net/mlx5: HWS, added send engine and context handling

Added implementation of send engine and handling of HWS context.

Reviewed-by: Itamar Gozlan <igozlan@nvidia.com>
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
This commit is contained in:
Yevgeny Kliteynik 2024-06-20 02:35:50 +03:00 committed by Saeed Mahameed
parent d4a605e968
commit 2ca62599aa
4 changed files with 1803 additions and 0 deletions

View File

@ -0,0 +1,260 @@
// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
/* Copyright (c) 2024 NVIDIA CORPORATION. All rights reserved. */
#include "mlx5hws_internal.h"
bool mlx5hws_context_cap_dynamic_reparse(struct mlx5hws_context *ctx)
{
return IS_BIT_SET(ctx->caps->rtc_reparse_mode, MLX5_IFC_RTC_REPARSE_BY_STC);
}
u8 mlx5hws_context_get_reparse_mode(struct mlx5hws_context *ctx)
{
/* Prefer to use dynamic reparse, reparse only specific actions */
if (mlx5hws_context_cap_dynamic_reparse(ctx))
return MLX5_IFC_RTC_REPARSE_NEVER;
/* Otherwise use less efficient static */
return MLX5_IFC_RTC_REPARSE_ALWAYS;
}
static int hws_context_pools_init(struct mlx5hws_context *ctx)
{
struct mlx5hws_pool_attr pool_attr = {0};
u8 max_log_sz;
int ret;
int i;
ret = mlx5hws_pat_init_pattern_cache(&ctx->pattern_cache);
if (ret)
return ret;
ret = mlx5hws_definer_init_cache(&ctx->definer_cache);
if (ret)
goto uninit_pat_cache;
/* Create an STC pool per FT type */
pool_attr.pool_type = MLX5HWS_POOL_TYPE_STC;
pool_attr.flags = MLX5HWS_POOL_FLAGS_FOR_STC_POOL;
max_log_sz = min(MLX5HWS_POOL_STC_LOG_SZ, ctx->caps->stc_alloc_log_max);
pool_attr.alloc_log_sz = max(max_log_sz, ctx->caps->stc_alloc_log_gran);
for (i = 0; i < MLX5HWS_TABLE_TYPE_MAX; i++) {
pool_attr.table_type = i;
ctx->stc_pool[i] = mlx5hws_pool_create(ctx, &pool_attr);
if (!ctx->stc_pool[i]) {
mlx5hws_err(ctx, "Failed to allocate STC pool [%d]", i);
ret = -ENOMEM;
goto free_stc_pools;
}
}
return 0;
free_stc_pools:
for (i = 0; i < MLX5HWS_TABLE_TYPE_MAX; i++)
if (ctx->stc_pool[i])
mlx5hws_pool_destroy(ctx->stc_pool[i]);
mlx5hws_definer_uninit_cache(ctx->definer_cache);
uninit_pat_cache:
mlx5hws_pat_uninit_pattern_cache(ctx->pattern_cache);
return ret;
}
static void hws_context_pools_uninit(struct mlx5hws_context *ctx)
{
int i;
for (i = 0; i < MLX5HWS_TABLE_TYPE_MAX; i++) {
if (ctx->stc_pool[i])
mlx5hws_pool_destroy(ctx->stc_pool[i]);
}
mlx5hws_definer_uninit_cache(ctx->definer_cache);
mlx5hws_pat_uninit_pattern_cache(ctx->pattern_cache);
}
static int hws_context_init_pd(struct mlx5hws_context *ctx)
{
int ret = 0;
ret = mlx5_core_alloc_pd(ctx->mdev, &ctx->pd_num);
if (ret) {
mlx5hws_err(ctx, "Failed to allocate PD\n");
return ret;
}
ctx->flags |= MLX5HWS_CONTEXT_FLAG_PRIVATE_PD;
return 0;
}
static int hws_context_uninit_pd(struct mlx5hws_context *ctx)
{
if (ctx->flags & MLX5HWS_CONTEXT_FLAG_PRIVATE_PD)
mlx5_core_dealloc_pd(ctx->mdev, ctx->pd_num);
return 0;
}
static void hws_context_check_hws_supp(struct mlx5hws_context *ctx)
{
struct mlx5hws_cmd_query_caps *caps = ctx->caps;
/* HWS not supported on device / FW */
if (!caps->wqe_based_update) {
mlx5hws_err(ctx, "Required HWS WQE based insertion cap not supported\n");
return;
}
if (!caps->eswitch_manager) {
mlx5hws_err(ctx, "HWS is not supported for non eswitch manager port\n");
