Commit 01e035fd authored by Yevgeny Kliteynik's avatar Yevgeny Kliteynik Committed by Jakub Kicinski
Browse files

net/mlx5: HWS, handle modify header actions dependency



Having adjacent accelerated modify header actions (so-called
pattern-argument actions) may result in inconsistent outcome.
These inconsistencies can take the form of writes to the same
field or a read coupled with a write to the same field. The
solution is to detect such dependencies and insert nops between
the offending actions.

The existing implementation had a few issues, which pretty much
required a complete rewrite of the code that handles these
dependencies.

In the new implementation we're doing the following:

* Checking any two adjacent actions for conflicts (not just
  odd-even pairs).
* Marking 'set' and 'add' action fields as destination, rather
  than source, for the purposes of checking for conflicts.
* Checking all types of actions ('add', 'set', 'copy') for
  dependencies.
* Managing offsets of the args in the buffer - copy the action
  args to the right place in the buffer.
* Checking that after inserting nops we're still within the number
  of supported actions - return an error otherwise.

Signed-off-by: default avatarVlad Dogaru <vdogaru@nvidia.com>
Signed-off-by: default avatarYevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: default avatarMark Bloch <mbloch@nvidia.com>
Signed-off-by: default avatarTariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/1747766802-958178-5-git-send-email-tariqt@nvidia.com


Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
parent 0b6e452c
Loading
Loading
Loading
Loading
+14 −7
Original line number Diff line number Diff line
@@ -1190,14 +1190,15 @@ hws_action_create_modify_header_hws(struct mlx5hws_action *action,
				    struct mlx5hws_action_mh_pattern *pattern,
				    u32 log_bulk_size)
{
	u16 num_actions, max_mh_actions = 0, hw_max_actions;
	struct mlx5hws_context *ctx = action->ctx;
	u16 num_actions, max_mh_actions = 0;
	int i, ret, size_in_bytes;
	u32 pat_id, arg_id = 0;
	__be64 *new_pattern;
	size_t pat_max_sz;

	pat_max_sz = MLX5HWS_ARG_CHUNK_SIZE_MAX * MLX5HWS_ARG_DATA_SIZE;
	hw_max_actions = pat_max_sz / MLX5HWS_MODIFY_ACTION_SIZE;
	size_in_bytes = pat_max_sz * sizeof(__be64);
	new_pattern = kcalloc(num_of_patterns, size_in_bytes, GFP_KERNEL);
	if (!new_pattern)
@@ -1211,10 +1212,14 @@ hws_action_create_modify_header_hws(struct mlx5hws_action *action,

		cur_num_actions = pattern[i].sz / MLX5HWS_MODIFY_ACTION_SIZE;

		mlx5hws_pat_calc_nop(pattern[i].data, cur_num_actions,
				     pat_max_sz / MLX5HWS_MODIFY_ACTION_SIZE,
				     &new_num_actions, &nop_locations,
		ret = mlx5hws_pat_calc_nop(pattern[i].data, cur_num_actions,
					   hw_max_actions, &new_num_actions,
					   &nop_locations,
					   &new_pattern[i * pat_max_sz]);
		if (ret) {
			mlx5hws_err(ctx, "Too many actions after nop insertion\n");
			goto free_new_pat;
		}

		action[i].modify_header.nop_locations = nop_locations;
		action[i].modify_header.num_of_actions = new_num_actions;
@@ -2116,10 +2121,12 @@ static void hws_action_modify_write(struct mlx5hws_send_engine *queue,
		if (unlikely(!new_arg_data))
			return;

		for (i = 0, j = 0; i < num_of_actions; i++, j++) {
			memcpy(&new_arg_data[j], arg_data, MLX5HWS_MODIFY_ACTION_SIZE);
		for (i = 0, j = 0; j < num_of_actions; i++, j++) {
			if (BIT(i) & nop_locations)
				j++;
			memcpy(&new_arg_data[j * MLX5HWS_MODIFY_ACTION_SIZE],
			       &arg_data[i * MLX5HWS_MODIFY_ACTION_SIZE],
			       MLX5HWS_MODIFY_ACTION_SIZE);
		}
	}

