DPDK patches and discussions
 help / color / mirror / Atom feed
From: Usama Nadeem <usama.nadeem@emumba.com>
To: "Medvedkin, Vladimir" <vladimir.medvedkin@intel.com>
Cc: "Walsh, Conor" <conor.walsh@intel.com>,
	"thomas@monjalon.net" <thomas@monjalon.net>,
	"dev@dpdk.org" <dev@dpdk.org>
Subject: Re: [dpdk-dev] [PATCH v4] examples/l3fwd: ipv4 and udp/tcp cksum verification through software
Date: Tue, 16 Nov 2021 10:20:35 +0500	[thread overview]
Message-ID: <CABDMEvnD3vqSm1i9xBWV7s3LnrPAFNe+RBbHQmZd1sVuZ7sBVQ@mail.gmail.com> (raw)
In-Reply-To: <acedeacc-2cec-f1e3-e925-7dab429597e3@intel.com>

[-- Attachment #1: Type: text/plain, Size: 11620 bytes --]

Hi Medvedkin, Vladimir,


Thank you for your suggestions. Two counters aren't really necessary. We
also don't need the "dropped" variable. "continue" can also be used to
implement logic. We will update the logic in the next patch version.

This patch supports only IPV4 packets in LPM. It does not support other
lookup methods, neither does it support IPV6 packets for now. If the
current patch is satisfactory, we intend to begin work on those as well.

Regarding perf drop, I will submit a new version of the patch, containing
the callback.


On Thu, Nov 4, 2021 at 9:19 PM Medvedkin, Vladimir <
vladimir.medvedkin@intel.com> wrote:

