Browse zstd specs (44 of 44)

Proofs / zstd / Earlier, longer properties

x_stable_out

ZSTD_d_stableOutBuffer: memory safe on any bytes; on compliant frames, every byte written is the content.

spec
89 lines of C
with its headers
1,383 lines
status
Proved
44 of 44
On this page

The spec

The spec is the C program harness/x_stable_out.c, 89 lines, shown whole as it is in the repository at 2e46017. It includes harness/writer.h (1,231 lines) and harness/zlimits.h (63 lines). Its first comment states the property in full, with every bound.

C
// Law stable_out (PROMISES.md: S7, V12; zstd's experimental API): ZSTD_d_stableOutBuffer, decoding into one output
// buffer the caller never changes ("the ZSTD_outBuffer will ALWAYS be the same between calls").
//
// Claim (QuickProof.Contract.harness_safe of this program): no execution has undefined behaviour, and every gp_assert
// holds: (a) on ANY bytes, over ANY input chunks, into a buffer of ANY size: "this flag is ALWAYS memory safe, and will
// never access out-of-bounds memory", and "0 <= pos <= size" (after an error the stream is reset); (b) on ANY series
// of compliant frames, into a buffer that holds its content ("MUST be large enough to fit the entire decompressed
// frame"): every byte written is the content, no call fails unless zlimits.h (window, memory, or the 16th call in a
// row without progress: stall_error), and a 0 return at the end of the series comes with the whole content.  Draw order for (a): n, bytes, cap (env.c).
#include <stdlib.h>
#include "quickproof.h"
#include "writer.h"
#include "zlimits.h"

#define SRC_MAX 4096
#define DST_MAX 65536
#define CALLS_MAX 4096

static wseries_t sr;

int main(void)
{
	int valid = choose(2);
	size_t n, cap, ipos = 0, calls, i, r, stalled = 0;   /* stalled: calls in a row that read and wrote nothing */
	unsigned char *src, *dst;
	ZSTD_DStream *ds;
	ZSTD_outBuffer out;
	if (valid) {
		w_series(&sr, 0);
		n = sr.n;
		src = exact_copy(sr.bytes, n);
		cap = choose_in(sr.content_n, sr.content_n + 16);
	} else {
		n = nondet_size();
		gp_assume(n <= SRC_MAX);
		src = malloc(n > 0 ? n : 1);
		if (src != NULL)
			nondet_bytes(src, n);
		cap = nondet_size();
		gp_assume(cap <= DST_MAX);
	}
	dst = malloc(cap > 0 ? cap : 1);
	ds = ZSTD_createDStream();
	if (src == NULL || dst == NULL || ds == NULL || ZSTD_isError(ZSTD_DCtx_setParameter(ds, ZSTD_d_stableOutBuffer, 1)))
		n = 0;                          /* out of memory: end the run */
	out.dst = dst;
	out.size = cap;
	out.pos = 0;
	for (calls = 0; n > 0 && calls < CALLS_MAX; calls++) {
		size_t ilen = choose_in(0, n - ipos), pos0 = out.pos;
		unsigned char *ib = exact_copy(src + ipos, ilen);
		ZSTD_inBuffer in;
		if (ib == NULL)
			break;
		in.src = ib;
		in.size = ilen;
		in.pos = 0;
		r = ZSTD_decompressStream(ds, &out, &in);
		gp_assert(in.pos <= in.size);
		gp_assert(out.pos <= out.size);
		ipos += in.pos;
		free(ib);
		if (valid) {
			for (i = pos0; i < out.pos; i++) {
				gp_assert(i < sr.content_n);
				gp_assert(dst[i] == sr.content[i]);
			}
			if (ZSTD_isError(r)) {
				gp_assert(!within_stream(&sr) || ZSTD_getErrorCode(r) == ZSTD_error_memory_allocation
				          || stall_error(ZSTD_getErrorCode(r), stalled, &in, &out, pos0));
				break;
			}
			stalled = in.pos == 0 && out.pos == pos0 ? stalled + 1 : 0;
			if (r == 0 && ipos == n) {
				gp_assert(out.pos == sr.content_n);
				break;
			}
		} else if (ZSTD_isError(r)) {
			ZSTD_DCtx_reset(ds, ZSTD_reset_session_only);
			out.pos = 0;
		} else if (ipos == n && (r == 0 || ilen == 0)) {
			break;
		}
	}
	ZSTD_freeDStream(ds);
	free(src);
	free(dst);
	return 0;
}

harness/writer.h (1,231 lines) is too long to show here; it is in the repository.

