On 5/14/2019 3:07 PM, Jiri Pirko wrote:
> Sun, May 12, 2019 at 10:37:35AM CEST, a...@mellanox.com wrote:
>>
>>
>> On 5/9/2019 11:23 AM, Jiri Pirko wrote:
>>> Tue, May 07, 2019 at 02:58:32PM CEST, a...@mellanox.com wrote:
>>>>
>>>>
>>>> On 5/7/2019 3:41 PM, Jiri Pirko wrote:
>>>>> Mon, Apr 29, 2019 at 04:17:39PM CEST, a...@mellanox.com wrote:
>>>>>> TX reporter reports an error on two scenarios:
>>>>>> - TX timeout on a specific tx queue
>>>>>> - TX completion error on a specific send queue
>>>>>> Prior to this patch, no dump data was supported by the tx reporter. This
>>>>>> patch adds support for SW data dump of the related SQ context. The dump
>>>>>> is simply the SQ's raw memory snapshot taken right after the error was
>>>>>> reported, before any recovery procedure was launched. With this
>>>>>> approach, no maintenance is needed as the driver fetch the actual data
>>>>>> according to the layout on which the SQ was compiled with. By providing
>>>>>> a SW context, one can easily debug error on a given SQ.
>>>>>>
>>>>>> In order to offline translate the raw memory into a human readable
>>>>>> format, the user can use some out-of-kernel scripts which receives as an
>>>>>> input the following:
>>>>>> - Object raw memory
>>>>>> - Driver object compiled with debug info (can be taken/generated at any
>>>>>> time from the machine)
>>>>>> - Object name
>>>>>>
>>>>>> An example of such script output can be seen below.
>>>>>> Note: the script is not offered as part of this patch as it do not
>>>>>> belong to the kernel, I just described it in order to grasp the general
>>>>>> idea of how/what can be fetched from SW dump via devlink health.
>>>>>>
>>>>>> The output of the SW dump can be extracted by devlink health command:
>>>>>> $ sudo devlink health dump show pci/0000:00:0b.0 reporter tx.
>>>>>> mlx5e_txqsq: sqn: 6336
>>>>>> memory:
>>>>>> 00 00 00 00 00 00 00 00
>>>>>> 01 00 00 00 00 00 00 00
>>>>>> 00 00 00 00 00 00 00 00
>>>>>> 45 f4 88 cb 09 00 00 00
>>>>>> 00 00 00 00 00 00 00 00
>>>>>> 00 00 00 00 00 00 00 00
>>>>>> c0 ff ff ff 1f 00 00 00
>>>>>> f8 18 1e 89 81 88 ff ff
>>>>>> ...
>>>>>>
>>>>>> script output below, with struct members names and actual values:
>>>>>>
>>>>>> struct mlx5e_txqsq {
>>>>>> short unsigned int cc 0x5 ;
>>>>>> unsigned int dma_fifo_cc 0x5 ;
>>>>>> struct net_dim {
>>>>>> unsigned char state 0x1 ;
>>>>>> struct net_dim_stats {
>>>>>> int ppms 0x0 ;
>>>>>> int bpms 0x0 ;
>>>>>> int epms 0x0 ;
>>>>>> } prev_stats;
>>>>>> struct net_dim_sample {
>>>>>> long long int time 0x90766ef9d ;
>>>>>> unsigned int pkt_ctr 0x0 ;
>>>>>> unsigned int byte_ctr 0x0 ;
>>>>>> short unsigned int event_ctr 0x0 ;
>>>>>> } start_sample;
>>>>>> struct work_struct {
>>>>>> struct {
>>>>>> long int counter 0x1fffffffc0 ;
>>>>>> } data;
>>>>>> struct list_head {
>>>>>> struct list_head * next
>>>>>> 0xffff8881b08998f8 ;
>>>>>> struct list_head * prev
>>>>>> 0xffff8881b08998f8 ;
>>>>>> } entry;
>>>>>> void (*func)(struct work_struct *)
