[v2] parser: don't keep alloca()ing in a loop for substitutions

Commit Message

Ahelenia Ziemiańska Dec. 15, 2022, 5:02 p.m. UTC

When encountering
  printf %010000d | tr 0 \` | sh -n
  printf %09999d  | tr 0 \` | sh -n
you want no output and "Syntax error: EOF in backquote substitution",
respectively; instead, current dash segfaults.

This is because the alloca for the save buffer is run, naturally,
in the same function, so first it allocates one byte, then two,
then ..., then appx. 4000 (for me, depends on the binary),
then it segfaults on the memcpy (it's even worse, since due to
alignment, it usually allocates much more for the early stuff).

Nevertheless, the stack frame grows unboundedly, until we completely
destroy the stack. Instead of squirreling the out block away, then
letting subsequent allocations override the original, mark it used,
and just re-copy it to the top of the dash stack. This increases peak
memory usage somewhat
(in the most pathological case ‒ the above but with three nines ‒
 from 23.26 to 173.7KiB according to massif,
 in parsing a regular program (ratrun from ratrun 0c)
 from 28.68 to 29.19;
 a simpler program (ibid., rat) stays at 5.422;
 parsing libtoolize, debootstrap, and dkms
 (the biggest shell programs in my /[s]bin by size + by `/$( count)
 likewise stay the same at 12.02, 41.48, and 6.438)
but it's barely measurable outside of truly pathological conditions
that were a step away from a segfault previously.

Fixes: https://bugs.debian.org/966156
---
Naturally, I hadn't considered that.

This version I've run through valgrind in a good few configurations and
am happy to conclude there are no leaks (and the memory usage bump is
imperceptible unless you were almost-crashing anyway).

 src/parser.c | 13 ++++---------
 1 file changed, 4 insertions(+), 9 deletions(-)

Comments

Herbert Xu Jan. 5, 2023, 9:02 a.m. UTC | #1

наб <[e-mail]> wrote:
>
> Naturally, I hadn't considered that.
> 
> This version I've run through valgrind in a good few configurations and
> am happy to conclude there are no leaks (and the memory usage bump is
> imperceptible unless you were almost-crashing anyway).

Nice work!

> @@ -1446,10 +1443,8 @@ done:
>        if (oldstyle)
>                tokpushback = 0;
>        out = growstackto(savelen + 1);
> -       if (str) {
> -               memcpy(out, str, savelen);
> -               STADJUST(savelen, out);
> -       }
> +       memcpy(out, str, savelen);
> +       STADJUST(savelen, out);

Minor nit but these three lines can be combined into:

	out = stnputs(stackblock(), str, savelen);

Thanks,

Patch

diff --git a/src/parser.c b/src/parser.c
index a552c47..89698cb 100644
--- a/src/parser.c
+++ b/src/parser.c
@@ -1360,12 +1360,9 @@  parsebackq: {
 	struct heredoc *saveheredoclist;
 	int uninitialized_var(saveprompt);
 
-	str = NULL;
+	str = stackblock();
 	savelen = out - (char *)stackblock();
-	if (savelen > 0) {
-		str = alloca(savelen);
-		memcpy(str, stackblock(), savelen);
-	}
+	grabstackblock(savelen);
         if (oldstyle) {
                 /* We must read until the closing backquote, giving special
                    treatment to some slashes, and then push the string and
@@ -1446,10 +1443,8 @@  done:
 	if (oldstyle)
 		tokpushback = 0;
 	out = growstackto(savelen + 1);
-	if (str) {
-		memcpy(out, str, savelen);
-		STADJUST(savelen, out);
-	}
+	memcpy(out, str, savelen);
+	STADJUST(savelen, out);
 	USTPUTC(CTLBACKQ, out);
 	if (oldstyle)
 		goto parsebackq_oldreturn;