/* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright 2006 Sun Microsystems, Inc. All rights reserved. * Use is subject to license terms. */ #pragma ident "%Z%%M% %I% %E% SMI" #if !defined(lint) #include "assym.h" #endif /* !lint */ /* * General assembly language routines. * It is the intent of this file to contain routines that are * specific to cpu architecture. */ /* * WARNING: If you add a fast trap handler which can be invoked by a * non-privileged user, you may have to use the FAST_TRAP_DONE macro * instead of "done" instruction to return back to the user mode. See * comments for the "fast_trap_done" entry point for more information. */ #define FAST_TRAP_DONE \ ba,a fast_trap_done /* * Override GET_NATIVE_TIME for the cpu module code. This is not * guaranteed to be exactly one instruction, be careful of using * the macro in delay slots. * * Do not use any instruction that modifies condition codes as the * caller may depend on these to remain unchanged across the macro. */ #if defined(CHEETAH) || defined(OLYMPUS_C) #define GET_NATIVE_TIME(out, scr1, scr2) \ rd STICK, out #define DELTA_NATIVE_TIME(delta, reg, scr1, scr2, scr3) \ rd STICK, reg; \ add reg, delta, reg; \ wr reg, STICK #define RD_TICKCMPR(out, scr) \ rd STICK_COMPARE, out #define WR_TICKCMPR(in, scr1, scr2, label) \ wr in, STICK_COMPARE #elif defined(HUMMINGBIRD) #include /* * the current hummingbird version of %stick and %stick_cmp * were both implemented as (2) 32-bit locations in ASI_IO space; * the hdwr should support atomic r/w; meanwhile: ugly alert! ... * * 64-bit opcodes are required, but move only 32-bits: * * ldxa [phys]ASI_IO, %dst reads the low 32-bits from phys into %dst * stxa %src, [phys]ASI_IO writes the low 32-bits from %src into phys * * reg equivalent [phys]ASI_IO * ------------------ --------------- * %stick_cmp low-32 0x1FE.0000.F060 * %stick_cmp high-32 0x1FE.0000.F068 * %stick low-32 0x1FE.0000.F070 * %stick high-32 0x1FE.0000.F078 */ #define HSTC_LOW 0x60 /* stick_cmp low 32-bits */ #define HSTC_HIGH 0x68 /* stick_cmp high 32-bits */ #define HST_LOW 0x70 /* stick low 32-bits */ #define HST_HIGH 0x78 /* stick high 32-bits */ #define HST_DIFF 0x08 /* low<-->high diff */ /* * Any change in the number of instructions in SETL41() * will affect SETL41_OFF */ #define SETL41(reg, byte) \ sethi %hi(0x1FE00000), reg; /* 0000.0000.1FE0.0000 */ \ or reg, 0xF, reg; /* 0000.0000.1FE0.000F */ \ sllx reg, 12, reg; /* 0000.01FE.0000.F000 */ \ or reg, byte, reg; /* 0000.01FE.0000.F0xx */ /* * SETL41_OFF is used to calulate the relative PC value when a * branch instruction needs to go over SETL41() macro */ #define SETL41_OFF 16 /* * reading stick requires 2 loads, and there could be an intervening * low-to-high 32-bit rollover resulting in a return value that is * off by about (2 ^ 32); this rare case is prevented by re-reading * the low-32 bits after the high-32 and verifying the "after" value * is >= the "before" value; if not, increment the high-32 value. * * this method is limited to 1 rollover, and based on the fixed * stick-frequency (5555555), requires the loads to complete within * 773 seconds; incrementing the high-32 value will not overflow for * about 52644 years. * * writing stick requires 2 stores; if the old/new low-32 value is * near 0xffffffff, there could be another rollover (also rare). * to prevent this, we first write a 0 to the low-32, then write * new values to the high-32 then the low-32. * * When we detect a carry in the lower %stick register, we need to * read HST_HIGH again. However at the point where we detect this, * we need to rebuild the register address HST_HIGH.This involves more * than one instructions and a branch is unavoidable. However, most of * the time, there is no carry. So we take the penalty of a branch * instruction only when there is carry (less frequent). * * For GET_NATIVE_TIME(), we start afresh and branch to SETL41(). * For DELTA_NATIVE_TIME(), we branch to just after SETL41() since * addr already points to HST_LOW. * * NOTE: this method requires disabling interrupts before using * DELTA_NATIVE_TIME. */ #define GET_NATIVE_TIME(out, scr, tmp) \ SETL41(scr, HST_LOW); \ ldxa [scr]ASI_IO, tmp; \ inc HST_DIFF, scr; \ ldxa [scr]ASI_IO, out; \ dec HST_DIFF, scr; \ ldxa [scr]ASI_IO, scr; \ sub scr, tmp, tmp; \ brlz,pn tmp, .