From: Tomas Hruby Date: Wed, 15 Sep 2010 14:11:25 +0000 (+0000) Subject: SMP - lazy FPU X-Git-Tag: v3.2.0~839 X-Git-Url: http://zhaoyanbai.com/repos/%22http:/www.isc.org/icons/zpipe.c?a=commitdiff_plain;h=5b8b62376589f683f89c4513db3cda3d21731b34;p=minix.git SMP - lazy FPU - when a process is migrated to a different CPU it may have an active FPU context in the processor registers. We must save it and migrate it together with the process. --- diff --git a/kernel/arch/i386/arch_smp.c b/kernel/arch/i386/arch_smp.c index 1d4a59bf8..de7f942de 100644 --- a/kernel/arch/i386/arch_smp.c +++ b/kernel/arch/i386/arch_smp.c @@ -60,6 +60,7 @@ PRIVATE phys_bytes copy_trampoline(void) tramp_size = (unsigned) &__trampoline_end - (unsigned)&trampoline; s = env_get("memory"); + s = (char *) get_value(params_buffer, "memory"); if (!s) return 0; @@ -238,6 +239,7 @@ PRIVATE void ap_finish_booting(void) printf("CPU %d paging is on\n", cpu); lapic_enable(cpu); + fpu_init(); if (app_cpu_init_timer(system_hz)) { panic("FATAL : failed to initialize timer interrupts CPU %d, " diff --git a/kernel/arch/i386/arch_system.c b/kernel/arch/i386/arch_system.c index 971a189c8..d9bc31cab 100644 --- a/kernel/arch/i386/arch_system.c +++ b/kernel/arch/i386/arch_system.c @@ -205,7 +205,7 @@ PUBLIC void arch_get_aout_headers(const int i, struct exec *h) phys_copy(aout + i * A_MINHDR, vir2phys(h), (phys_bytes) A_MINHDR); } -PRIVATE void fpu_init(void) +PUBLIC void fpu_init(void) { unsigned short cw, sw; @@ -219,10 +219,8 @@ PRIVATE void fpu_init(void) * Set CR0_NE and CR0_MP to handle fpu exceptions * in native mode. */ write_cr0(read_cr0() | CR0_MP_NE); - fpu_presence = 1; + get_cpulocal_var(fpu_presence) = 1; if(_cpufeature(_CPUF_I386_FXSR)) { - register struct proc *rp; - phys_bytes aligned_fp_area; u32_t cr4 = read_cr4() | CR4_OSFXSR; /* Enable FXSR. */ /* OSXMMEXCPT if supported @@ -233,35 +231,18 @@ PRIVATE void fpu_init(void) write_cr4(cr4); osfxsr_feature = 1; - - for (rp = BEG_PROC_ADDR; rp < END_PROC_ADDR; ++rp) { - /* FXSR requires 16-byte alignment of memory - * image, but unfortunately some old tools - * (probably linker) ignores ".balign 16" - * applied to our memory image. - * Thus we have to do manual alignment. - */ - aligned_fp_area = - (phys_bytes) &rp->p_fpu_state.fpu_image; - if(aligned_fp_area % FPUALIGN) { - aligned_fp_area += FPUALIGN - - (aligned_fp_area % FPUALIGN); - } - rp->p_fpu_state.fpu_save_area_p = - (void *) aligned_fp_area; - } } else { osfxsr_feature = 0; } } else { /* No FPU presents. */ - fpu_presence = 0; + get_cpulocal_var(fpu_presence) = 0; osfxsr_feature = 0; return; } } -PUBLIC void save_fpu(struct proc *pr) +PUBLIC void save_local_fpu(struct proc *pr) { if(!fpu_presence) return; @@ -275,6 +256,34 @@ PUBLIC void save_fpu(struct proc *pr) } } +PUBLIC void save_fpu(struct proc *pr) +{ +#if CONFIG_SMP + if (cpuid == pr->p_cpu) { + save_local_fpu(pr); + } + else { + int stopped; + + /* remember if the process was already stopped */ + stopped = RTS_ISSET(pr, RTS_PROC_STOP); + + /* stop the remote process and force it's context to be saved */ + smp_schedule_stop_proc_save_ctx(pr); + + /* + * If the