[PATCH] Add the vgetcpu vsyscall

This patch adds a vgetcpu vsyscall, which depending on the CPU RDTSCP
capability uses either the RDTSCP or CPUID to obtain a CPU and node
numbers and pass them to the program.

AK: Lots of changes over Vojtech's original code:
Better prototype for vgetcpu()
It's better to pass the cpu / node numbers as separate arguments
to avoid mistakes when going from SMP to NUMA.
Also add a fast time stamp based cache using a user supplied
argument to speed things more up.
Use fast method from Chuck Ebbert to retrieve node/cpu from
GDT limit instead of CPUID
Made sure RDTSCP init is always executed after node is known.
Drop printk

Signed-off-by: Vojtech Pavlik <vojtech@suse.cz>
Signed-off-by: Andi Kleen <ak@suse.de>

4 files changed, 93 insertions, 9 deletions

diff --git a/arch/x86_64/kernel/head.S b/arch/x86_64/kernel/head.S
index c9739ca81d06..505ec4a57506 100644
--- a/arch/x86_64/kernel/head.S
+++ b/arch/x86_64/kernel/head.S
@@ -371,7 +371,7 @@ ENTRY(cpu_gdt_table)
         .quad   0,0                     /* TSS */
         .quad   0,0                     /* LDT */
         .quad   0,0,0                   /* three TLS descriptors */ 
-        .quad   0                       /* unused */
+        .quad   0x0000f40000000000      /* node/CPU stored in limit */
 gdt_end:        
         /* asm/segment.h:GDT_ENTRIES must match this */ 
         /* This should be a multiple of the cache line size */
diff --git a/arch/x86_64/kernel/time.c b/arch/x86_64/kernel/time.c
index 97b9e46d1992..560ed944dc0e 100644
--- a/arch/x86_64/kernel/time.c
+++ b/arch/x86_64/kernel/time.c
@@ -899,12 +899,8 @@ static int __cpuinit
 time_cpu_notifier(struct notifier_block *nb, unsigned long action, void *hcpu)
 {
         unsigned cpu = (unsigned long) hcpu;
-        if (action == CPU_ONLINE &&
-                cpu_has(&cpu_data[cpu], X86_FEATURE_RDTSCP)) {
-                unsigned p;
-                p = smp_processor_id() | (cpu_to_node(smp_processor_id())<<12);
-                write_rdtscp_aux(p);
-        }
+        if (action == CPU_ONLINE)
+                vsyscall_set_cpu(cpu);
         return NOTIFY_DONE;
 }
 
@@ -993,6 +989,11 @@ void time_init_gtod(void)
         if (unsynchronized_tsc())
                 notsc = 1;
 
+        if (cpu_has(&boot_cpu_data, X86_FEATURE_RDTSCP))
+                vgetcpu_mode = VGETCPU_RDTSCP;
+        else
+                vgetcpu_mode = VGETCPU_LSL;
+
         if (vxtime.hpet_address && notsc) {
                 timetype = hpet_use_timer ? "HPET" : "PIT/HPET";
                 if (hpet_use_timer)
diff --git a/arch/x86_64/kernel/vmlinux.lds.S b/arch/x86_64/kernel/vmlinux.lds.S
index 7c4de31471d4..8d5a5149bb3a 100644
--- a/arch/x86_64/kernel/vmlinux.lds.S
+++ b/arch/x86_64/kernel/vmlinux.lds.S
@@ -99,6 +99,9 @@ SECTIONS
   .vxtime : AT(VLOAD(.vxtime)) { *(.vxtime) }
   vxtime = VVIRT(.vxtime);
 
+  .vgetcpu_mode : AT(VLOAD(.vgetcpu_mode)) { *(.vgetcpu_mode) }
+  vgetcpu_mode = VVIRT(.vgetcpu_mode);
+
   .wall_jiffies : AT(VLOAD(.wall_jiffies)) { *(.wall_jiffies) }
   wall_jiffies = VVIRT(.wall_jiffies);
 
diff --git a/arch/x86_64/kernel/vsyscall.c b/arch/x86_64/kernel/vsyscall.c
index f603037df162..902783bc4d53 100644
--- a/arch/x86_64/kernel/vsyscall.c
+++ b/arch/x86_64/kernel/vsyscall.c
@@ -26,6 +26,7 @@
 #include <linux/seqlock.h>
 #include <linux/jiffies.h>
 #include <linux/sysctl.h>
+#include <linux/getcpu.h>
 
 #include <asm/vsyscall.h>
 #include <asm/pgtable.h>
@@ -33,11 +34,15 @@
 #include <asm/fixmap.h>
 #include <asm/errno.h>
 #include <asm/io.h>
+#include <asm/segment.h>
+#include <asm/desc.h>
+#include <asm/topology.h>
 
