One gc thread moves a delay list to a tofree list

Each core's gc thread free elements on the tofree list

One gc thread moves a delay list to a tofree list
1a129849 · Frans Kaashoek · cdaa456b · 1a129849 · 1a129849 · 1a129849
--- a/gc.c
+++ b/gc.c
@@ -9,20 +9,25 @@
 #include "cpu.h"
 #include "kmtrace.h"
-#define NGC 1
+// GC scheme based on Fraser's:
+// a machine has a global_epoch
-// 1 worker in total. more workers makes sense, if we have per-core process
+// a process maintain an epoch (>= global_epoch)
-// lists.
+// one gc thread and state (e.g., NEPOCH delaylists and one tofreelists) per core
-#define NWORKER 1
+// a process add to its core epoch's delayed freelist on delayed_free
-static struct { struct condvar cv __mpalign__; } rcu_cv[NWORKER];
+// a gcc performs two jobs:
-static struct { struct spinlock l __mpalign__; } gc_lock;
+// 1. one gcc thread perform step 1:
+//   updates a thread's epoch, when not in an epoch
+//   compute min over all process's epochs, and sets global_epoch to min
+//   move a core's (global_epoch-2)'s delayed list to a core's tofree list
+//   (costs linear in the number of processes.)
+// 2. in parallel gc threads free the elements on the tofree list (till global_epoach)
+//   (costs linear in the number of elements to be freed)
 enum { gc_debug = 0 };
 struct gc {
  u64 epoch;
  struct gc *next;  
-  struct gc *free;
  union {
    struct {
      void (*dofree)(void *);
@@ -38,16 +43,23 @@ struct gc {
  int type;
 } __mpalign__;
-u64 global_epoch __mpalign__;
+static struct gc_state { 
+  struct condvar cv;
+  struct gc delayed[NEPOCH];
+  struct gc tofree[NEPOCH];
+  int ndelayed;
+  int min_epoch;
+} __mpalign__ gc_state[NCPU] __mpalign__;
-enum { rcu_debug = 0 };
+static struct { struct spinlock l __mpalign__; } gc_lock;
+u64 global_epoch __mpalign__;
 struct gc *
 gc_alloc()
 {
  struct gc *r = kmalloc(sizeof(struct gc));
  assert(r);
-  myproc()->ndelayed++;
+  gc_state[mycpu()->id].ndelayed++;
  return r;
 }
@@ -55,6 +67,11 @@ static void *
 gc_min(void *vkey, void *v, void *arg){
  u64 *min_epoch_p = arg;
  struct proc *p = (struct proc *) v;
+  acquire(&p->gc_epoch_lock);
+  if (p->epoch_depth == 0) {
+    p->epoch = global_epoch;  
+  }
+  release(&p->gc_epoch_lock);
  if (*min_epoch_p > p->epoch) {
      *min_epoch_p = p->epoch;
  }
@@ -78,62 +95,70 @@ gc_free_elem(struct gc *r)
 }
 static int
-gc_free_list(struct gc *head, u64 epoch)
+gc_free_tofreelist(struct gc **head, u64 epoch)
 {
  int nfree = 0;
  struct gc *r, *nr;
-  for (r = head; r != NULL; r = nr) {
+  for (r = *head; r != NULL; r = nr) {
    if (r->epoch > epoch) {
-      cprintf("%lu %lu\n", r->epoch, epoch);
+      cprintf("gc_free_tofreelist: r->epoch %ld > epoch %ld\n", r->epoch, epoch);
      assert(0);
    }
    nr = r->next;
    gc_free_elem(r);
    nfree++;
  }
+  *head = r;
  return nfree;
 }
-// move to free delayed list to free list so that a process can do its own freeing
+// move to free delayed list to tofreelist so that a process can do its own freeing
 void *
-gc_move_to_free_proc(void *vkey, void *v, void *arg){
+gc_move_to_tofree_cpu(int c, u64 epoch)
-  u64 *epoch = arg;
+{
-  struct proc *p = (struct proc *) v;
  struct gc *head;
-  uint32 fe = (*epoch - (NEPOCH-2)) % NEPOCH;
+  uint32 fe = (epoch - (NEPOCH-2)) % NEPOCH;
  int cas;
-  assert(p->gc_epoch[fe].epoch == *epoch-(NEPOCH-2));   // XXX race with setting epoch = 0
+  assert(gc_state[c].delayed[fe].epoch == epoch-(NEPOCH-2));   // XXX race with setting epoch = 0
-  // unhook list for fe epoch atomically
+  // unhook list for fe epoch atomically; this shouldn't fail
-  head = p->gc_epoch[fe].next;
+  head = gc_state[c].delayed[fe].next;
-  // this shouldn't fail, because no core is modifying it.
