On Fri, Jun 22, 2012 at 01:23:42PM +0300, Konstantin Belousov wrote: > On Fri, Jun 22, 2012 at 11:54:28AM +0300, Alexander Motin wrote: > > On 22.06.2012 11:25, Konstantin Belousov wrote: > > >On Fri, Jun 22, 2012 at 11:08:50AM +0300, Alexander Motin wrote: > > >>On 06/22/12 10:06, Konstantin Belousov wrote: > > >>>Author: kib > > >>>Date: Fri Jun 22 07:06:40 2012 > > >>>New Revision: 237433 > > >>>URL: http://svn.freebsd.org/changeset/base/237433 > > >>> > > >>>Log: > > >>> Implement mechanism to export some kernel timekeeping data to > > >>> usermode, using shared page. The structures and functions have vdso > > >>> prefix, to indicate the intended location of the code in some future. > > >>> > > >>> The versioned per-algorithm data is exported in the format of struct > > >>> vdso_timehands, which mostly repeats the content of in-kernel struct > > >>> timehands. Usermode reading of the structure can be lockless. > > >>> Compatibility export for 32bit processes on 64bit host is also > > >>> provided. Kernel also provides usermode with indication about > > >>> currently used timecounter, so that libc can fall back to syscall if > > >>> configured timecounter is unknown to usermode code. > > >>> > > >>> The shared data updates are initiated both from the tc_windup(), where > > >>> a fast task is queued to do the update, and from sysctl handlers which > > >>> change timecounter. A manual override switch > > >>> kern.timecounter.fast_gettime allows to turn off the mechanism. > > >>> > > >>> Only x86 architectures export the real algorithm data, and there, only > > >>> for tsc timecounter. HPET counters page could be exported as well, but > > >>> I prefer to not further glue the kernel and libc ABI there until > > >>> proper vdso-based solution is developed. > > >>> > > >>> Minimal stubs neccessary for non-x86 architectures to still compile > > >>> are provided. > > >>> > > >>> Discussed with: bde > > >>> Reviewed by: jhb > > >>> Tested by: flo > > >>> MFC after: 1 month > > >> > > >> > > >>>@@ -1360,6 +1367,7 @@ tc_windup(void) > > >>> #endif > > >>> > > >>> timehands = th; > > >>>+ > > >>>taskqueue_enqueue_fast(taskqueue_fast,&tc_windup_push_vdso_task); > > >>> } > > >>> > > >>> /* Report or change the active timecounter hardware. */ > > >> > > >>This taskqueue_enqueue_fast() will schedule extra thread to run each > > >>time hardclock() fires. One thread may be not a big problem, but > > >>together with callout swi and possible other threads woken up there it > > >>will wake up several other CPU cores from sleep just to put them back in > > >>few microseconds. Now davide@ and me are trying to fix that by avoiding > > >>callout SWI use for simple tasks. Please, let's not create another > > >>problem same time. > > > > > >The patch was public for quite a time. If you have some comments about > > >it, it would be much more productive to let me know about them before > > >the commit, not after. > > > > I'm sorry, I haven't seen it. My mad. > > > > >Anyway, what is your proposal for 'let's not create another problem > > >same time' part of the message ? It was discussed, as a possibility, > > >to have permanent mapping for the shared page in the KVA and to perform > > >lock-less update of the struct vdso_timehands directly from hardclock > > >handler. My opinion was that amount of work added by tc_windup > > >eventhandler is not trivial, so it is better to be postponed to > > >less critical context. It also slightly more safe to not perform > > >lockless update for vdso_timehands, since otherwise module load which > > >register exec handler could cause transient gettimeofday() failure > > >in usermode. > > > > > >This might boil down to the fact that tc_windup function is called > > >too often, in fact. Also, packing execution of tc_windup eventhandler > > >together with the clock swi is fine from my POV. > > > > I have nothing against using shared pages. On UP system I would probably > > have not so much against several threads. But on SMP system it will > > cause at least one, but in many cases two extra CPUs to be woken up. > > There are two or more threads to run on hardclock(): this taskqueue, > > callout swi and some thread(s) woken from callouts. Scheduler has no > > idea how heavy they are. So it will try to move each of them to separate > > idle CPU. Does the amount of work done in event handlers worth extra > > Watts consumed by rapidly waking CPUs? As quite rare person running > > FreeBSD on laptop, I am not sure. I am not sure even that on > > desktop/server this won't kill all benefit of fast clocks by limiting > > TurboBoost. > > Patch below would probably work, but I cannot test it right now on real > hardware due to ACPI issue. It worked for me in qemu. > > commit 4f2ffd93b36d20eae61495776fc6b0855745fd7f > Author: Konstantin Belousov <k...@freebsd.org> > Date: Fri Jun 22 13:19:22 2012 +0300 > > Use persistent kernel mapping of the shared page, and update the > vdso_timehands from hardclock, instead of scheduling task.
