Friday, May 29, 2015

Random notes on the Go 1.4 runtime.

These are some notes on the Go 1.4 Run-Time written while working on the new Clive kernel. I just copied the text from the TR, which will be linked at the Lsub page and updated in the future. They should be easy to follow using a Go 1.4 tree because the package and function names are indicated. The line numbers are useful mostly for our local use in Acme at Lsub, using the text output from wr(1).

Francisco J. Ballesteros
TR LSUB-15-2 draft


Abstract

This TR contains notes about the Go run time as of Go 1.4. They are intended to aid in the construction of a kernel for Clive.

1. Initialization

  • rt0_darwin_amd64.s:13 main() is the entry point for the binary. A direct jump to...

  • asm_amd64.s:9: runtime·rt0_go which is the actual crt0. This initializes the stack calls _cgo_init if needed, updates stack guards, setups TLS, sets m to m0 and g to g0 and continues calling...

    • runtime·args to save the arguments from the OS
    • os_darwin.c:54: runtime·osinit to initialize threading (they don't use their libc)
    • proc.c:122: runtime·schedinit(void) to create a G that calls runtime·main.
      • traceback.go:48: func tracebackinit()
      • symtab.go:46: func symtabinit
      • stack.c:42: runtime·stackinit(void) calls this function for the stacks in the pool:
        • mheap.c:663: runtime·MSpanList_Init
      • malloc.c:109: runtime·mallocinit
        • mheap.c:57: runtime·MHeap_Init initializes the heap using these:
          • mfixalloc.c:16: runtime·FixAlloc_Init, for several allocs.
          • mheap.c:663: runtime·MSpanList_Init, for free and busy lists.
          • mcentral.c:21: runtime·MCentral_Init, for mcentral lists.
        • mcache.c:19: runtime·allocmcache sets g->m->mcache, a per-P malloc cache for small objects.
      • proc.c:190: mcommoninit calls os-specific mpreinit and links m into the sched list (allm).
      • goargs and goenvs save the UNIX arguments.
      • mgc0.c:1345: runtime·gcinit sets GC masks for data and bss.
      • proc.c:2655: procresize is called to create Ps for GOMAXPROCS.
    • proc.c:2154: runtime·newproc is called to run runtime·main using
      • asm_amd64.s:218: runtime·onM to call newproc_m) to run runtime·main. This switches to g0, if not on m stack (i.e., no switch this time).
        • proc.c:2128: newproc_m is called by g0 and copies the args and then calls:
          • proc.c:2182: runtime·newproc1. This creates a G to run a function (runtime·main). Using:
            • proc.c:2295: gfget to get a G from the free list, or
            • proc.c:2102: runtime·malg to allocate a G, which ends up doing a new(g) in go.
            Then it copies arguments to the new G stack, sets the sched label to call the function, using
            • stack.c:806: runtime·gostartcallfn, similar to set-label.
            Sets the state to Grunnable and calls
            • proc.c:3284: runqput to put the new G on a local runq or the global one.
            • proc.c:1276: wakep, which is a nop or a call to sched the M for P or make a new M for P.
              • proc.c:1200: startm
    • proc.c:818: runtime·mstart starts the M. It calls:
      • asminit to do per-thread initialization (nop for our arch).
      • os_darwin.c:144: minit initializes signal handling for our arch.
      • proc.c:1537: schedule runs the scheduller and calls this when gets something to run:
        • proc.c:1358: execute is mostly a call to a goto-label:
          • asm_amd64.s:143: gogo

    At the end, the new G calls
  • proc.go:16: runtime·main(), which calls
    • runtime_init
    • main_init
    • main_main
    • exit(0)

    These are the initialization code and main program for the binary. The runtime·init function in proc.go spawns the goroutine to call gogc(1) in a loop.

    2. Labels

    A Gobuf is similar to a Label:

          struct    Gobuf
          {
              uintptr    sp;    // the actual label.
              uintptr    pc;    // the actual label.
              G*    g;    // the actual label.
              void*    ctxt;    // aux context pointer.
              uintreg    ret;    // return value
              uintptr    lr;    // link register used in ARM.
          };
          

    The main operations are:

    • runtime·gosave, i.e., set label. Clears ctxt and ret.
    • runtime·gogo, i.e., goto label. Sets ctxt in DX, used by runtime·morestack and returns ret after the actual jump to the label.

    Here, pc, sp, and g are the actual context saved and restored. That is the label.

