1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* Copyright (c) 2019 Mellanox Technologies. */ 3 4 #include <net/devlink.h> 5 #include "reporter.h" 6 #include "lib/eq.h" 7 8 #define MLX5E_TX_REPORTER_PER_SQ_MAX_LEN 256 9 10 struct mlx5e_tx_err_ctx { 11 int (*recover)(struct mlx5e_txqsq *sq); 12 struct mlx5e_txqsq *sq; 13 }; 14 15 static int mlx5e_wait_for_sq_flush(struct mlx5e_txqsq *sq) 16 { 17 unsigned long exp_time = jiffies + msecs_to_jiffies(2000); 18 19 while (time_before(jiffies, exp_time)) { 20 if (sq->cc == sq->pc) 21 return 0; 22 23 msleep(20); 24 } 25 26 netdev_err(sq->channel->netdev, 27 "Wait for SQ 0x%x flush timeout (sq cc = 0x%x, sq pc = 0x%x)\n", 28 sq->sqn, sq->cc, sq->pc); 29 30 return -ETIMEDOUT; 31 } 32 33 static void mlx5e_reset_txqsq_cc_pc(struct mlx5e_txqsq *sq) 34 { 35 WARN_ONCE(sq->cc != sq->pc, 36 "SQ 0x%x: cc (0x%x) != pc (0x%x)\n", 37 sq->sqn, sq->cc, sq->pc); 38 sq->cc = 0; 39 sq->dma_fifo_cc = 0; 40 sq->pc = 0; 41 } 42 43 static int mlx5e_sq_to_ready(struct mlx5e_txqsq *sq, int curr_state) 44 { 45 struct mlx5_core_dev *mdev = sq->channel->mdev; 46 struct net_device *dev = sq->channel->netdev; 47 struct mlx5e_modify_sq_param msp = {0}; 48 int err; 49 50 msp.curr_state = curr_state; 51 msp.next_state = MLX5_SQC_STATE_RST; 52 53 err = mlx5e_modify_sq(mdev, sq->sqn, &msp); 54 if (err) { 55 netdev_err(dev, "Failed to move sq 0x%x to reset\n", sq->sqn); 56 return err; 57 } 58 59 memset(&msp, 0, sizeof(msp)); 60 msp.curr_state = MLX5_SQC_STATE_RST; 61 msp.next_state = MLX5_SQC_STATE_RDY; 62 63 err = mlx5e_modify_sq(mdev, sq->sqn, &msp); 64 if (err) { 65 netdev_err(dev, "Failed to move sq 0x%x to ready\n", sq->sqn); 66 return err; 67 } 68 69 return 0; 70 } 71 72 static int mlx5e_tx_reporter_err_cqe_recover(struct mlx5e_txqsq *sq) 73 { 74 struct mlx5_core_dev *mdev = sq->channel->mdev; 75 struct net_device *dev = sq->channel->netdev; 76 u8 state; 77 int err; 78 79 if (!test_bit(MLX5E_SQ_STATE_RECOVERING, &sq->state)) 80 return 0; 81 82 err = mlx5_core_query_sq_state(mdev, sq->sqn, &state); 83 if (err) { 84 netdev_err(dev, "Failed to query SQ 0x%x state. err = %d\n", 85 sq->sqn, err); 86 return err; 87 } 88 89 if (state != MLX5_SQC_STATE_ERR) { 90 netdev_err(dev, "SQ 0x%x not in ERROR state\n", sq->sqn); 91 return -EINVAL; 92 } 93 94 mlx5e_tx_disable_queue(sq->txq); 95 96 err = mlx5e_wait_for_sq_flush(sq); 97 if (err) 98 return err; 99 100 /* At this point, no new packets will arrive from the stack as TXQ is 101 * marked with QUEUE_STATE_DRV_XOFF. In addition, NAPI cleared all 102 * pending WQEs. SQ can safely reset the SQ. 103 */ 104 105 err = mlx5e_sq_to_ready(sq, state); 106 if (err) 107 return err; 108 109 mlx5e_reset_txqsq_cc_pc(sq); 110 sq->stats->recover++; 111 