, except
 473          *        that we also have:
 474          *      - r18: address of syscall entry point
 475          *      - r21: ar.fpsr
 476          *      - r26: ar.pfs
 477          *      - r27: ar.rsc
 478          *      - r29: psr
 479          *
 480          * We used to clear some PSR bits here but that requires slow
 481          * serialization.  Fortuntely, that isn't really necessary.
 482          * The rationale is as follows: we used to clear bits
 483          * ~PSR_PRESERVED_BITS in PSR.L.  Since
 484          * PSR_PRESERVED_BITS==PSR.{UP,MFL,MFH,PK,DT,PP,SP,RT,IC}, we
 485          * ended up clearing PSR.{BE,AC,I,DFL,DFH,DI,DB,SI,TB}.
 486          * However,
 487          *
 488          * PSR.BE : already is turned off in __kernel_syscall_via_epc()
 489          * PSR.AC : don't care (kernel normally turns PSR.AC on)
 490          * PSR.I  : already turned off by the time paravirt_fsys_bubble_down gets
 491          *          invoked
 492          * PSR.DFL: always 0 (kernel never turns it on)
 493          * PSR.DFH: don't care --- kernel never touches f32-f127 on its own
 494          *          initiative
 495          * PSR.DI : always 0 (kernel never turns it on)
 496          * PSR.SI : always 0 (kernel never turns it on)
 497          * PSR.DB : don't care --- kernel never enables kernel-level
 498          *          breakpoints
 499          * PSR.TB : must be 0 already; if it wasn't zero on entry to
 500          *          __kernel_syscall_via_epc, the branch to paravirt_fsys_bubble_down
 501          *          will trigger a taken branch; the taken-trap-handler then
 502          *          converts the syscall into a break-based system-call.
 503          */
 504         /*
 505          * Reading psr.l gives us only bits 0-31, psr.it, and psr.mc.
 506          * The rest we have to synthesize.
 507          */
 508 #       define PSR_ONE_BITS             ((3 << IA64_PSR_CPL0_BIT)       \
 509                                          | (0x1 << IA64_PSR_RI_BIT)     \
 510                                          | IA64_PSR_BN | IA64_PSR_I)
 511
 512         invala                                  // M0|1
 513         movl r14=ia64_ret_from_syscall          // X
 514
 515         nop.m 0
 516         movl r28=__kernel_syscall_via_break     // X    create cr.iip
 517         ;;
 518
 519         mov r2=r16                              // A    get task addr to addl-addressable register
 520         adds r16=IA64_TASK_THREAD_ON_USTACK_OFFSET,r16 // A
 521         mov r31=pr                              // I0   save pr (2 cyc)
 522         ;;
 523         st1 [r16]=r0                            // M2|3 clear current->thread.on_ustack flag
 524         addl r22=IA64_RBS_OFFSET,r2             // A    compute base of RBS
 525         add r3=TI_FLAGS+IA64_TASK_SIZE,r2       // A
 526         ;;
 527         ld4 r3=[r3]                             // M0|1 r3 = current_thread_info()->flags
 528         lfetch.fault.excl.nt1 [r22]             // M0|1 prefetch register backing-store
 529         nop.i 0
 530         ;;
 531         mov ar.rsc=0                            // M2   set enforced lazy mode, pl 0, LE, loadrs=0
 532 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
 533         MOV_FROM_ITC(p0, p6, r30, r23)          // M    get cycle for accounting
 534 #else
 535         nop.m 0
 536 #endif
 537         nop.i 0
 538         ;;
 539         mov r23=ar.bspstore                     // M2 (12 cyc) save ar.bspstore
 540         mov.m r24=ar.rnat                       // M2 (5 cyc) read ar.rnat (dual-issues!)
