Source file src/runtime/os_linux.go

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  package runtime
     6  
     7  import (
     8  	"internal/abi"
     9  	"internal/goarch"
    10  	"internal/runtime/atomic"
    11  	"internal/runtime/strconv"
    12  	"internal/runtime/syscall"
    13  	"unsafe"
    14  )
    15  
    16  // sigPerThreadSyscall is the same signal (SIGSETXID) used by glibc for
    17  // per-thread syscalls on Linux. We use it for the same purpose in non-cgo
    18  // binaries.
    19  const sigPerThreadSyscall = _SIGRTMIN + 1
    20  
    21  type mOS struct {
    22  	// profileTimer holds the ID of the POSIX interval timer for profiling CPU
    23  	// usage on this thread.
    24  	//
    25  	// It is valid when the profileTimerValid field is true. A thread
    26  	// creates and manages its own timer, and these fields are read and written
    27  	// only by this thread. But because some of the reads on profileTimerValid
    28  	// are in signal handling code, this field should be atomic type.
    29  	profileTimer      int32
    30  	profileTimerValid atomic.Bool
    31  
    32  	// needPerThreadSyscall indicates that a per-thread syscall is required
    33  	// for doAllThreadsSyscall.
    34  	needPerThreadSyscall atomic.Uint8
    35  
    36  	// This is a pointer to a chunk of memory allocated with a special
    37  	// mmap invocation in vgetrandomGetState().
    38  	vgetrandomState uintptr
    39  
    40  	waitsema uint32 // semaphore for parking on locks
    41  }
    42  
    43  //go:noescape
    44  func futex(addr unsafe.Pointer, op int32, val uint32, ts, addr2 unsafe.Pointer, val3 uint32) int32
    45  
    46  // Linux futex.
    47  //
    48  //	futexsleep(uint32 *addr, uint32 val)
    49  //	futexwakeup(uint32 *addr)
    50  //
    51  // Futexsleep atomically checks if *addr == val and if so, sleeps on addr.
    52  // Futexwakeup wakes up threads sleeping on addr.
    53  // Futexsleep is allowed to wake up spuriously.
    54  
    55  const (
    56  	_FUTEX_PRIVATE_FLAG = 128
    57  	_FUTEX_WAIT_PRIVATE = 0 | _FUTEX_PRIVATE_FLAG
    58  	_FUTEX_WAKE_PRIVATE = 1 | _FUTEX_PRIVATE_FLAG
    59  )
    60  
    61  // Atomically,
    62  //
    63  //	if(*addr == val) sleep
    64  //
    65  // Might be woken up spuriously; that's allowed.
    66  // Don't sleep longer than ns; ns < 0 means forever.
    67  //
    68  //go:nosplit
    69  func futexsleep(addr *uint32, val uint32, ns int64) {
    70  	// Some Linux kernels have a bug where futex of
    71  	// FUTEX_WAIT returns an internal error code
    72  	// as an errno. Libpthread ignores the return value
    73  	// here, and so can we: as it says a few lines up,
    74  	// spurious wakeups are allowed.
    75  	if ns < 0 {
    76  		futex(unsafe.Pointer(addr), _FUTEX_WAIT_PRIVATE, val, nil, nil, 0)
    77  		return
    78  	}
    79  
    80  	var ts timespec
    81  	ts.setNsec(ns)
    82  	futex(unsafe.Pointer(addr), _FUTEX_WAIT_PRIVATE, val, unsafe.Pointer(&ts), nil, 0)
    83  }
    84  
    85  // If any procs are sleeping on addr, wake up at most cnt.
    86  //
    87  //go:nosplit
    88  func futexwakeup(addr *uint32, cnt uint32) {
    89  	ret := futex(unsafe.Pointer(addr), _FUTEX_WAKE_PRIVATE, cnt, nil, nil, 0)
    90  	if ret >= 0 {
    91  		return
    92  	}
    93  
    94  	// I don't know that futex wakeup can return
    95  	// EAGAIN or EINTR, but if it does, it would be
    96  	// safe to loop and call futex again.
    97  	systemstack(func() {
    98  		print("futexwakeup addr=", addr, " returned ", ret, "\n")
    99  	})
   100  
   101  	*(*int32)(unsafe.Pointer(uintptr(0x1006))) = 0x1006
   102  }
   103  
   104  func getCPUCount() int32 {
   105  	// This buffer is huge (8 kB) but we are on the system stack
   106  	// and there should be plenty of space (64 kB).
   107  	// Also this is a leaf, so we're not holding up the memory for long.
   108  	// See golang.org/issue/11823.
   109  	// The suggested behavior here is to keep trying with ever-larger
   110  	// buffers, but we don't have a dynamic memory allocator at the
   111  	// moment, so that's a bit tricky and seems like overkill.