> Hi Usama,
>
> On 04/11/2021 12:11, Walsh, Conor wrote:
> >> From: dev <dev-bounces@dpdk.org> On Behalf Of Usama Nadeem
> >> Sent: Thursday 14 October 2021 19:43
> >> To: thomas@monjalon.net
> >> Cc: dev@dpdk.org; Usama Nadeem <usama.nadeem@emumba.com>
> >> Subject: [dpdk-dev] [PATCH v4] examples/l3fwd: ipv4 and udp/tcp cksum
> >> verification through software
> >>
> >> checks if ipv4 and udptcp cksum offload capability available
> >> If not available, cksum is verified through software
> >> If cksum is corrupt, packet is dropped, rest of the packets
> >> are forwarded back.
> >>
> >> Bugzilla ID:545
> >> Signed-off-by: Usama Nadeem <usama.nadeem@emumba.com>
> >> ---
> >
> > Hi Usama,
> >
> > This should be done in a generic way that allows all the lookup methods
> to support it not just LPM.
> > check_software_cksum should go in a common file and be called from LPM,
> FIB and possibly EM.
> >
> > Thanks,
> > Conor.
> >
> >>   examples/l3fwd/l3fwd.h     |  6 ++++
> >>   examples/l3fwd/l3fwd_lpm.c | 72
> >> ++++++++++++++++++++++++++++++++++++--
> >>   examples/l3fwd/main.c      | 33 +++++++++++++++--
> >>   3 files changed, 105 insertions(+), 6 deletions(-)
> >>
> >> diff --git a/examples/l3fwd/l3fwd.h b/examples/l3fwd/l3fwd.h
> >> index a808d60247..c2c21a91fb 100644
> >> --- a/examples/l3fwd/l3fwd.h
> >> +++ b/examples/l3fwd/l3fwd.h
> >> @@ -55,6 +55,8 @@
> >>   #define L3FWD_HASH_ENTRIES         (1024*1024*1)
> >>   #endif
> >>   #define HASH_ENTRY_NUMBER_DEFAULT  4
> >> +extern bool l3_sft_cksum;
> >> +extern bool l4_sft_cksum;
> >>
> >>   struct mbuf_table {
> >>      uint16_t len;
> >> @@ -210,6 +212,10 @@ em_main_loop(__rte_unused void *dummy);
> >>   int
> >>   lpm_main_loop(__rte_unused void *dummy);
> >>
> >> +int
> >> +check_software_cksum(struct rte_mbuf **pkts_burst,
> >> +struct rte_mbuf **pkts_burst_to_send, int nb_rx);
> >> +
> >>   int
> >>   fib_main_loop(__rte_unused void *dummy);
> >>
> >> diff --git a/examples/l3fwd/l3fwd_lpm.c b/examples/l3fwd/l3fwd_lpm.c
> >> index 232b606b54..ecaf323943 100644
> >> --- a/examples/l3fwd/l3fwd_lpm.c
> >> +++ b/examples/l3fwd/l3fwd_lpm.c
> >> @@ -26,6 +26,7 @@
> >>   #include <rte_udp.h>
> >>   #include <rte_lpm.h>
> >>   #include <rte_lpm6.h>
> >> +#include <rte_net.h>
> >>
> >>   #include "l3fwd.h"
> >>   #include "l3fwd_event.h"
> >> @@ -139,16 +140,65 @@ lpm_get_dst_port_with_ipv4(const struct
> >> lcore_conf *qconf, struct rte_mbuf *pkt,
> >>   #include "l3fwd_lpm.h"
> >>   #endif
> >>
> >> +
> >> +int check_software_cksum(struct rte_mbuf **pkts_burst,
> >> +struct rte_mbuf **pkts_burst_to_send, int nb_rx)
> >> +{
> >> +    int j;
> >> +    int i = 0;
> >> +    struct rte_net_hdr_lens hdr_lens;
> >> +    struct rte_ipv4_hdr *ipv4_hdr;
> >> +    void *l3_hdr;
> >> +    void *l4_hdr;
> >> +    rte_be16_t prev_cksum;
> >> +    int dropped_pkts_udp_tcp = 0;
> >> +    int dropped_pkts_ipv4 = 0;
>
> Why do you need two separate counters if you eventually summing them up?
>
> >> +    bool dropped;
> >> +    for (j = 0; j < nb_rx; j++) {
> >> +            dropped = false;
> >> +            rte_net_get_ptype(pkts_burst[j], &hdr_lens,
> >> RTE_PTYPE_ALL_MASK);
> >> +            l3_hdr = rte_pktmbuf_mtod_offset(pkts_burst[j],
> >> +            void *, hdr_lens.l2_len);
> >> +            l4_hdr = rte_pktmbuf_mtod_offset(pkts_burst[j],
> >> +            void *, hdr_lens.l2_len + hdr_lens.l3_len);
>
> here hdr_lens.l3_len could be non initialized, for example in case of
> MPLS packet.
>
> >> +            ipv4_hdr = l3_hdr;
> >> +            prev_cksum = ipv4_hdr->hdr_checksum;
>
> it could be non IPv4 packet.
>
> >> +            ipv4_hdr->hdr_checksum = 0;
> >> +            ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
>
> same here and below, it can be IPv6 for example.
>
> >> +
> >> +            if (l3_sft_cksum && prev_cksum != ipv4_hdr-
> >>> hdr_checksum) {
> >> +                    rte_pktmbuf_free(pkts_burst[j]);
> >> +                    dropped_pkts_ipv4++;
> >> +                    dropped = true;
>
> Do you need "dropped" value + the the final if statement at all? Maybe
> it's better to just
> ...
>      continue;
> }
> here...
>
> >> +            } else if (l4_sft_cksum &&
> >> +                            rte_ipv4_udptcp_cksum_verify
> >> +                            (l3_hdr, l4_hdr) != 0) {
> >> +
> >> +                    rte_pktmbuf_free(pkts_burst[j]);
> >> +                    dropped_pkts_udp_tcp++;
> >> +                    dropped = true;
> >> +            }
> >> +            if (dropped == false) { >> +
> pkts_burst_to_send[i] = pkts_burst[j];
> >> +                    i++;
>
> ...and execute this code unconditionally?
>
> >> +            }
> >> +
> >> +    }
> >> +    return dropped_pkts_udp_tcp+dropped_pkts_ipv4;
> >> +}
> >> +
> >>   /* main processing loop */
> >>   int
> >>   lpm_main_loop(__rte_unused void *dummy)
> >>   {
> >>      struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
> >> +    struct rte_mbuf *pkts_burst_to_send[MAX_PKT_BURST];
> >>      unsigned lcore_id;
> >>      uint64_t prev_tsc, diff_tsc, cur_tsc;
> >>      int i, nb_rx;
> >>      uint16_t portid;
> >>      uint8_t queueid;
> >> +    int dropped;
> >>      struct lcore_conf *qconf;
> >>      const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) /
> >>              US_PER_S * BURST_TX_DRAIN_US;
> >> @@ -209,19 +259,35 @@ lpm_main_loop(__rte_unused void *dummy)
> >>                      if (nb_rx == 0)
> >>                              continue;