 541         nop.i 0
 542         ;;
 543         mov ar.bspstore=r22                     // M2 (6 cyc) switch to kernel RBS
 544         movl r8=PSR_ONE_BITS                    // X
 545         ;;
 546         mov r25=ar.unat                         // M2 (5 cyc) save ar.unat
 547         mov r19=b6                              // I0   save b6 (2 cyc)
 548         mov r20=r1                              // A    save caller's gp in r20
 549         ;;
 550         or r29=r8,r29                           // A    construct cr.ipsr value to save
 551         mov b6=r18                              // I0   copy syscall entry-point to b6 (7 cyc)
 552         addl r1=IA64_STK_OFFSET-IA64_PT_REGS_SIZE,r2 // A compute base of memory stack
 553
 554         mov r18=ar.bsp                          // M2   save (kernel) ar.bsp (12 cyc)
 555         cmp.ne pKStk,pUStk=r0,r0                // A    set pKStk <- 0, pUStk <- 1
 556         br.call.sptk.many b7=ia64_syscall_setup // B
 557         ;;
 558 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
 559         // mov.m r30=ar.itc is called in advance
 560         add r16=TI_AC_STAMP+IA64_TASK_SIZE,r2
 561         add r17=TI_AC_LEAVE+IA64_TASK_SIZE,r2
 562         ;;
 563         ld8 r18=[r16],TI_AC_STIME-TI_AC_STAMP   // time at last check in kernel
 564         ld8 r19=[r17],TI_AC_UTIME-TI_AC_LEAVE   // time at leave kernel
 565         ;;
 566         ld8 r20=[r16],TI_AC_STAMP-TI_AC_STIME   // cumulated stime
 567         ld8 r21=[r17]                           // cumulated utime
 568         sub r22=r19,r18                         // stime before leave kernel
 569         ;;
 570         st8 [r16]=r30,TI_AC_STIME-TI_AC_STAMP   // update stamp
 571         sub r18=r30,r19                         // elapsed time in user mode
 572         ;;
 573         add r20=r20,r22                         // sum stime
 574         add r21=r21,r18                         // sum utime
 575         ;;
 576         st8 [r16]=r20                           // update stime
 577         st8 [r17]=r21                           // update utime
 578         ;;
 579 #endif
 580         mov ar.rsc=0x3                          // M2   set eager mode, pl 0, LE, loadrs=0
 581         mov rp=r14                              // I0   set the real return addr
 582         and r3=_TIF_SYSCALL_TRACEAUDIT,r3       // A
 583         ;;
 584         SSM_PSR_I(p0, p6, r22)                  // M2   we're on kernel stacks now, reenable irqs
 585         cmp.eq p8,p0=r3,r0                      // A
 586 (p10)   br.cond.spnt.many ia64_ret_from_syscall // B    return if bad call-frame or r15 is a NaT
 587
 588         nop.m 0
 589 (p8)    br.call.sptk.many b6=b6                 // B    (ignore return address)
 590         br.cond.spnt ia64_trace_syscall         // B
 591 END(paravirt_fsys_bubble_down)
 592
 593         .rodata
 594         .align 8
 595         .globl paravirt_fsyscall_table
 596
 597         data8 paravirt_fsys_bubble_down
 598 paravirt_fsyscall_table:
 599         data8 fsys_ni_syscall
 600         data8 0                         // exit                 // 1025
 601         data8 0                         // read
 602         data8 0                         // write
 603         data8 0                         // open
 604         data8 0                         // close
 605         data8 0                         // creat                // 1030
 606         data8 0                         // link
 607         data8 0                         // unlink
 608         data8 0                         // execve
 609         data8 0                         // chdir
 610         data8 0                         // fchdir               // 1035
 611         data8 0                         // utimes
 612         data8 0                         // mknod
 613         data8 0                         // chmod
 614         data8 0                         // chown
 615         data8 0                         // lseek                // 1040
 616         data8 fsys_getpid               // getpid
 617         data8 fsys_getppid              // getppid
 618         data8 0                         // mount
 619         data8 0                         // umount
 620         data8 0                         // setuid               // 1045
 621         data8 0                         // getuid
 622         data8 0                         // geteuid
 623         data8 0                         // ptrace
 624         data8 0                         // access
 625         data8 0                         // sync                 // 1050
 626         data8 0                         // fsync
 627         data8 0                         // fdatasync
 628         data8 0                         // kill
 629         data8 0                         // rename
 630         data8 0                         // mkdir                // 1055
 631         data8 0                         // rmdir
 632         data8 0                         // dup
 633         data8 0                         // pipe
 634         data8 0                         // times
 635         data8 0                         // brk                  // 1060
 636         data8 0                         // setgid
 637         data8 0                         // getgid
 638         data8 0                         // getegid
 639         data8 0                         // acct
 640         data8 0                         // ioctl                // 1065
 641         data8 0                         // fcntl
 642         data8 0                         // umask
 643         data8 0                         // chroot
 644         data8 0                         // ustat
 645         data8 0                         // dup2                 // 1070
 646         data8 0                         // setreuid
 647         data8 0                         // setregid
 648         data8 0                         // getresuid
 649         data8 0                         // setresuid
 650         data8 0                         // getresgid            // 1075
 651         data8 0                         // setresgid
 652         data8 0                         // getgroups
 653         data8 0                         // setgroups
 654         data8 0                         // getpgid
 655         data8 0                         // setpgid              // 1080
 656         data8 0                         // setsid
 657         data8 0                         // getsid
 658         data8 0                         // sethostname
 659         data8 0                         // setrlimit
 660         data8 0                         // getrlimit            // 1085