   112  	const maxCPUs = 64 * 1024
   113  	var buf [maxCPUs / 8]byte
   114  	r := sched_getaffinity(0, unsafe.Sizeof(buf), &buf[0])
   115  	if r < 0 {
   116  		return 1
   117  	}
   118  	n := int32(0)
   119  	for _, v := range buf[:r] {
   120  		for v != 0 {
   121  			n += int32(v & 1)
   122  			v >>= 1
   123  		}
   124  	}
   125  	if n == 0 {
   126  		n = 1
   127  	}
   128  	return n
   129  }
   130  
   131  // Clone, the Linux rfork.
   132  const (
   133  	_CLONE_VM             = 0x100
   134  	_CLONE_FS             = 0x200
   135  	_CLONE_FILES          = 0x400
   136  	_CLONE_SIGHAND        = 0x800
   137  	_CLONE_PTRACE         = 0x2000
   138  	_CLONE_VFORK          = 0x4000
   139  	_CLONE_PARENT         = 0x8000
   140  	_CLONE_THREAD         = 0x10000
   141  	_CLONE_NEWNS          = 0x20000
   142  	_CLONE_SYSVSEM        = 0x40000
   143  	_CLONE_SETTLS         = 0x80000
   144  	_CLONE_PARENT_SETTID  = 0x100000
   145  	_CLONE_CHILD_CLEARTID = 0x200000
   146  	_CLONE_UNTRACED       = 0x800000
   147  	_CLONE_CHILD_SETTID   = 0x1000000
   148  	_CLONE_STOPPED        = 0x2000000
   149  	_CLONE_NEWUTS         = 0x4000000
   150  	_CLONE_NEWIPC         = 0x8000000
   151  
   152  	// As of QEMU 2.8.0 (5ea2fc84d), user emulation requires all six of these
   153  	// flags to be set when creating a thread; attempts to share the other
   154  	// five but leave SYSVSEM unshared will fail with -EINVAL.
   155  	//
   156  	// In non-QEMU environments CLONE_SYSVSEM is inconsequential as we do not
   157  	// use System V semaphores.
   158  
   159  	cloneFlags = _CLONE_VM | /* share memory */
   160  		_CLONE_FS | /* share cwd, etc */
   161  		_CLONE_FILES | /* share fd table */
   162  		_CLONE_SIGHAND | /* share sig handler table */
   163  		_CLONE_SYSVSEM | /* share SysV semaphore undo lists (see issue #20763) */
   164  		_CLONE_THREAD /* revisit - okay for now */
   165  )
   166  
   167  //go:noescape
   168  func clone(flags int32, stk, mp, gp, fn unsafe.Pointer) int32
   169  
   170  // May run with m.p==nil, so write barriers are not allowed.
   171  //
   172  //go:nowritebarrier
   173  func newosproc(mp *m) {
   174  	stk := unsafe.Pointer(mp.g0.stack.hi)
   175  	/*
   176  	 * note: strace gets confused if we use CLONE_PTRACE here.
   177  	 */
   178  	if false {
   179  		print("newosproc stk=", stk, " m=", mp, " g=", mp.g0, " clone=", abi.FuncPCABI0(clone), " id=", mp.id, " ostk=", &mp, "\n")
   180  	}
   181  
   182  	// Disable signals during clone, so that the new thread starts
   183  	// with signals disabled. It will enable them in minit.
   184  	var oset sigset
   185  	sigprocmask(_SIG_SETMASK, &sigset_all, &oset)
   186  	ret := retryOnEAGAIN(func() int32 {
   187  		r := clone(cloneFlags, stk, unsafe.Pointer(mp), unsafe.Pointer(mp.g0), unsafe.Pointer(abi.FuncPCABI0(mstart)))
   188  		// clone returns positive TID, negative errno.
   189  		// We don't care about the TID.
   190  		if r >= 0 {
   191  			return 0
   192  		}
   193  		return -r
   194  	})
   195  	sigprocmask(_SIG_SETMASK, &oset, nil)
   196  
   197  	if ret != 0 {
   198  		print("runtime: failed to create new OS thread (have ", mcount(), " already; errno=", ret, ")\n")
   199  		if ret == _EAGAIN {
   200  			println("runtime: may need to increase max user processes (ulimit -u)")
   201  		}
   202  		throw("newosproc")
   203  	}
   204  }
   205  
   206  // Version of newosproc that doesn't require a valid G.
