How an asynchronous cryptographic callback race condition in net/tls/tls_sw.c freed network pages while still queued for zero-copy transmission.
Imagine you order a meal for delivery. The chef starts cooking, but you call and cancel the order. Because the kitchen staff wasn't notified properly, the delivery driver picks up an empty plate, drives to a new customer's house, and serves them whatever was left on the counter, contaminating the new customer's food.
The attacker creates a TLS socket and enables kernel encryption offload via setsockopt(TCP_ULP, "tls").
The attacker sends data using MSG_ZEROCOPY, queuing pages for hardware crypto.
The attacker closes the socket before the crypto accelerator completes its async callback.
The cleanup handler frees memory pages that the delayed hardware callback subsequently overwrites.
Disajikan dalam bahasa pemrograman tingkat tinggi yang mudah dibaca (tanpa assembly rumit).
// VULNERABLE: net/tls/tls_sw.c before patch
static void tls_encrypt_done(void *data, int err) {
struct tls_context *ctx = data;
struct tls_sw_context_tx *ctx_tx = tls_sw_ctx_tx(ctx);
// ROOT CAUSE:
// Asynchronous completion callback assumes socket context is still locked!
// If the socket was closed while crypto was in flight, ctx_tx is already freed!
clear_bit(TLS_TX_SYNC_MORE, &ctx_tx->tx_bitmask);
tls_free_open_rec(ctx);
}
// SECURE: net/tls/tls_sw.c patch
static void tls_encrypt_done(void *data, int err) {
struct tls_context *ctx = data;
// 1. Verify context reference counter before dereferencing context pointers
if (!refcount_inc_not_zero(&ctx->refcount)) {
return; // Socket is already dying, abort callback safely!
}
struct tls_sw_context_tx *ctx_tx = tls_sw_ctx_tx(ctx);
clear_bit(TLS_TX_SYNC_MORE, &ctx_tx->tx_bitmask);
tls_free_open_rec(ctx);
// 2. Drop reference cleanly
refcount_dec(&ctx->refcount);
}