-(SETL41_OFF+24); \ sllx out, 32, out; \ or out, scr, out #define DELTA_NATIVE_TIME(delta, addr, high, low, tmp) \ SETL41(addr, HST_LOW); \ ldxa [addr]ASI_IO, tmp; \ inc HST_DIFF, addr; \ ldxa [addr]ASI_IO, high; \ dec HST_DIFF, addr; \ ldxa [addr]ASI_IO, low; \ sub low, tmp, tmp; \ brlz,pn tmp, .-24; \ sllx high, 32, high; \ or high, low, high; \ add high, delta, high; \ srl high, 0, low; \ srlx high, 32, high; \ stxa %g0, [addr]ASI_IO; \ inc HST_DIFF, addr; \ stxa high, [addr]ASI_IO; \ dec HST_DIFF, addr; \ stxa low, [addr]ASI_IO #define RD_TICKCMPR(out, scr) \ SETL41(scr, HSTC_LOW); \ ldxa [scr]ASI_IO, out; \ inc HST_DIFF, scr; \ ldxa [scr]ASI_IO, scr; \ sllx scr, 32, scr; \ or scr, out, out #define WR_TICKCMPR(in, scra, scrd, label) \ SETL41(scra, HSTC_HIGH); \ srlx in, 32, scrd; \ stxa scrd, [scra]ASI_IO; \ dec HST_DIFF, scra; \ stxa in, [scra]ASI_IO #else /* !CHEETAH && !HUMMINGBIRD */ #define GET_NATIVE_TIME(out, scr1, scr2) \ rdpr %tick, out #define DELTA_NATIVE_TIME(delta, reg, scr1, scr2, scr3) \ rdpr %tick, reg; \ add reg, delta, reg; \ wrpr reg, %tick #define RD_TICKCMPR(out, scr) \ rd TICK_COMPARE, out #ifdef BB_ERRATA_1 /* writes to TICK_COMPARE may fail */ /* * Writes to the TICK_COMPARE register sometimes fail on blackbird modules. * The failure occurs only when the following instruction decodes to wr or * wrpr. The workaround is to immediately follow writes to TICK_COMPARE * with a read, thus stalling the pipe and keeping following instructions * from causing data corruption. Aligning to a quadword will ensure these * two instructions are not split due to i$ misses. */ #define WR_TICKCMPR(cmpr,scr1,scr2,label) \ ba,a .bb_errata_1.label ;\ .align 64 ;\ .bb_errata_1.label: ;\ wr cmpr, TICK_COMPARE ;\ rd TICK_COMPARE, %g0 #else /* BB_ERRATA_1 */ #define WR_TICKCMPR(in,scr1,scr2,label) \ wr in, TICK_COMPARE #endif /* BB_ERRATA_1 */ #endif /* !CHEETAH && !HUMMINGBIRD */ #include #if defined(lint) #include #include #include #include #include #include #endif /* lint */ #include #include #include /* To get SYSBASE and PAGESIZE */ #include #include #include #include #include #include #include #if !defined(lint) #include "assym.h" #endif /* !lint */ #if defined(lint) uint_t get_impl(void) { return (0); } #else /* lint */ ENTRY(get_impl) GET_CPU_IMPL(%o0) retl nop SET_SIZE(get_impl) #endif /* lint */ #if defined(lint) /* * Softint generated when counter field of tick reg matches value field * of tick_cmpr reg */ /*ARGSUSED*/ void tickcmpr_set(uint64_t clock_cycles) {} #else /* lint */ ENTRY_NP(tickcmpr_set) ! get 64-bit clock_cycles interval mov %o0, %o2 mov 8, %o3 ! A reasonable initial step size 1: WR_TICKCMPR(%o2,%o4,%o5,__LINE__) ! Write to TICK_CMPR GET_NATIVE_TIME(%o0, %o4, %o5) ! Read %tick to confirm the sllx %o0, 1, %o0 ! value we wrote was in the future. srlx %o0, 1, %o0 cmp %o2, %o0 ! If the value we wrote was in the bg,pt %xcc, 2f ! future, then blow out of here. sllx %o3, 1, %o3 ! If not, then double our step size, ba,pt %xcc, 1b ! and take another lap. add %o0, %o3, %o2 ! 2: retl nop SET_SIZE(tickcmpr_set) #endif /* lint */ #if defined(lint) void tickcmpr_disable(void) {} #else /* lint */ ENTRY_NP(tickcmpr_disable) mov 1, %g1 sllx %g1, TICKINT_DIS_SHFT, %o0 WR_TICKCMPR(%o0,%o4,%o5,__LINE__) ! Write to TICK_CMPR retl nop SET_SIZE(tickcmpr_disable) #endif /* lint */ #if defined(lint) /* * tick_write_delta() increments %tick by the specified delta. This should * only be called after a CPR event to assure that gethrtime() continues to * increase monotonically. Obviously, writing %tick needs to de done very * carefully to avoid introducing unnecessary %tick skew across CPUs. For * this reason, we make sure we're i-cache hot before actually writing to * %tick. */ /*ARGSUSED*/ void tick_write_delta(uint64_t delta) {} #else /* lint */ #ifdef DEBUG .seg ".text" tick_write_panic: .asciz "tick_write_delta: interrupts already disabled