 #define __vsyscall(nr) __attribute__ ((unused,__section__(".vsyscall_" #nr)))
 
 int __sysctl_vsyscall __section_sysctl_vsyscall = 1;
 seqlock_t __xtime_lock __section_xtime_lock = SEQLOCK_UNLOCKED;
+int __vgetcpu_mode __section_vgetcpu_mode;
 
 #include <asm/unistd.h>
 
@@ -127,9 +132,46 @@ time_t __vsyscall(1) vtime(time_t *t)
         return __xtime.tv_sec;
 }
 
-long __vsyscall(2) venosys_0(void)
+/* Fast way to get current CPU and node.
+   This helps to do per node and per CPU caches in user space.
+   The result is not guaranteed without CPU affinity, but usually
+   works out because the scheduler tries to keep a thread on the same
+   CPU.
+
+   tcache must point to a two element sized long array.
+   All arguments can be NULL. */
+long __vsyscall(2)
+vgetcpu(unsigned *cpu, unsigned *node, struct getcpu_cache *tcache)
 {
-        return -ENOSYS;
+        unsigned int dummy, p;
+        unsigned long j = 0;
+
+        /* Fast cache - only recompute value once per jiffies and avoid
+           relatively costly rdtscp/cpuid otherwise.
+           This works because the scheduler usually keeps the process
+           on the same CPU and this syscall doesn't guarantee its
+           results anyways.
+           We do this here because otherwise user space would do it on
+           its own in a likely inferior way (no access to jiffies).
+           If you don't like it pass NULL. */
+        if (tcache && tcache->t0 == (j = __jiffies)) {
+                p = tcache->t1;
+        } else if (__vgetcpu_mode == VGETCPU_RDTSCP) {
+                /* Load per CPU data from RDTSCP */
+                rdtscp(dummy, dummy, p);
+        } else {
+                /* Load per CPU data from GDT */
+                asm("lsl %1,%0" : "=r" (p) : "r" (__PER_CPU_SEG));
+        }
+        if (tcache) {
+                tcache->t0 = j;
+                tcache->t1 = p;
+        }
+        if (cpu)
+                *cpu = p & 0xfff;
+        if (node)
+                *node = p >> 12;
+        return 0;
 }
 
 long __vsyscall(3) venosys_1(void)
@@ -200,6 +242,43 @@ static ctl_table kernel_root_table2[] = {
 
 #endif
 
+static void __cpuinit write_rdtscp_cb(void *info)
+{
+        write_rdtscp_aux((unsigned long)info);
+}
+
+void __cpuinit vsyscall_set_cpu(int cpu)
+{
+        unsigned long *d;
+        unsigned long node = 0;
+#ifdef CONFIG_NUMA
+        node = cpu_to_node[cpu];
+#endif
+        if (cpu_has(&cpu_data[cpu], X86_FEATURE_RDTSCP)) {
+                void *info = (void *)((node << 12) | cpu);
+                /* Can happen on preemptive kernel */
+                if (get_cpu() == cpu)
+                        write_rdtscp_cb(info);
+#ifdef CONFIG_SMP
+                else {
+                        /* the notifier is unfortunately not executed on the
+                           target CPU */
+                        smp_call_function_single(cpu,write_rdtscp_cb,info,0,1);
+                }
+#endif
+                put_cpu();
+        }
+
+        /* Store cpu number in limit so that it can be loaded quickly
+           in user space in vgetcpu.
+           12 bits for the CPU and 8 bits for the node. */
+        d = (unsigned long *)(cpu_gdt(cpu) + GDT_ENTRY_PER_CPU);
+        *d = 0x0f40000000000ULL;
+        *d |= cpu;
+        *d |= (node & 0xf) << 12;
+        *d |= (node >> 4) << 48;
+}
+
 static void __init map_vsyscall(void)
 {
         extern char __vsyscall_0;
@@ -214,6 +293,7 @@ static int __init vsyscall_init(void)
                         VSYSCALL_ADDR(__NR_vgettimeofday)));
         BUG_ON((unsigned long) &vtime != VSYSCALL_ADDR(__NR_vtime));
         BUG_ON((VSYSCALL_ADDR(0) != __fix_to_virt(VSYSCALL_FIRST_PAGE)));
+        BUG_ON((unsigned long) &vgetcpu != VSYSCALL_ADDR(__NR_vgetcpu));
         map_vsyscall();
 #ifdef CONFIG_SYSCTL
         register_sysctl_table(kernel_root_table2, 0);