process wasn't stopped let the process run again. The + * process is kept block by the fact that the kernel cannot run + * on its cpu + */ + if (!stopped) + RTS_UNSET(pr, RTS_PROC_STOP); + } +#else + save_local_fpu(pr); +#endif +} + PUBLIC void restore_fpu(struct proc *pr) { if(!proc_used_fpu(pr)) { @@ -328,8 +337,6 @@ PUBLIC void arch_init(void) BOOT_VERBOSE(printf("APIC not present, using legacy PIC\n")); } #endif - - fpu_init(); } PUBLIC void ser_putc(char c) diff --git a/kernel/cpulocals.h b/kernel/cpulocals.h index 3153d3675..6c0437cf3 100644 --- a/kernel/cpulocals.h +++ b/kernel/cpulocals.h @@ -78,6 +78,9 @@ DECLARE_CPULOCAL(int, cpu_is_idle); /* let the others know that you are idle */ DECLARE_CPULOCAL(u64_t ,tsc_ctr_switch); /* when did we switched time accounting */ +DECLARE_CPULOCAL(char ,fpu_presence); /* whether the cpu has FPU or not */ +DECLARE_CPULOCAL(struct proc * ,fpu_owner); /* who owns the FPU of the local cpu */ + DECLARE_CPULOCAL_END #endif /* __ASSEMBLY__ */ diff --git a/kernel/glo.h b/kernel/glo.h index 3c790acda..fcbd73a94 100644 --- a/kernel/glo.h +++ b/kernel/glo.h @@ -44,7 +44,6 @@ EXTERN time_t boottime; EXTERN char params_buffer[512]; /* boot monitor parameters */ EXTERN int minix_panicing; EXTERN char fpu_presence; -EXTERN struct proc * fpu_owner; EXTERN int verboseboot; /* verbose boot, init'ed in cstart */ #define MAGICTEST 0xC0FFEE23 EXTERN u32_t magictest; /* global magic number */ diff --git a/kernel/main.c b/kernel/main.c index 3c049c29a..d191ee9e0 100644 --- a/kernel/main.c +++ b/kernel/main.c @@ -32,7 +32,6 @@ /* Prototype declarations for PRIVATE functions. */ FORWARD _PROTOTYPE( void announce, (void)); -void ser_dump_queues(void); PUBLIC void bsp_finish_booting(void) { int i; @@ -69,6 +68,19 @@ PUBLIC void bsp_finish_booting(void) "cannot continue without any clock source!"); } + fpu_init(); + +#ifdef CONFIG_WATCHDOG + if (watchdog_enabled) { + if (arch_watchdog_init()) { + printf("WARNING watchdog initialization failed! Disabled\n"); + watchdog_enabled = 0; + } + else + BOOT_VERBOSE(printf("Watchdog enabled\n");); + } +#endif + /* Warnings for sanity checks that take time. These warnings are printed * so it's a clear warning no full release should be done with them * enabled. diff --git a/kernel/proc.c b/kernel/proc.c index 3728a3c55..f0289ea78 100644 --- a/kernel/proc.c +++ b/kernel/proc.c @@ -155,6 +155,23 @@ PUBLIC void proc_init(void) ip->p_rts_flags |= RTS_PROC_STOP; set_idle_name(ip->p_name, i); } + + for (rp = BEG_PROC_ADDR; rp < END_PROC_ADDR; ++rp) { + /* + * FXSR requires 16-byte alignment of memory image, but + * unfortunately a.out does not preserve the alignment while + * linking. Thus we have to do manual alignment. + */ + phys_bytes aligned_fp_area; + aligned_fp_area = + (phys_bytes) &rp->p_fpu_state.fpu_image; + if(aligned_fp_area % FPUALIGN) { + aligned_fp_area += FPUALIGN - + (aligned_fp_area % FPUALIGN); + } + rp->p_fpu_state.fpu_save_area_p = + (void *) aligned_fp_area; + } } PRIVATE void switch_address_space_idle(void) @@ -346,7 +363,7 @@ check_misc_flags: context_stop(proc_addr(KERNEL)); /* If the process isn't the owner of FPU, enable