   207  //
   208  //go:nosplit
   209  func newosproc0(stacksize uintptr, fn unsafe.Pointer) {
   210  	stack := sysAlloc(stacksize, &memstats.stacks_sys, "OS thread stack")
   211  	if stack == nil {
   212  		writeErrStr(failallocatestack)
   213  		exit(1)
   214  	}
   215  	ret := clone(cloneFlags, unsafe.Pointer(uintptr(stack)+stacksize), nil, nil, fn)
   216  	if ret < 0 {
   217  		writeErrStr(failthreadcreate)
   218  		exit(1)
   219  	}
   220  }
   221  
   222  const (
   223  	_AT_NULL     = 0  // End of vector
   224  	_AT_PAGESZ   = 6  // System physical page size
   225  	_AT_PLATFORM = 15 // string identifying platform
   226  	_AT_HWCAP    = 16 // hardware capability bit vector
   227  	_AT_SECURE   = 23 // secure mode boolean
   228  	_AT_RANDOM   = 25 // introduced in 2.6.29
   229  	_AT_HWCAP2   = 26 // hardware capability bit vector 2
   230  )
   231  
   232  var procAuxv = []byte("/proc/self/auxv\x00")
   233  
   234  var addrspace_vec [1]byte
   235  
   236  func mincore(addr unsafe.Pointer, n uintptr, dst *byte) int32
   237  
   238  var auxvreadbuf [128]uintptr
   239  
   240  func sysargs(argc int32, argv **byte) {
   241  	n := argc + 1
   242  
   243  	// skip over argv, envp to get to auxv
   244  	for argv_index(argv, n) != nil {
   245  		n++
   246  	}
   247  
   248  	// skip NULL separator
   249  	n++
   250  
   251  	// now argv+n is auxv
   252  	auxvp := (*[1 << 28]uintptr)(add(unsafe.Pointer(argv), uintptr(n)*goarch.PtrSize))
   253  
   254  	if pairs := sysauxv(auxvp[:]); pairs != 0 {
   255  		auxv = auxvp[: pairs*2 : pairs*2]
   256  		return
   257  	}
   258  	// In some situations we don't get a loader-provided
   259  	// auxv, such as when loaded as a library on Android.
   260  	// Fall back to /proc/self/auxv.
   261  	fd := open(&procAuxv[0], 0 /* O_RDONLY */, 0)
   262  	if fd < 0 {
   263  		// On Android, /proc/self/auxv might be unreadable (issue 9229), so we fallback to
   264  		// try using mincore to detect the physical page size.
   265  		// mincore should return EINVAL when address is not a multiple of system page size.
   266  		const size = 256 << 10 // size of memory region to allocate
   267  		p, err := mmap(nil, size, _PROT_READ|_PROT_WRITE, _MAP_ANON|_MAP_PRIVATE, -1, 0)
   268  		if err != 0 {
   269  			return
   270  		}
   271  		var n uintptr
   272  		for n = 4 << 10; n < size; n <<= 1 {
   273  			err := mincore(unsafe.Pointer(uintptr(p)+n), 1, &addrspace_vec[0])
   274  			if err == 0 {
   275  				physPageSize = n
   276  				break
   277  			}
   278  		}
   279  		if physPageSize == 0 {
   280  			physPageSize = size
   281  		}
   282  		munmap(p, size)
   283  		return
   284  	}
   285  
   286  	n = read(fd, noescape(unsafe.Pointer(&auxvreadbuf[0])), int32(unsafe.Sizeof(auxvreadbuf)))
   287  	closefd(fd)
   288  	if n < 0 {
   289  		return
   290  	}
   291  	// Make sure buf is terminated, even if we didn't read
   292  	// the whole file.
   293  	auxvreadbuf[len(auxvreadbuf)-2] = _AT_NULL
   294  	pairs := sysauxv(auxvreadbuf[:])
   295  	auxv = auxvreadbuf[: pairs*2 : pairs*2]
   296  }
   297  
   298  // secureMode holds the value of AT_SECURE passed in the auxiliary vector.
   299  var secureMode bool
   300  
   301  func sysauxv(auxv []uintptr) (pairs int) {
   302  	// Process the auxiliary vector entries provided by the kernel when the
   303  	// program is executed. See getauxval(3).
   304  	var i int
   305  	for ; auxv[i] != _AT_NULL; i += 2 {
   306  		tag, val := auxv[i], auxv[i+1]
   307  		switch tag {
   308  		case _AT_RANDOM:
   309  			// The kernel provides a pointer to 16 bytes of cryptographically
   310  			// random data. Note that in cgo programs this value may have
   311  			// already been used by libc at this point, and in particular glibc
   312  			// and musl use the value as-is for stack and pointer protector
   313  			// cookies from libc_start_main and/or dl_start. Also, cgo programs
   314  			// may use the value after we do.
