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.
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..2950380 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,
@@ -1921,10 +1916,3 @@ tc_fill_vdso_timehands32(struct vdso_timehands32
*vdso_th32)
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
pgp7MRKhGk8Mu.pgp
Description: PGP signature