>>>>>> 0xffffffffa02d0e30 ;
>>>>>> } work;
>>>>>> unsigned char profile_ix 0x60 ;
>>>>>> unsigned char mode 0x72 ;
>>>>>> unsigned char tune_state 0x35 ;
>>>>>> unsigned char steps_right 0xa0 ;
>>>>>> unsigned char steps_left 0xff ;
>>>>>> unsigned char tired 0xff ;
>>>>>> } dim;
>>>>>> short unsigned int pc 0x0 ;
>>>>>> unsigned int dma_fifo_pc 0x0 ;
>>>>>> struct mlx5e_cq {
>>>>>> struct mlx5_cqwq {
>>>>>> struct mlx5_frag_buf_ctrl {
>>>>>> struct mlx5_buf_list * frags 0x500000005 ;
>>>>>> unsigned int sz_m1 0x0 ;
>>>>>> short unsigned int frag_sz_m1 0x0 ;
>>>>>> short unsigned int strides_offset 0x0 ;
>>>>>> unsigned char log_sz 0x0 ;
>>>>>> unsigned char log_stride 0x0 ;
>>>>>> unsigned char log_frag_strides 0x0 ;
>>>>>> } fbc;
>>>>>> __be32 * db 0x0 ;
>>>>>> unsigned int cc 0x0 ;
>>>>>> } wq;
>>>>>> short unsigned int event_ctr 0x0 ;
>>>>>> struct napi_struct * napi 0x0 ;
>>>>>> struct mlx5_core_cq {
>>>>>> unsigned int cqn 0x0 ;
>>>>>> int cqe_sz 0x0 ;
>>>>>> __be32 * set_ci_db 0xffff8881b1aa4988 ;
>>>>>> __be32 * arm_db 0x3f000003ff ;
>>>>>> struct mlx5_uars_page * uar 0x6060a ;
>>>>>> struct refcount_struct {
>>>>>> struct {
>>>>>> int counter 0xa1814500 ;
>>>>>> } refs;
>>>>>> } refcount;
>>>>>> struct completion {
>>>>>> unsigned int done 0x5 ;
>>>>>> struct wait_queue_head {
>>>>>> struct spinlock {
>>>>>> union {
>>>>>> struct raw_spinlock {
>>>>>> struct
>>>>>> qspinlock {
>>>>>> union
>>>>>> {
>>>>>>
>>>>>> struct {
>>>>>>
>>>>>> int counter
>>>>>> 0x5 ;
>>>>>>
>>>>>> } val;
>>>>>>
>>>>>> struct {
>>>>>>
>>>>>> unsigned char locked
>>>>>> 0x5 ;
>>>>>>
>>>>>> unsigned char pending
>>>>>> 0x0 ;
>>>>>>
>>>>>> } ;
>>>>>>
>>>>>> struct {
>>>>>>
>>>>>> short unsigned int
>>>>>> locked_pending 0x5 ;
>>>>>>
>>>>>> short unsigned int tail
>>>>>> 0x0 ;
>>>>>>
>>>>>> } ;
>>>>>> } ;
>>>>>> } raw_lock;
>>>>>> } rlock;
>>>>>> } ;
>>>>>> } lock;
>>>>>> struct list_head {
>>>>>> struct list_head * next
>>>>>> 0xffff8881b089bb88 ;
>>>>>> struct list_head * prev
>>>>>> 0x4000000c0a ;
>>>>>> } head;
>>>>>> } wait;
>>>>>> } free;
>>>>>> unsigned int vector 0xa1814500 ;
>>>>>> unsigned int irqn 0xffff8881 ;
>>>>>> void (*comp)(struct mlx5_core_cq *)
>>>>>> 0xffff8881a1814504 ;
>>>>>> void (*event)(struct mlx5_core_cq *, enum
>>>>>> mlx5_event) 0xffff8881a2cdea08 ;
>>>>>> unsigned int cons_index 0x1 ;
>>>>>> unsigned int arm_sn 0x0 ;
>>>>>> struct mlx5_rsc_debug * dbg 0x0 ;
>>>>>> int pid 0x0 ;
>>>>>> struct {
>>>>>> struct list_head {
>>>>>> struct list_head * next
>>>>>> 0xffffffff ;
>>>>>> struct list_head * prev
>>>>>> 0xffffffffffffffff ;
>>>>>> } list;
>>>>>> void (*comp)(struct mlx5_core_cq *)
>>>>>> 0xffffffffa0356940 ;
>>>>>> void * priv 0x0 ;
>>>>>> } tasklet_ctx;
>>>>>> int reset_notify_added 0x0 ;
>>>>>> struct list_head {
>>>>>> struct list_head * next
>>>>>> 0xffffffffa0300700 ;
>>>>>> struct list_head * prev 0xd ;
>>>>>> } reset_notify;
>>>>>> struct mlx5_eq_comp * eq 0x0 ;
>>>>>> short unsigned int uid 0x9a70 ;
>>>>>> } mcq;
