Re: [PATCH v10 perf, bpf-next 1/9] perf, bpf: Introduce PERF_RECORD_KSYMBOL

[Peter Zijlstra]

On Thu, Jan 17, 2019 at 01:56:53PM +0100, Peter Zijlstra wrote:
> +static __always_inline struct latch_tree_node *
> +latch_tree_first(struct latch_tree_root *root)
> +{
> +	struct latch_tree_node *ltn = NULL;
> +	struct rb_node *node;
> +	unsigned int seq;
> +
> +	do {
> +		struct rb_root *rbr;
> +
> +		seq = raw_read_seqcount_latch(&root->seq);
> +		rbr = &root->tree[seq & 1];
> +		node = rb_first(rbr);
> +	} while (read_seqcount_retry(&root->seq, seq));
> +
> +	if (node)
> +		ltn = __lt_from_rb(node, seq & 1);
> +
> +	return ltn;
> +}
> +
> +/**
> + * latch_tree_next() - find the next @ltn in @root per sort order
> + * @root: trees to which @ltn belongs
> + * @ltn: nodes to start from
> + *
> + * Does a lockless lookup in the trees @root for the next node starting at
> + * @ltn.
> + *
> + * Using this function outside of the write side lock is rather dodgy but given
> + * latch_tree_erase() doesn't re-init the nodes and the whole iteration is done
> + * under a single RCU critical section, it should be non-fatal and generate some
> + * semblance of order - albeit possibly missing chunks of the tree.
> + */
> +static __always_inline struct latch_tree_node *
> +latch_tree_next(struct latch_tree_root *root, struct latch_tree_node *ltn)
> +{
> +	struct rb_node *node;
> +	unsigned int seq;
> +
> +	do {
> +		seq = raw_read_seqcount_latch(&root->seq);
> +		node = rb_next(&ltn->node[seq & 1]);
> +	} while (read_seqcount_retry(&root->seq, seq));
> +
> +	return __lt_from_rb(node, seq & 1);
> +}

> +static int kallsym_tree_kallsym(unsigned int symnum, unsigned long *value, char *type,
> +				char *sym, char *modname, int *exported)
> +{
> +	struct latch_tree_node *ltn;
> +	int i, ret = -ERANGE;
> +
> +	rcu_read_lock();
> +	for (i = 0, ltn = latch_tree_first(&kallsym_tree); i < symnum && ltn;
> +	     i++, ltn = latch_tree_next(&kallsym_tree, ltn))
> +		;

On second thought; I don't think this will be good enough after all.
Missing entire subtrees is too much.

The rcu-list iteration will only miss newly added symbols, and for those
we'll get the events, combined we'll still have a complete picture. Not
so when a whole subtree goes missing.

I thought I could avoid the list this way, but alas, not so.

> +
> +	if (ltn) {
> +		struct kallsym_node *kn;
> +		char *mod;
> +
> +		kn = container_of(ltn, struct kallsym_node, kn_node);
> +
> +		kn->kn_names(kn, sym, &mod);
> +		if (mod)
> +			strlcpy(modname, mod, MODULE_NAME_LEN);
> +		else
> +			modname[0] = '\0';
> +
> +		*type = 't';
> +		*exported = 0;
> +		ret = 0;
> +	}
> +	rcu_read_unlock();
> +
> +	return ret;
> +}