on entry" #endif /* DEBUG */ ENTRY_NP(tick_write_delta) rdpr %pstate, %g1 #ifdef DEBUG andcc %g1, PSTATE_IE, %g0 ! If DEBUG, check that interrupts bnz 0f ! aren't already disabled. sethi %hi(tick_write_panic), %o1 save %sp, -SA(MINFRAME), %sp ! get a new window to preserve caller call panic or %i1, %lo(tick_write_panic), %o0 #endif /* DEBUG */ 0: wrpr %g1, PSTATE_IE, %pstate ! Disable interrupts mov %o0, %o2 ba 0f ! Branch to cache line-aligned instr. nop .align 16 0: nop ! The next 3 instructions are now hot. DELTA_NATIVE_TIME(%o2, %o3, %o4, %o5, %g2) ! read/inc/write %tick retl ! Return wrpr %g0, %g1, %pstate ! delay: Re-enable interrupts #endif /* lint */ #if defined(lint) /* * return 1 if disabled */ int tickcmpr_disabled(void) { return (0); } #else /* lint */ ENTRY_NP(tickcmpr_disabled) RD_TICKCMPR(%g1, %o0) retl srlx %g1, TICKINT_DIS_SHFT, %o0 SET_SIZE(tickcmpr_disabled) #endif /* lint */ /* * Get current tick */ #if defined(lint) u_longlong_t gettick(void) { return (0); } #else /* lint */ ENTRY(gettick) GET_NATIVE_TIME(%o0, %o2, %o3) retl nop SET_SIZE(gettick) #endif /* lint */ /* * Return the counter portion of the tick register. */ #if defined(lint) uint64_t gettick_counter(void) { return(0); } #else /* lint */ ENTRY_NP(gettick_counter) rdpr %tick, %o0 sllx %o0, 1, %o0 retl srlx %o0, 1, %o0 ! shake off npt bit SET_SIZE(gettick_counter) #endif /* lint */ /* * Provide a C callable interface to the trap that reads the hi-res timer. * Returns 64-bit nanosecond timestamp in %o0 and %o1. */ #if defined(lint) hrtime_t gethrtime(void) { return ((hrtime_t)0); } hrtime_t gethrtime_unscaled(void) { return ((hrtime_t)0); } hrtime_t gethrtime_max(void) { return ((hrtime_t)0); } void scalehrtime(hrtime_t *hrt) { *hrt = 0; } void gethrestime(timespec_t *tp) { tp->tv_sec = 0; tp->tv_nsec = 0; } time_t gethrestime_sec(void) { return (0); } void gethrestime_lasttick(timespec_t *tp) { tp->tv_sec = 0; tp->tv_nsec = 0; } /*ARGSUSED*/ void hres_tick(void) { } void panic_hres_tick(void) { } #else /* lint */ ENTRY_NP(gethrtime) GET_HRTIME(%g1, %o0, %o1, %o2, %o3, %o4, %o5, %g2) ! %g1 = hrtime retl mov %g1, %o0 SET_SIZE(gethrtime) ENTRY_NP(gethrtime_unscaled) GET_NATIVE_TIME(%g1, %o2, %o3) ! %g1 = native time retl mov %g1, %o0 SET_SIZE(gethrtime_unscaled) ENTRY_NP(gethrtime_waitfree) ALTENTRY(dtrace_gethrtime) GET_NATIVE_TIME(%g1, %o2, %o3) ! %g1 = native time NATIVE_TIME_TO_NSEC(%g1, %o2, %o3) retl mov %g1, %o0 SET_SIZE(dtrace_gethrtime) SET_SIZE(gethrtime_waitfree) ENTRY(gethrtime_max) NATIVE_TIME_MAX(%g1) NATIVE_TIME_TO_NSEC(%g1, %o0, %o1) ! hrtime_t's are signed, max hrtime_t must be positive mov -1, %o2 brlz,a %g1, 1f srlx %o2, 1, %g1 1: retl mov %g1, %o0 SET_SIZE(gethrtime_max) ENTRY(scalehrtime) ldx [%o0], %o1 NATIVE_TIME_TO_NSEC(%o1, %o2, %o3) retl stx %o1, [%o0] SET_SIZE(scalehrtime) /* * Fast trap to return a timestamp, uses trap window, leaves traps * disabled. Returns a 64-bit nanosecond timestamp in %o0 and %o1. * * This is the handler for the ST_GETHRTIME trap. */ ENTRY_NP(get_timestamp) GET_HRTIME(%g1, %g2, %g3, %g4, %g5, %o0, %o1, %o2) ! %g1 = hrtime srlx %g1, 32, %o0 ! %o0 = hi32(%g1) srl %g1, 0, %o1 ! %o1 = lo32(%g1) FAST_TRAP_DONE SET_SIZE(get_timestamp) /* * Macro to convert GET_HRESTIME() bits into a timestamp. * * We use two separate macros so that the platform-dependent GET_HRESTIME() * can be as small as possible; CONV_HRESTIME() implements the generic part. */ #define CONV_HRESTIME(hrestsec, hrestnsec, adj, nslt, nano) \ brz,pt adj, 3f; /* no adjustments, it's easy */ \ add hrestnsec, nslt, hrestnsec; /* hrest.tv_nsec += nslt */ \ brlz,pn adj, 2f; /* if hrestime_adj negative */ \ srlx nslt, ADJ_SHIFT, nslt; /* delay: nslt >>= 4 */ \ subcc adj, nslt, %g0; /* hrestime_adj - nslt/16 */ \ movg %xcc, nslt, adj; /* adj by min(adj, nslt/16) */ \ ba 3f; /* go convert to sec/nsec */ \ add hrestnsec, adj, hrestnsec; /* delay: apply adjustment */ \ 2: addcc adj, nslt, %g0; /* hrestime_adj + nslt/16 */ \ bge,a,pt %xcc, 3f; /* is adj less negative? */ \ add hrestnsec, adj, hrestnsec; /* yes: hrest.nsec += adj */ \ sub hrestnsec, nslt, hrestnsec; /* no: hrest.nsec -= nslt/16 */ \ 3: cmp hrestnsec, nano; /* more than a billion? */ \ bl,pt %xcc, 4f; /* if not, we're done */ \ nop; /* delay: do nothing :( */ \ add hrestsec, 1, hrestsec; /* hrest.tv_sec++; */ \ sub hrestnsec, nano, hrestnsec; /* hrest.tv_nsec -= NANOSEC; */ \ ba,a 3b; /* check >= billion again */ \ 4: ENTRY_NP(gethrestime) GET_HRESTIME(%o1, %o2, %o3, %o4, %o5, %g1, %g2, %g3, %g4) CONV_HRESTIME(%o1, %o2, %o3, %o4, %o5) stn %o1, [%o0] retl stn %o2, [%o0 + CLONGSIZE] SET_SIZE(gethrestime) /* * Similar to gethrestime(), but gethrestime_sec() returns current hrestime * seconds. */ ENTRY_NP(gethrestime_sec) GET_HRESTIME(%o0, %o2, %o3, %o4, %o5, %g1, %g2, %g3, %g4) CONV_HRESTIME(%o0, %o2, %o3, %o4, %o5) retl ! %o0 current hrestime seconds nop SET_SIZE(gethrestime_sec) /* * Returns the hrestime on the last tick. This is simpler than gethrestime() * and gethrestime_sec(): no conversion is required. gethrestime_lasttick() * follows the same locking algorithm as GET_HRESTIME and GET_HRTIME, * outlined in detail in clock.h. (Unlike GET_HRESTIME/GET_HRTIME, we don't * rely on load dependencies to effect the membar #LoadLoad, instead declaring * it explicitly.) */ ENTRY_NP(gethrestime_lasttick) sethi %hi(hres_lock), %o1 0: lduw [%o1 + %lo(hres_lock)], %o2 ! Load lock value membar #LoadLoad ! Load of lock must complete andn %o2, 1, %o2 ! Mask off lowest bit ldn [%o1 + %lo(hrestime)], %g1 ! Seconds. add %o1, %lo(hrestime), %o4 ldn [%o4 + CLONGSIZE], %g2 ! Nanoseconds. membar #LoadLoad ! All loads must complete lduw [%o1 + %lo(hres_lock)], %o3 ! Reload lock value cmp %o3, %o2 ! If lock is locked or has bne 0b ! changed, retry. stn %g1, [%o0] ! Delay: store seconds retl stn %g2, [%o0 + CLONGSIZE] ! Delay: store nanoseconds SET_SIZE(gethrestime_lasttick) /* * Fast trap for gettimeofday(). Returns a timestruc_t in %o0 and %o1. * * This is the handler for the ST_GETHRESTIME trap. */ ENTRY_NP(get_hrestime) GET_HRESTIME(%o0, %o1, %g1, %g2, %g3, %g4, %g5, %o2, %o3) CONV_HRESTIME(%o0, %o1, %g1, %g2, %g3) FAST_TRAP_DONE SET_SIZE(get_hrestime) /* * Fast trap to return lwp virtual time, uses trap window, leaves traps * disabled. Returns a 64-bit number in %o0:%o1, which is the number * of nanoseconds consumed. * * This is the handler for the ST_GETHRVTIME trap. * * Register usage: * %o0, %o1 = return lwp virtual time * %o2 = CPU/thread * %o3 = lwp * %g1 = scratch * %g5 = scratch */ ENTRY_NP(get_virtime) GET_NATIVE_TIME(%g5, %g1, %g2) ! %g5 = native time in ticks CPU_ADDR(%g2, %g3) ! CPU struct ptr to %g2 ldn [%g2 + CPU_THREAD], %g2 ! thread pointer to %g2 ldn [%g2 + T_LWP], %g3 ! lwp pointer to %g3 /* * Subtract start time of current microstate from time * of day to get increment for lwp virtual time. */ ldx [%g3 + LWP_STATE_START], %g1 ! ms_state_start sub %g5, %g1, %g5 /* * Add current value of ms_acct[LMS_USER] */ ldx [%g3 + LWP_ACCT_USER], %g1 ! ms_acct[LMS_USER] add %g5, %g1, %g5 NATIVE_TIME_TO_NSEC(%g5, %g1, %o0) srl %g5, 0, %o1 ! %o1 = lo32(%g5) srlx %g5, 32, %o0 ! %o0 = hi32(%g5) FAST_TRAP_DONE SET_SIZE(get_virtime) .seg ".text" hrtime_base_panic: .asciz "hrtime_base stepping back" ENTRY_NP(hres_tick) save %sp, -SA(MINFRAME), %sp ! get a new window sethi %hi(hrestime), %l4 ldstub [%l4 + %lo(hres_lock + HRES_LOCK_OFFSET)], %l5 ! try locking 7: tst %l5 bz,pt %xcc, 8f ! if we got it, drive on ld [%l4 + %lo(nsec_scale)], %l5 ! delay: %l5 = scaling factor ldub [%l4 + %lo(hres_lock + HRES_LOCK_OFFSET)], %l5 9: tst %l5 bz,a,pn %xcc, 7b ldstub [%l4 + %lo(hres_lock + HRES_LOCK_OFFSET)], %l5 ba,pt %xcc, 9b ldub [%l4 + %lo(hres_lock + HRES_LOCK_OFFSET)], %l5 8: membar #StoreLoad|#StoreStore ! ! update hres_last_tick. %l5 has the scaling factor (nsec_scale). ! ldx [%l4 + %lo(hrtime_base)], %g1 ! load current hrtime_base GET_NATIVE_TIME(%l0, %l3, %l6) ! current native time stx %l0, [%l4 + %lo(hres_last_tick)]! prev = current ! convert native time to nsecs NATIVE_TIME_TO_NSEC_SCALE(%l0, %l5, %l2, NSEC_SHIFT) sub %l0, %g1, %i1 ! get accurate nsec delta ldx [%l4 + %lo(hrtime_base)], %l1 cmp %l1, %l0 bg,pn %xcc, 9f nop stx %l0, [%l4 + %lo(hrtime_base)] ! update