   315  			startupRand = (*[16]byte)(unsafe.Pointer(val))[:]
   316  
   317  		case _AT_PAGESZ:
   318  			physPageSize = val
   319  
   320  		case _AT_SECURE:
   321  			secureMode = val == 1
   322  		}
   323  
   324  		archauxv(tag, val)
   325  		vdsoauxv(tag, val)
   326  	}
   327  	return i / 2
   328  }
   329  
   330  var sysTHPSizePath = []byte("/sys/kernel/mm/transparent_hugepage/hpage_pmd_size\x00")
   331  
   332  func getHugePageSize() uintptr {
   333  	var numbuf [20]byte
   334  	fd := open(&sysTHPSizePath[0], 0 /* O_RDONLY */, 0)
   335  	if fd < 0 {
   336  		return 0
   337  	}
   338  	ptr := noescape(unsafe.Pointer(&numbuf[0]))
   339  	n := read(fd, ptr, int32(len(numbuf)))
   340  	closefd(fd)
   341  	if n <= 0 {
   342  		return 0
   343  	}
   344  	n-- // remove trailing newline
   345  	v, ok := strconv.Atoi(slicebytetostringtmp((*byte)(ptr), int(n)))
   346  	if !ok || v < 0 {
   347  		v = 0
   348  	}
   349  	if v&(v-1) != 0 {
   350  		// v is not a power of 2
   351  		return 0
   352  	}
   353  	return uintptr(v)
   354  }
   355  
   356  func osinit() {
   357  	numCPUStartup = getCPUCount()
   358  	physHugePageSize = getHugePageSize()
   359  	osArchInit()
   360  	vgetrandomInit()
   361  }
   362  
   363  var urandom_dev = []byte("/dev/urandom\x00")
   364  
   365  func readRandom(r []byte) int {
   366  	// Note that all supported Linux kernels should provide AT_RANDOM which
   367  	// populates startupRand, so this fallback should be unreachable.
   368  	fd := open(&urandom_dev[0], 0 /* O_RDONLY */, 0)
   369  	n := read(fd, unsafe.Pointer(&r[0]), int32(len(r)))
   370  	closefd(fd)
   371  	return int(n)
   372  }
   373  
   374  func goenvs() {
   375  	goenvs_unix()
   376  }
   377  
   378  // Called to do synchronous initialization of Go code built with
   379  // -buildmode=c-archive or -buildmode=c-shared.
   380  // None of the Go runtime is initialized.
   381  //
   382  //go:nosplit
   383  //go:nowritebarrierrec
   384  func libpreinit() {
   385  	initsig(true)
   386  }
   387  
   388  // Called to initialize a new m (including the bootstrap m).
   389  // Called on the parent thread (main thread in case of bootstrap), can allocate memory.
   390  func mpreinit(mp *m) {
   391  	mp.gsignal = malg(32 * 1024) // Linux wants >= 2K
   392  	mp.gsignal.m = mp
   393  }
   394  
   395  func gettid() uint32
   396  
   397  // Called to initialize a new m (including the bootstrap m).
   398  // Called on the new thread, cannot allocate memory.
   399  func minit() {
   400  	minitSignals()
   401  
   402  	// Cgo-created threads and the bootstrap m are missing a
   403  	// procid. We need this for asynchronous preemption and it's
   404  	// useful in debuggers.
   405  	getg().m.procid = uint64(gettid())
   406  }
   407  
   408  // Called from dropm to undo the effect of an minit.
   409  //
   410  //go:nosplit
   411  func unminit() {
   412  	unminitSignals()
   413  	getg().m.procid = 0
   414  }
   415  
   416  // Called from mexit, but not from dropm, to undo the effect of thread-owned
   417  // resources in minit, semacreate, or elsewhere. Do not take locks after calling this.
   418  //
   419  // This always runs without a P, so //go:nowritebarrierrec is required.
   420  //
   421  //go:nowritebarrierrec
   422  func mdestroy(mp *m) {
   423  }
   424  
   425  // #ifdef GOARCH_386
   426  // #define sa_handler k_sa_handler
   427  // #endif
   428  
   429  func sigreturn__sigaction()
   430  func sigtramp() // Called via C ABI
   431  func cgoSigtramp()
   432  
   433  //go:noescape
   434  func sigaltstack(new, old *stackt)
   435  
   436  //go:noescape
   437  func setitimer(mode int32, new, old *itimerval)
   438  
   439  //go:noescape
   440  func timer_create(clockid int32, sevp *sigevent, timerid *int32) int32
   441  
   442  //go:noescape
   443  func timer_settime(timerid int32, flags int32, new, old *itimerspec) int32
   444  
   445  //go:noescape
   446  func timer_delete(timerid int32) int32
   447  
   448  //go:noescape
   449  func rtsigprocmask(how int32, new, old *sigset, size int32)
   450  
   451  //go:nosplit
   452  //go:nowritebarrierrec
   453  func sigprocmask(how int32, new, old *sigset) {
   454  	rtsigprocmask(how, new, old, int32(unsafe.Sizeof(*new)))
   455  }
   456  
   457  func raise(sig uint32)
   458  func raiseproc(sig uint32)
   459  
   460  //go:noescape
   461  func sched_getaffinity(pid, len uintptr, buf *byte) int32
   462  func osyield()
   463  
   464  //go:nosplit
   465  func osyield_no_g() {
   466  	osyield()
   467  }
   468  
   469  func pipe2(flags int32) (r, w int32, errno int32)
   470  
   471  //go:nosplit
   472  func fcntl(fd, cmd, arg int32) (ret int32, errno int32) {
   473  	r, _, err := syscall.Syscall6(syscall.SYS_FCNTL, uintptr(fd), uintptr(cmd), uintptr(arg), 0, 0, 0)
   474  	return int32(r), int32(err)
   475  }
   476  
   477  const (
   478  	_si_max_size    = 128
   479  	_sigev_max_size = 64
   480  )
   481  
   482  //go:nosplit
   483  //go:nowritebarrierrec
   484  func setsig(i uint32, fn uintptr) {
   485  	var sa sigactiont
   486  	sa.sa_flags = _SA_SIGINFO | _SA_ONSTACK | _SA_RESTORER | _SA_RESTART
   487  	sigfillset(&sa.sa_mask)
   488  	// Although Linux manpage says "sa_restorer element is obsolete and
   489  	// should not be used". x86_64 kernel requires it. Only use it on
   490  	// x86.