return;
}
/* Current solution requires all rules to set reparse bit */
if ((!caps->nic_ft.reparse ||
(!caps->fdb_ft.reparse && caps->eswitch_manager)) ||
!IS_BIT_SET(caps->rtc_reparse_mode, MLX5_IFC_RTC_REPARSE_ALWAYS)) {
mlx5hws_err(ctx, "Required HWS reparse cap not supported\n");
return;
}
/* FW/HW must support 8DW STE */
if (!IS_BIT_SET(caps->ste_format, MLX5_IFC_RTC_STE_FORMAT_8DW)) {
mlx5hws_err(ctx, "Required HWS STE format not supported\n");
return;
}
/* Adding rules by hash and by offset are requirements */
if (!IS_BIT_SET(caps->rtc_index_mode, MLX5_IFC_RTC_STE_UPDATE_MODE_BY_HASH) ||
!IS_BIT_SET(caps->rtc_index_mode, MLX5_IFC_RTC_STE_UPDATE_MODE_BY_OFFSET)) {
mlx5hws_err(ctx, "Required HWS RTC update mode not supported\n");
return;
}
/* Support for SELECT definer ID is required */
if (!IS_BIT_SET(caps->definer_format_sup, MLX5_IFC_DEFINER_FORMAT_ID_SELECT)) {
mlx5hws_err(ctx, "Required HWS Dynamic definer not supported\n");
return;
}
ctx->flags |= MLX5HWS_CONTEXT_FLAG_HWS_SUPPORT;
}
static int hws_context_init_hws(struct mlx5hws_context *ctx,
struct mlx5hws_context_attr *attr)
{
int ret;
hws_context_check_hws_supp(ctx);
if (!(ctx->flags & MLX5HWS_CONTEXT_FLAG_HWS_SUPPORT))
return 0;
ret = hws_context_init_pd(ctx);
if (ret)
return ret;
ret = hws_context_pools_init(ctx);
if (ret)
goto uninit_pd;
if (attr->bwc)
ctx->flags |= MLX5HWS_CONTEXT_FLAG_BWC_SUPPORT;
ret = mlx5hws_send_queues_open(ctx, attr->queues, attr->queue_size);
if (ret)
goto pools_uninit;
INIT_LIST_HEAD(&ctx->tbl_list);
return 0;
pools_uninit:
hws_context_pools_uninit(ctx);
uninit_pd:
hws_context_uninit_pd(ctx);
return ret;
}
static void hws_context_uninit_hws(struct mlx5hws_context *ctx)
{
if (!(ctx->flags & MLX5HWS_CONTEXT_FLAG_HWS_SUPPORT))
return;
mlx5hws_send_queues_close(ctx);
hws_context_pools_uninit(ctx);
hws_context_uninit_pd(ctx);
}
struct mlx5hws_context *mlx5hws_context_open(struct mlx5_core_dev *mdev,
struct mlx5hws_context_attr *attr)
{
struct mlx5hws_context *ctx;
int ret;
ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
if (!ctx)
return NULL;
ctx->mdev = mdev;
mutex_init(&ctx->ctrl_lock);
xa_init(&ctx->peer_ctx_xa);
ctx->caps = kzalloc(sizeof(*ctx->caps), GFP_KERNEL);
if (!ctx->caps)
goto free_ctx;
ret = mlx5hws_cmd_query_caps(mdev, ctx->caps);
if (ret)
goto free_caps;
ret = mlx5hws_vport_init_vports(ctx);
if (ret)
goto free_caps;
ret = hws_context_init_hws(ctx, attr);
if (ret)
goto uninit_vports;
mlx5hws_debug_init_dump(ctx);
return ctx;
uninit_vports:
mlx5hws_vport_uninit_vports(ctx);
free_caps:
kfree(ctx->caps);
free_ctx:
xa_destroy(&ctx->peer_ctx_xa);
mutex_destroy(&ctx->ctrl_lock);
kfree(ctx);
return NULL;
}
int mlx5hws_context_close(struct mlx5hws_context *ctx)
{
mlx5hws_debug_uninit_dump(ctx);
hws_context_uninit_hws(ctx);
mlx5hws_vport_uninit_vports(ctx);
kfree(ctx->caps);
xa_destroy(&ctx->peer_ctx_xa);
mutex_destroy(&ctx->ctrl_lock);
kfree(ctx);
return 0;
}
void mlx5hws_context_set_peer(struct mlx5hws_context *ctx,
struct mlx5hws_context *peer_ctx,
u16 peer_vhca_id)
{
mutex_lock(&ctx->ctrl_lock);
if (xa_err(xa_store(&ctx->peer_ctx_xa, peer_vhca_id, peer_ctx, GFP_KERNEL)))
pr_warn("HWS: failed storing peer vhca ID in peer xarray\n");
mutex_unlock(&ctx->ctrl_lock);
}

