Quan Tian <tianquan23@xxxxxxxxx> wrote:
> - if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
> - nft_trans_destroy(trans);
> - break;
> + if (trans->ctx.table->flags & __NFT_TABLE_F_UPDATE) {
> + if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
> + nf_tables_table_disable(net, trans->ctx.table);
> + trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
> }
> - if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
> - nf_tables_table_disable(net, trans->ctx.table);
> -
> - trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
> + kfree(trans->ctx.table->udata);
This kfree() needs to happen in nft_commit_release().
Otherwise netlink dumps (which run without a lock) results in
use-after-free.
Second issue is that ->udata and ->udlen would have to be
updated atomically so that async dump doesn't observe old/new
pointer with a larger udlen.
I suggest a preparation patch that replaces
u16 udlen;
u8 *udata;
in struct nft_table with
struct nlattr *udata;
so its enough to swap() ctx.table->udata with nft_trans_table_udata(),
then kfree() the old udata from nft_commit_release().
At _release time we can be sure no other cpu is referencing the old
udata.