the FPU exception */ - if(fpu_owner != p) + if(get_cpulocal_var(fpu_owner) != p) enable_fpu_exception(); else disable_fpu_exception(); @@ -1573,6 +1590,7 @@ PUBLIC void proc_no_time(struct proc * p) PUBLIC void copr_not_available_handler(void) { struct proc * p; + struct proc ** local_fpu_owner; /* * Disable the FPU exception (both for the kernel and for the process * once it's scheduled), and initialize or restore the FPU state. @@ -1583,9 +1601,10 @@ PUBLIC void copr_not_available_handler(void) p = get_cpulocal_var(proc_ptr); /* if FPU is not owned by anyone, do not store anything */ - if (fpu_owner != NULL) { - assert(fpu_owner != p); - save_fpu(fpu_owner); + local_fpu_owner = get_cpulocal_var_ptr(fpu_owner); + if (*local_fpu_owner != NULL) { + assert(*local_fpu_owner != p); + save_local_fpu(*local_fpu_owner); } /* @@ -1593,12 +1612,17 @@ PUBLIC void copr_not_available_handler(void) * schedule! */ restore_fpu(p); - fpu_owner = p; + *local_fpu_owner = p; context_stop(proc_addr(KERNEL)); restore_user_context(p); NOT_REACHABLE; } -PUBLIC void release_fpu(void) { - fpu_owner = NULL; +PUBLIC void release_fpu(struct proc * p) { + struct proc ** fpu_owner_ptr; + + fpu_owner_ptr = get_cpu_var_ptr(p->p_cpu, fpu_owner); + + if (*fpu_owner_ptr == p) + *fpu_owner_ptr = NULL; } diff --git a/kernel/proto.h b/kernel/proto.h index aa2d4413d..76725ca7c 100644 --- a/kernel/proto.h +++ b/kernel/proto.h @@ -34,6 +34,7 @@ _PROTOTYPE( void context_stop, (struct proc * p) ); _PROTOTYPE( void context_stop_idle, (void) ); _PROTOTYPE( void restore_fpu, (struct proc *) ); _PROTOTYPE( void save_fpu, (struct proc *) ); +_PROTOTYPE( void save_local_fpu, (struct proc *) ); _PROTOTYPE( void fpu_sigcontext, (struct proc *, struct sigframe *fr, struct sigcontext *sc) ); /* main.c */ @@ -178,6 +179,10 @@ _PROTOTYPE( vir_bytes alloc_remote_segment, (u32_t *, segframe_t *, _PROTOTYPE( int intr_init, (int, int) ); _PROTOTYPE( void halt_cpu, (void) ); _PROTOTYPE( void arch_init, (void) ); +/* arch dependent FPU initialization per CPU */ +_PROTOTYPE( void fpu_init, (void) ); +/* returns true if pfu is present and initialized */ +_PROTOTYPE( int is_fpu, (void) ); _PROTOTYPE( void ser_putc, (char) ); _PROTOTYPE( __dead void arch_shutdown, (int) ); _PROTOTYPE( __dead void arch_monitor, (void) ); @@ -213,7 +218,7 @@ _PROTOTYPE(void release_address_space, (struct proc *pr)); _PROTOTYPE(void enable_fpu_exception, (void)); _PROTOTYPE(void disable_fpu_exception, (void)); -_PROTOTYPE(void release_fpu, (void)); +_PROTOTYPE(void release_fpu, (struct proc * p)); /* utility.c */ _PROTOTYPE( void cpu_print_freq, (unsigned cpu)); diff --git a/kernel/smp.c b/kernel/smp.c index 2dd730491..1fc322012 100644 --- a/kernel/smp.c +++ b/kernel/smp.c @@ -19,6 +19,7 @@ PRIVATE struct sched_ipi_data sched_ipi_data[CONFIG_MAX_CPUS]; #define SCHED_IPI_STOP_PROC 1 #define SCHED_IPI_VM_INHIBIT 2 +#define SCHED_IPI_SAVE_CTX 4 static volatile unsigned ap_cpus_booted; @@ -116,6 +117,30 @@ PUBLIC void smp_schedule_vminhibit(struct proc * p) assert(RTS_ISSET(p, RTS_VMINHIBIT)); } +PUBLIC void smp_schedule_stop_proc_save_ctx(struct proc * p) +{ + /* + * stop