arch/x86/kernel/vsyscall_64.c

   1 /*
   2  *  Copyright (C) 2001 Andrea Arcangeli <andrea@suse.de> SuSE
   3  *  Copyright 2003 Andi Kleen, SuSE Labs.
   4  *
   5  *  Thanks to hpa@transmeta.com for some useful hint.
   6  *  Special thanks to Ingo Molnar for his early experience with
   7  *  a different vsyscall implementation for Linux/IA32 and for the name.
   8  *
   9  *  vsyscall 1 is located at -10Mbyte, vsyscall 2 is located
  10  *  at virtual address -10Mbyte+1024bytes etc... There are at max 4
  11  *  vsyscalls. One vsyscall can reserve more than 1 slot to avoid
  12  *  jumping out of line if necessary. We cannot add more with this
  13  *  mechanism because older kernels won't return -ENOSYS.
  14  *  If we want more than four we need a vDSO.
  15  *
  16  *  Note: the concept clashes with user mode linux. If you use UML and
  17  *  want per guest time just set the kernel.vsyscall64 sysctl to 0.
  18  */
  19
  20 #include <linux/time.h>
  21 #include <linux/init.h>
  22 #include <linux/kernel.h>
  23 #include <linux/timer.h>
  24 #include <linux/seqlock.h>
  25 #include <linux/jiffies.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/clocksource.h>
  28 #include <linux/getcpu.h>
  29 #include <linux/cpu.h>
  30 #include <linux/smp.h>
  31 #include <linux/notifier.h>
  32
  33 #include <asm/vsyscall.h>
  34 #include <asm/pgtable.h>
  35 #include <asm/page.h>
  36 #include <asm/unistd.h>
  37 #include <asm/fixmap.h>
  38 #include <asm/errno.h>
  39 #include <asm/io.h>
  40 #include <asm/segment.h>
  41 #include <asm/desc.h>
  42 #include <asm/topology.h>
  43 #include <asm/vgtod.h>
  44
  45 #define __vsyscall(nr) \
  46                 __attribute__ ((unused, __section__(".vsyscall_" #nr))) notrace
  47 #define __syscall_clobber "r11","cx","memory"
  48
  49 /*
  50  * vsyscall_gtod_data contains data that is :
  51  * - readonly from vsyscalls
  52  * - written by timer interrupt or systcl (/proc/sys/kernel/vsyscall64)
  53  * Try to keep this structure as small as possible to avoid cache line ping pongs
  54  */
  55 int __vgetcpu_mode __section_vgetcpu_mode;
  56
  57 struct vsyscall_gtod_data __vsyscall_gtod_data __section_vsyscall_gtod_data =
  58 {
  59         .lock = SEQLOCK_UNLOCKED,
  60         .sysctl_enabled = 1,
  61 };
  62
  63 void update_vsyscall_tz(void)
  64 {
  65         unsigned long flags;
  66
  67         write_seqlock_irqsave(&vsyscall_gtod_data.lock, flags);
  68         /* sys_tz has changed */
  69         vsyscall_gtod_data.sys_tz = sys_tz;
  70         write_sequnlock_irqrestore(&vsyscall_gtod_data.lock, flags);
  71 }
  72
  73 void update_vsyscall(struct timespec *wall_time, struct clocksource *clock)
  74 {
  75         unsigned long flags;
  76
  77         write_seqlock_irqsave(&vsyscall_gtod_data.lock, flags);
  78         /* copy vsyscall data */
  79         vsyscall_gtod_data.clock.vread = clock->vread;
  80         vsyscall_gtod_data.clock.cycle_last = clock->cycle_last;
  81         vsyscall_gtod_data.clock.mask = clock->mask;
  82         vsyscall_gtod_data.clock.mult = clock->mult;
  83         vsyscall_gtod_data.clock.shift = clock->shift;
  84         vsyscall_gtod_data.wall_time_sec = wall_time->tv_sec;
  85         vsyscall_gtod_data.wall_time_nsec = wall_time->tv_nsec;
  86         vsyscall_gtod_data.wall_to_monotonic = wall_to_monotonic;
  87         write_sequnlock_irqrestore(&vsyscall_gtod_data.lock, flags);
  88 }
  89
  90 /* RED-PEN may want to readd seq locking, but then the variable should be
  91  * write-once.