>>>>>> struct mlx5e_channel * channel 0xffff8881b0899a70 ;
>>>>>> struct mlx5_core_dev * mdev 0x4800000001 ;
>>>>>> struct mlx5_wq_ctrl {
>>>>>> struct mlx5_core_dev * mdev 0xffffffffa02d5350 ;
>>>>>> struct mlx5_frag_buf {
>>>>>> struct mlx5_buf_list * frags
>>>>>> 0xffffffffa02d5460 ;
>>>>>> int npages 0x0 ;
>>>>>> int size 0x5 ;
>>>>>> unsigned char page_shift 0x8 ;
>>>>>> } buf;
>>>>>> struct mlx5_db {
>>>>>> __be32 * db 0x1c6 ;
>>>>>> union {
>>>>>> struct mlx5_db_pgdir * pgdir 0x0 ;
>>>>>> struct mlx5_ib_user_db_page * user_page
>>>>>> 0x0 ;
>>>>>> } u;
>>>>>> long long unsigned int dma
>>>>>> 0xffff8881b0899ab0 ;
>>>>>> int index 0x0 ;
>>>>>> } db;
>>>>>> } wq_ctrl;
>>>>>> } cq;
>>>>>> struct mlx5_wq_cyc {
>>>>>> struct mlx5_frag_buf_ctrl {
>>>>>> struct mlx5_buf_list * frags 0xffff8881a7600160 ;
>>>>>> unsigned int sz_m1 0xa7600160 ;
>>>>>> short unsigned int frag_sz_m1 0x8881 ;
>>>>>> short unsigned int strides_offset 0xffff ;
>>>>>> unsigned char log_sz 0x88 ;
>>>>>> unsigned char log_stride 0x49 ;
>>>>>> unsigned char log_frag_strides 0xaa ;
>>>>>> } fbc;
>>>>>> __be32 * db 0x1000000000010 ;
>>>>>> short unsigned int sz 0xc ;
>>>>>> short unsigned int wqe_ctr 0x0 ;
>>>>>> short unsigned int cur_sz 0x0 ;
>>>>>> } wq;
>>>>>> unsigned int dma_fifo_mask 0xa1814500 ;
>>>>>> struct mlx5e_sq_stats * stats 0xffff8881a33a0348 ;
>>>>>> struct {
>>>>>> struct mlx5e_sq_dma * dma_fifo 0x1a1814500 ;
>>>>>> struct mlx5e_tx_wqe_info * wqe_info 0x14 ;
>>>>>> } db;
>>>>>> void * uar_map 0x0 ;
>>>>>> struct netdev_queue * txq 0x0 ;
>>>>>> unsigned int sqn 0x18c0 ;
>>>>>> unsigned char min_inline_mode 0x0 ;
>>>>>> struct device * pdev 0x0 ;
>>>>>> unsigned int mkey_be 0x0 ;
>>>>>> long unsigned int state 0x0 ;
>>>>>> struct hwtstamp_config * tstamp 0x0 ;
>>>>>> struct mlx5_clock * clock 0xffff8881b1aa6f88 ;
>>>>>> struct mlx5_wq_ctrl {
>>>>>> struct mlx5_core_dev * mdev 0x3f000003ff ;
>>>>>> struct mlx5_frag_buf {
>>>>>> struct mlx5_buf_list * frags 0x6060a ;
>>>>>> int npages 0xa1814604 ;
>>>>>> int size 0xffff8881 ;
>>>>>> unsigned char page_shift 0x0 ;
>>>>>> } buf;
>>>>>> struct mlx5_db {
>>>>>> __be32 * db 0xfff ;
>>>>>> union {
>>>>>> struct mlx5_db_pgdir * pgdir 0x0 ;
>>>>>> struct mlx5_ib_user_db_page * user_page
>>>>>> 0x0 ;
>>>>>> } u;
>>>>>> long long unsigned int dma 0xffff888188440000 ;
>>>>>> int index 0x8b074000 ;
>>>>>> } db;
>>>>>> } wq_ctrl;
>>>>>> struct mlx5e_channel * channel 0xffffc9000010d800 ;
>>>>>> int txq_ix 0xa0020180 ;
>>>>>> unsigned int rate_limit 0xffff8881 ;
>>>>>> struct work_struct {
>>>>>> struct {
>>>>>> long int counter 0x1000018c0 ;
>>>>>> } data;
>>>>>> struct list_head {
>>>>>> struct list_head * next 0xffff8881c32b68e8 ;
>>>>>> struct list_head * prev 0x800 ;
>>>>>> } entry;
>>>>>> void (*func)(struct work_struct *) 0x9 ;
>>>>>> } recover_work;
>>>>>> } ;
>>>>>
>>>>> I don't get it. You are dumping live kernel memory? There are already
>>>>> facilities to do that in place. Why to replicate it?
