"Implement" mbind(2).

We still only advertise a single NUMA node, and ignore mempolicy
accordingly, but mbind() at least now succeeds and has effects reflected
by get_mempolicy().

Also fix handling of nodemasks: round sizes to unsigned long (as
documented and done by Linux), and zero trailing bits when copying them
out.

PiperOrigin-RevId: 251950859

3 files changed, 62 insertions, 0 deletions

diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go
index 0a026ff8c..604866d04 100644
--- a/pkg/sentry/mm/mm.go
+++ b/pkg/sentry/mm/mm.go
@@ -276,6 +276,12 @@ type vma struct {
 
 	mlockMode memmap.MLockMode
 
+	// numaPolicy is the NUMA policy for this vma set by mbind().
+	numaPolicy int32
+
+	// numaNodemask is the NUMA nodemask for this vma set by mbind().
+	numaNodemask uint64
+
 	// If id is not nil, it controls the lifecycle of mappable and provides vma
 	// metadata shown in /proc/[pid]/maps, and the vma holds a reference.
 	id memmap.MappingIdentity
diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go
index af1e53f5d..9cf136532 100644
--- a/pkg/sentry/mm/syscalls.go
+++ b/pkg/sentry/mm/syscalls.go
@@ -973,6 +973,59 @@ func (mm *MemoryManager) MLockAll(ctx context.Context, opts MLockAllOpts) error
 	return nil
 }
 
+// NumaPolicy implements the semantics of Linux's get_mempolicy(MPOL_F_ADDR).
+func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (int32, uint64, error) {
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	vseg := mm.vmas.FindSegment(addr)
+	if !vseg.Ok() {
+		return 0, 0, syserror.EFAULT
+	}
+	vma := vseg.ValuePtr()
+	return vma.numaPolicy, vma.numaNodemask, nil
+}
+
+// SetNumaPolicy implements the semantics of Linux's mbind().
+func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy int32, nodemask uint64) error {
+	if !addr.IsPageAligned() {
+		return syserror.EINVAL
+	}
+	// Linux allows this to overflow.
+	la, _ := usermem.Addr(length).RoundUp()
+	ar, ok := addr.ToRange(uint64(la))
+	if !ok {
+		return syserror.EINVAL
+	}
+	if ar.Length() == 0 {
+		return nil
+	}
+
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+	defer func() {
+		mm.vmas.MergeRange(ar)
+		mm.vmas.MergeAdjacent(ar)
+	}()
+	vseg := mm.vmas.LowerBoundSegment(ar.Start)
+	lastEnd := ar.Start
+	for {
+		if !vseg.Ok() || lastEnd < vseg.Start() {
+			// "EFAULT: ... there was an unmapped hole in the specified memory
+			// range specified [sic] by addr and len." - mbind(2)
+			return syserror.EFAULT
+		}
+		vseg = mm.vmas.Isolate(vseg, ar)
+		vma := vseg.ValuePtr()
+		vma.numaPolicy = policy
+		vma.numaNodemask = nodemask
+		lastEnd = vseg.End()
+		if ar.End <= lastEnd {
+			return nil
+		}
+		vseg, _ = vseg.NextNonEmpty()
+	}
+}
+
 // Decommit implements the semantics of Linux's madvise(MADV_DONTNEED).
 func (mm *MemoryManager) Decommit(addr usermem.Addr, length uint64) error {
 	ar, ok := addr.ToRange(length)
diff --git a/pkg/sentry/mm/vma.go b/pkg/sentry/mm/vma.go
index 02203f79f..0af8de5b0 100644
--- a/pkg/sentry/mm/vma.go
+++ b/pkg/sentry/mm/vma.go
@@ -107,6 +107,7 @@ func (mm *MemoryManager) createVMALocked(ctx context.Context, opts memmap.MMapOp
 		private:        opts.Private,
 		growsDown:      opts.GrowsDown,
 		mlockMode:      opts.MLockMode,
+		numaPolicy:     linux.MPOL_DEFAULT,
 		id:             opts.MappingIdentity,
 		hint:           opts.Hint,
 	}
@@ -436,6 +437,8 @@ func (vmaSetFunctions) Merge(ar1 usermem.AddrRange, vma1 vma, ar2 usermem.AddrRa
 		vma1.private != vma2.private ||
 		vma1.growsDown != vma2.growsDown ||
 		vma1.mlockMode != vma2.mlockMode ||
+		vma1.numaPolicy != vma2.numaPolicy ||
+		vma1.numaNodemask != vma2.numaNodemask ||
 		vma1.id != vma2.id ||
 		vma1.hint != vma2.hint {
 		return vma{}, false