Implement madvise(MADV_DONTFORK)

PiperOrigin-RevId: 254253777

5 files changed, 72 insertions, 1 deletions

diff --git a/pkg/sentry/mm/lifecycle.go b/pkg/sentry/mm/lifecycle.go
index 06e4372ff..4e9ca1de6 100644
--- a/pkg/sentry/mm/lifecycle.go
+++ b/pkg/sentry/mm/lifecycle.go
@@ -86,10 +86,22 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) {
 	}
 
 	// Copy vmas.
+	dontforks := false
 	dstvgap := mm2.vmas.FirstGap()
 	for srcvseg := mm.vmas.FirstSegment(); srcvseg.Ok(); srcvseg = srcvseg.NextSegment() {
 		vma := srcvseg.Value() // makes a copy of the vma
 		vmaAR := srcvseg.Range()
+
+		if vma.dontfork {
+			length := uint64(vmaAR.Length())
+			mm2.usageAS -= length
+			if vma.isPrivateDataLocked() {
+				mm2.dataAS -= length
+			}
+			dontforks = true
+			continue
+		}
+
 		// Inform the Mappable, if any, of the new mapping.
 		if vma.mappable != nil {
 			if err := vma.mappable.AddMapping(ctx, mm2, vmaAR, vma.off, vma.canWriteMappableLocked()); err != nil {
@@ -118,6 +130,10 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) {
 	defer mm2.activeMu.Unlock()
 	mm.activeMu.Lock()
 	defer mm.activeMu.Unlock()
+	if dontforks {
+		defer mm.pmas.MergeRange(mm.applicationAddrRange())
+	}
+	srcvseg := mm.vmas.FirstSegment()
 	dstpgap := mm2.pmas.FirstGap()
 	var unmapAR usermem.AddrRange
 	for srcpseg := mm.pmas.FirstSegment(); srcpseg.Ok(); srcpseg = srcpseg.NextSegment() {
@@ -125,6 +141,27 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) {
 		if !pma.private {
 			continue
 		}
+
+		if dontforks {
+			// Find the 'vma' that contains the starting address
+			// associated with the 'pma' (there must be one).
+			srcvseg = srcvseg.seekNextLowerBound(srcpseg.Start())
+			if checkInvariants {
+				if !srcvseg.Ok() {
+					panic(fmt.Sprintf("no vma covers pma range %v", srcpseg.Range()))
+				}
+				if srcpseg.Start() < srcvseg.Start() {
+					panic(fmt.Sprintf("vma %v ran ahead of pma %v", srcvseg.Range(), srcpseg.Range()))
+				}
+			}
+
+			srcpseg = mm.pmas.Isolate(srcpseg, srcvseg.Range())
+			if srcvseg.ValuePtr().dontfork {
+				continue
+			}
+			pma = srcpseg.ValuePtr()
+		}
+
 		if !pma.needCOW {
 			pma.needCOW = true
 			if pma.effectivePerms.Write {
diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go
index 2ec2ad99b..7bb96b159 100644
--- a/pkg/sentry/mm/mm.go
+++ b/pkg/sentry/mm/mm.go
@@ -274,6 +274,9 @@ type vma struct {
 	// metag, none of which we currently support.
 	growsDown bool `state:"manual"`
 
+	// dontfork is the MADV_DONTFORK setting for this vma configured by madvise().
+	dontfork bool
+
 	mlockMode memmap.MLockMode
 
 	// numaPolicy is the NUMA policy for this vma set by mbind().
diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go
index 9aa39e31d..c2466c988 100644
--- a/pkg/sentry/mm/syscalls.go
+++ b/pkg/sentry/mm/syscalls.go
@@ -1026,6 +1026,32 @@ func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy
 	}
 }
 
+// SetDontFork implements the semantics of madvise MADV_DONTFORK.
+func (mm *MemoryManager) SetDontFork(addr usermem.Addr, length uint64, dontfork bool) error {
+	ar, ok := addr.ToRange(length)
+	if !ok {
+		return syserror.EINVAL
+	}
+
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+	defer func() {
+		mm.vmas.MergeRange(ar)
+		mm.vmas.MergeAdjacent(ar)
+	}()
+
+	for vseg := mm.vmas.LowerBoundSegment(ar.Start); vseg.Ok() && vseg.Start() < ar.End; vseg = vseg.NextSegment() {
+		vseg = mm.vmas.Isolate(vseg, ar)
+		vma := vseg.ValuePtr()
+		vma.dontfork = dontfork
+	}
+
+	if mm.vmas.SpanRange(ar) != ar.Length() {
+		return syserror.ENOMEM
+	}
+	return nil
+}
+
 // 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 9f846cdb8..074e2b141 100644
--- a/pkg/sentry/mm/vma.go
+++ b/pkg/sentry/mm/vma.go
@@ -439,6 +439,7 @@ func (vmaSetFunctions) Merge(ar1 usermem.AddrRange, vma1 vma, ar2 usermem.AddrRa
 		vma1.mlockMode != vma2.mlockMode ||
 		vma1.numaPolicy != vma2.numaPolicy ||
 		vma1.numaNodemask != vma2.numaNodemask ||
+		vma1.dontfork != vma2.dontfork ||
 		vma1.id != vma2.id ||
 		vma1.hint != vma2.hint {
 		return vma{}, false
diff --git a/pkg/sentry/syscalls/linux/sys_mmap.go b/pkg/sentry/syscalls/linux/sys_mmap.go
index 8a45dceeb..d831833bc 100644
--- a/pkg/sentry/syscalls/linux/sys_mmap.go
+++ b/pkg/sentry/syscalls/linux/sys_mmap.go
@@ -180,6 +180,10 @@ func Madvise(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca
 	switch adv {
 	case linux.MADV_DONTNEED:
 		return 0, nil, t.MemoryManager().Decommit(addr, length)
+	case linux.MADV_DOFORK:
+		return 0, nil, t.MemoryManager().SetDontFork(addr, length, false)
+	case linux.MADV_DONTFORK:
+		return 0, nil, t.MemoryManager().SetDontFork(addr, length, true)
 	case linux.MADV_HUGEPAGE, linux.MADV_NOHUGEPAGE:
 		fallthrough
 	case linux.MADV_MERGEABLE, linux.MADV_UNMERGEABLE:
@@ -191,7 +195,7 @@ func Madvise(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca
 	case linux.MADV_NORMAL, linux.MADV_RANDOM, linux.MADV_SEQUENTIAL, linux.MADV_WILLNEED:
 		// Do nothing, we totally ignore the suggestions above.
 		return 0, nil, nil
-	case linux.MADV_REMOVE, linux.MADV_DOFORK, linux.MADV_DONTFORK:
+	case linux.MADV_REMOVE:
 		// These "suggestions" have application-visible side effects, so we
 		// have to indicate that we don't support them.
 		return 0, nil, syserror.ENOSYS