hrtime_base ! ! apply adjustment, if any ! ldx [%l4 + %lo(hrestime_adj)], %l0 ! %l0 = hrestime_adj brz %l0, 2f ! hrestime_adj == 0 ? ! yes, skip adjustments clr %l5 ! delay: set adj to zero tst %l0 ! is hrestime_adj >= 0 ? bge,pt %xcc, 1f ! yes, go handle positive case srl %i1, ADJ_SHIFT, %l5 ! delay: %l5 = adj addcc %l0, %l5, %g0 ! hrestime_adj < -adj ? bl,pt %xcc, 2f ! yes, use current adj neg %l5 ! delay: %l5 = -adj ba,pt %xcc, 2f mov %l0, %l5 ! no, so set adj = hrestime_adj 1: subcc %l0, %l5, %g0 ! hrestime_adj < adj ? bl,a,pt %xcc, 2f ! yes, set adj = hrestime_adj mov %l0, %l5 ! delay: adj = hrestime_adj 2: ldx [%l4 + %lo(timedelta)], %l0 ! %l0 = timedelta sub %l0, %l5, %l0 ! timedelta -= adj stx %l0, [%l4 + %lo(timedelta)] ! store new timedelta stx %l0, [%l4 + %lo(hrestime_adj)] ! hrestime_adj = timedelta or %l4, %lo(hrestime), %l2 ldn [%l2], %i2 ! %i2:%i3 = hrestime sec:nsec ldn [%l2 + CLONGSIZE], %i3 add %i3, %l5, %i3 ! hrestime.nsec += adj add %i3, %i1, %i3 ! hrestime.nsec += nslt set NANOSEC, %l5 ! %l5 = NANOSEC cmp %i3, %l5 bl,pt %xcc, 5f ! if hrestime.tv_nsec < NANOSEC sethi %hi(one_sec), %i1 ! delay add %i2, 0x1, %i2 ! hrestime.tv_sec++ sub %i3, %l5, %i3 ! hrestime.tv_nsec - NANOSEC mov 0x1, %l5 st %l5, [%i1 + %lo(one_sec)] 5: stn %i2, [%l2] stn %i3, [%l2 + CLONGSIZE] ! store the new hrestime membar #StoreStore ld [%l4 + %lo(hres_lock)], %i1 inc %i1 ! release lock st %i1, [%l4 + %lo(hres_lock)] ! clear hres_lock ret restore 9: ! ! release hres_lock ! ld [%l4 + %lo(hres_lock)], %i1 inc %i1 st %i1, [%l4 + %lo(hres_lock)] sethi %hi(hrtime_base_panic), %o0 call panic or %o0, %lo(hrtime_base_panic), %o0 SET_SIZE(hres_tick) #endif /* lint */ #if !defined(lint) && !defined(__lint) .seg ".text" kstat_q_panic_msg: .asciz "kstat_q_exit: qlen == 0" ENTRY(kstat_q_panic) save %sp, -SA(MINFRAME), %sp sethi %hi(kstat_q_panic_msg), %o0 call panic or %o0, %lo(kstat_q_panic_msg), %o0 /*NOTREACHED*/ SET_SIZE(kstat_q_panic) #define BRZPN brz,pn #define BRZPT brz,pt #define KSTAT_Q_UPDATE(QOP, QBR, QZERO, QRETURN, QTYPE) \ ld [%o0 + QTYPE/**/CNT], %o1; /* %o1 = old qlen */ \ QOP %o1, 1, %o2; /* %o2 = new qlen */ \ QBR %o1, QZERO; /* done if qlen == 0 */ \ st %o2, [%o0 + QTYPE/**/CNT]; /* delay: save qlen */ \ ldx [%o0 + QTYPE/**/LASTUPDATE], %o3; \ ldx [%o0 + QTYPE/**/TIME], %o4; /* %o4 = old time */ \ ldx [%o0 + QTYPE/**/LENTIME], %o5; /* %o5 = old lentime */ \ sub %g1, %o3, %o2; /* %o2 = time delta */ \ mulx %o1, %o2, %o3; /* %o3 = cur lentime */ \ add %o4, %o2, %o4; /* %o4 = new time */ \ add %o5, %o3, %o5; /* %o5 = new lentime */ \ stx %o4, [%o0 + QTYPE/**/TIME]; /* save time */ \ stx %o5, [%o0 + QTYPE/**/LENTIME]; /* save lentime */ \ QRETURN; \ stx %g1, [%o0 + QTYPE/**/LASTUPDATE]; /* lastupdate = now */ .align 16 ENTRY(kstat_waitq_enter) GET_NATIVE_TIME(%g1, %g2, %g3) KSTAT_Q_UPDATE(add, BRZPT, 1f, 1:retl, KSTAT_IO_W) SET_SIZE(kstat_waitq_enter) .align 16 ENTRY(kstat_waitq_exit) GET_NATIVE_TIME(%g1, %g2, %g3) KSTAT_Q_UPDATE(sub, BRZPN, kstat_q_panic, retl, KSTAT_IO_W) SET_SIZE(kstat_waitq_exit) .align 16 ENTRY(kstat_runq_enter) GET_NATIVE_TIME(%g1, %g2, %g3) KSTAT_Q_UPDATE(add, BRZPT, 1f, 1:retl, KSTAT_IO_R) SET_SIZE(kstat_runq_enter) .align 16 ENTRY(kstat_runq_exit) GET_NATIVE_TIME(%g1, %g2, %g3) KSTAT_Q_UPDATE(sub, BRZPN, kstat_q_panic, retl, KSTAT_IO_R) SET_SIZE(kstat_runq_exit) .align 16 ENTRY(kstat_waitq_to_runq) GET_NATIVE_TIME(%g1, %g2, %g3) KSTAT_Q_UPDATE(sub, BRZPN, kstat_q_panic, 1:, KSTAT_IO_W) KSTAT_Q_UPDATE(add, BRZPT, 1f, 1:retl, KSTAT_IO_R) SET_SIZE(kstat_waitq_to_runq) .align 16 ENTRY(kstat_runq_back_to_waitq) GET_NATIVE_TIME(%g1, %g2, %g3) KSTAT_Q_UPDATE(sub, BRZPN, kstat_q_panic, 1:, KSTAT_IO_R) KSTAT_Q_UPDATE(add, BRZPT, 1f, 1:retl, KSTAT_IO_W) SET_SIZE(kstat_runq_back_to_waitq) #endif /* !