   491  	if GOARCH == "386" || GOARCH == "amd64" {
   492  		sa.sa_restorer = abi.FuncPCABI0(sigreturn__sigaction)
   493  	}
   494  	if fn == abi.FuncPCABIInternal(sighandler) { // abi.FuncPCABIInternal(sighandler) matches the callers in signal_unix.go
   495  		if iscgo {
   496  			fn = abi.FuncPCABI0(cgoSigtramp)
   497  		} else {
   498  			fn = abi.FuncPCABI0(sigtramp)
   499  		}
   500  	}
   501  	sa.sa_handler = fn
   502  	sigaction(i, &sa, nil)
   503  }
   504  
   505  //go:nosplit
   506  //go:nowritebarrierrec
   507  func setsigstack(i uint32) {
   508  	var sa sigactiont
   509  	sigaction(i, nil, &sa)
   510  	if sa.sa_flags&_SA_ONSTACK != 0 {
   511  		return
   512  	}
   513  	sa.sa_flags |= _SA_ONSTACK
   514  	sigaction(i, &sa, nil)
   515  }
   516  
   517  //go:nosplit
   518  //go:nowritebarrierrec
   519  func getsig(i uint32) uintptr {
   520  	var sa sigactiont
   521  	sigaction(i, nil, &sa)
   522  	return sa.sa_handler
   523  }
   524  
   525  // setSignalstackSP sets the ss_sp field of a stackt.
   526  //
   527  //go:nosplit
   528  func setSignalstackSP(s *stackt, sp uintptr) {
   529  	*(*uintptr)(unsafe.Pointer(&s.ss_sp)) = sp
   530  }
   531  
   532  //go:nosplit
   533  func (c *sigctxt) fixsigcode(sig uint32) {
   534  }
   535  
   536  // sysSigaction calls the rt_sigaction system call.
   537  //
   538  //go:nosplit
   539  func sysSigaction(sig uint32, new, old *sigactiont) {
   540  	if rt_sigaction(uintptr(sig), new, old, unsafe.Sizeof(sigactiont{}.sa_mask)) != 0 {
   541  		// Workaround for bugs in QEMU user mode emulation.
   542  		//
   543  		// QEMU turns calls to the sigaction system call into
   544  		// calls to the C library sigaction call; the C
   545  		// library call rejects attempts to call sigaction for
   546  		// SIGCANCEL (32) or SIGSETXID (33).
   547  		//
   548  		// QEMU rejects calling sigaction on SIGRTMAX (64).
   549  		//
   550  		// Just ignore the error in these case. There isn't
   551  		// anything we can do about it anyhow.
   552  		if sig != 32 && sig != 33 && sig != 64 {
   553  			// Use system stack to avoid split stack overflow on ppc64/ppc64le.
   554  			systemstack(func() {
   555  				throw("sigaction failed")
   556  			})
   557  		}
   558  	}
   559  }
   560  
   561  // rt_sigaction is implemented in assembly.
   562  //
   563  //go:noescape
   564  func rt_sigaction(sig uintptr, new, old *sigactiont, size uintptr) int32
   565  
   566  func getpid() int
   567  func tgkill(tgid, tid, sig int)
   568  
   569  // signalM sends a signal to mp.
   570  func signalM(mp *m, sig int) {
   571  	tgkill(getpid(), int(mp.procid), sig)
   572  }
   573  
   574  // validSIGPROF compares this signal delivery's code against the signal sources
   575  // that the profiler uses, returning whether the delivery should be processed.
   576  // To be processed, a signal delivery from a known profiling mechanism should
   577  // correspond to the best profiling mechanism available to this thread. Signals
   578  // from other sources are always considered valid.
   579  //
   580  //go:nosplit
   581  func validSIGPROF(mp *m, c *sigctxt) bool {
   582  	code := int32(c.sigcode())
   583  	setitimer := code == _SI_KERNEL
   584  	timer_create := code == _SI_TIMER
   585  
   586  	if !(setitimer || timer_create) {
   587  		// The signal doesn't correspond to a profiling mechanism that the
   588  		// runtime enables itself. There's no reason to process it, but there's
   589  		// no reason to ignore it either.