+  cas = __sync_bool_compare_and_swap(&(gc_state[c].delayed[fe].next), head, 0);
-  cas = __sync_bool_compare_and_swap(&(p->gc_epoch[fe].next), head, 0);
  assert(cas);
-  // insert list into local free list so that each core can do its own frees
-  assert (p->gc_epoch[fe].free == 0);
+  // insert list into tofree list so that each core can free in parallel and free its elements
-  cas = __sync_bool_compare_and_swap(&(p->gc_epoch[fe].free), 0, head);
+  if(gc_state[c].tofree[fe].epoch != gc_state[c].delayed[fe].epoch) {
+    cprintf("%d: tofree epoch %lu delayed epoch %lu\n", c, gc_state[c].tofree[fe].epoch,
+	    gc_state[c].delayed[fe].epoch);
+    assert(0);
+  }
+  cas = __sync_bool_compare_and_swap(&(gc_state[c].tofree[fe].next), 0, head);
  assert(cas);
-  assert(p->gc_epoch[fe].next == 0);
+  // move delayed NEPOCH's adhead
+  gc_state[c].delayed[fe].epoch += NEPOCH;
+  assert(gc_state[c].delayed[fe].next == 0);
  return 0;
 }
-// Fraser's reclaimation scheme: free all delayed-free items in global_epoch-2
 // only one thread should call this function
 static void
-gc_move_to_free(u64 epoch)
+gc_move_to_tofree(u64 epoch)
 {
  if (gc_debug)
-    cprintf("%d: free epoch %ld\n", myproc()->pid, epoch);
+    cprintf("%d: free epoch %ld\n", mycpu()->id, epoch);
-  myproc()->rcu_read_depth++;  // ensure ns_enumate's call to gc_begin_epoch doesn't call gc()
+  for (int c = 0; c < NCPU; c++) {
-  ns_enumerate(nspid, gc_move_to_free_proc, &epoch);
+    gc_move_to_tofree_cpu(c, epoch);
-  myproc()->rcu_read_depth--;
+  }
  int ok  = __sync_bool_compare_and_swap(&global_epoch, epoch, epoch+1);
  assert(ok);
 }
-// If all threads have seen global_epoch, we can free elements in global_epoch-2
+// If all threads have seen global_epoch, we can move elements in global_epoch-2 to tofreelist
 static void
-gc(void)
+gc_delayfreelist(void)
 {
  int r = tryacquire(&gc_lock.l);
  if (r == 0) return;
@@ -141,11 +166,18 @@ gc(void)
  u64 global = global_epoch;
  u64 min = global;
-  myproc()->rcu_read_depth++;  // ensure ns_enumate's call to gc_begin_epoch doesn't call gc()
+  // make that global_epoch doesn't run into a core's min_epoch
+  for (int c = 0; c < NCPU; c++) { 
+    int w = gc_state[c].min_epoch + NEPOCH-1;
+    if (w < min) {
+      min = w;
+    }
+  }
+  myproc()->epoch_depth++;// ensure ns_enumate's call to gc_begin_epoch doesn't have sideeffects
  ns_enumerate(nspid, gc_min, &min);
-  myproc()->rcu_read_depth--;
+  myproc()->epoch_depth--;
  if (min >= global) {
-    gc_move_to_free(min);
+    gc_move_to_tofree(min);
  }
  release(&gc_lock.l);
 }
@@ -154,18 +186,19 @@ static void
 gc_delayed_int(struct gc *r)