View File

@ -0,0 +1,64 @@
/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
/* Copyright (c) 2024 NVIDIA Corporation & Affiliates */
#ifndef MLX5HWS_CONTEXT_H_
#define MLX5HWS_CONTEXT_H_
enum mlx5hws_context_flags {
MLX5HWS_CONTEXT_FLAG_HWS_SUPPORT = 1 << 0,
MLX5HWS_CONTEXT_FLAG_PRIVATE_PD = 1 << 1,
MLX5HWS_CONTEXT_FLAG_BWC_SUPPORT = 1 << 2,
};
enum mlx5hws_context_shared_stc_type {
MLX5HWS_CONTEXT_SHARED_STC_DECAP_L3 = 0,
MLX5HWS_CONTEXT_SHARED_STC_DOUBLE_POP = 1,
MLX5HWS_CONTEXT_SHARED_STC_MAX = 2,
};
struct mlx5hws_context_common_res {
struct mlx5hws_action_default_stc *default_stc;
struct mlx5hws_action_shared_stc *shared_stc[MLX5HWS_CONTEXT_SHARED_STC_MAX];
struct mlx5hws_cmd_forward_tbl *default_miss;
};
struct mlx5hws_context_debug_info {
struct dentry *steering_debugfs;
struct dentry *fdb_debugfs;
};
struct mlx5hws_context_vports {
u16 esw_manager_gvmi;
u16 uplink_gvmi;
struct xarray vport_gvmi_xa;
};
struct mlx5hws_context {
struct mlx5_core_dev *mdev;
struct mlx5hws_cmd_query_caps *caps;
u32 pd_num;
struct mlx5hws_pool *stc_pool[MLX5HWS_TABLE_TYPE_MAX];
struct mlx5hws_context_common_res common_res[MLX5HWS_TABLE_TYPE_MAX];
struct mlx5hws_pattern_cache *pattern_cache;
struct mlx5hws_definer_cache *definer_cache;
struct mutex ctrl_lock; /* control lock to protect the whole context */
enum mlx5hws_context_flags flags;
struct mlx5hws_send_engine *send_queue;
size_t queues;
struct mutex *bwc_send_queue_locks; /* protect BWC queues */
struct list_head tbl_list;
struct mlx5hws_context_debug_info debug_info;
struct xarray peer_ctx_xa;
struct mlx5hws_context_vports vports;
};
static inline bool mlx5hws_context_bwc_supported(struct mlx5hws_context *ctx)
{
return ctx->flags & MLX5HWS_CONTEXT_FLAG_BWC_SUPPORT;
}
bool mlx5hws_context_cap_dynamic_reparse(struct mlx5hws_context *ctx);
u8 mlx5hws_context_get_reparse_mode(struct mlx5hws_context *ctx);
#endif /* MLX5HWS_CONTEXT_H_ */