   590  		return true
   591  	}
   592  
   593  	if mp == nil {
   594  		// Since we don't have an M, we can't check if there's an active
   595  		// per-thread timer for this thread. We don't know how long this thread
   596  		// has been around, and if it happened to interact with the Go scheduler
   597  		// at a time when profiling was active (causing it to have a per-thread
   598  		// timer). But it may have never interacted with the Go scheduler, or
   599  		// never while profiling was active. To avoid double-counting, process
   600  		// only signals from setitimer.
   601  		//
   602  		// When a custom cgo traceback function has been registered (on
   603  		// platforms that support runtime.SetCgoTraceback), SIGPROF signals
   604  		// delivered to a thread that cannot find a matching M do this check in
   605  		// the assembly implementations of runtime.cgoSigtramp.
   606  		return setitimer
   607  	}
   608  
   609  	// Having an M means the thread interacts with the Go scheduler, and we can
   610  	// check whether there's an active per-thread timer for this thread.
   611  	if mp.profileTimerValid.Load() {
   612  		// If this M has its own per-thread CPU profiling interval timer, we
   613  		// should track the SIGPROF signals that come from that timer (for
   614  		// accurate reporting of its CPU usage; see issue 35057) and ignore any
   615  		// that it gets from the process-wide setitimer (to not over-count its
   616  		// CPU consumption).
   617  		return timer_create
   618  	}
   619  
   620  	// No active per-thread timer means the only valid profiler is setitimer.
   621  	return setitimer
   622  }
   623  
   624  func setProcessCPUProfiler(hz int32) {
   625  	setProcessCPUProfilerTimer(hz)
   626  }
   627  
   628  func setThreadCPUProfiler(hz int32) {
   629  	mp := getg().m
   630  	mp.profilehz = hz
   631  
   632  	// destroy any active timer
   633  	if mp.profileTimerValid.Load() {
   634  		timerid := mp.profileTimer
   635  		mp.profileTimerValid.Store(false)
   636  		mp.profileTimer = 0
   637  
   638  		ret := timer_delete(timerid)
   639  		if ret != 0 {
   640  			print("runtime: failed to disable profiling timer; timer_delete(", timerid, ") errno=", -ret, "\n")
   641  			throw("timer_delete")
   642  		}
   643  	}
   644  
   645  	if hz == 0 {
   646  		// If the goal was to disable profiling for this thread, then the job's done.
   647  		return
   648  	}
   649  
   650  	// The period of the timer should be 1/Hz. For every "1/Hz" of additional
   651  	// work, the user should expect one additional sample in the profile.
   652  	//
   653  	// But to scale down to very small amounts of application work, to observe
   654  	// even CPU usage of "one tenth" of the requested period, set the initial
   655  	// timing delay in a different way: So that "one tenth" of a period of CPU
   656  	// spend shows up as a 10% chance of one sample (for an expected value of
   657  	// 0.1 samples), and so that "two and six tenths" periods of CPU spend show
   658  	// up as a 60% chance of 3 samples and a 40% chance of 2 samples (for an
   659  	// expected value of 2.6). Set the initial delay to a value in the uniform
   660  	// random distribution between 0 and the desired period. And because "0"
   661  	// means "disable timer", add 1 so the half-open interval [0,period) turns
   662  	// into (0,period].
   663  	//
   664  	// Otherwise, this would show up as a bias away from short-lived threads and
   665  	// from threads that are only occasionally active: for example, when the
   666  	// garbage collector runs on a mostly-idle system, the additional threads it
   667  	// activates may do a couple milliseconds of GC-related work and nothing
   668  	// else in the few seconds that the profiler observes.
   669  	spec := new(itimerspec)
   670  	spec.it_value.setNsec(1 + int64(cheaprandn(uint32(1e9/hz))))
   671  	spec.it_interval.setNsec(1e9 / int64(hz))
   672  
   673  	var timerid int32
   674  	var sevp sigevent
   675  	sevp.notify = _SIGEV_THREAD_ID
   676  	sevp.signo = _SIGPROF
   677  	sevp.sigev_notify_thread_id = int32(mp.procid)
   678  	ret := timer_create(_CLOCK_THREAD_CPUTIME_ID, &sevp, &timerid)
   679  	if ret != 0 {
   680  		// If we cannot create a timer for this M, leave profileTimerValid false
   681  		// to fall back to the process-wide setitimer profiler.
   682  		return
   683  	}
   684  
   685  	ret = timer_settime(timerid, 0, spec, nil)
   686  	if ret != 0 {
   687  		print("runtime: failed to configure profiling timer; timer_settime(", timerid,
   688  			", 0, {interval: {",
   689  			spec.it_interval.tv_sec, "s + ", spec.it_interval.tv_nsec, "ns} value: {",
   690  			spec.it_value.tv_sec, "s + ", spec.it_value.tv_nsec, "ns}}, nil) errno=", -ret, "\n")
   691  		throw("timer_settime")
   692  	}
   693  
   694  	mp.profileTimer = timerid
   695  	mp.profileTimerValid.Store(true)
   696  }
   697  
   698  // perThreadSyscallArgs contains the system call number, arguments, and
   699  // expected return values for a system call to be executed on all threads.