    The field ctxt is used to point to a frame in the stack and seems to be the user context in other places. Seems to be an auxiliary pointer but it is not clear (yet) how it is used. The field lr seems to be the link register for the ARM and is not used by our architecture.

    3. Sleep/Wakeup

    These are giant locks:

          void    runtime·stoptheworld(void);
          void    runtime·starttheworld(void);
          extern uint32 runtime·worldsema;
          

    used for example to dump all the stacks.

    These are the usual locks with a user-level fast path:

          void    runtime·lock(Mutex*);
          void    runtime·unlock(Mutex*);
          

    These are sleep/wakeup like structures, and timed out variants for them with the usual conventions of at most one calling sleep and at most one calling wakeup for it. Plus, there is a clear operation to reset the thing for a further use:

          void    runtime·noteclear(Note*);
          void    runtime·notesleep(Note*);
          void    runtime·notewakeup(Note*);
          bool    runtime·notetsleep(Note*, int64);  // false - timeout
          bool    runtime·notetsleepg(Note*, int64);  // false - timeout
          

    And these are futexes or semaphores (eg., on Darwin) to implement the ones above:

          uintptr    runtime·semacreate(void);
          int32    runtime·semasleep(int64);
          void    runtime·semawakeup(M*);
          

    In many cases lock-free data structures are used; or at least parts of them are handled lock-free by using atomic and CAS-like operations, eg. to change statuses of processes and to link them to a list.

    4. Scheduling

    Most of the interesting code is in proc.c. There are three central structures:
    G
    A goroutine. g refers to the current and g0 is the idle process. This is a user-level thread. g is a register on the ARM and a slot in TLS everywhere else.
    M
    A machine, actually a UNIX process. m refers to the current one. This is a kernel-level thread that runs Gs or is idle or is in a syscall.
    P
    A processor, actually a per GOMAXPROCS sched.

    The idea is that M takes a P when it must run Gs. Ps were introduced to do job stealing.

    The interesting bits of G are:

          struct    G
          {
              Stack    stack;    // [stack.lo, stack.hi).
              uintptr    stackguard0;    // stack.lo+StackGuard or StackPreempt
              uintptr    stackguard1;    // stack.lo+StackGuard on g0 or ~0 on others
      
              Panic*    panic;    // innermost panic - offset known to liblink
              Defer*    defer;    // innermost defer
              Gobuf    sched;
              uintptr    syscallsp;    // if status==Gsyscall, syscallsp = sched.sp to use during gc
              uintptr    syscallpc;    // if status==Gsyscall, syscallpc = sched.pc to use during gc
              void*    param;    // passed parameter on wakeup
              int64    goid;
              G*    schedlink;
              bool    preempt;    // preemption signal, dup of stackguard0 = StackPreempt
              M*    m;    // for debuggers, but offset not hard-coded
              M*    lockedm;
              SudoG*    waiting;    // sudog structures this G is waiting on
              ...
          };
          

    The interesting bits of M are:

          struct    M
          {
              G*    g0;        // goroutine with scheduling stack
              G*    gsignal;    // signal-handling G
              uintptr    tls[4];        // thread-local storage (for x86 extern register)
              void    (*mstartfn)(void);
              G*    curg;        // current running goroutine
              G*    caughtsig;    // goroutine running during fatal signal
              P*    p;        // attached P for executing Go code (nil if not executing Go code)
              int32    mallocing,throwing,gcing;
              int32    locks;
              bool    spinning;    // M is out of work and is actively looking for work
              bool    blocked;    // M is blocked on a Note
              Note    park;
              M*    alllink;    // on allm
              M*    schedlink;
              MCache*    mcache;
              G*    lockedg;
              uint32    locked;        // tracking for LockOSThread
              M*    nextwaitm;    // next M waiting for lock
              uintptr    waitsema;    // semaphore for parking on locks
              uint32    waitsemacount;
              uint32    waitsemalock;
              bool    (*waitunlockf)(G*, void*);
              void*    waitlock;
              uintptr scalararg[4];    // scalar argument/return for mcall
              void*   ptrarg[4];    // pointer argument/return for mcall
              ...
          };
      
          

    A value of m->locks greater than zero prevents preemption and also prevents garbage colloection.