harness/zlimits.h, 63 lines
C
// The limits under which zstd may refuse a compliant frame, as its documents state them (PROMISES.md V4): "a decoder
// is allowed to reject a compressed frame which requests a memory size beyond decoder's authorized range"
// (format), zstd's range being ZSTD_WINDOWLOG_MAX (31 on 64-bit: zstd.h) for any decoding and, for streaming,
// "By default, a decompression context accepts window sizes <= (1 << ZSTD_WINDOWLOG_LIMIT_DEFAULT)" (zstd.h);
// the memory allocation of its contexts; and, for streaming, zstd's deliberate busy-loop guard (stall_error).
#ifndef ZSTD_HARNESS_ZLIMITS_H
#define ZSTD_HARNESS_ZLIMITS_H
#include "writer.h"
#ifndef ZSTD_STATIC_LINKING_ONLY
#define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_WINDOWLOG_MAX, ZSTD_WINDOWLOG_LIMIT_DEFAULT */
#endif
#include "zstd.h"
#include "zstd_errors.h"

/* every Zstandard frame of the series is within zstd's window range (single-segment: Window_Size = content size) */
static int within_oneshot(const wseries_t *sr)
{
	unsigned k;
	for (k = 0; k < sr->nframes; k++)
		if (!sr->f[k].skippable && !sr->f[k].single && sr->f[k].window_log > ZSTD_WINDOWLOG_MAX)
			return 0;
	return 1;
}

/* ... and within the streaming default */
static int within_stream(const wseries_t *sr)
{
	unsigned k;
	for (k = 0; k < sr->nframes; k++)
		if (!sr->f[k].skippable && sr->f[k].window > (1ULL << ZSTD_WINDOWLOG_LIMIT_DEFAULT))
			return 0;
	return within_oneshot(sr);
}

/* zstd's busy-loop guard, deliberate (commit 31769ce70 "error on no forward progress": a caller "may be caught in an
   infinite loop, calling the streaming decompression function again and again, without making any progress";
   db18a62f8 gave it its error codes): ZSTD_decompressStream fails on the ZSTD_NO_FORWARD_PROGRESS_MAX-th (16 by
   default) call in a row that reads no input and writes no output, with noForwardProgress_destFull if the output is
   full, else noForwardProgress_inputEmpty (the input is then empty).  zstd.h does not document it.  True iff error
   code e is that error, legitimately: this call (fresh input, output position opos0 before it) read and wrote
   nothing, and so did the `stalled` calls in a row before it. */
#define NO_PROGRESS_MAX 16
static int stall_error(ZSTD_ErrorCode e, size_t stalled, const ZSTD_inBuffer *in, const ZSTD_outBuffer *out,
                       size_t opos0)
{
	if (stalled + 1 < NO_PROGRESS_MAX || in->pos != 0 || out->pos != opos0)
		return 0;
	if (e == ZSTD_error_noForwardProgress_destFull)
		return out->pos == out->size;
	return e == ZSTD_error_noForwardProgress_inputEmpty && out->pos < out->size && in->pos == in->size;
}

/* a copy of b[0..n) in a buffer of exactly n bytes (NULL: out of memory) */
static unsigned char *exact_copy(const unsigned char *b, size_t n)
{
	unsigned char *p = malloc(n > 0 ? n : 1);
	size_t i;
	if (p != NULL)
		for (i = 0; i < n; i++)
			p[i] = b[i];
	return p;
}
#endif

The theorem

QuickProof generated this statement from the spec, in src/contracts/x_stable_out.v:

Rocq
Theorem x_stable_out : harness_safe whole_prog.

src/contracts/laws.v states it on the spec's own program, whole_prog (the spec linked with the code under test), with its proof:

Rocq
Theorem x_stable_out : harness_safe Zstd.contracts.x_stable_out.whole_prog.
Proof. exact (proj1 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 Zstd.contracts.suite.suite_x))))))))))). Qed.

Read as a statement about runs, it says that no run of the program has undefined behaviour or a failing check, for every value the spec can draw. That reading rests on VST's adequacy theorem, which is not yet instantiated in Rocq (formerly Coq) for QuickProof's environment.

The C files (units) its program links
  • patch/decoder.c
  • quickproof/libc/libc.c
  • quickproof/libc/malloc_unit.c

What it leaves out

  • What the first comment of the spec excludes, and its bounds.
  • Like every spec of this proof: termination (a run that never returns is not a counterexample), one thread, and binaries built by a compiler (the theorem is about the code as CompCert defines C for x86_64, LP64).
  • Other configurations of zstd (flags, legacy formats, assembly) and the compressor.
  • The edits to zstd's code, listed in the repository, are checked against the original by testing, not proved.

The summary page says what the proof as a whole covers.

How it was proved

AI agents wrote the proof, and Rocq checks it. qp verify passes on this repository: this statement is the one generated from the spec at the contract commit, and its proof uses only Rocq's standard axioms and VST's abstract malloc library.