   700  type perThreadSyscallArgs struct {
   701  	trap uintptr
   702  	a1   uintptr
   703  	a2   uintptr
   704  	a3   uintptr
   705  	a4   uintptr
   706  	a5   uintptr
   707  	a6   uintptr
   708  	r1   uintptr
   709  	r2   uintptr
   710  }
   711  
   712  // perThreadSyscall is the system call to execute for the ongoing
   713  // doAllThreadsSyscall.
   714  //
   715  // perThreadSyscall may only be written while mp.needPerThreadSyscall == 0 on
   716  // all Ms.
   717  var perThreadSyscall perThreadSyscallArgs
   718  
   719  // syscall_runtime_doAllThreadsSyscall and executes a specified system call on
   720  // all Ms.
   721  //
   722  // The system call is expected to succeed and return the same value on every
   723  // thread. If any threads do not match, the runtime throws.
   724  //
   725  //go:linkname syscall_runtime_doAllThreadsSyscall syscall.runtime_doAllThreadsSyscall
   726  //go:uintptrescapes
   727  func syscall_runtime_doAllThreadsSyscall(trap, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, err uintptr) {
   728  	if iscgo {
   729  		// In cgo, we are not aware of threads created in C, so this approach will not work.
   730  		panic("doAllThreadsSyscall not supported with cgo enabled")
   731  	}
   732  
   733  	// STW to guarantee that user goroutines see an atomic change to thread
   734  	// state. Without STW, goroutines could migrate Ms while change is in
   735  	// progress and e.g., see state old -> new -> old -> new.
   736  	//
   737  	// N.B. Internally, this function does not depend on STW to
   738  	// successfully change every thread. It is only needed for user
   739  	// expectations, per above.
   740  	stw := stopTheWorld(stwAllThreadsSyscall)
   741  
   742  	// This function depends on several properties:
   743  	//
   744  	// 1. All OS threads that already exist are associated with an M in
   745  	//    allm. i.e., we won't miss any pre-existing threads.
   746  	// 2. All Ms listed in allm will eventually have an OS thread exist.
   747  	//    i.e., they will set procid and be able to receive signals.
   748  	// 3. OS threads created after we read allm will clone from a thread
   749  	//    that has executed the system call. i.e., they inherit the
   750  	//    modified state.
   751  	//
   752  	// We achieve these through different mechanisms:
   753  	//
   754  	// 1. Addition of new Ms to allm in allocm happens before clone of its
   755  	//    OS thread later in newm.
   756  	// 2. newm does acquirem to avoid being preempted, ensuring that new Ms
   757  	//    created in allocm will eventually reach OS thread clone later in
   758  	//    newm.
   759  	// 3. We take allocmLock for write here to prevent allocation of new Ms
   760  	//    while this function runs. Per (1), this prevents clone of OS
   761  	//    threads that are not yet in allm.
   762  	allocmLock.lock()
   763  
   764  	// Disable preemption, preventing us from changing Ms, as we handle
   765  	// this M specially.
   766  	//
   767  	// N.B. STW and lock() above do this as well, this is added for extra
   768  	// clarity.
   769  	acquirem()
   770  
   771  	// N.B. allocmLock also prevents concurrent execution of this function,
   772  	// serializing use of perThreadSyscall, mp.needPerThreadSyscall, and
   773  	// ensuring all threads execute system calls from multiple calls in the
   774  	// same order.
   775  
   776  	r1, r2, errno := syscall.Syscall6(trap, a1, a2, a3, a4, a5, a6)
   777  	if GOARCH == "ppc64" || GOARCH == "ppc64le" {
   778  		// TODO(https://go.dev/issue/51192 ): ppc64 doesn't use r2.
   779  		r2 = 0
   780  	}
   781  	if errno != 0 {
   782  		releasem(getg().m)
   783  		allocmLock.unlock()
   784  		startTheWorld(stw)
   785  		return r1, r2, errno
   786  	}
   787  
   788  	perThreadSyscall = perThreadSyscallArgs{
   789  		trap: trap,
   790  		a1:   a1,
   791  		a2:   a2,
   792  		a3:   a3,
   793  		a4:   a4,
   794  		a5:   a5,
   795  		a6:   a6,
   796  		r1:   r1,
   797  		r2:   r2,
   798  	}
   799  
   800  	// Wait for all threads to start.
   801  	//
   802  	// As described above, some Ms have been added to allm prior to
   803  	// allocmLock, but not yet completed OS clone and set procid.
   804  	//
   805  	// At minimum we must wait for a thread to set procid before we can
   806  	// send it a signal.