+38 −36
Original line number Diff line number Diff line
@@ -490,8 +490,8 @@ hws_action_modify_get_target_fields(u8 action_type, __be64 *pattern,
	switch (action_type) {
	case MLX5_ACTION_TYPE_SET:
	case MLX5_ACTION_TYPE_ADD:
		*src_field = MLX5_GET(set_action_in, pattern, field);
		*dst_field = INVALID_FIELD;
		*src_field = INVALID_FIELD;
		*dst_field = MLX5_GET(set_action_in, pattern, field);
		break;
	case MLX5_ACTION_TYPE_COPY:
		*src_field = MLX5_GET(copy_action_in, pattern, src_field);
@@ -522,57 +522,59 @@ bool mlx5hws_pat_verify_actions(struct mlx5hws_context *ctx, __be64 pattern[], s
	return true;
}

void mlx5hws_pat_calc_nop(__be64 *pattern, size_t num_actions,
int mlx5hws_pat_calc_nop(__be64 *pattern, size_t num_actions,
			 size_t max_actions, size_t *new_size,
			 u32 *nop_locations, __be64 *new_pat)
{
	u16 prev_src_field = 0, prev_dst_field = 0;
	u16 prev_src_field = INVALID_FIELD, prev_dst_field = INVALID_FIELD;
	u16 src_field, dst_field;
	u8 action_type;
	bool dependent;
	size_t i, j;

	*new_size = num_actions;
	*nop_locations = 0;

	if (num_actions == 1)
		return;
		return 0;

	for (i = 0, j = 0; i < num_actions; i++, j++) {
		action_type = MLX5_GET(set_action_in, &pattern[i], action_type);
		if (j >= max_actions)
			return -EINVAL;

		action_type = MLX5_GET(set_action_in, &pattern[i], action_type);
		hws_action_modify_get_target_fields(action_type, &pattern[i],
						    &src_field, &dst_field);
		if (i % 2) {
			if (action_type == MLX5_ACTION_TYPE_COPY &&
			    (prev_src_field == src_field ||
			     prev_dst_field == dst_field)) {
				/* need Nop */

		/* For every action, look at it and the previous one. The two
		 * actions are dependent if:
		 */
		dependent =
			(i > 0) &&
			/* At least one of the actions is a write and */
			(dst_field != INVALID_FIELD ||
			 prev_dst_field != INVALID_FIELD) &&
			/* One reads from the other's source */
			(dst_field == prev_src_field ||
			 src_field == prev_dst_field ||
			 /* Or both write to the same destination */
			 dst_field == prev_dst_field);

		if (dependent) {
			*new_size += 1;
			*nop_locations |= BIT(i);
			memset(&new_pat[j], 0, MLX5HWS_MODIFY_ACTION_SIZE);
				MLX5_SET(set_action_in, &new_pat[j],
					 action_type,
					 MLX5_MODIFICATION_TYPE_NOP);
				j++;
			} else if (prev_src_field == src_field) {
				/* need Nop */
				*new_size += 1;
				*nop_locations |= BIT(i);
				MLX5_SET(set_action_in, &new_pat[j],
					 action_type,
			MLX5_SET(set_action_in, &new_pat[j], action_type,
				 MLX5_MODIFICATION_TYPE_NOP);
			j++;
			}
		}
		memcpy(&new_pat[j], &pattern[i], MLX5HWS_MODIFY_ACTION_SIZE);
		/* check if no more space */
		if (j > max_actions) {
			*new_size = num_actions;
			*nop_locations = 0;
			return;
			if (j >= max_actions)
				return -EINVAL;
		}

		memcpy(&new_pat[j], &pattern[i], MLX5HWS_MODIFY_ACTION_SIZE);
		prev_src_field = src_field;
		prev_dst_field = dst_field;
	}

	return 0;
}
+3 −3
Original line number Diff line number Diff line
@@ -96,7 +96,7 @@ int mlx5hws_arg_write_inline_arg_data(struct mlx5hws_context *ctx,
				      u8 *arg_data,
				      size_t data_size);

void mlx5hws_pat_calc_nop(__be64 *pattern, size_t num_actions,
int mlx5hws_pat_calc_nop(__be64 *pattern, size_t num_actions,
			 size_t max_actions, size_t *new_size,
			 u32 *nop_locations, __be64 *new_pat);
#endif /* MLX5HWS_PAT_ARG_H_ */