    The interesting bits of P are:

          struct P
          {
              Mutex    lock;
              uint32    status;        // Pidle, Prunning, Psyscall, Pgcstop, Pdead
              P*    link;
              uint32    schedtick;        // incremented on every scheduler call
              uint32    syscalltick;    // incremented on every system call
              M*    m;        // back-link to associated M (nil if idle)
              MCache*    mcache;
              Defer*    deferpool[5];    // pool of available Defers
      
              // Cache of goroutine ids, amortizes accesses to runtime·sched.goidgen.
              uint64    goidcache;
              uint64    goidcacheend;
      
              // Queue of runnable goroutines.
              uint32    runqhead;
              uint32    runqtail;
              G*    runq[256];
      
              // Available G's (status == Gdead)
              G*    gfree;
              int32    gfreecnt;
          };
          

    This is a local scheduler plus cached structures per UNIX process used to run Go code. And then there is a global scheduler structure that also caches some structures.

          struct    SchedT
          {
              Mutex    lock;
              uint64    goidgen;
              M*    midle;     // idle m's waiting for work
              P*    pidle;      // idle P's
              G*    runqhead; // Global runnable queue.
              Mutex    gflock;    // global cache go dead Gs
              G*    gfree;
              uint32    gcwaiting;    // gc is waiting to run
              int32    stopwait;
              Note    stopnote;
              ...
          };
          

    There is a single runtime·sched, and runtime·sched.lock protects it all.

    4.1. Process creation

    Code like go fn() is translated as a call to proc.c:2154: runtime·newproc as shown in the first section for initialization. This records the function and a pointer to the ... arguments yn g->m->ptrarg[] and then calls newproc_m using runtime·onM.
    • There's a switch to M's g0
    • newproc_m runs there.
    • And there's a switch back when done.

    Then newproc_m takes the function and arguments from m->ptrarg and calls runtime·newproc1:
    • This adds to m->locks to avoid preemption and takes P from m->p.
    • A G is taken from the cache at P or allocated (with StackMin stack bytes).

    At this point G is Gdead. Then arguments are copied into G's stack and it's prepared:
    • It's label pc, sp, and g are set to the function pointer and the new stack and g, and then adjusted to pretend that such function did call gosave, so we can gogo (gotolabel) it. In this case, the label ctxt is set to the FuncVal value going. The return value is set to call goexit+PCQuantum which is a call to goexit1.
    • Its state is changed to Grunnable (cas op.)
    • A new id is taken from the id cache at P (perhaps refilling the cache).

    Now it's put in the scheduler queue by calling runqput(p, newg). This uses atomic load/store to put newg at p->runq[tail], using p->runqtail as a increasing counter for the tail that is copied to tail and truncated as an index for runq. This happens if p->runq is not full.

    If p->runq is full, runqputslow is called to take half of the local queue at P and place them at the global runtime·sched.runqhead/tail queue while holding the global scheduler lock. if the CAS operation to operate on the local queue fails, the whole runqput is retried.

    4.2. Process termination

    Performed by a call to
    • proc.c:1682: runtime·goexit1, that does an mcall to run goexit0 at g0.
    • goexit0 (or any other mcall) never returns and calls gogo to sched to somewhere else, usually to g->sched to let it run later. The G state is set to Gdead. Then these are called:
      • proc.c:1598: dropg, to release m->curg->m and m->curg (g is now g0).
      • proc.c:2259: gfput, to put the dying G at p->gfree, linked through g->schedlink. If there are too many free Gs cached, they are moved to the global list at runtime·sched.gfree.
      • schedule, runs one scheduler loop and jumps to a new G.

    Note that returning from the main function of a goroutine returns to
    • asm_amd64.s:2237: runtime·goexit, which calls runtime·goexit1.

    4.3. Context switches

    Calls to schedule can be found at:
    • proc.c:661: mquiesce, used to run code at g0 and then schedule back a G.
    • proc.c:868: mstart, to jump to a new g0 for a new M.
    • proc.c:1655: runtime·park_m, used to make G wait for something.
    • proc.c:1675: runtime·gosched_m, an mcall from Gosched at proc.go.
    • proc.c:1718: goexit0, to run g0 or other Gs when exiting.
    • proc.c:2022: exitsyscall0, called after a system call at g0.
    • malloc.go:482: gogc calls Gosched when the GC is done.
    • mgc0.go:87: bgsweep calls Gosched.

    Plus, if g->stackguard0 is StackPreempt, then asm_amd64.s:824:runtime·stackcheck calls morestack and preemts if the goroutine seems to be running user code (not runtime code), by calling gosched_m like Gosched does. I have not checked out if the compiler inserts calls in loops that do not perform function calls (which check the stack), and those might never be preempted; which does not matter much for us now becase they would be bugs in the code and not the normal case. The stack checks are inserted silently by the linker unless NOSPLIT is indicated, and thus normal function calls are a source of possible preemptions.

