On 9/18/2020 10:23 AM, Igor Ryzhov wrote:
Hi,

Your code will work only if stats are updated at least once between two overflows.
>

In this case it will have problems in 'i40e_stat_update_48()' too.
It seems there is no way to detect if the increase in stats is N or MAX_48+N by the software. And obviously there is no way to detect if the overflow occurred more than once.

So it's still up to the application to handle this properly. I think it should be mentioned in the docs.


+1 to document.

Igor

On Fri, Sep 18, 2020 at 6:45 AM Jiang, JunyuX <junyux.ji...@intel.com <mailto:junyux.ji...@intel.com>> wrote:

    Hi Ferruh,

     > -----Original Message-----
     > From: Ferruh Yigit <ferruh.yi...@intel.com
    <mailto:ferruh.yi...@intel.com>>
     > Sent: Wednesday, September 16, 2020 8:31 PM
     > To: Jiang, JunyuX <junyux.ji...@intel.com
    <mailto:junyux.ji...@intel.com>>; dev@dpdk.org <mailto:dev@dpdk.org>
     > Cc: Guo, Jia <jia....@intel.com <mailto:jia....@intel.com>>;
    Xing, Beilei <beilei.x...@intel.com <mailto:beilei.x...@intel.com>>;
     > sta...@dpdk.org <mailto:sta...@dpdk.org>
     > Subject: Re: [dpdk-stable] [PATCH v2] net/i40e: fix incorrect
    byte counters
     >
     > On 9/16/2020 2:51 AM, Junyu Jiang wrote:
     > > This patch fixed the issue that rx/tx bytes overflowed
     >
     > "Rx/Tx statistics counters overflowed"?
     >
    Yes, the rx_bytes and tx_bytes counter in X710 cards is 48-bit long,
    if keep sending packets for a log time, the register will overflow.

     > > on 48 bit limitation by enlarging the limitation.
     > >
     > > Fixes: 4861cde46116 ("i40e: new poll mode driver")
     > > Cc: sta...@dpdk.org <mailto:sta...@dpdk.org>
     > >
     > > Signed-off-by: Junyu Jiang <junyux.ji...@intel.com
    <mailto:junyux.ji...@intel.com>>
     > > ---
     > >   drivers/net/i40e/i40e_ethdev.c | 47
     > ++++++++++++++++++++++++++++++++++
     > >   drivers/net/i40e/i40e_ethdev.h |  9 +++++++
     > >   2 files changed, 56 insertions(+)
     > >
     > > diff --git a/drivers/net/i40e/i40e_ethdev.c
     > > b/drivers/net/i40e/i40e_ethdev.c index 563f21d9d..4d4ea9861 100644
     > > --- a/drivers/net/i40e/i40e_ethdev.c
     > > +++ b/drivers/net/i40e/i40e_ethdev.c
     > > @@ -3073,6 +3073,13 @@ i40e_update_vsi_stats(struct i40e_vsi *vsi)
     > >     i40e_stat_update_48(hw, I40E_GLV_BPRCH(idx),
     > I40E_GLV_BPRCL(idx),
     > >                         vsi->offset_loaded, &oes->rx_broadcast,
     > >                         &nes->rx_broadcast);
     > > +   /* enlarge the limitation when rx_bytes overflowed */
     > > +   if (vsi->offset_loaded) {
     > > +           if (I40E_RXTX_BYTES_LOW(vsi->old_rx_bytes) > nes-
     > >rx_bytes)
     > > +                   nes->rx_bytes += (uint64_t)1 <<
    I40E_48_BIT_WIDTH;
     > > +           nes->rx_bytes += I40E_RXTX_BYTES_HIGH(vsi-
     > >old_rx_bytes);
     > > +   }
     > > +   vsi->old_rx_bytes = nes->rx_bytes;
     >
     >
     > Can you please describe this logic? (indeed better to describe it
    in the
     > commit log)
     >
     > 'nes->rx_bytes' is diff in the stats register since last read.
     > 'old_rx_bytes' is the previous stats diff.
     >
     > Why/how "I40E_RXTX_BYTES_LOW(vsi->old_rx_bytes) > nes->rx_bytes" has
     > a meaning? Isn't this very depends on the read frequency?
     >
     > I guess I am missing something but please help me understand.
     >
    This patch fixes the issue of rx/tx bytes counter register overflow:
    The counter register in i40e is 48-bit long, when overflow,
    nes->rx_bytes becomes less than old_rx_bytes, the correct value of
    nes->rx_bytes should be plused 1 << 48.
    Use I40E_RXTX_BYTES_HIGH() to remember the MSB, nes->rx_bytes plus
    the MSB is the correct value, So that using uint64_t to enlarge the
    48 bit  limitation of register .

     > Also can you please confirm the initial value of the
    "vsi->offset_loaded" is
     > correct.
     >
    offset_loaded will be true when get statistics of  port and
    offset_loaded will be false when reset or clear the statistics,
    so if  offset_loaded is false, shouldn't to calculate the value of
    nes->rx_bytes, it will be 0.

     > <....>
     >
     > > @@ -282,6 +282,9 @@ struct rte_flow {
     > >   #define I40E_ETH_OVERHEAD \
     > >     (RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN +
     > I40E_VLAN_TAG_SIZE * 2)
     > >
     > > +#define I40E_RXTX_BYTES_HIGH(bytes) ((bytes) & ~I40E_48_BIT_MASK)
     > > +#define I40E_RXTX_BYTES_LOW(bytes) ((bytes) & I40E_48_BIT_MASK)
     > > +
     >
     > HIGH/LOW is a little misleading, for 64Bits it sounds like it is
    getting low 32 bits
     > and high 32 bits, can you please clarify macro masks out
     > 48/16 bits.
     >
    Yes, I will change the macro name in V3.
     >
     > >   struct i40e_adapter;
     > >   struct rte_pci_driver;
     > >
     > > @@ -399,6 +402,8 @@ struct i40e_vsi {
     > >     uint8_t vlan_anti_spoof_on; /* The VLAN anti-spoofing
    enabled */
     > >     uint8_t vlan_filter_on; /* The VLAN filter enabled */
     > >     struct i40e_bw_info bw_info; /* VSI bandwidth information */
     > > +   uint64_t old_rx_bytes;
     > > +   uint64_t old_tx_bytes;
     >
     > 'prev' seems better naming than 'old', what do you think renaming
     > 'old_rx_bytes' -> 'prev_rx_bytes' (for all variables)?
    Yes, it's better, I will fix it in V3.


Reply via email to