mlx5e_activate_txqsq(sq); 112 113 return 0; 114 } 115 116 static int mlx5_tx_health_report(struct devlink_health_reporter *tx_reporter, 117 char *err_str, 118 struct mlx5e_tx_err_ctx *err_ctx) 119 { 120 if (IS_ERR_OR_NULL(tx_reporter)) { 121 netdev_err(err_ctx->sq->channel->netdev, err_str); 122 return err_ctx->recover(err_ctx->sq); 123 } 124 125 return devlink_health_report(tx_reporter, err_str, err_ctx); 126 } 127 128 void mlx5e_tx_reporter_err_cqe(struct mlx5e_txqsq *sq) 129 { 130 char err_str[MLX5E_TX_REPORTER_PER_SQ_MAX_LEN]; 131 struct mlx5e_tx_err_ctx err_ctx = {0}; 132 133 err_ctx.sq = sq; 134 err_ctx.recover = mlx5e_tx_reporter_err_cqe_recover; 135 sprintf(err_str, "ERR CQE on SQ: 0x%x", sq->sqn); 136 137 mlx5_tx_health_report(sq->channel->priv->tx_reporter, err_str, 138 &err_ctx); 139 } 140 141 static int mlx5e_tx_reporter_timeout_recover(struct mlx5e_txqsq *sq) 142 { 143 struct mlx5_eq_comp *eq = sq->cq.mcq.eq; 144 u32 eqe_count; 145 146 netdev_err(sq->channel->netdev, "EQ 0x%x: Cons = 0x%x, irqn = 0x%x\n", 147 eq->core.eqn, eq->core.cons_index, eq->core.irqn); 148 149 eqe_count = mlx5_eq_poll_irq_disabled(eq); 150 if (!eqe_count) { 151 clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state); 152 return -EIO; 153 } 154 155 netdev_err(sq->channel->netdev, "Recover %d eqes on EQ 0x%x\n", 156 eqe_count, eq->core.eqn); 157 sq->channel->stats->eq_rearm++; 158 return 0; 159 } 160 161 int mlx5e_tx_reporter_timeout(struct mlx5e_txqsq *sq) 162 { 163 char err_str[MLX5E_TX_REPORTER_PER_SQ_MAX_LEN]; 164 struct mlx5e_tx_err_ctx err_ctx; 165 166 err_ctx.sq = sq; 167 err_ctx.recover = mlx5e_tx_reporter_timeout_recover; 168 sprintf(err_str, 169 "TX timeout on queue: %d, SQ: 0x%x, CQ: 0x%x, SQ Cons: 0x%x SQ Prod: 0x%x, usecs since last trans: %u\n", 170 sq->channel->ix, sq->sqn, sq->cq.mcq.cqn, sq->cc, sq->pc, 171 jiffies_to_usecs(jiffies - sq->txq->trans_start)); 172 173 return mlx5_tx_health_report(sq->channel->priv->tx_reporter, err_str, 174 &err_ctx); 175 } 176 177 /* state lock cannot be grabbed within this function. 178 * It can cause a dead lock or a read-after-free. 179 */ 180 static int mlx5e_tx_reporter_recover_from_ctx(struct mlx5e_tx_err_ctx *err_ctx) 181 { 182 return err_ctx->recover(err_ctx->sq); 183 } 184 185 static int mlx5e_tx_reporter_recover_all(struct mlx5e_priv *priv) 186 { 187 int err = 0; 188 189 rtnl_lock(); 190 mutex_lock(&priv->state_lock); 191 192 if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) 193 goto out; 194 195 err = mlx5e_safe_reopen_channels(priv); 196 197 out: 198 mutex_unlock(&priv->state_lock); 199 rtnl_unlock(); 200 201 return err; 202 } 203 204 static int mlx5e_tx_reporter_recover(struct devlink_health_reporter *reporter, 205 void *context) 206 { 207 struct