    As a side-efect, a NOSPLIT function call is not preemptible by this mechanism.

    The code in schedule runs the scheduler:

    • There can be no m->locks, or it's a bug.
    • If m->lockedg then
      • proc.c:1287: stoplockedm is called, to stop the current G at M until G can run again. This is done by lending m->p to another M, if we have a P using handoffp(releasep()), and then calling runtime·notesleep(&g->m->park) to sleep. Later acquirep(m->nextp) is used to get a new P before returning and...
      • proc.c:1358: execute is called for m->lockedg.

    • If runtime·sched.gcwaiting then gcstopm is called, which calls releasep and stopm; This happens when stoptheworld is called and, once we are done schedule restarts again.

    At this point schedule picks a ready G. Once in a while from the global pool, and usually from m->p.

    • proc.c:3333: runqget picks one G from the queue using p->runqhead and CAS.
      • When there is no G ready to run, proc.c:1381: findrunnable tries to steal calling globrunqget and, runtime·netpoll (!!).
      • Then if there's no work, m->spinning is set and it tries to steal from others.
      • When everything fails, it calls stopm and blocks.

    Once it gets a G to run:
    • proc.c:1358: execute is called, which is mostly a call to a goto-label:
      • asm_amd64.s:143: gogo

    5. System calls

    All system calls call
  • asm_darwin_amd64.s:19 syscall.Syscall or Syscall6 that
    • calls proc.c:1761 runtime·reentersyscall (from runtime·entersyscall)

        to prepare for a system call

      • This increments m->locks to avoid preemption, and tricks g->stackguard0 to make sure the no split-stack happens.
      • Sets G status to Gsyscall
      • Saves the PC and SP in g->sched.
      • Releases p->m and m->mcache
      • Sets P status to Psyscall
      • Calls entersyscall_gcwait when runtime·sched.gcwaiting
      • Decrements m->locks before proceding further.
    • moves parameters into registers and executes
    • executes SYSCALL, and, upon return from the system call, calls
    • calls proc.c:1893 runtime·exitsyscall calls directly
      • exitsyscallfast to either re-acquire the last P or or get another idle P.
      • or runs exitsyscall0 as an mcall to run the scheduler and then return to continue when the scheduler returns and G can run again.
        • This calls pidleget and then acquirep and execute or stopm and schedule, depending.
    • and returns to the caller.

    There are other wrappers for system calls but in the end they call the above entry points.

    Two other functions, syscall·RawSyscall and ·RawSyscall6 are wrappers that issue the system call without doing anything: no calls to runtime·entersyscall and no calls to runtime·exitsyscall. I don't know why they are not called from the entry points above.

    6. Signals

    There is a global runtime·sigtab[sig] table with entries of this data type:

          struct    SigTab
          {
              int32    flags;
              int8    *name;
          };
          enum
          {
              SigNotify = 1<<0,    // let signal.Notify have signal, even if from kernel
              SigKill = 1<<1,        // if signal.Notify doesn't take it, exit quietly
              SigThrow = 1<<2,    // if signal.Notify doesn't take it, exit loudly
              SigPanic = 1<<3,    // if the signal is from the kernel, panic
              SigDefault = 1<<4,    // if the signal isn't explicitly requested, don't monitor it
              SigHandling = 1<<5,    // our signal handler is registered
              SigIgnored = 1<<6,    // the signal was ignored before we registered for it
              SigGoExit = 1<<7,    // cause all runtime procs to exit (only used on Plan 9).
          };
          

    Signals are handled in their own signal stack. When a new M is initialized,

    • runtime·minit is called, eg., for Darwin, it calls
      • sys_darwin_amd64.s:265: runtime·sigaltstack to set the signal stack context, and
      • sys_darwin_386.s:218: runtime·sigprocmask to set the signal mask to none.

    Later on, when dropm is called to release an M,
    • runtime·unminit is called and it calls runtime·signalstack to unset the signal stack.

    The signal package init function:
    • calls runtime·signal_enable, which when called for the first time sets sig.inuse and clears the sig.note note, to prepare for handling signals.

    • loops calling process(syscall.Signal(signal_recv())). Here,
      • sig.s:21 signal_recv is just a jump to
      • sigqueue.go:91 runtime·signal_recv, which returns a signal number when it's posted by the OS.