> >>
> >> +                    if (l3_sft_cksum || l4_sft_cksum) {
> >> +                            dropped =
> >> check_software_cksum(pkts_burst,
>
> You are calling this function fight after rte_eth_rx_burst(), so
> pkts_burst[] can have any possible packet proto, but current
> check_software_cksum() implementation if purely IPv4.
>
> >> +                            pkts_burst_to_send,     nb_rx);
> >> +
> >> +                            nb_rx = nb_rx-dropped;
> >> +                    }
> >> +
> >> +
> >>   #if defined RTE_ARCH_X86 || defined __ARM_NEON \
> >>                       || defined RTE_ARCH_PPC_64
> >> +            if (l3_sft_cksum == false && l4_sft_cksum == false)
> >>                      l3fwd_lpm_send_packets(nb_rx, pkts_burst,
> >>                                              portid, qconf);
> >> +            else
> >> +                    l3fwd_lpm_send_packets(nb_rx,
> >> pkts_burst_to_send,
> >> +                                            portid, qconf);
> >> +
> >>   #else
> >> -                    l3fwd_lpm_no_opt_send_packets(nb_rx,
> >> pkts_burst,
> >> +                    if (l3_sft_cksum == false && l4_sft_cksum == false)
>
> While those if statements are perfectly predictable, it is still better
> to avoid branching in hot path if possible. You can implement
> check_software_cksum() to work with a single pkts_burst[] modifying it
> in place and throw away pkts_burst_to_send[] and corresponding branches.
>
> >> +                            l3fwd_lpm_no_opt_send_packets(nb_rx,
> >> pkts_burst,
> >>                                                      portid, qconf);
> >> +                    else
> >> +                            l3fwd_lpm_no_opt_send_packets(nb_rx,
> >> +                            pkts_burst_to_send, portid, qconf);
> >> +
> >>   #endif /* X86 */
> >>              }
> >> -
> >>              cur_tsc = rte_rdtsc();
> >>      }
> >> -
> >>      return 0;
> >>   }
> >>
> >> diff --git a/examples/l3fwd/main.c b/examples/l3fwd/main.c
> >> index 00ac267af1..a54aca070d 100644
> >> --- a/examples/l3fwd/main.c
> >> +++ b/examples/l3fwd/main.c
> >> @@ -61,6 +61,9 @@ static uint16_t nb_txd = RTE_TEST_TX_DESC_DEFAULT;
> >>   /**< Ports set in promiscuous mode off by default. */
> >>   static int promiscuous_on;
> >>
> >> +bool l3_sft_cksum;
> >> +bool l4_sft_cksum;
> >> +
> >>   /* Select Longest-Prefix, Exact match or Forwarding Information Base.
> */
> >>   enum L3FWD_LOOKUP_MODE {
> >>      L3FWD_LOOKUP_DEFAULT,
> >> @@ -123,7 +126,6 @@ static struct rte_eth_conf port_conf = {
> >>              .mq_mode = ETH_MQ_RX_RSS,
> >>              .max_rx_pkt_len = RTE_ETHER_MAX_LEN,
> >>              .split_hdr_size = 0,
> >> -            .offloads = DEV_RX_OFFLOAD_CHECKSUM,
> >>      },
> >>      .rx_adv_conf = {
> >>              .rss_conf = {
> >> @@ -981,6 +983,7 @@ prepare_ptype_parser(uint16_t portid, uint16_t
> >> queueid)
> >>      return 0;
> >>   }
> >>
> >> +
> >>   static void
> >>   l3fwd_poll_resource_setup(void)
> >>   {
> >> @@ -993,7 +996,8 @@ l3fwd_poll_resource_setup(void)
> >>      unsigned int nb_ports;
> >>      unsigned int lcore_id;
> >>      int ret;
> >> -
> >> +    l3_sft_cksum = false;
> >> +    l4_sft_cksum = false;
> >>      if (check_lcore_params() < 0)
> >>              rte_exit(EXIT_FAILURE, "check_lcore_params failed\n");
> >>
> >> @@ -1034,11 +1038,34 @@ l3fwd_poll_resource_setup(void)
> >>                      rte_exit(EXIT_FAILURE,
> >>                              "Error during getting device (port %u)
> info:
> >> %s\n",
> >>                              portid, strerror(-ret));
> >> -
> >>              if (dev_info.tx_offload_capa &
> >> DEV_TX_OFFLOAD_MBUF_FAST_FREE)
> >>                      local_port_conf.txmode.offloads |=
> >>                              DEV_TX_OFFLOAD_MBUF_FAST_FREE;
> >>
> >> +            if (dev_info.rx_offload_capa &
> >> DEV_RX_OFFLOAD_IPV4_CKSUM)
> >> +                    local_port_conf.rxmode.offloads |=
> >> +                    DEV_RX_OFFLOAD_IPV4_CKSUM;
> >> +            else {
> >> +                    l3_sft_cksum = true;
> >> +                    printf("WARNING: IPV4 checksum offload not
> >> available.\n");
> >> +                    }
> >> +
> >> +            if (dev_info.rx_offload_capa &
> >> DEV_RX_OFFLOAD_UDP_CKSUM)
> >> +                    local_port_conf.rxmode.offloads |=
> >> +                            DEV_RX_OFFLOAD_UDP_CKSUM;
> >> +            else {
> >> +                    l4_sft_cksum = true;
> >> +                    printf("WARNING: UDP checksum offload not
> >> available.\n");
> >> +            }
> >> +
> >> +            if (dev_info.rx_offload_capa &
> >> DEV_RX_OFFLOAD_TCP_CKSUM)
> >> +                    local_port_conf.rxmode.offloads |=
> >> +                            DEV_RX_OFFLOAD_TCP_CKSUM;
> >> +            else {
> >> +                    l4_sft_cksum = true;
> >> +                    printf("WARNING: TCP checksum offload not
> >> available.\n");
> >> +            }
> >> +
> >>              local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
> >>                      dev_info.flow_type_rss_offloads;
> >>
> >> --
> >> 2.25.1
> >
>
> generalizing:
> 1. As Conor said earlier, make it generic, not only LPM and not only for
> run-to-completion mode, it should be done for event mode as well.
> 2. Function must work not only with IPv4.
> 3. There should be no performance degradation if NIC supports CSUM offload.
>
> --
> Regards,
> Vladimir
>