>>>> I am dumping the driver's memory under a lock so I can ensure it's
>>>> consistency (as appose to /dev/mem)
>>>> vmcore cannot be taken from a live kernel (without crashing).
>>>> I need the memory's snapshot right after the error from the driver's
>>>> context.
>>>
>>> Got it. However, this sounds like a generic problem not specific to
>>> nic drivers. How other subsystems resolve this (if they do at all)?
>>>
>>>
>> Correct, this is a suggested debugging solution for a generic problem:
>> enabling the user of a run time memory snapshot for kernel modules (at a
>> given error event). My research shows that other subsystems deal with
>> errors either by panicking (too much) or by debug/log prints (too little).
>> This solution is (a) low in maintenance (b) consistent in memory (c) has
>> small performance impact (d) use an existing infra-structure between the
>> kernel module and the user space.
>
> I'm still convinced that dumping kernel memory over devlink health dump
> is a good idea :/
>
>
>> It might be ported to other subsystems using their own user-space vs.
>> kernel tools. Regardless of how the memory output was generated to the
>> user, the parsing script can work on it.
>
> Could you share the script? How is it going to be distributed?
I thought that the script should be in a available on Mellanox website.
The script is still pending review but I will be happy to share it when
its ready.
>
>
>>
>>>
>>>> Which other tools do you mean?
>>>>>
>>>>>
>>>>>>
>>>>>> Signed-off-by: Aya Levin <a...@mellanox.com>
>>>>>> ---
>>>>>> .../ethernet/mellanox/mlx5/core/en/reporter_tx.c | 100
>>>>>> +++++++++++++++++++++
>>>>>> 1 file changed, 100 insertions(+)
>>>>>>
>>>>>> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c
>>>>>> b/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c
>>>>>> index 476dd97f7f2f..8a39f5525e57 100644
>>>>>> --- a/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c
>>>>>> +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c
>>>>>> @@ -9,6 +9,7 @@
>>>>>>
>>>>>> struct mlx5e_tx_err_ctx {
>>>>>> int (*recover)(struct mlx5e_txqsq *sq);
>>>>>> + int (*dump)(struct mlx5e_txqsq *sq);
>>>>>> struct mlx5e_txqsq *sq;
>>>>>> };
>>>>>>
>>>>>> @@ -281,10 +282,109 @@ static int mlx5e_tx_reporter_diagnose(struct
>>>>>> devlink_health_reporter *reporter,
>>>>>> return err;
>>>>>> }
>>>>>>
>>>>>> +static int mlx5e_tx_reporter_sw_dump_from_ctx(struct mlx5e_priv *priv,
>>>>>> + struct mlx5e_txqsq *sq,
>>>>>> + struct devlink_fmsg *fmsg)
>>>>>> +{
>>>>>> + u64 *ptr = (u64 *)sq;
>>>>>> + int copy, err;
>>>>>> + int i = 0;
>>>>>> +
>>>>>> + if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
>>>>>> + return 0;
>>>>>> +
>>>>>> + err = devlink_fmsg_pair_nest_start(fmsg, "mlx5e_txqsq");
>>>>>> + if (err)