    When signal.Notify is called to install a handler:
    • signal.go:49 Notify records in a handlers table for the signal given or for all the signals, and calls
      • signal_unix.go:47 signal·enableSignal, which is just a call to
      • sigqueue.go:128 runtime·signal_enable. This sets the signal in sig.wanted and calls sigenable_go which is just a call, using onM to this:
        • signal.c:8 runtime·sigenable_m enables a signal calling with the signal number kept at m->scalararg[0]. It is just a call to:
        • signal_unix.c:44 runtime·sigenable enables a signal, making a call to the next with runtime·sighandler as the handler function.
          • os_darwin.c:519: runtime·setsig calls runtime·sigaction specifying runtime·sigtramp as the handler and supplying the handler function (runtime·sighandler) in the sigaction structure.

    Later, when a signal arrives:
    • sys_darwin_386.s:240 runtime·sigtramp
      • saves g and sets m->gsignal as the current G, and then calls the actual handler:
        • signal_amd64x.c:44 runtime·sighandler looks into the global runtime·sigtab and might panic if not handled or deliver the signal by calling the next.
          • sigqueue.go:48 runtime·sigsend is called with a signal number to send the signal to receivers. The signal is queued in a global if there is no receiver ready but the signal is wanted and is not already in the queue.

    From here, the loop started in the singal·init function would send the signal to the users channel with a non-blocking send. That is, if there is not enough buffering in the channel signals are lost; but that's ok.

    7. Memory allocation

    Memory is allocated by calls to new. The implementation for this builtin is:
    • malloc.go:348 newobject, which calls to mallocgc using flagNoScan as a flag if the type is marked as having no pointers.
      • malloc.go:46 mallocgc is the main entry point for memory allocation.

    Slices are created by calling malloc.go:357 newarray, which does the same. Raw memory is allocated by calls to rawmem which also calls mallocgc with flags flagNoScan and flagNoZero; i.e., that's almost a direct malloc call, but for the GC.

    There are several allocation classes. Sizes are recorded in runtime·class_to_size[]. Alignments are:

    • 8, for sizes under 16 bytes and sizes that are not a power of 2.
    • 16, for sizes under 128 bytes,
    • size/8, for sizes under 2KiB,
    • 256, for sizes starting at 2KiB.
    Regarding sizes:
    • Sizes under MaxSmallSize (32KiB), use the larger size that keeps the same number of objects in each page. They are allocated from a per-P cache.
    • Sizes starting at 32KiB are rounded to PageSize. They are allocated from the global heap.

    The allocator used depends on the allocation:
    • Objects with no pointers and less than 16 bytes (maxTinySize) are allocated within a single allocation and share it. This is done in the P cache. The allocation is released when all such objects are collected. This is done in the P cache.
    • Objects up to 1024-8 bytes go into a size class rounded to multiples of 8 bytes.
    • Objects up to 32KiB are rounded to multiples of 128 bytes. This is done in the P cache.
    • Objects larger than 32KiB are allocated on the heap by calling largeAlloc_m onM.

    The heap operates using pages, but smaller allocators up in the hierarchy uses bytes. This is how it works:
    • The tiny allocations use the tiny allocation from the MCache structure at p->mcache. There is one per P and it requires no locks:
          struct MCache
          {
              byte*    tiny;
              uintptr    tinysize;
              // The rest is not accessed on every malloc.
              MSpan*    alloc[NumSizeClasses];    // spans to allocate from
      
              StackFreeList stackcache[NumStackOrders];
              SudoG*    sudogcache;
              ...
          };
          

      When tiny is exhausted, a new one is taken from Mcache.alloc, and, if that is exhausted, a new one is allocated by a call to runtime·MCache_Refill, using onM. This asks runtime·MCentral_CacheSpan to allocate a new span and places it into Mache.alloc.

    • The allocations with pointers or starting at 16 bytes, try first to use p->cache.alloc to get an allocation of the right size. If this fails, MCache_Refill is called as before, onM.

    • Large allocations starting at 32KiB call largeAlloc_m onM.
      • malloc.c:372: runtime·largeAlloc_m rounds the size to a multiple of PageSize and calls:
        • mheap.c:231: runtime·MHeap_Alloc calls mheap_alloc directly if it's a call made by g0, or mcalls it.
          • mheap.c:171 mheap_alloc allocates n memory pages.
        • mgc0.c:1815: runtime·markspan is called to tell the GC.

    Unless the allocation is flagNoScan, the GC bitmap is updated by looking at the type given to mallocgc to record which words are pointers. Also, the unused part at the end of the actual allocation is marked in the GC bitmap as dead.

    Looking at the MHeap now, for allocations under 1Mbyte (for 4KiB pages), there is a list of allocation with exactly that page size. For larger allocations there is a final large list used.


  • No comments:

    Post a Comment