the processes and force the complete context of the process to + * be saved (i.e. including FPU state and such) + */ + smp_schedule_sync(p, SCHED_IPI_STOP_PROC | SCHED_IPI_SAVE_CTX); + assert(RTS_ISSET(p, RTS_PROC_STOP)); +} + +PUBLIC void smp_schedule_migrate_proc(struct proc * p, unsigned dest_cpu) +{ + /* + * stop the processes and force the complete context of the process to + * be saved (i.e. including FPU state and such) + */ + smp_schedule_sync(p, SCHED_IPI_STOP_PROC | SCHED_IPI_SAVE_CTX); + assert(RTS_ISSET(p, RTS_PROC_STOP)); + + /* assign the new cpu and let the process run again */ + p->p_cpu = dest_cpu; + RTS_UNSET(p, RTS_PROC_STOP); +} + PUBLIC void smp_ipi_sched_handler(void) { struct proc * curr; @@ -134,6 +159,16 @@ PUBLIC void smp_ipi_sched_handler(void) if (flgs & SCHED_IPI_STOP_PROC) { RTS_SET(p, RTS_PROC_STOP); } + if (flgs & SCHED_IPI_SAVE_CTX) { + /* all context have been save already, FPU remains */ + if (proc_used_fpu(p) && + get_cpulocal_var(fpu_owner) == p) { + disable_fpu_exception(); + save_local_fpu(p); + /* we re preparing to migrate somewhere else */ + release_fpu(p); + } + } if (flgs & SCHED_IPI_VM_INHIBIT) { RTS_SET(p, RTS_VMINHIBIT); } diff --git a/kernel/smp.h b/kernel/smp.h index 10855a679..ad23ce9ce 100644 --- a/kernel/smp.h +++ b/kernel/smp.h @@ -63,6 +63,11 @@ _PROTOTYPE(void smp_schedule, (unsigned cpu)); _PROTOTYPE(void smp_schedule_stop_proc, (struct proc * p)); /* stop a process on a different cpu because its adress space is being changed */ _PROTOTYPE(void smp_schedule_vminhibit, (struct proc * p)); +/* stop the process and for saving its full context */ +_PROTOTYPE(void smp_schedule_stop_proc_save_ctx, (struct proc * p)); +/* migrate the full context of a process to the destination CPU */ +_PROTOTYPE(void smp_schedule_migrate_proc, + (struct proc * p, unsigned dest_cpu)); _PROTOTYPE(void arch_send_smp_schedule_ipi, (unsigned cpu)); _PROTOTYPE(void arch_smp_halt_cpu, (void)); diff --git a/kernel/system.c b/kernel/system.c index cf0fe9ba0..434abf1b6 100644 --- a/kernel/system.c +++ b/kernel/system.c @@ -669,9 +669,7 @@ PUBLIC int sched_proc(struct proc *p, if (proc_is_runnable(p)) { #ifdef CONFIG_SMP if (p->p_cpu != cpuid && cpu != -1 && cpu != p->p_cpu) { - printf("WARNING : changing cpu of a runnable process %d " - "on a different cpu!\n", p->p_endpoint); - return(EINVAL); + smp_schedule_migrate_proc(p, cpu); } #endif diff --git a/kernel/system/do_clear.c b/kernel/system/do_clear.c index 056bbf243..8592c98ef 100644 --- a/kernel/system/do_clear.c +++ b/kernel/system/do_clear.c @@ -54,6 +54,9 @@ PUBLIC int do_clear(struct proc * caller, message * m_ptr) * and mark slot as FREE. Also mark saved fpu contents as not significant. */ RTS_SETFLAGS(rc, RTS_SLOT_FREE); + + /* release FPU */ + release_fpu(rc); rc->p_misc_flags &= ~MF_FPU_INITIALIZED; /* Release the process table slot. If this is a system process, also @@ -70,10 +73,6 @@ PUBLIC int do_clear(struct proc * caller, message * m_ptr) } #endif - /* release FPU */ - if (fpu_owner == rc) - release_fpu(); - return OK; } diff --git a/kernel/system/do_exec.c b/kernel/system/do_exec.c index 85aaf1a3b..762e02c1a 