(lint || __lint) */ #ifdef lint int64_t timedelta; hrtime_t hres_last_tick; timestruc_t hrestime; int64_t hrestime_adj; int hres_lock; uint_t nsec_scale; hrtime_t hrtime_base; int traptrace_use_stick; #else /* lint */ /* * -- WARNING -- * * The following variables MUST be together on a 128-byte boundary. * In addition to the primary performance motivation (having them all * on the same cache line(s)), code here and in the GET*TIME() macros * assumes that they all have the same high 22 address bits (so * there's only one sethi). */ .seg ".data" .global timedelta, hres_last_tick, hrestime, hrestime_adj .global hres_lock, nsec_scale, hrtime_base, traptrace_use_stick .global nsec_shift, adj_shift /* XXX - above comment claims 128-bytes is necessary */ .align 64 timedelta: .word 0, 0 /* int64_t */ hres_last_tick: .word 0, 0 /* hrtime_t */ hrestime: .nword 0, 0 /* 2 longs */ hrestime_adj: .word 0, 0 /* int64_t */ hres_lock: .word 0 nsec_scale: .word 0 hrtime_base: .word 0, 0 traptrace_use_stick: .word 0 nsec_shift: .word NSEC_SHIFT adj_shift: .word ADJ_SHIFT #endif /* lint */ /* * drv_usecwait(clock_t n) [DDI/DKI - section 9F] * usec_delay(int n) [compatibility - should go one day] * Delay by spinning. * * delay for n microseconds. numbers <= 0 delay 1 usec * * With UltraSPARC-III the combination of supporting mixed-speed CPUs * and variable clock rate for power management requires that we * use %stick to implement this routine. * * For OPL platforms that support the "sleep" instruction, we * conditionally (ifdef'ed) insert a "sleep" instruction in * the loop. Note that theoritically we should have move (duplicated) * the code down to spitfire/us3/opl specific asm files - but this * is alot of code duplication just to add one "sleep" instruction. * We chose less code duplication for this. */ #if defined(lint) /*ARGSUSED*/ void drv_usecwait(clock_t n) {} /*ARGSUSED*/ void usec_delay(int n) {} #else /* lint */ ENTRY(drv_usecwait) ALTENTRY(usec_delay) brlez,a,pn %o0, 0f mov 1, %o0 0: sethi %hi(sticks_per_usec), %o1 lduw [%o1 + %lo(sticks_per_usec)], %o1 mulx %o1, %o0, %o1 ! Scale usec to ticks inc %o1 ! We don't start on a tick edge GET_NATIVE_TIME(%o2, %o3, %o4) add %o1, %o2, %o1 1: #ifdef _OPL .word 0x81b01060 ! insert "sleep" instruction #endif /* _OPL */ ! use byte code for now cmp %o1, %o2 GET_NATIVE_TIME(%o2, %o3, %o4) bgeu,pt %xcc, 1b nop retl nop SET_SIZE(usec_delay) SET_SIZE(drv_usecwait) #endif /* lint */ #if defined(lint) /* ARGSUSED */ void pil14_interrupt(int level) {} #else /* lint */ /* * Level-14 interrupt prologue. */ ENTRY_NP(pil14_interrupt) CPU_ADDR(%g1, %g2) rdpr %pil, %g6 ! %g6 = interrupted PIL stn %g6, [%g1 + CPU_PROFILE_PIL] ! record interrupted PIL rdpr %tstate, %g6 rdpr %tpc, %g5 btst TSTATE_PRIV, %g6 ! trap from supervisor mode? bnz,a,pt %xcc, 1f stn %g5, [%g1 + CPU_PROFILE_PC] ! if so, record kernel PC stn %g5, [%g1 + CPU_PROFILE_UPC] ! if not, record user PC ba pil_interrupt_common ! must be large-disp branch stn %g0, [%g1 + CPU_PROFILE_PC] ! zero kernel PC 1: ba pil_interrupt_common ! must be large-disp branch stn %g0, [%g1 + CPU_PROFILE_UPC] ! zero user PC SET_SIZE(pil14_interrupt) ENTRY_NP(tick_rtt) ! ! Load TICK_COMPARE into %o5; if bit 63 is set, then TICK_COMPARE is ! disabled. If TICK_COMPARE is enabled, we know that we need to ! reenqueue the interrupt request structure. We'll then check TICKINT ! in SOFTINT; if it's set, then we know that we were in a TICK_COMPARE ! interrupt. In this case, TICK_COMPARE may have been rewritten ! recently; we'll compare %o5 to the current time to verify that it's ! in the future. ! ! Note that %o5 is live until after 1f. ! XXX - there is a subroutine call while %o5 is live! ! RD_TICKCMPR(%o5, %g1) srlx %o5, TICKINT_DIS_SHFT, %g1 brnz,pt %g1, 2f nop rdpr %pstate, %g5 andn %g5, PSTATE_IE, %g1 wrpr %g0, %g1, %pstate ! Disable vec interrupts sethi %hi(cbe_level14_inum), %o1 ld [%o1 + %lo(cbe_level14_inum)], %o1 call intr_enqueue_req ! preserves %o5 and %g5 mov PIL_14, %o0 ! Check SOFTINT for TICKINT/STICKINT rd SOFTINT, %o4 set (TICK_INT_MASK | STICK_INT_MASK), %o0 andcc %o4, %o0, %g0 bz,a,pn %icc, 2f wrpr %g0, %g5, %pstate ! Enable vec interrupts ! clear TICKINT/STICKINT wr %o0, CLEAR_SOFTINT ! ! Now that we've cleared TICKINT, we can reread %tick and confirm ! that the value we programmed is still in the future. If it isn't, ! we need to reprogram TICK_COMPARE to fire as soon as possible. ! GET_NATIVE_TIME(%o0, %g1, %g2) ! %o0 = tick sllx %o0, 1, %o0 ! Clear the DIS bit srlx %o0, 1, %o0 cmp %o5, %o0 ! In the future? bg,a,pt %xcc, 2f ! Yes, drive on. wrpr %g0, %g5, %pstate ! delay: enable vec intr ! ! If we're here, then we have programmed TICK_COMPARE with a %tick ! which is in the past; we'll now load an initial step size, and loop ! until we've managed to program TICK_COMPARE to fire in the future. ! mov 8, %o4 ! 