  92  */
  93 static __always_inline void do_get_tz(struct timezone * tz)
  94 {
  95         *tz = __vsyscall_gtod_data.sys_tz;
  96 }
  97
  98 static __always_inline int gettimeofday(struct timeval *tv, struct timezone *tz)
  99 {
 100         int ret;
 101         asm volatile("syscall"
 102                 : "=a" (ret)
 103                 : "0" (__NR_gettimeofday),"D" (tv),"S" (tz)
 104                 : __syscall_clobber );
 105         return ret;
 106 }
 107
 108 static __always_inline long time_syscall(long *t)
 109 {
 110         long secs;
 111         asm volatile("syscall"
 112                 : "=a" (secs)
 113                 : "0" (__NR_time),"D" (t) : __syscall_clobber);
 114         return secs;
 115 }
 116
 117 static __always_inline void do_vgettimeofday(struct timeval * tv)
 118 {
 119         cycle_t now, base, mask, cycle_delta;
 120         unsigned seq;
 121         unsigned long mult, shift, nsec;
 122         cycle_t (*vread)(void);
 123         do {
 124                 seq = read_seqbegin(&__vsyscall_gtod_data.lock);
 125
 126                 vread = __vsyscall_gtod_data.clock.vread;
 127                 if (unlikely(!__vsyscall_gtod_data.sysctl_enabled || !vread)) {
 128                         gettimeofday(tv,NULL);
 129                         return;
 130                 }
 131                 now = vread();
 132                 base = __vsyscall_gtod_data.clock.cycle_last;
 133                 mask = __vsyscall_gtod_data.clock.mask;
 134                 mult = __vsyscall_gtod_data.clock.mult;
 135                 shift = __vsyscall_gtod_data.clock.shift;
 136
 137                 tv->tv_sec = __vsyscall_gtod_data.wall_time_sec;
 138                 nsec = __vsyscall_gtod_data.wall_time_nsec;
 139         } while (read_seqretry(&__vsyscall_gtod_data.lock, seq));
 140
 141         /* calculate interval: */
 142         cycle_delta = (now - base) & mask;
 143         /* convert to nsecs: */
 144         nsec += (cycle_delta * mult) >> shift;
 145
 146         while (nsec >= NSEC_PER_SEC) {
 147                 tv->tv_sec += 1;
 148                 nsec -= NSEC_PER_SEC;
 149         }
 150         tv->tv_usec = nsec / NSEC_PER_USEC;
 151 }
 152
 153 int __vsyscall(0) vgettimeofday(struct timeval * tv, struct timezone * tz)
 154 {
 155         if (tv)
 156                 do_vgettimeofday(tv);
 157         if (tz)
 158                 do_get_tz(tz);
 159         return 0;
 160 }
 161
 162 /* This will break when the xtime seconds get inaccurate, but that is
 163  * unlikely */
 164 time_t __vsyscall(1) vtime(time_t *t)
 165 {
 166         struct timeval tv;
 167         time_t result;
 168         if (unlikely(!__vsyscall_gtod_data.sysctl_enabled))
 169                 return time_syscall(t);
 170
 171         vgettimeofday(&tv, NULL);
 172         result = tv.tv_sec;
 173         if (t)
 174                 *t = result;
 175         return result;
 176 }
 177
 178 /* Fast way to get current CPU and node.
 179    This helps to do per node and per CPU caches in user space.
 180    The result is not guaranteed without CPU affinity, but usually
 181    works out because the scheduler tries to keep a thread on the same
 182    CPU.
 183
 184    tcache must point to a two element sized long array.
 185    All arguments can be NULL. */
 186 long __vsyscall(2)
 187 vgetcpu(unsigned *cpu, unsigned *node, struct getcpu_cache *tcache)
 188 {
 189         unsigned int p;
 190         unsigned long j = 0;
 191
 192         /* Fast cache - only recompute value once per jiffies and avoid
 193            relatively costly rdtscp/cpuid otherwise.
 194            This works because the scheduler usually keeps the process
 195            on the same CPU and this syscall doesn't guarantee its
 196            results anyways.
 197            We do this here because otherwise user space would do it on
 198            its own in a likely inferior way (no access to jiffies).