   807  	//
   808  	// We take this one step further and wait for all threads to start
   809  	// before sending any signals. This prevents system calls from getting
   810  	// applied twice: once in the parent and once in the child, like so:
   811  	//
   812  	//          A                     B                  C
   813  	//                         add C to allm
   814  	// doAllThreadsSyscall
   815  	//   allocmLock.lock()
   816  	//   signal B
   817  	//                         <receive signal>
   818  	//                         execute syscall
   819  	//                         <signal return>
   820  	//                         clone C
   821  	//                                             <thread start>
   822  	//                                             set procid
   823  	//   signal C
   824  	//                                             <receive signal>
   825  	//                                             execute syscall
   826  	//                                             <signal return>
   827  	//
   828  	// In this case, thread C inherited the syscall-modified state from
   829  	// thread B and did not need to execute the syscall, but did anyway
   830  	// because doAllThreadsSyscall could not be sure whether it was
   831  	// required.
   832  	//
   833  	// Some system calls may not be idempotent, so we ensure each thread
   834  	// executes the system call exactly once.
   835  	for mp := allm; mp != nil; mp = mp.alllink {
   836  		for atomic.Load64(&mp.procid) == 0 {
   837  			// Thread is starting.
   838  			osyield()
   839  		}
   840  	}
   841  
   842  	// Signal every other thread, where they will execute perThreadSyscall
   843  	// from the signal handler.
   844  	gp := getg()
   845  	tid := gp.m.procid
   846  	for mp := allm; mp != nil; mp = mp.alllink {
   847  		if atomic.Load64(&mp.procid) == tid {
   848  			// Our thread already performed the syscall.
   849  			continue
   850  		}
   851  		mp.needPerThreadSyscall.Store(1)
   852  		signalM(mp, sigPerThreadSyscall)
   853  	}
   854  
   855  	// Wait for all threads to complete.
   856  	for mp := allm; mp != nil; mp = mp.alllink {
   857  		if mp.procid == tid {
   858  			continue
   859  		}
   860  		for mp.needPerThreadSyscall.Load() != 0 {
   861  			osyield()
   862  		}
   863  	}
   864  
   865  	perThreadSyscall = perThreadSyscallArgs{}
   866  
   867  	releasem(getg().m)
   868  	allocmLock.unlock()
   869  	startTheWorld(stw)
   870  
   871  	return r1, r2, errno
   872  }
   873  
   874  // runPerThreadSyscall runs perThreadSyscall for this M if required.
   875  //
   876  // This function throws if the system call returns with anything other than the
   877  // expected values.
   878  //
   879  //go:nosplit
   880  func runPerThreadSyscall() {
   881  	gp := getg()
   882  	if gp.m.needPerThreadSyscall.Load() == 0 {
   883  		return
   884  	}
   885  
   886  	args := perThreadSyscall
   887  	r1, r2, errno := syscall.Syscall6(args.trap, args.a1, args.a2, args.a3, args.a4, args.a5, args.a6)
   888  	if GOARCH == "ppc64" || GOARCH == "ppc64le" {
   889  		// TODO(https://go.dev/issue/51192 ): ppc64 doesn't use r2.
   890  		r2 = 0
   891  	}
   892  	if errno != 0 || r1 != args.r1 || r2 != args.r2 {
   893  		print("trap:", args.trap, ", a123456=[", args.a1, ",", args.a2, ",", args.a3, ",", args.a4, ",", args.a5, ",", args.a6, "]\n")
   894  		print("results: got {r1=", r1, ",r2=", r2, ",errno=", errno, "}, want {r1=", args.r1, ",r2=", args.r2, ",errno=0}\n")
   895  		fatal("AllThreadsSyscall6 results differ between threads; runtime corrupted")
   896  	}
   897  
   898  	gp.m.needPerThreadSyscall.Store(0)
   899  }
   900  
   901  const (
   902  	_SI_USER     = 0
   903  	_SI_TKILL    = -6
   904  	_SYS_SECCOMP = 1
   905  )
   906  
   907  // sigFromUser reports whether the signal was sent because of a call
   908  // to kill or tgkill.
   909  //
   910  //go:nosplit
   911  func (c *sigctxt) sigFromUser() bool {
   912  	code := int32(c.sigcode())
   913  	return code == _SI_USER || code == _SI_TKILL
   914  }
   915  
   916  // sigFromSeccomp reports whether the signal was sent from seccomp.
   917  //
   918  //go:nosplit
   919  func (c *sigctxt) sigFromSeccomp() bool {
   920  	code := int32(c.sigcode())
   921  	return code == _SYS_SECCOMP
   922  }
   923  
   924  //go:nosplit
   925  func mprotect(addr unsafe.Pointer, n uintptr, prot int32) (ret int32, errno int32) {
   926  	r, _, err := syscall.Syscall6(syscall.SYS_MPROTECT, uintptr(addr), n, uintptr(prot), 0, 0, 0)
   927  	return int32(r), int32(err)
   928  }
   929  

View as plain text