8 = arbitrary inital step 1: add %o0, %o4, %o5 ! Add the step WR_TICKCMPR(%o5,%g1,%g2,__LINE__) ! Write to TICK_CMPR GET_NATIVE_TIME(%o0, %g1, %g2) ! %o0 = tick sllx %o0, 1, %o0 ! Clear the DIS bit srlx %o0, 1, %o0 cmp %o5, %o0 ! In the future? bg,a,pt %xcc, 2f ! Yes, drive on. wrpr %g0, %g5, %pstate ! delay: enable vec intr ba 1b ! No, try again. sllx %o4, 1, %o4 ! delay: double step size 2: ba current_thread_complete nop SET_SIZE(tick_rtt) #endif /* lint */ #if defined(lint) || defined(__lint) /* ARGSUSED */ uint64_t find_cpufrequency(volatile uchar_t *clock_ptr) { return (0); } #else /* lint */ #ifdef DEBUG .seg ".text" find_cpufreq_panic: .asciz "find_cpufrequency: interrupts already disabled on entry" #endif /* DEBUG */ ENTRY_NP(find_cpufrequency) rdpr %pstate, %g1 #ifdef DEBUG andcc %g1, PSTATE_IE, %g0 ! If DEBUG, check that interrupts bnz 0f ! are currently enabled sethi %hi(find_cpufreq_panic), %o1 call panic or %o1, %lo(find_cpufreq_panic), %o0 #endif /* DEBUG */ 0: wrpr %g1, PSTATE_IE, %pstate ! Disable interrupts 3: ldub [%o0], %o1 ! Read the number of seconds mov %o1, %o2 ! remember initial value in %o2 1: GET_NATIVE_TIME(%o3, %g4, %g5) cmp %o1, %o2 ! did the seconds register roll over? be,pt %icc, 1b ! branch back if unchanged ldub [%o0], %o2 ! delay: load the new seconds val brz,pn %o2, 3b ! if the minutes just rolled over, ! the last second could have been ! inaccurate; try again. mov %o2, %o4 ! delay: store init. val. in %o2 2: GET_NATIVE_TIME(%o5, %g4, %g5) cmp %o2, %o4 ! did the seconds register roll over? be,pt %icc, 2b ! branch back if unchanged ldub [%o0], %o4 ! delay: load the new seconds val brz,pn %o4, 0b ! if the minutes just rolled over, ! the last second could have been ! inaccurate; try again. wrpr %g0, %g1, %pstate ! delay: re-enable interrupts retl sub %o5, %o3, %o0 ! return the difference in ticks SET_SIZE(find_cpufrequency) #endif /* lint */ #if defined(lint) /* * Prefetch a page_t for write or read, this assumes a linear * scan of sequential page_t's. */ /*ARGSUSED*/ void prefetch_page_w(void *pp) {} /*ARGSUSED*/ void prefetch_page_r(void *pp) {} #else /* lint */ #if defined(CHEETAH) || defined(CHEETAH_PLUS) || defined(JALAPENO) || \ defined(SERRANO) ! ! On US-III, the prefetch instruction queue is 8 entries deep. ! Also, prefetches for write put data in the E$, which has ! lines of 512 bytes for an 8MB cache. Each E$ line is further ! subblocked into 64 byte chunks. ! ! Since prefetch can only bring in 64 bytes at a time (See Sparc ! v9 Architecture Manual pp.204) and a page_t is 128 bytes, ! then 2 prefetches are required in order to bring an entire ! page into the E$. ! ! Since the prefetch queue is 8 entries deep, we currently can ! only have 4 prefetches for page_t's outstanding. Thus, we ! prefetch n+4 ahead of where we are now: ! ! 4 * sizeof(page_t) -> 512 ! 4 * sizeof(page_t) +64 -> 576 ! ! Example ! ======= ! contiguous page array in memory... ! ! |AAA1|AAA2|BBB1|BBB2|CCC1|CCC2|DDD1|DDD2|XXX1|XXX2|YYY1|YYY2|... ! ^ ^ ^ ^ ^ ^ ! pp | pp+4*sizeof(page)+64 ! | ! pp+4*sizeof(page) ! ! Prefetch ! Queue ! +-------+<--- In this iteration, we're working with pp (AAA1), ! |Preftch| but we enqueue prefetch for addr = XXX1 ! | XXX1 | ! +-------+<--- this queue slot will be a prefetch instruction for ! |Preftch| for addr = pp + 4*sizeof(page_t) + 64 (or second ! | XXX2 | half of page XXX) ! +-------+ ! |Preftch|<-+- The next time around this function, we'll be ! | YYY1 | | working with pp = BBB1, but will be enqueueing ! +-------+ | prefetches to for both halves of page YYY, ! |Preftch| | while both halves of page XXX are in transit ! | YYY2 |<-+ make their way into the E$. ! +-------+ ! |Preftch| ! | ZZZ1 | ! +-------+ ! . . ! : : ! ! E$ ! +============================================... ! | XXX1 | XXX2 | YYY1 | YYY2 | ZZZ1 | ZZZ2 | ! +============================================... ! | | | | | | | ! +============================================... ! . ! : ! ! So we should expect the first four page accesses to stall ! while we warm up the cache, afterwhich, most of the pages ! will have their pp ready in the E$. ! ! Also note that if sizeof(page_t) grows beyond 128, then ! we'll need an additional prefetch to get an entire page ! into the E$, thus reducing the number of outstanding page ! prefetches to 2 (ie. 3 prefetches/page = 6 queue slots) ! etc. ! ! Cheetah+ ! ======== ! On Cheetah+ we use "#n_write" prefetches as these avoid ! unnecessary RTS->RTO bus transaction state change, and ! just issues RTO transaction. (See pp.77 of Cheetah+ Delta ! PRM). On Cheetah, #n_write prefetches are reflected with ! RTS->RTO state transition regardless. ! #define STRIDE1 512 #define STRIDE2 576 #if STRIDE1 != (PAGE_SIZE * 4) #error "STRIDE1 != (PAGE_SIZE * 4)" #endif /* STRIDE1 != (PAGE_SIZE * 4) */ ENTRY(prefetch_page_w) prefetch [%o0+STRIDE1], #n_writes retl prefetch [%o0+STRIDE2], #n_writes SET_SIZE(prefetch_page_w) ! ! Note on CHEETAH to prefetch for read, we really use #one_write. ! This fetches to E$ (general use) rather than P$ (floating point use). ! ENTRY(prefetch_page_r) prefetch [%o0+STRIDE1], #one_write retl prefetch [%o0+STRIDE2], #one_write SET_SIZE(prefetch_page_r) #elif defined(SPITFIRE) || defined(HUMMINGBIRD) ! ! UltraSparcII can have up to 3 prefetches outstanding. ! A page_t is 128 bytes (2 prefetches of 64 bytes each) ! So prefetch for pp + 1, which is ! ! pp + sizeof(page_t) ! and ! pp + sizeof(page_t) + 64 ! #define STRIDE1 128 #define STRIDE2 192 #if STRIDE1 != PAGE_SIZE #error "STRIDE1 != PAGE_SIZE" #endif /* STRIDE1 != PAGE_SIZE */ ENTRY(prefetch_page_w) prefetch [%o0+STRIDE1], #n_writes retl prefetch [%o0+STRIDE2], #n_writes SET_SIZE(prefetch_page_w) ENTRY(prefetch_page_r) prefetch [%o0+STRIDE1], #n_reads retl prefetch [%o0+STRIDE2], #n_reads SET_SIZE(prefetch_page_r) #elif defined(OLYMPUS_C) ! ! Prefetch strides for Olympus-C ! #define STRIDE1 0x440 #define STRIDE2 0x640 ENTRY(prefetch_page_w) prefetch [%o0+STRIDE1], #n_writes retl prefetch [%o0+STRIDE2], #n_writes SET_SIZE(prefetch_page_w) ENTRY(prefetch_page_r) prefetch [%o0+STRIDE1], #n_writes retl prefetch [%o0+STRIDE2], #n_writes SET_SIZE(prefetch_page_r) #else /* OLYMPUS_C */ #error "You need to fix this for your new cpu type." #endif /* OLYMPUS_C */ #endif /* lint */ #if defined(lint) /* * Prefetch struct smap for write. */ /*ARGSUSED*/ void prefetch_smap_w(void *smp) {} #else /* lint */ #if defined(CHEETAH) || defined(CHEETAH_PLUS) || defined(JALAPENO) || \ defined(SERRANO) #define PREFETCH_Q_LEN 8 #elif defined(SPITFIRE) || defined(HUMMINGBIRD) #define PREFETCH_Q_LEN 3 #elif defined(OLYMPUS_C) ! ! Use length of one for now. ! #define PREFETCH_Q_LEN 1 #else /* OLYMPUS_C */ #error You need to fix this for your new cpu type. #endif /* OLYMPUS_C */ #include #ifdef SEGKPM_SUPPORT #define SMAP_SIZE 72 #define SMAP_STRIDE (((PREFETCH_Q_LEN * 64) / SMAP_SIZE) * 64) #else /* SEGKPM_SUPPORT */ ! ! The hardware will prefetch the 64 byte cache aligned block ! that contains the address specified in the prefetch instruction. ! Since the size of the smap struct is 48 bytes, issuing 1 prefetch ! per pass will suffice as long as we prefetch far enough ahead to ! make sure we don't stall for the cases where the smap object ! spans multiple hardware prefetch blocks. Let's prefetch as far ! ahead as the hardware will allow. ! ! The smap array is processed with decreasing address pointers. ! #define SMAP_SIZE 48 #define SMAP_STRIDE (PREFETCH_Q_LEN * SMAP_SIZE) #endif /* SEGKPM_SUPPORT */ ENTRY(prefetch_smap_w) retl prefetch [%o0-SMAP_STRIDE], #n_writes SET_SIZE(prefetch_smap_w) #endif /* lint */ #if defined(lint) || defined(__lint) /* ARGSUSED */ uint64_t getidsr(void) { return 0; } #else /* lint */ ENTRY_NP(getidsr) retl ldxa [%g0]ASI_INTR_DISPATCH_STATUS, %o0 SET_SIZE(getidsr) #endif /* lint */