mlx5e_priv *priv = devlink_health_reporter_priv(reporter); 208 struct mlx5e_tx_err_ctx *err_ctx = context; 209 210 return err_ctx ? mlx5e_tx_reporter_recover_from_ctx(err_ctx) : 211 mlx5e_tx_reporter_recover_all(priv); 212 } 213 214 static int 215 mlx5e_tx_reporter_build_diagnose_output(struct devlink_fmsg *fmsg, 216 u32 sqn, u8 state, bool stopped) 217 { 218 int err; 219 220 err = devlink_fmsg_obj_nest_start(fmsg); 221 if (err) 222 return err; 223 224 err = devlink_fmsg_u32_pair_put(fmsg, "sqn", sqn); 225 if (err) 226 return err; 227 228 err = devlink_fmsg_u8_pair_put(fmsg, "HW state", state); 229 if (err) 230 return err; 231 232 err = devlink_fmsg_bool_pair_put(fmsg, "stopped", stopped); 233 if (err) 234 return err; 235 236 err = devlink_fmsg_obj_nest_end(fmsg); 237 if (err) 238 return err; 239 240 return 0; 241 } 242 243 static int mlx5e_tx_reporter_diagnose(struct devlink_health_reporter *reporter, 244 struct devlink_fmsg *fmsg) 245 { 246 struct mlx5e_priv *priv = devlink_health_reporter_priv(reporter); 247 int i, err = 0; 248 249 mutex_lock(&priv->state_lock); 250 251 if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) 252 goto unlock; 253 254 err = devlink_fmsg_arr_pair_nest_start(fmsg, "SQs"); 255 if (err) 256 goto unlock; 257 258 for (i = 0; i < priv->channels.num * priv->channels.params.num_tc; 259 i++) { 260 struct mlx5e_txqsq *sq = priv->txq2sq[i]; 261 u8 state; 262 263 err = mlx5_core_query_sq_state(priv->mdev, sq->sqn, &state); 264 if (err) 265 goto unlock; 266 267 err = mlx5e_tx_reporter_build_diagnose_output(fmsg, sq->sqn, 268 state, 269 netif_xmit_stopped(sq->txq)); 270 if (err) 271 goto unlock; 272 } 273 err = devlink_fmsg_arr_pair_nest_end(fmsg); 274 if (err) 275 goto unlock; 276 277 unlock: 278 mutex_unlock(&priv->state_lock); 279 return err; 280 } 281 282 static const struct devlink_health_reporter_ops mlx5_tx_reporter_ops = { 283 .name = "tx", 284 .recover = mlx5e_tx_reporter_recover, 285 .diagnose = mlx5e_tx_reporter_diagnose, 286 }; 287 288 #define MLX5_REPORTER_TX_GRACEFUL_PERIOD 500 289 290 int mlx5e_tx_reporter_create(struct mlx5e_priv *priv) 291 { 292 struct mlx5_core_dev *mdev = priv->mdev; 293 struct devlink *devlink = priv_to_devlink(mdev); 294 295 priv->tx_reporter = 296 devlink_health_reporter_create(devlink, &mlx5_tx_reporter_ops, 297 MLX5_REPORTER_TX_GRACEFUL_PERIOD, 298 true, priv); 299 if (IS_ERR(priv->tx_reporter)) 300 netdev_warn(priv->netdev, 301 "Failed to create tx reporter, err = %ld\n", 302 PTR_ERR(priv->tx_reporter)); 303 return IS_ERR_OR_NULL(priv->tx_reporter); 304 } 305 306 void mlx5e_tx_reporter_destroy(struct mlx5e_priv *priv) 307 { 308 if (IS_ERR_OR_NULL(priv->tx_reporter)) 309 return; 310 311 devlink_health_reporter_destroy(priv->tx_reporter); 312 } 313