 661         data8 0                         // getrusage
 662         data8 fsys_gettimeofday         // gettimeofday
 663         data8 0                         // settimeofday
 664         data8 0                         // select
 665         data8 0                         // poll                 // 1090
 666         data8 0                         // symlink
 667         data8 0                         // readlink
 668         data8 0                         // uselib
 669         data8 0                         // swapon
 670         data8 0                         // swapoff              // 1095
 671         data8 0                         // reboot
 672         data8 0                         // truncate
 673         data8 0                         // ftruncate
 674         data8 0                         // fchmod
 675         data8 0                         // fchown               // 1100
 676         data8 0                         // getpriority
 677         data8 0                         // setpriority
 678         data8 0                         // statfs
 679         data8 0                         // fstatfs
 680         data8 0                         // gettid               // 1105
 681         data8 0                         // semget
 682         data8 0                         // semop
 683         data8 0                         // semctl
 684         data8 0                         // msgget
 685         data8 0                         // msgsnd               // 1110
 686         data8 0                         // msgrcv
 687         data8 0                         // msgctl
 688         data8 0                         // shmget
 689         data8 0                         // shmat
 690         data8 0                         // shmdt                // 1115
 691         data8 0                         // shmctl
 692         data8 0                         // syslog
 693         data8 0                         // setitimer
 694         data8 0                         // getitimer
 695         data8 0                                                 // 1120
 696         data8 0
 697         data8 0
 698         data8 0                         // vhangup
 699         data8 0                         // lchown
 700         data8 0                         // remap_file_pages     // 1125
 701         data8 0                         // wait4
 702         data8 0                         // sysinfo
 703         data8 0                         // clone
 704         data8 0                         // setdomainname
 705         data8 0                         // newuname             // 1130
 706         data8 0                         // adjtimex
 707         data8 0
 708         data8 0                         // init_module
 709         data8 0                         // delete_module
 710         data8 0                                                 // 1135
 711         data8 0
 712         data8 0                         // quotactl
 713         data8 0                         // bdflush
 714         data8 0                         // sysfs
 715         data8 0                         // personality          // 1140
 716         data8 0                         // afs_syscall
 717         data8 0                         // setfsuid
 718         data8 0                         // setfsgid
 719         data8 0                         // getdents
 720         data8 0                         // flock                // 1145
 721         data8 0                         // readv
 722         data8 0                         // writev
 723         data8 0                         // pread64
 724         data8 0                         // pwrite64
 725         data8 0                         // sysctl               // 1150
 726         data8 0                         // mmap
 727         data8 0                         // munmap
 728         data8 0                         // mlock
 729         data8 0                         // mlockall
 730         data8 0                         // mprotect             // 1155
 731         data8 0                         // mremap
 732         data8 0                         // msync
 733         data8 0                         // munlock
 734         data8 0                         // munlockall
 735         data8 0                         // sched_getparam       // 1160
 736         data8 0                         // sched_setparam
 737         data8 0                         // sched_getscheduler
 738         data8 0                         // sched_setscheduler
 739         data8 0                         // sched_yield
 740         data8 0                         // sched_get_priority_max       // 1165
 741         data8 0                         // sched_get_priority_min
 742         data8 0                         // sched_rr_get_interval
 743         data8 0                         // nanosleep
 744         data8 0                         // nfsservctl
 745         data8 0                         // prctl                // 1170
 746         data8 0                         // getpagesize
 747         data8 0                         // mmap2
 748         data8 0                         // pciconfig_read
 749         data8 0                         // pciconfig_write
 750         data8 0                         // perfmonctl           // 1175
 751         data8 0                         // sigaltstack
 752         data8 0                         // rt_sigaction
 753         data8 0                         // rt_sigpending
 754         data8 0                         // rt_sigprocmask
 755         data8 0                         // rt_sigqueueinfo      // 1180
 756         data8 0                         // rt_sigreturn
 757         data8 0                         // rt_sigsuspend
 758         data8 0                         // rt_sigtimedwait
 759         data8 0                         // getcwd
 760         data8 0                         // capget               // 1185
 761         data8 0                         // capset
 762         data8 0                         // sendfile
 763         data8 0
 764         data8 0
 765         data8 0                         // socket               // 1190
 766         data8 0                         // bind
 767         data8 0                         // connect
 768         data8 0                         // listen
 769         data8 0                         // accept
 770         data8 0                         // getsockname          // 1195
 771         data8 0                         // getpeername
 772         data8 0                         // socketpair
 773         data8 0                         // send