-- 
-Usama

[-- Attachment #2: Type: text/html, Size: 17599 bytes --]

  reply	other threads:[~2021-11-16  5:20 UTC|newest]

Thread overview: 19+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-09-13 12:09 [dpdk-dev] [PATCH] examples: ipv4, udp and tcp checksum offload warning usamanadeem321
2021-09-13 15:11 ` Stephen Hemminger
2021-09-14 18:08 ` [dpdk-dev] [PATCH v2] Warns if IPv4, UDP or TCP checksum offload not available Usama Nadeem
2021-09-14 18:28   ` Stephen Hemminger
2021-09-14 22:22   ` Ananyev, Konstantin
2021-09-14 23:44     ` Stephen Hemminger
2021-09-15  8:43       ` Ananyev, Konstantin
2021-10-08 15:51   ` [dpdk-dev] [PATCH v3] ipv4 and udp/tcp cksum verification through software Usama Nadeem
2021-10-14 18:43     ` [dpdk-dev] [PATCH v4] examples/l3fwd: " Usama Nadeem
2021-11-01  8:33       ` Usama Nadeem
2021-11-04 11:11       ` Walsh, Conor
2021-11-04 16:19         ` Medvedkin, Vladimir
2021-11-16  5:20           ` Usama Nadeem [this message]
2021-11-16  5:21         ` Usama Nadeem
2023-06-30 21:50         ` Stephen Hemminger
2021-11-04 13:19       ` Ananyev, Konstantin
2021-11-16  5:18         ` Usama Nadeem
2022-01-14  9:30         ` Usama Nadeem
2022-01-14 12:05           ` Ananyev, Konstantin

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CABDMEvnD3vqSm1i9xBWV7s3LnrPAFNe+RBbHQmZd1sVuZ7sBVQ@mail.gmail.com \
    --to=usama.nadeem@emumba.com \
    --cc=conor.walsh@intel.com \
    --cc=dev@dpdk.org \
    --cc=thomas@monjalon.net \
    --cc=vladimir.medvedkin@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).