 {
  pushcli();
+  int c = mycpu()->id;
  u64 myepoch = myproc()->epoch;
-  u64 minepoch = myproc()->gc_epoch[myepoch % NEPOCH].epoch;
+  u64 minepoch = gc_state[c].delayed[myepoch % NEPOCH].epoch;
  if (gc_debug) 
-    cprintf("%d: gc_delayed: %lu ndelayed %d\n", myproc()->pid, global_epoch, myproc()->ndelayed);
+    cprintf("(%d, %d): gc_delayed: %lu ndelayed %d\n", c, myproc()->pid, global_epoch, gc_state[c].ndelayed);
  if (myepoch != minepoch) {
    cprintf("%d: myepoch %lu minepoch %lu\n", myproc()->pid, myepoch, minepoch);
    panic("gc_delayed_int");
  }
  r->epoch = myepoch;
  do {
-    r->next = myproc()->gc_epoch[myepoch % NEPOCH].next;
+    r->next = gc_state[c].delayed[myepoch % NEPOCH].next;
-  } while (!__sync_bool_compare_and_swap(&(myproc()->gc_epoch[myepoch % NEPOCH].next), r->next, r));
+  } while (!__sync_bool_compare_and_swap(&(gc_state[c].delayed[myepoch % NEPOCH].next), r->next, r));
  popcli();
 }
@@ -194,65 +227,32 @@ gc_delayed2(int a1, u64 a2, void (*dofree)(int,u64))
  gc_delayed_int(r);
 }
-static void*
-gc_free(void *vkey, void *v, void *arg)
-{
-  struct proc *p = (struct proc *) v;
-  acquire(&p->gc_lock);
-  u64 global = global_epoch;
-  for (u64 epoch = p->epoch; epoch < global; epoch++) {
-    int j = (epoch - (NEPOCH - 2)) % NEPOCH;
-    assert(p->gc_epoch[j].epoch == epoch-2);
-    struct gc *free = p->gc_epoch[j].free;
-    int ok = __sync_bool_compare_and_swap(&(p->gc_epoch[j].free), free, NULL);
-    assert(ok);
-    int nfree = gc_free_list(free, epoch - 2);
-    p->ndelayed -= nfree;
-    if (gc_debug && nfree > 0)
-      cprintf("%d: epoch %d freed %d\n", p->pid, epoch - 2, nfree);
-    p->gc_epoch[j].epoch = p->gc_epoch[j].epoch + NEPOCH;
-  }
-  p->epoch = global;  // not atomic, but it never goes backwards
-  __sync_synchronize();
-  release(&p->gc_lock);
-  return NULL;
-}
 void
 gc_start(void)
 {
-  cv_wakeup(&rcu_cv[0].cv);   // NWORKER = 1
+  cv_wakeup(&gc_state[mycpu()->id].cv);
-  //  cv_wakeup(&rcu_cv[mycpu()->id].cv);
 }
 void
 gc_begin_epoch(void)
 {
  if (myproc() == NULL) return;
-  if (myproc()->rcu_read_depth++ > 0)
+  acquire(&myproc()->gc_epoch_lock);
-    return;
+  if (myproc()->epoch_depth++ > 0)
-  gc_free(NULL, (void *) myproc(), NULL);
+    goto done;
+  myproc()->epoch = global_epoch;  // not atomic, but it never goes backwards
+  // __sync_synchronize();
+ done:
+  release(&myproc()->gc_epoch_lock);
 }
 void
 gc_end_epoch(void)
 {
  if (myproc() == NULL) return;
-  if (--myproc()->rcu_read_depth > 0)
+  acquire(&myproc()->gc_epoch_lock);
-    return;
+  --myproc()->epoch_depth;
+  release(&myproc()->gc_epoch_lock);
-#if 0