100644 --- a/kernel/system/do_exec.c +++ b/kernel/system/do_exec.c @@ -46,8 +46,7 @@ PUBLIC int do_exec(struct proc * caller, message * m_ptr) * will be initialized, when it's used next time. */ rp->p_misc_flags &= ~MF_FPU_INITIALIZED; /* force reloading FPU if the current process is the owner */ - if (rp == fpu_owner) - release_fpu(); + release_fpu(rp); return(OK); } #endif /* USE_EXEC */ diff --git a/kernel/system/do_fork.c b/kernel/system/do_fork.c index 49815e5ac..e73f75487 100644 --- a/kernel/system/do_fork.c +++ b/kernel/system/do_fork.c @@ -53,10 +53,7 @@ PUBLIC int do_fork(struct proc * caller, message * m_ptr) map_ptr= (struct mem_map *) m_ptr->PR_MEM_PTR; /* make sure that the FPU context is saved in parent before copy */ - if (fpu_owner == rpp) { - disable_fpu_exception(); - save_fpu(rpp); - } + save_fpu(rpp); /* Copy parent 'proc' struct to child. And reinitialize some fields. */ gen = _ENDPOINT_G(rpc->p_endpoint); #if (_MINIX_CHIP == _CHIP_INTEL) diff --git a/kernel/system/do_mcontext.c b/kernel/system/do_mcontext.c index 2b1e49650..ec93c86f0 100644 --- a/kernel/system/do_mcontext.c +++ b/kernel/system/do_mcontext.c @@ -43,10 +43,7 @@ PUBLIC int do_getmcontext(struct proc * caller, message * m_ptr) mc.mc_fpu_flags = 0; if (proc_used_fpu(rp)) { /* make sure that the FPU context is saved into proc structure first */ - if (fpu_owner == rp) { - disable_fpu_exception(); - save_fpu(rp); - } + save_fpu(rp); mc.mc_fpu_flags = 0 | rp->p_misc_flags & MF_FPU_INITIALIZED; memcpy(&(mc.mc_fpu_state), rp->p_fpu_state.fpu_save_area_p, FPU_XFP_SIZE); @@ -92,8 +89,7 @@ PUBLIC int do_setmcontext(struct proc * caller, message * m_ptr) } else rp->p_misc_flags &= ~MF_FPU_INITIALIZED; /* force reloading FPU in either case */ - if (fpu_owner == rp) - release_fpu(); + release_fpu(rp); #endif return(OK); diff --git a/kernel/system/do_sigreturn.c b/kernel/system/do_sigreturn.c index 7f91a0a34..592c8e2a2 100644 --- a/kernel/system/do_sigreturn.c +++ b/kernel/system/do_sigreturn.c @@ -60,8 +60,7 @@ PUBLIC int do_sigreturn(struct proc * caller, message * m_ptr) FPU_XFP_SIZE); rp->p_misc_flags |= MF_FPU_INITIALIZED; /* Restore math usage flag. */ /* force reloading FPU */ - if (fpu_owner == rp) - release_fpu(); + release_fpu(rp); } #endif diff --git a/kernel/system/do_sigsend.c b/kernel/system/do_sigsend.c index 3346ae10e..6cabc10df 100644 --- a/kernel/system/do_sigsend.c +++ b/kernel/system/do_sigsend.c @@ -46,10 +46,7 @@ PUBLIC int do_sigsend(struct proc * caller, message * m_ptr) #if (_MINIX_CHIP == _CHIP_INTEL) if(proc_used_fpu(rp)) { /* save the FPU context before saving it to the sig context */ - if (fpu_owner == rp) { - disable_fpu_exception(); - save_fpu(rp); - } + save_fpu(rp); memcpy(&sc.sc_fpu_state, rp->p_fpu_state.fpu_save_area_p, FPU_XFP_SIZE); } diff --git a/kernel/utility.c b/kernel/utility.c index 3a080a9aa..7d6521846 100644 --- a/kernel/utility.c +++ b/kernel/utility.c @@ -80,3 +80,8 @@ PUBLIC void cpu_print_freq(unsigned cpu) freq = cpu_get_freq(cpu); printf("CPU %d freq %lu MHz\n", cpu, div64u(freq, 1000000)); } + +PUBLIC int is_fpu(void) +{ + return get_cpulocal_var(fpu_presence); +}