Currently, wbBufFlush does nothing if the goroutine is dying on the
assumption that the system is crashing anyway and running the write
barrier may crash it even more. However, it fails to reset the
buffer's "next" pointer. As a result, if there are later write
barriers on the same P, the write barrier will overflow the write
barrier buffer and start corrupting other fields in the P or other
heap objects. Often, this corrupts fields in the next allocated P
since they tend to be together in the heap.
Fix this by always resetting the buffer's "next" pointer, even if
we're not doing anything with the pointers in the buffer.
Updates #22987 and #22988. (May fix; it's hard to say.)
Change-Id: I82c11ea2d399e1658531c3e8065445a66b7282b2
Reviewed-on: https://go-review.googlesource.com/83016
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
}
}
+// discard resets b's next pointer, but not its end pointer.
+//
+// This must be nosplit because it's called by wbBufFlush.
+//
+//go:nosplit
+func (b *wbBuf) discard() {
+ b.next = uintptr(unsafe.Pointer(&b.buf[0]))
+}
+
// putFast adds old and new to the write barrier buffer and returns
// false if a flush is necessary. Callers should use this as:
//
//go:nowritebarrierrec
//go:nosplit
func wbBufFlush(dst *uintptr, src uintptr) {
+ // Note: Every possible return from this function must reset
+ // the buffer's next pointer to prevent buffer overflow.
+
if getg().m.dying > 0 {
// We're going down. Not much point in write barriers
// and this way we can allow write barriers in the
// panic path.
+ getg().m.p.ptr().wbBuf.discard()
return
}
cgoCheckWriteBarrier(dst, src)
if !writeBarrier.needed {
// We were only called for cgocheck.
- b := &getg().m.p.ptr().wbBuf
- b.next = uintptr(unsafe.Pointer(&b.buf[0]))
+ getg().m.p.ptr().wbBuf.discard()
return
}
}