Slightly improved version. Since tc_fill_vdso_timehands is now called from hardclock context, thee is no need to spin waiting for valid current generation of timehands. diff --git a/sys/kern/kern_exec.c b/sys/kern/kern_exec.c index 80502e3..9365223 100644 --- a/sys/kern/kern_exec.c +++ b/sys/kern/kern_exec.c @@ -1517,42 +1517,13 @@ exec_unregister(execsw_arg) static struct sx shared_page_alloc_sx; static vm_object_t shared_page_obj; static int shared_page_free; - -struct sf_buf * -shared_page_write_start(int base) -{ - vm_page_t m; - struct sf_buf *s; - - VM_OBJECT_LOCK(shared_page_obj); - m = vm_page_grab(shared_page_obj, OFF_TO_IDX(base), VM_ALLOC_RETRY); - VM_OBJECT_UNLOCK(shared_page_obj); - s = sf_buf_alloc(m, SFB_DEFAULT); - return (s); -} - -void -shared_page_write_end(struct sf_buf *sf) -{ - vm_page_t m; - - m = sf_buf_page(sf); - sf_buf_free(sf); - VM_OBJECT_LOCK(shared_page_obj); - vm_page_wakeup(m); - VM_OBJECT_UNLOCK(shared_page_obj); -} +char *shared_page_mapping; void shared_page_write(int base, int size, const void *data) { - struct sf_buf *sf; - vm_offset_t sk; - sf = shared_page_write_start(base); - sk = sf_buf_kva(sf); - bcopy(data, (void *)(sk + (base & PAGE_MASK)), size); - shared_page_write_end(sf); + bcopy(data, shared_page_mapping + base, size); } static int @@ -1596,6 +1567,7 @@ static void shared_page_init(void *dummy __unused) { vm_page_t m; + vm_offset_t addr; sx_init(&shared_page_alloc_sx, "shpsx"); shared_page_obj = vm_pager_allocate(OBJT_PHYS, 0, PAGE_SIZE, @@ -1605,25 +1577,24 @@ shared_page_init(void *dummy __unused) VM_ALLOC_ZERO); m->valid = VM_PAGE_BITS_ALL; VM_OBJECT_UNLOCK(shared_page_obj); + addr = kmem_alloc_nofault(kernel_map, PAGE_SIZE); + pmap_qenter(addr, &m, 1); + shared_page_mapping = (char *)addr; } SYSINIT(shp, SI_SUB_EXEC, SI_ORDER_FIRST, (sysinit_cfunc_t)shared_page_init, NULL); static void -timehands_update(void *arg) +timehands_update(struct sysentvec *sv) { - struct sysentvec *sv; - struct sf_buf *sf; struct vdso_timehands th; struct vdso_timekeep *tk; uint32_t enabled, idx; - sv = arg; - sx_xlock(&shared_page_alloc_sx); enabled = tc_fill_vdso_timehands(&th); - sf = shared_page_write_start(sv->sv_timekeep_off); - tk = (void *)(sf_buf_kva(sf) + (sv->sv_timekeep_off & PAGE_MASK)); + tk = (struct vdso_timekeep *)(shared_page_mapping + + sv->sv_timekeep_off); idx = sv->sv_timekeep_curr; atomic_store_rel_32(&tk->tk_th[idx].th_gen, 0); if (++idx >= VDSO_TH_NUM) @@ -1637,25 +1608,19 @@ timehands_update(void *arg) tk->tk_enabled = enabled; atomic_store_rel_32(&tk->tk_th[idx].th_gen, sv->sv_timekeep_gen); tk->tk_current = idx; - shared_page_write_end(sf); - sx_xunlock(&shared_page_alloc_sx); } #ifdef COMPAT_FREEBSD32 static void -timehands_update32(void *arg) +timehands_update32(struct sysentvec *sv) { - struct sysentvec *sv; - struct sf_buf *sf; struct vdso_timekeep32 *tk; struct vdso_timehands32 th; uint32_t enabled, idx; - sv = arg; - sx_xlock(&shared_page_alloc_sx); enabled = tc_fill_vdso_timehands32(&th); - sf = shared_page_write_start(sv->sv_timekeep_off); - tk = (void *)(sf_buf_kva(sf) + (sv->sv_timekeep_off & PAGE_MASK)); + tk = (struct vdso_timekeep32 *)(shared_page_mapping + + sv->sv_timekeep_off); idx = sv->sv_timekeep_curr; atomic_store_rel_32(&tk->tk_th[idx].th_gen, 0); if (++idx >= VDSO_TH_NUM) @@ -1669,11 +1634,32 @@ timehands_update32(void *arg) tk->tk_enabled = enabled; atomic_store_rel_32(&tk->tk_th[idx].th_gen, sv->sv_timekeep_gen); tk->tk_current = idx; - shared_page_write_end(sf); - sx_xunlock(&shared_page_alloc_sx); } #endif +/* + * This is hackish, but easiest way to avoid creating list structures + * that needs to be iterated over from the hardclock interrupt + * context. + */ +static struct sysentvec *host_sysentvec; +#ifdef COMPAT_FREEBSD32 +static struct sysentvec *compat32_sysentvec; +#endif + +void +timekeep_push_vdso(void) +{ + + if (host_sysentvec != NULL && host_sysentvec->sv_timekeep_base != 0) + timehands_update(host_sysentvec); +#ifdef COMPAT_FREEBSD32 + if (compat32_sysentvec != NULL && + compat32_sysentvec->sv_timekeep_base != 0) + timehands_update32(compat32_sysentvec); +#endif +} + void exec_sysvec_init(void *param) { @@ -1688,29 +1674,32 @@ exec_sysvec_init(void *param) sv->sv_shared_page_obj = shared_page_obj; sv->sv_sigcode_base = sv->sv_shared_page_base + shared_page_fill(*(sv->sv_szsigcode), 16, sv->sv_sigcode); + if ((sv->sv_flags & SV_ABI_MASK) != SV_ABI_FREEBSD) + return; tk_ver = VDSO_TK_VER_CURR; #ifdef COMPAT_FREEBSD32 if ((sv->sv_flags & SV_ILP32) != 0) { tk_base = shared_page_alloc(sizeof(struct vdso_timekeep32) + sizeof(struct vdso_timehands32) * VDSO_TH_NUM, 16); KASSERT(tk_base != -1, ("tk_base -1 for 32bit")); - EVENTHANDLER_REGISTER(tc_windup, timehands_update32, sv, - EVENTHANDLER_PRI_ANY); shared_page_write(tk_base + offsetof(struct vdso_timekeep32, tk_ver), sizeof(uint32_t), &tk_ver); + KASSERT(compat32_sysentvec == 0, + ("Native compat32 already registered")); + compat32_sysentvec = sv; } else { #endif tk_base = shared_page_alloc(sizeof(struct vdso_timekeep) + sizeof(struct vdso_timehands) * VDSO_TH_NUM, 16); KASSERT(tk_base != -1, ("tk_base -1 for native")); - EVENTHANDLER_REGISTER(tc_windup, timehands_update, sv, - EVENTHANDLER_PRI_ANY); shared_page_write(tk_base + offsetof(struct vdso_timekeep, tk_ver), sizeof(uint32_t), &tk_ver); + KASSERT(host_sysentvec == 0, ("Native already registered")); + host_sysentvec = sv; #ifdef COMPAT_FREEBSD32 } #endif sv->sv_timekeep_base = sv->sv_shared_page_base + tk_base; sv->sv_timekeep_off = tk_base; - EVENTHANDLER_INVOKE(tc_windup); + timekeep_push_vdso(); } diff --git a/sys/kern/kern_tc.c b/sys/kern/kern_tc.c index 0b8fefe..4a75af5 100644 --- a/sys/kern/kern_tc.c +++ b/sys/kern/kern_tc.c @@ -31,7 +31,6 @@ __FBSDID("$FreeBSD$"); #include <sys/systm.h> #include <sys/timeffc.h> #include <sys/timepps.h> -#include <sys/taskqueue.h> #include <sys/timetc.h> #include <sys/timex.h> #include <sys/vdso.h> @@ -121,12 +120,8 @@ SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW, ×tepwarnings, 0, "Log time steps"); static void tc_windup(void); -static void tc_windup_push_vdso(void *ctx, int pending); static void cpu_tick_calibrate(int); -static struct task tc_windup_push_vdso_task = TASK_INITIALIZER(0, - tc_windup_push_vdso, 0); - static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS) { @@ -1367,7 +1362,7 @@ tc_windup(void) #endif timehands = th; - taskqueue_enqueue_fast(taskqueue_fast, &tc_windup_push_vdso_task); + timekeep_push_vdso(); } /* Report or change the active timecounter hardware. */ @@ -1394,7 +1389,7 @@ sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS) (void)newtc->tc_get_timecount(newtc); timecounter = newtc; - EVENTHANDLER_INVOKE(tc_windup); + timekeep_push_vdso(); return (0); } return (EINVAL); @@ -1865,7 +1860,7 @@ sysctl_fast_gettime(SYSCTL_HANDLER_ARGS) if (error != 0) return (error); vdso_th_enable = old_vdso_th_enable; - EVENTHANDLER_INVOKE(tc_windup); + timekeep_push_vdso(); return (0); } SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime, @@ -1877,19 +1872,15 @@ tc_fill_vdso_timehands(struct vdso_timehands *vdso_th) { struct timehands *th; uint32_t enabled; - int gen; - do { - th = timehands; - gen = th->th_generation; - vdso_th->th_algo = VDSO_TH_ALGO_1; - vdso_th->th_scale = th->th_scale; - vdso_th->th_offset_count = th->th_offset_count; - vdso_th->th_counter_mask = th->th_counter->tc_counter_mask; - vdso_th->th_offset = th->th_offset; - vdso_th->th_boottime = boottimebin; - enabled = cpu_fill_vdso_timehands(vdso_th); - } while (gen == 0 || timehands->th_generation != gen); + th = timehands; + vdso_th->th_algo = VDSO_TH_ALGO_1; + vdso_th->th_scale = th->th_scale; + vdso_th->th_offset_count = th->th_offset_count; + vdso_th->th_counter_mask = th->th_counter->tc_counter_mask; + vdso_th->th_offset = th->th_offset; + vdso_th->th_boottime = boottimebin; + enabled = cpu_fill_vdso_timehands(vdso_th); if (!vdso_th_enable) enabled = 0; return (enabled); @@ -1901,30 +1892,19 @@ tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32) { struct timehands *th; uint32_t enabled; - int gen; - do { - th = timehands; - gen = th->th_generation; - vdso_th32->th_algo = VDSO_TH_ALGO_1; - *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale; - vdso_th32->th_offset_count = th->th_offset_count; - vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask; - vdso_th32->th_offset.sec = th->th_offset.sec; - *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac; - vdso_th32->th_boottime.sec = boottimebin.sec; - *(uint64_t *)&vdso_th32->th_boottime.frac[0] = boottimebin.frac; - enabled = cpu_fill_vdso_timehands32(vdso_th32); - } while (gen == 0 || timehands->th_generation != gen); + th = timehands; + vdso_th32->th_algo = VDSO_TH_ALGO_1; + *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale; + vdso_th32->th_offset_count = th->th_offset_count; + vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask; + vdso_th32->th_offset.sec = th->th_offset.sec; + *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac; + vdso_th32->th_boottime.sec = boottimebin.sec; + *(uint64_t *)&vdso_th32->th_boottime.frac[0] = boottimebin.frac; + enabled = cpu_fill_vdso_timehands32(vdso_th32); if (!vdso_th_enable) enabled = 0; return (enabled); } #endif - -static void -tc_windup_push_vdso(void *ctx, int pending) -{ - - EVENTHANDLER_INVOKE(tc_windup); -} diff --git a/sys/sys/sysent.h b/sys/sys/sysent.h index 22769c2..6de72d9 100644 --- a/sys/sys/sysent.h +++ b/sys/sys/sysent.h @@ -265,8 +265,6 @@ int shared_page_alloc(int size, int align); int shared_page_fill(int size, int align, const void *data); void shared_page_write(int base, int size, const void *data); void exec_sysvec_init(void *param); -struct sf_buf *shared_page_write_start(int base); -void shared_page_write_end(struct sf_buf *sf); #define INIT_SYSENTVEC(name, sv) \ SYSINIT(name, SI_SUB_EXEC, SI_ORDER_ANY, \ diff --git a/sys/sys/vdso.h b/sys/sys/vdso.h index 9f3f3af..653a606 100644 --- a/sys/sys/vdso.h +++ b/sys/sys/vdso.h @@ -29,7 +29,6 @@ #define _SYS_VDSO_H #include <sys/types.h> -#include <sys/eventhandler.h> #include <machine/vdso.h> struct vdso_timehands { @@ -74,6 +73,8 @@ u_int __vdso_gettc(const struct vdso_timehands *vdso_th); #ifdef _KERNEL +void timekeep_push_vdso(void); + uint32_t tc_fill_vdso_timehands(struct vdso_timehands *vdso_th); /* @@ -86,9 +87,6 @@ uint32_t tc_fill_vdso_timehands(struct vdso_timehands *vdso_th); */ uint32_t cpu_fill_vdso_timehands(struct vdso_timehands *vdso_th); -typedef void (*tc_windup_fn)(void *); -EVENTHANDLER_DECLARE(tc_windup, tc_windup_fn); - #define VDSO_TH_NUM 4 #ifdef COMPAT_FREEBSD32
pgpX8MbTeqUKZ.pgp
Description: PGP signature