From 73dd3a4839c1d27c36d4dcc92e1ff44225ecbeb7 Mon Sep 17 00:00:00 2001 From: Mohamad Haj Yahia Date: Thu, 23 Feb 2017 11:19:36 +0200 Subject: net/mlx5: Avoid using pending command interface slots Currently when firmware command gets stuck or it takes long time to complete, the driver command will get timeout and the command slot is freed and can be used for new commands, and if the firmware receive new command on the old busy slot its behavior is unexpected and this could be harmful. To fix this when the driver command gets timeout we return failure, but we don't free the command slot and we wait for the firmware to explicitly respond to that command. Once all the entries are busy we will stop processing new firmware commands. Fixes: 9cba4ebcf374 ('net/mlx5: Fix potential deadlock in command mode change') Signed-off-by: Mohamad Haj Yahia Cc: kernel-team@fb.com Signed-off-by: Saeed Mahameed --- drivers/net/ethernet/mellanox/mlx5/core/cmd.c | 41 +++++++++++++++++++++--- drivers/net/ethernet/mellanox/mlx5/core/eq.c | 2 +- drivers/net/ethernet/mellanox/mlx5/core/health.c | 2 +- 3 files changed, 38 insertions(+), 7 deletions(-) (limited to 'drivers') diff --git a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c index 5bdaf3d545b2..10d282841f5b 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c @@ -774,7 +774,7 @@ static void cb_timeout_handler(struct work_struct *work) mlx5_core_warn(dev, "%s(0x%x) timeout. Will cause a leak of a command resource\n", mlx5_command_str(msg_to_opcode(ent->in)), msg_to_opcode(ent->in)); - mlx5_cmd_comp_handler(dev, 1UL << ent->idx); + mlx5_cmd_comp_handler(dev, 1UL << ent->idx, true); } static void cmd_work_handler(struct work_struct *work) @@ -804,6 +804,7 @@ static void cmd_work_handler(struct work_struct *work) } cmd->ent_arr[ent->idx] = ent; + set_bit(MLX5_CMD_ENT_STATE_PENDING_COMP, &ent->state); lay = get_inst(cmd, ent->idx); ent->lay = lay; memset(lay, 0, sizeof(*lay)); @@ -825,6 +826,20 @@ static void cmd_work_handler(struct work_struct *work) if (ent->callback) schedule_delayed_work(&ent->cb_timeout_work, cb_timeout); + /* Skip sending command to fw if internal error */ + if (pci_channel_offline(dev->pdev) || + dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) { + u8 status = 0; + u32 drv_synd; + + ent->ret = mlx5_internal_err_ret_value(dev, msg_to_opcode(ent->in), &drv_synd, &status); + MLX5_SET(mbox_out, ent->out, status, status); + MLX5_SET(mbox_out, ent->out, syndrome, drv_synd); + + mlx5_cmd_comp_handler(dev, 1UL << ent->idx, true); + return; + } + /* ring doorbell after the descriptor is valid */ mlx5_core_dbg(dev, "writing 0x%x to command doorbell\n", 1 << ent->idx); wmb(); @@ -835,7 +850,7 @@ static void cmd_work_handler(struct work_struct *work) poll_timeout(ent); /* make sure we read the descriptor after ownership is SW */ rmb(); - mlx5_cmd_comp_handler(dev, 1UL << ent->idx); + mlx5_cmd_comp_handler(dev, 1UL << ent->idx, (ent->ret == -ETIMEDOUT)); } } @@ -879,7 +894,7 @@ static int wait_func(struct mlx5_core_dev *dev, struct mlx5_cmd_work_ent *ent) wait_for_completion(&ent->done); } else if (!wait_for_completion_timeout(&ent->done, timeout)) { ent->ret = -ETIMEDOUT; - mlx5_cmd_comp_handler(dev, 1UL << ent->idx); + mlx5_cmd_comp_handler(dev, 1UL << ent->idx, true); } err = ent->ret; @@ -1375,7 +1390,7 @@ static void free_msg(struct mlx5_core_dev *dev, struct mlx5_cmd_msg *msg) } } -void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, u64 vec) +void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, u64 vec, bool forced) { struct mlx5_cmd *cmd = &dev->cmd; struct mlx5_cmd_work_ent *ent; @@ -1395,6 +1410,19 @@ void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, u64 vec) struct semaphore *sem; ent = cmd->ent_arr[i]; + + /* if we already completed the command, ignore it */ + if (!test_and_clear_bit(MLX5_CMD_ENT_STATE_PENDING_COMP, + &ent->state)) { + /* only real completion can free the cmd slot */ + if (!forced) { + mlx5_core_err(dev, "Command completion arrived after timeout (entry idx = %d).\n", + ent->idx); + free_ent(cmd, ent->idx); + } + continue; + } + if (ent->callback) cancel_delayed_work(&ent->cb_timeout_work); if (ent->page_queue) @@ -1417,7 +1445,10 @@ void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, u64 vec) mlx5_core_dbg(dev, "command completed. ret 0x%x, delivery status %s(0x%x)\n", ent->ret, deliv_status_to_str(ent->status), ent->status); } - free_ent(cmd, ent->idx); + + /* only real completion will free the entry slot */ + if (!forced) + free_ent(cmd, ent->idx); if (ent->callback) { ds = ent->ts2 - ent->ts1; diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eq.c b/drivers/net/ethernet/mellanox/mlx5/core/eq.c index ea5d8d37a75c..33eae5ad2fb0 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/eq.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/eq.c @@ -422,7 +422,7 @@ static irqreturn_t mlx5_eq_int(int irq, void *eq_ptr) break; case MLX5_EVENT_TYPE_CMD: - mlx5_cmd_comp_handler(dev, be32_to_cpu(eqe->data.cmd.vector)); + mlx5_cmd_comp_handler(dev, be32_to_cpu(eqe->data.cmd.vector), false); break; case MLX5_EVENT_TYPE_PORT_CHANGE: diff --git a/drivers/net/ethernet/mellanox/mlx5/core/health.c b/drivers/net/ethernet/mellanox/mlx5/core/health.c index d0515391d33b..44f59b1d6f0f 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/health.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/health.c @@ -90,7 +90,7 @@ static void trigger_cmd_completions(struct mlx5_core_dev *dev) spin_unlock_irqrestore(&dev->cmd.alloc_lock, flags); mlx5_core_dbg(dev, "vector 0x%llx\n", vector); - mlx5_cmd_comp_handler(dev, vector); + mlx5_cmd_comp_handler(dev, vector, true); return; no_trig: -- cgit v1.2.3