-  // kick gcc early if under memory pressure
-  int free = 0;
-  for (int j = 0; j < NEPOCH; j++) {
-    if (myproc()->gc_epoch[j].free)
-      free = 1;
-  }
-  u64 nd = myproc()->ndelayed;
-  if (!free && nd > NGC) {
-    gc_start();
-  }
-#endif
 }
 static void
@@ -260,18 +260,25 @@ gc_worker(void *x)
 {
  struct spinlock wl;
+  cprintf("gc_worker: %d\n", mycpu()->id);
  initlock(&wl, "rcu_gc_worker dummy");   // dummy lock
  for (;;) {
+    u64 i;
    acquire(&wl);
+    cv_sleep(&gc_state[mycpu()->id].cv, &wl);  
-    myproc()->rcu_read_depth++;  // call gc_free once for gc_worker
-    ns_enumerate(nspid, gc_free, NULL);
-    myproc()->rcu_read_depth--;
-    gc();
-    cv_sleep(&rcu_cv[0].cv, &wl);   // NWORKER = 1
    release(&wl);
+    u64 global = global_epoch;
+    for (i = gc_state[mycpu()->id].min_epoch; i < global-2; i++) {
+      int nfree = gc_free_tofreelist(&(gc_state[mycpu()->id].tofree[i%NEPOCH].next), i);
+      gc_state[mycpu()->id].tofree[i%NEPOCH].epoch += NEPOCH;
+      if (gc_debug && nfree > 0) {
+	cprintf("%d: epoch %d freed %d\n", mycpu()->id, i, nfree);
+      }
+    }
+    gc_state[mycpu()->id].min_epoch = i;
+    gc_delayfreelist();
  }
 }
@@ -279,13 +286,7 @@ void
 initprocgc(struct proc *p)
 {
  p->epoch = global_epoch;
-  p->gc_epoch = kmalloc(sizeof(struct gc) * NEPOCH);
+  initlock(&p->gc_epoch_lock, "per process gc_lock");
-  initlock(&p->gc_lock, "per process gc_lock");
-  for (u64 i = global_epoch-2; i < global_epoch+2; i++)  {
-    p->gc_epoch[i % NEPOCH].epoch = i;
-    p->gc_epoch[i % NEPOCH].free = NULL;
-    p->gc_epoch[i % NEPOCH].next = NULL;
-  }
 }
@@ -295,12 +296,15 @@ initgc(void)
  initlock(&gc_lock.l, "gc");
  global_epoch = NEPOCH-2;
-  for (int i = 0; i < NWORKER; i++) {
+  for (int i = 0; i < NCPU; i++) {
-    initcondvar(&rcu_cv[i].cv, "rcu_gc_cv");
+    for (int j = 0; j < NEPOCH; j++) {
+      gc_state[i].delayed[j].epoch = j;
+      gc_state[i].tofree[j].epoch = j;
+    }
+    initcondvar(&gc_state[i].cv, "gc_cv");
  }
-  // one worker for now
+  for (u32 c = 0; c < NCPU; c++) {
-  for (u32 c = 0; c < NWORKER; c++) {
    struct proc *gcp; 
    gcp = threadalloc(gc_worker, NULL);

--- a/kernel.h
+++ b/kernel.h
@@ -23,7 +23,6 @@ struct stat;
 struct proc;
 struct vmap;
 struct pipe;
-struct gc;
 // bio.c
 void            binit(void);

--- a/proc.h
+++ b/proc.h
@@ -54,10 +54,8 @@ struct proc {
  SLIST_ENTRY(proc) child_next;
  struct condvar cv;
  u64 epoch;
-  u64 ndelayed;
+  struct spinlock gc_epoch_lock;
-  struct gc *gc_epoch;
+  u64 epoch_depth;
-  struct spinlock gc_lock;
-  u64 rcu_read_depth;
  char lockname[16];
  int on_runq;
  int cpu_pin;