2
0
mirror of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git synced 2025-09-04 20:19:47 +08:00

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: Vlad Dogaru <vdogaru@nvidia.com>
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/1747766802-958178-5-git-send-email-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This commit is contained in:
Yevgeny Kliteynik 2025-05-20 21:46:42 +03:00 committed by Jakub Kicinski
parent 0b6e452caf
commit 01e035fd03
3 changed files with 55 additions and 46 deletions

View File

@ -1190,14 +1190,15 @@ hws_action_create_modify_header_hws(struct mlx5hws_action *action,
struct mlx5hws_action_mh_pattern *pattern, struct mlx5hws_action_mh_pattern *pattern,
u32 log_bulk_size) u32 log_bulk_size)
{ {
u16 num_actions, max_mh_actions = 0, hw_max_actions;
struct mlx5hws_context *ctx = action->ctx; struct mlx5hws_context *ctx = action->ctx;
u16 num_actions, max_mh_actions = 0;
int i, ret, size_in_bytes; int i, ret, size_in_bytes;
u32 pat_id, arg_id = 0; u32 pat_id, arg_id = 0;
__be64 *new_pattern; __be64 *new_pattern;
size_t pat_max_sz; size_t pat_max_sz;
pat_max_sz = MLX5HWS_ARG_CHUNK_SIZE_MAX * MLX5HWS_ARG_DATA_SIZE; 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); size_in_bytes = pat_max_sz * sizeof(__be64);
new_pattern = kcalloc(num_of_patterns, size_in_bytes, GFP_KERNEL); new_pattern = kcalloc(num_of_patterns, size_in_bytes, GFP_KERNEL);
if (!new_pattern) 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; cur_num_actions = pattern[i].sz / MLX5HWS_MODIFY_ACTION_SIZE;
mlx5hws_pat_calc_nop(pattern[i].data, cur_num_actions, ret = mlx5hws_pat_calc_nop(pattern[i].data, cur_num_actions,
pat_max_sz / MLX5HWS_MODIFY_ACTION_SIZE, hw_max_actions, &new_num_actions,
&new_num_actions, &nop_locations, &nop_locations,
&new_pattern[i * pat_max_sz]); &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.nop_locations = nop_locations;
action[i].modify_header.num_of_actions = new_num_actions; 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)) if (unlikely(!new_arg_data))
return; return;
for (i = 0, j = 0; i < num_of_actions; i++, j++) { for (i = 0, j = 0; j < num_of_actions; i++, j++) {
memcpy(&new_arg_data[j], arg_data, MLX5HWS_MODIFY_ACTION_SIZE);
if (BIT(i) & nop_locations) if (BIT(i) & nop_locations)
j++; j++;
memcpy(&new_arg_data[j * MLX5HWS_MODIFY_ACTION_SIZE],
&arg_data[i * MLX5HWS_MODIFY_ACTION_SIZE],
MLX5HWS_MODIFY_ACTION_SIZE);
} }
} }

View File

@ -490,8 +490,8 @@ hws_action_modify_get_target_fields(u8 action_type, __be64 *pattern,
switch (action_type) { switch (action_type) {
case MLX5_ACTION_TYPE_SET: case MLX5_ACTION_TYPE_SET:
case MLX5_ACTION_TYPE_ADD: case MLX5_ACTION_TYPE_ADD:
*src_field = MLX5_GET(set_action_in, pattern, field); *src_field = INVALID_FIELD;
*dst_field = INVALID_FIELD; *dst_field = MLX5_GET(set_action_in, pattern, field);
break; break;
case MLX5_ACTION_TYPE_COPY: case MLX5_ACTION_TYPE_COPY:
*src_field = MLX5_GET(copy_action_in, pattern, src_field); *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; 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, size_t max_actions, size_t *new_size,
u32 *nop_locations, __be64 *new_pat) 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; u16 src_field, dst_field;
u8 action_type; u8 action_type;
bool dependent;
size_t i, j; size_t i, j;
*new_size = num_actions; *new_size = num_actions;
*nop_locations = 0; *nop_locations = 0;
if (num_actions == 1) if (num_actions == 1)
return; return 0;
for (i = 0, j = 0; i < num_actions; i++, j++) { 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], hws_action_modify_get_target_fields(action_type, &pattern[i],
&src_field, &dst_field); &src_field, &dst_field);
if (i % 2) {
if (action_type == MLX5_ACTION_TYPE_COPY && /* For every action, look at it and the previous one. The two
(prev_src_field == src_field || * actions are dependent if:
prev_dst_field == dst_field)) { */
/* need Nop */ dependent =
*new_size += 1; (i > 0) &&
*nop_locations |= BIT(i); /* At least one of the actions is a write and */
memset(&new_pat[j], 0, MLX5HWS_MODIFY_ACTION_SIZE); (dst_field != INVALID_FIELD ||
MLX5_SET(set_action_in, &new_pat[j], prev_dst_field != INVALID_FIELD) &&
action_type, /* One reads from the other's source */
MLX5_MODIFICATION_TYPE_NOP); (dst_field == prev_src_field ||
j++; src_field == prev_dst_field ||
} else if (prev_src_field == src_field) { /* Or both write to the same destination */
/* need Nop */ dst_field == prev_dst_field);
*new_size += 1;
*nop_locations |= BIT(i); if (dependent) {
MLX5_SET(set_action_in, &new_pat[j], *new_size += 1;
action_type, *nop_locations |= BIT(i);
MLX5_MODIFICATION_TYPE_NOP); memset(&new_pat[j], 0, MLX5HWS_MODIFY_ACTION_SIZE);
j++; 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); if (j >= max_actions)
/* check if no more space */ return -EINVAL;
if (j > max_actions) {
*new_size = num_actions;
*nop_locations = 0;
return;
} }
memcpy(&new_pat[j], &pattern[i], MLX5HWS_MODIFY_ACTION_SIZE);
prev_src_field = src_field; prev_src_field = src_field;
prev_dst_field = dst_field; prev_dst_field = dst_field;
} }
return 0;
} }

View File

@ -96,7 +96,7 @@ int mlx5hws_arg_write_inline_arg_data(struct mlx5hws_context *ctx,
u8 *arg_data, u8 *arg_data,
size_t data_size); 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, size_t max_actions, size_t *new_size,
u32 *nop_locations, __be64 *new_pat); u32 *nop_locations, __be64 *new_pat);
#endif /* MLX5HWS_PAT_ARG_H_ */ #endif /* MLX5HWS_PAT_ARG_H_ */