>>>>>> + return err;
>>>>>> +
>>>>>> + err = devlink_fmsg_obj_nest_start(fmsg);
>>>>>> + if (err)
>>>>>> + return err;
>>>>>> +
>>>>>> + err = devlink_fmsg_arr_pair_nest_start(fmsg, "memory");
>>>>>> + if (err)
>>>>>> + return err;
>>>>>> +
>>>>>> + while (i < sizeof(struct mlx5e_txqsq)) {
>>>>>> + copy = sizeof(u64);
>>>>>> +
>>>>>> + if (i + copy > sizeof(struct mlx5e_txqsq))
>>>>>> + copy = sizeof(struct mlx5e_txqsq) - i;
>>>>>> +
>>>>>> + err = devlink_fmsg_binary_put(fmsg, ptr, copy);
>>>>>> + if (err)
>>>>>> + return err;
>>>>>> + ptr++;
>>>>>> + i += copy;
>>>>>> + }
>>>>>> +
>>>>>> + err = devlink_fmsg_arr_pair_nest_end(fmsg);
>>>>>> + if (err)
>>>>>> + return err;
>>>>>> +
>>>>>> + err = devlink_fmsg_obj_nest_end(fmsg);
>>>>>> + if (err)
>>>>>> + return err;
>>>>>> +
>>>>>> + err = devlink_fmsg_pair_nest_end(fmsg);
>>>>>> +
>>>>>> + return err;
>>>>>> +}
>>>>>> +
>>>>>> +static int mlx5e_tx_reporter_sw_dump_all(struct mlx5e_priv *priv,
>>>>>> + struct devlink_fmsg *fmsg)
>>>>>> +{
>>>>>> + int i, err = 0;
>>>>>> +
>>>>>> + mutex_lock(&priv->state_lock);
>>>>>> +
>>>>>> + if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
>>>>>> + goto unlock;
>>>>>> +
>>>>>> + err = devlink_fmsg_arr_pair_nest_start(fmsg, "SQs");
>>>>>> + if (err)
>>>>>> + goto unlock;
>>>>>> +
>>>>>> + for (i = 0; i < priv->channels.num *
>>>>>> priv->channels.params.num_tc;
>>>>>> + i++) {
>>>>>> + err = devlink_fmsg_obj_nest_start(fmsg);
>>>>>> + if (err)
>>>>>> + goto unlock;
>>>>>> +
>>>>>> + err = mlx5e_tx_reporter_sw_dump_from_ctx(priv,
>>>>>> priv->txq2sq[i],
>>>>>> + fmsg);
>>>>>> + if (err)
>>>>>> + goto unlock;
>>>>>> +
>>>>>> + err = devlink_fmsg_pair_nest_end(fmsg);
>>>>>> + if (err)
>>>>>> + goto unlock;
>>>>>> + }
>>>>>> + err = devlink_fmsg_arr_pair_nest_end(fmsg);
>>>>>> + if (err)
>>>>>> + goto unlock;
>>>>>> +
>>>>>> +unlock:
>>>>>> + mutex_unlock(&priv->state_lock);
>>>>>> + return err;
>>>>>> +}
>>>>>> +
>>>>>> +static int mlx5e_tx_reporter_sw_dump(struct devlink_health_reporter
>>>>>> *reporter,
>>>>>> + struct devlink_fmsg *fmsg, void
>>>>>> *context)
>>>>>> +{
>>>>>> + struct mlx5e_priv *priv =
>>>>>> devlink_health_reporter_priv(reporter);
>>>>>> + struct mlx5e_tx_err_ctx *err_ctx = context;
>>>>>> +
>>>>>> + return err_ctx ? mlx5e_tx_reporter_sw_dump_from_ctx(priv,
>>>>>> err_ctx->sq,
>>>>>> + fmsg) :
>>>>>> + mlx5e_tx_reporter_sw_dump_all(priv, fmsg);
>>>>>> +}
>>>>>> +
>>>>>> static const struct devlink_health_reporter_ops mlx5_tx_reporter_ops = {
>>>>>> .name = "tx",
>>>>>> .recover = mlx5e_tx_reporter_recover,
>>>>>> .diagnose = mlx5e_tx_reporter_diagnose,
>>>>>> + .dump = mlx5e_tx_reporter_sw_dump,
>>>>>> };
>>>>>>
>>>>>> #define MLX5_REPORTER_TX_GRACEFUL_PERIOD 500
>>>>>> --
>>>>>> 2.14.1
>>>>>>