File diff suppressed because it is too large Load Diff

View File

@ -0,0 +1,270 @@
/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
/* Copyright (c) 2024 NVIDIA Corporation & Affiliates */
#ifndef MLX5HWS_SEND_H_
#define MLX5HWS_SEND_H_
/* As a single operation requires at least two WQEBBS.
* This means a maximum of 16 such operations per rule.
*/
#define MAX_WQES_PER_RULE 32
enum mlx5hws_wqe_opcode {
MLX5HWS_WQE_OPCODE_TBL_ACCESS = 0x2c,
};
enum mlx5hws_wqe_opmod {
MLX5HWS_WQE_OPMOD_GTA_STE = 0,
MLX5HWS_WQE_OPMOD_GTA_MOD_ARG = 1,
};
enum mlx5hws_wqe_gta_opcode {
MLX5HWS_WQE_GTA_OP_ACTIVATE = 0,
MLX5HWS_WQE_GTA_OP_DEACTIVATE = 1,
};
enum mlx5hws_wqe_gta_opmod {
MLX5HWS_WQE_GTA_OPMOD_STE = 0,
MLX5HWS_WQE_GTA_OPMOD_MOD_ARG = 1,
};
enum mlx5hws_wqe_gta_sz {
MLX5HWS_WQE_SZ_GTA_CTRL = 48,
MLX5HWS_WQE_SZ_GTA_DATA = 64,
};
/* WQE Control segment. */
struct mlx5hws_wqe_ctrl_seg {
__be32 opmod_idx_opcode;
__be32 qpn_ds;
__be32 flags;
__be32 imm;
};
struct mlx5hws_wqe_gta_ctrl_seg {
__be32 op_dirix;
__be32 stc_ix[5];
__be32 rsvd0[6];
};
struct mlx5hws_wqe_gta_data_seg_ste {
__be32 rsvd0_ctr_id;
__be32 rsvd1_definer;
__be32 rsvd2[3];
union {
struct {
__be32 action[3];
__be32 tag[8];
};
__be32 jumbo[11];
};
};
struct mlx5hws_wqe_gta_data_seg_arg {
__be32 action_args[8];
};
struct mlx5hws_wqe_gta {
struct mlx5hws_wqe_gta_ctrl_seg gta_ctrl;
union {
struct mlx5hws_wqe_gta_data_seg_ste seg_ste;
struct mlx5hws_wqe_gta_data_seg_arg seg_arg;
};
};
struct mlx5hws_send_ring_cq {
struct mlx5_core_dev *mdev;
struct mlx5_cqwq wq;
struct mlx5_wq_ctrl wq_ctrl;
struct mlx5_core_cq mcq;
u16 poll_wqe;
};
struct mlx5hws_send_ring_priv {
struct mlx5hws_rule *rule;
void *user_data;
u32 num_wqebbs;
u32 id;
u32 retry_id;
u32 *used_id;
};
struct mlx5hws_send_ring_dep_wqe {
struct mlx5hws_wqe_gta_ctrl_seg wqe_ctrl;
struct mlx5hws_wqe_gta_data_seg_ste wqe_data;
struct mlx5hws_rule *rule;
u32 rtc_0;
u32 rtc_1;
u32 retry_rtc_0;
u32 retry_rtc_1;
u32 direct_index;
void *user_data;
};
struct mlx5hws_send_ring_sq {
struct mlx5_core_dev *mdev;
u16 cur_post;
u16 buf_mask;
struct mlx5hws_send_ring_priv *wr_priv;
unsigned int last_idx;
struct mlx5hws_send_ring_dep_wqe *dep_wqe;
unsigned int head_dep_idx;
unsigned int tail_dep_idx;
u32 sqn;
struct mlx5_wq_cyc wq;
struct mlx5_wq_ctrl wq_ctrl;
void __iomem *uar_map;
};
struct mlx5hws_send_ring {
struct mlx5hws_send_ring_cq send_cq;
struct mlx5hws_send_ring_sq send_sq;
};
struct mlx5hws_completed_poll_entry {
void *user_data;
enum mlx5hws_flow_op_status status;
};
struct mlx5hws_completed_poll {
struct mlx5hws_completed_poll_entry *entries;
u16 ci;
u16 pi;
u16 mask;
};
struct mlx5hws_send_engine {
struct mlx5hws_send_ring send_ring;
struct mlx5_uars_page *uar; /* Uar is shared between rings of a queue */
struct mlx5hws_completed_poll completed;
u16 used_entries;
u16 num_entries;
bool err;
struct mutex lock; /* Protects the send engine */
};
struct mlx5hws_send_engine_post_ctrl {
struct mlx5hws_send_engine *queue;
struct mlx5hws_send_ring *send_ring;
size_t num_wqebbs;
};
struct mlx5hws_send_engine_post_attr {
u8 opcode;