 774         data8 0                         // sendto
 775         data8 0                         // recv                 // 1200
 776         data8 0                         // recvfrom
 777         data8 0                         // shutdown
 778         data8 0                         // setsockopt
 779         data8 0                         // getsockopt
 780         data8 0                         // sendmsg              // 1205
 781         data8 0                         // recvmsg
 782         data8 0                         // pivot_root
 783         data8 0                         // mincore
 784         data8 0                         // madvise
 785         data8 0                         // newstat              // 1210
 786         data8 0                         // newlstat
 787         data8 0                         // newfstat
 788         data8 0                         // clone2
 789         data8 0                         // getdents64
 790         data8 0                         // getunwind            // 1215
 791         data8 0                         // readahead
 792         data8 0                         // setxattr
 793         data8 0                         // lsetxattr
 794         data8 0                         // fsetxattr
 795         data8 0                         // getxattr             // 1220
 796         data8 0                         // lgetxattr
 797         data8 0                         // fgetxattr
 798         data8 0                         // listxattr
 799         data8 0                         // llistxattr
 800         data8 0                         // flistxattr           // 1225
 801         data8 0                         // removexattr
 802         data8 0                         // lremovexattr
 803         data8 0                         // fremovexattr
 804         data8 0                         // tkill
 805         data8 0                         // futex                // 1230
 806         data8 0                         // sched_setaffinity
 807         data8 0                         // sched_getaffinity
 808         data8 fsys_set_tid_address      // set_tid_address
 809         data8 0                         // fadvise64_64
 810         data8 0                         // tgkill               // 1235
 811         data8 0                         // exit_group
 812         data8 0                         // lookup_dcookie
 813         data8 0                         // io_setup
 814         data8 0                         // io_destroy
 815         data8 0                         // io_getevents         // 1240
 816         data8 0                         // io_submit
 817         data8 0                         // io_cancel
 818         data8 0                         // epoll_create
 819         data8 0                         // epoll_ctl
 820         data8 0                         // epoll_wait           // 1245
 821         data8 0                         // restart_syscall
 822         data8 0                         // semtimedop
 823         data8 0                         // timer_create
 824         data8 0                         // timer_settime
 825         data8 0                         // timer_gettime        // 1250
 826         data8 0                         // timer_getoverrun
 827         data8 0                         // timer_delete
 828         data8 0                         // clock_settime
 829         data8 fsys_clock_gettime        // clock_gettime
 830         data8 0                         // clock_getres         // 1255
 831         data8 0                         // clock_nanosleep
 832         data8 0                         // fstatfs64
 833         data8 0                         // statfs64
 834         data8 0                         // mbind
 835         data8 0                         // get_mempolicy        // 1260
 836         data8 0                         // set_mempolicy
 837         data8 0                         // mq_open
 838         data8 0                         // mq_unlink
 839         data8 0                         // mq_timedsend
 840         data8 0                         // mq_timedreceive      // 1265
 841         data8 0                         // mq_notify
 842         data8 0                         // mq_getsetattr
 843         data8 0                         // kexec_load
 844         data8 0                         // vserver
 845         data8 0                         // waitid               // 1270
 846         data8 0                         // add_key
 847         data8 0                         // request_key
 848         data8 0                         // keyctl
 849         data8 0                         // ioprio_set
 850         data8 0                         // ioprio_get           // 1275
 851         data8 0                         // move_pages
 852         data8 0                         // inotify_init
 853         data8 0                         // inotify_add_watch
 854         data8 0                         // inotify_rm_watch
 855         data8 0                         // migrate_pages        // 1280
 856         data8 0                         // openat
 857         data8 0                         // mkdirat
 858         data8 0                         // mknodat
 859         data8 0                         // fchownat
 860         data8 0                         // futimesat            // 1285
 861         data8 0                         // newfstatat
 862         data8 0                         // unlinkat
 863         data8 0                         // renameat
 864         data8 0                         // linkat
 865         data8 0                         // symlinkat            // 1290
 866         data8 0                         // readlinkat
 867         data8 0                         // fchmodat
 868         data8 0                         // faccessat
 869         data8 0
 870         data8 0                                                 // 1295
 871         data8 0                         // unshare
 872         data8 0                         // splice
 873         data8 0                         // set_robust_list
 874         data8 0                         // get_robust_list
 875         data8 0                         // sync_file_range      // 1300
 876         data8 0                         // tee
 877         data8 0                         // vmsplice
 878         data8 0
 879         data8 fsys_getcpu               // getcpu               // 1304
 880
 881         // fill in zeros for the remaining entries
 882         .zero:
 883         .space paravirt_fsyscall_table + 8*NR_syscalls - .zero, 0