 199            If you don't like it pass NULL. */
 200         if (tcache && tcache->blob[0] == (j = __jiffies)) {
 201                 p = tcache->blob[1];
 202         } else if (__vgetcpu_mode == VGETCPU_RDTSCP) {
 203                 /* Load per CPU data from RDTSCP */
 204                 native_read_tscp(&p);
 205         } else {
 206                 /* Load per CPU data from GDT */
 207                 asm("lsl %1,%0" : "=r" (p) : "r" (__PER_CPU_SEG));
 208         }
 209         if (tcache) {
 210                 tcache->blob[0] = j;
 211                 tcache->blob[1] = p;
 212         }
 213         if (cpu)
 214                 *cpu = p & 0xfff;
 215         if (node)
 216                 *node = p >> 12;
 217         return 0;
 218 }
 219
 220 static long __vsyscall(3) venosys_1(void)
 221 {
 222         return -ENOSYS;
 223 }
 224
 225 #ifdef CONFIG_SYSCTL
 226
 227 static int
 228 vsyscall_sysctl_change(ctl_table *ctl, int write, struct file * filp,
 229                        void __user *buffer, size_t *lenp, loff_t *ppos)
 230 {
 231         return proc_dointvec(ctl, write, filp, buffer, lenp, ppos);
 232 }
 233
 234 static ctl_table kernel_table2[] = {
 235         { .procname = "vsyscall64",
 236           .data = &vsyscall_gtod_data.sysctl_enabled, .maxlen = sizeof(int),
 237           .mode = 0644,
 238           .proc_handler = vsyscall_sysctl_change },
 239         {}
 240 };
 241
 242 static ctl_table kernel_root_table2[] = {
 243         { .ctl_name = CTL_KERN, .procname = "kernel", .mode = 0555,
 244           .child = kernel_table2 },
 245         {}
 246 };
 247 #endif
 248
 249 /* Assume __initcall executes before all user space. Hopefully kmod
 250    doesn't violate that. We'll find out if it does. */
 251 static void __cpuinit vsyscall_set_cpu(int cpu)
 252 {
 253         unsigned long d;
 254         unsigned long node = 0;
 255 #ifdef CONFIG_NUMA
 256         node = cpu_to_node(cpu);
 257 #endif
 258         if (cpu_has(&cpu_data(cpu), X86_FEATURE_RDTSCP))
 259                 write_rdtscp_aux((node << 12) | cpu);
 260
 261         /* Store cpu number in limit so that it can be loaded quickly
 262            in user space in vgetcpu.
 263            12 bits for the CPU and 8 bits for the node. */
 264         d = 0x0f40000000000ULL;
 265         d |= cpu;
 266         d |= (node & 0xf) << 12;
 267         d |= (node >> 4) << 48;
 268         write_gdt_entry(get_cpu_gdt_table(cpu), GDT_ENTRY_PER_CPU, &d, DESCTYPE_S);
 269 }
 270
 271 static void __cpuinit cpu_vsyscall_init(void *arg)
 272 {
 273         /* preemption should be already off */
 274         vsyscall_set_cpu(raw_smp_processor_id());
 275 }
 276
 277 static int __cpuinit
 278 cpu_vsyscall_notifier(struct notifier_block *n, unsigned long action, void *arg)
 279 {
 280         long cpu = (long)arg;
 281         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
 282                 smp_call_function_single(cpu, cpu_vsyscall_init, NULL, 0, 1);
 283         return NOTIFY_DONE;
 284 }
 285
 286 void __init map_vsyscall(void)
 287 {
 288         extern char __vsyscall_0;
 289         unsigned long physaddr_page0 = __pa_symbol(&__vsyscall_0);
 290
 291         /* Note that VSYSCALL_MAPPED_PAGES must agree with the code below. */
 292         __set_fixmap(VSYSCALL_FIRST_PAGE, physaddr_page0, PAGE_KERNEL_VSYSCALL);
 293 }
 294
 295 static int __init vsyscall_init(void)
 296 {
 297         BUG_ON(((unsigned long) &vgettimeofday !=
 298                         VSYSCALL_ADDR(__NR_vgettimeofday)));
 299         BUG_ON((unsigned long) &vtime != VSYSCALL_ADDR(__NR_vtime));
 300         BUG_ON((VSYSCALL_ADDR(0) != __fix_to_virt(VSYSCALL_FIRST_PAGE)));
 301         BUG_ON((unsigned long) &vgetcpu != VSYSCALL_ADDR(__NR_vgetcpu));
 302 #ifdef CONFIG_SYSCTL
 303         register_sysctl_table(kernel_root_table2);
 304 #endif
 305         on_each_cpu(cpu_vsyscall_init, NULL, 0, 1);
 306         hotcpu_notifier(cpu_vsyscall_notifier, 0);
 307         return 0;
 308 }
 309
 310 __initcall(vsyscall_init);