u8 opmod;
u8 notify_hw;
u8 fence;
u8 match_definer_id;
u8 range_definer_id;
size_t len;
struct mlx5hws_rule *rule;
u32 id;
u32 retry_id;
u32 *used_id;
void *user_data;
};
struct mlx5hws_send_ste_attr {
u32 rtc_0;
u32 rtc_1;
u32 retry_rtc_0;
u32 retry_rtc_1;
u32 *used_id_rtc_0;
u32 *used_id_rtc_1;
bool wqe_tag_is_jumbo;
u8 gta_opcode;
u32 direct_index;
struct mlx5hws_send_engine_post_attr send_attr;
struct mlx5hws_rule_match_tag *wqe_tag;
struct mlx5hws_rule_match_tag *range_wqe_tag;
struct mlx5hws_wqe_gta_ctrl_seg *wqe_ctrl;
struct mlx5hws_wqe_gta_data_seg_ste *wqe_data;
struct mlx5hws_wqe_gta_data_seg_ste *range_wqe_data;
};
struct mlx5hws_send_ring_dep_wqe *
mlx5hws_send_add_new_dep_wqe(struct mlx5hws_send_engine *queue);
void mlx5hws_send_abort_new_dep_wqe(struct mlx5hws_send_engine *queue);
void mlx5hws_send_all_dep_wqe(struct mlx5hws_send_engine *queue);
void mlx5hws_send_queue_close(struct mlx5hws_send_engine *queue);
int mlx5hws_send_queue_open(struct mlx5hws_context *ctx,
struct mlx5hws_send_engine *queue,
u16 queue_size);
void mlx5hws_send_queues_close(struct mlx5hws_context *ctx);
int mlx5hws_send_queues_open(struct mlx5hws_context *ctx,
u16 queues,
u16 queue_size);
int mlx5hws_send_queue_action(struct mlx5hws_context *ctx,
u16 queue_id,
u32 actions);
int mlx5hws_send_test(struct mlx5hws_context *ctx,
u16 queues,
u16 queue_size);
struct mlx5hws_send_engine_post_ctrl
mlx5hws_send_engine_post_start(struct mlx5hws_send_engine *queue);
void mlx5hws_send_engine_post_req_wqe(struct mlx5hws_send_engine_post_ctrl *ctrl,
char **buf, size_t *len);
void mlx5hws_send_engine_post_end(struct mlx5hws_send_engine_post_ctrl *ctrl,
struct mlx5hws_send_engine_post_attr *attr);
void mlx5hws_send_ste(struct mlx5hws_send_engine *queue,
struct mlx5hws_send_ste_attr *ste_attr);
void mlx5hws_send_stes_fw(struct mlx5hws_context *ctx,
struct mlx5hws_send_engine *queue,
struct mlx5hws_send_ste_attr *ste_attr);
void mlx5hws_send_engine_flush_queue(struct mlx5hws_send_engine *queue);
static inline bool mlx5hws_send_engine_empty(struct mlx5hws_send_engine *queue)
{
struct mlx5hws_send_ring_sq *send_sq = &queue->send_ring.send_sq;
struct mlx5hws_send_ring_cq *send_cq = &queue->send_ring.send_cq;
return ((send_sq->cur_post & send_sq->buf_mask) == send_cq->poll_wqe);
}
static inline bool mlx5hws_send_engine_full(struct mlx5hws_send_engine *queue)
{
return queue->used_entries >= queue->num_entries;
}
static inline void mlx5hws_send_engine_inc_rule(struct mlx5hws_send_engine *queue)
{
queue->used_entries++;
}
static inline void mlx5hws_send_engine_dec_rule(struct mlx5hws_send_engine *queue)
{
queue->used_entries--;
}
static inline void mlx5hws_send_engine_gen_comp(struct mlx5hws_send_engine *queue,
void *user_data,
int comp_status)
{
struct mlx5hws_completed_poll *comp = &queue->completed;
comp->entries[comp->pi].status = comp_status;
comp->entries[comp->pi].user_data = user_data;
comp->pi = (comp->pi + 1) & comp->mask;
}
static inline bool mlx5hws_send_engine_err(struct mlx5hws_send_engine *queue)
{
return queue->err;
}
#endif /* MLX5HWS_SEND_H_ */