| Message ID | 1535127870-91663-1-git-send-email-tiago.lam@intel.com |
|---|---|
| Headers | show |
| Series | Support multi-segment mbufs | expand |
Hi Ilya, Since you've shown some concerns on the latest revision, does this address you concerns? More specifically, does patch 09/14 address the points you made on v7 of the series, where dp_packet_data() was being used to traverse the whole data, even if segmented? Thanks, Tiago. On 24/08/2018 17:24, Tiago Lam wrote: > Overview > ======== > This patchset introduces support for multi-segment mbufs to OvS-DPDK. > Multi-segment mbufs are typically used when the size of an mbuf is > insufficient to contain the entirety of a packet's data. Instead, the > data is split across numerous mbufs, each carrying a portion, or > 'segment', of the packet data. Mbufs are chained via their 'next' > attribute (an mbuf pointer). > > The main motivation behind the support for multi-segment mbufs is to > later introduce TSO (use case i. below) / GRO in OvS-DPDK, which is > planned to be introduced after this series. > > Use Cases > ========= > i. Handling oversized (guest-originated) frames, which are marked > for hardware accelration/offload (TSO, for example). > > Packets which originate from a non-DPDK source may be marked for > offload; as such, they may be larger than the permitted ingress > interface's MTU, and may be stored in an oversized dp-packet. In > order to transmit such packets over a DPDK port, their contents > must be copied to a DPDK mbuf (via dpdk_do_tx_copy). However, in > its current implementation, that function only copies data into > a single mbuf; if the space available in the mbuf is exhausted, > but not all packet data has been copied, then it is lost. > Similarly, when cloning a DPDK mbuf, it must be considered > whether that mbuf contains multiple segments. Both issues are > resolved within this patchset. > > ii. Handling jumbo frames. > > While OvS already supports jumbo frames, it does so by increasing > mbuf size, such that the entirety of a jumbo frame may be handled > in a single mbuf. This is certainly the preferred, and most > performant approach (and remains the default). > > Enabling multi-segment mbufs > ============================ > Multi-segment and single-segment mbufs are mutually exclusive, and the > user must decide on which approach to adopt on init. The introduction > of a new OVSDB field, 'dpdk-multi-seg-mbufs', facilitates this. > > This is a global boolean value, which determines how jumbo frames are > represented across all DPDK ports. In the absence of a user-supplied > value, 'dpdk-multi-seg-mbufs' defaults to false, i.e. multi-segment > mbufs must be explicitly enabled / single-segment mbufs remain the > default. > > Setting the field is identical to setting existing DPDK-specific OVSDB > fields: > > ovs-vsctl set Open_vSwitch . other_config:dpdk-init=true > ovs-vsctl set Open_vSwitch . other_config:dpdk-lcore-mask=0x10 > ovs-vsctl set Open_vSwitch . other_config:dpdk-socket-mem=4096,0 > ==> ovs-vsctl set Open_vSwitch . other_config:dpdk-multi-seg-mbufs=true > > Performance notes (based on v8, 1st non-RFC) > ================= > In order to test for regressions in performance, tests were run on top > of master 88125d6 and v8 of this patchset, both with the multi-segment > mbufs option enabled and disabled. > > VSperf was used to run the phy2phy_cont and pvp_cont tests with varying > packet sizes of 64B, 1500B and 7000B, on a 10Gbps interface. > > Test | Size | Master | Multi-seg disabled | Multi-seg enabled > ------------------------------------------------------------- > p2p | 64 | ~22.7 | ~22.65 | ~18.3 > p2p | 1500 | ~1.6 | ~1.6 | ~1.6 > p2p | 7000 | ~0.36 | ~0.36 | ~0.36 > pvp | 64 | ~6.7 | ~6.7 | ~6.3 > pvp | 1500 | ~1.6 | ~1.6 | ~1.6 > pvp | 7000 | ~0.36 | ~0.36 | ~0.36 > > Packet size is in bytes, while all packet rates are reported in mpps > (aggregated). > > No noticeable regression has been observed (certainly everything is > within the ± 5% margin of existing performance), aside from the 64B > packet size case when multi-segment mbuf is enabled. This is > expected, however, because of how Tx vectoriszed functions are > incompatible with multi-segment mbufs on some PMDs. The PMD under > use during these tests was the i40e (on a Intel X710 NIC), which > indeed doesn't support vectorized Tx functions with multi-segment > mbufs. > > --- > v9: - Rebase on master e4e2009 ("tunnel, tests: Sort flow output in ERSPAN > v1/v2 metadata"); > - Simplify patch 09/14. The functions introduced in packets.c were dropped > so the code in netdev-native-tnl.c remains largely the same. These can be > introduced at a later time, if needed (maybe when csum across segmented > data is introduced); > > v8: - Rebase on master 1dd218a ("ovsdb-idl: Fix recently introduced Python 3 > tests."); > - Address Ian's comment: > - Fix sparse warnings on patch 07/14 and 12/14 by allocating memory > dynamically. > - Address Ilya's comments: > - netdev_linux_tap_batch_send() and udp_extract_tnl_md() now linearize > the data before hand, beforing write()'ing or performing the checksums > in the data; > - Some other cases have been found and adapted; The new patch 09/14 > introduced in the series is where the "linearization" logic is > introduced and, as a consequence, some users of the dp_packet API, > which were assuming the data is held contiguously in memory, are > changed to use the new APIs. > - Add support for multi-segment mbufs to dp_packet_equal() (patch 06/14); > - Fix a bug in patch 08/14 where the call to dp_packet_copy_mbuf_flags() in > dp_packet_clone_with_headroom() was setting incorrectly the nb_segs field > on the destination mbuf; > - Add unit-tests for dp_packet_equal() and (new) dp_packet_linear_data() to > patch 12/14; > - Add a comment to jumbo-frames.rst under topics/dpdk/ to warn how > multi-segments mbufs may affect performance when using large packets > across DPDK and non-DPDK ports. > > v7: - Rebase on master 024810c ("Prepare for post-2.10.0 (2.10.90)."); > - Add Ben's proposed fix for automake's warning; > - Add a note to cover letter to explain this is preperatory work for TSO / > GRO. > > v6: - Rebase on master d1b235d ("tests: Add test for ovn-nbctl's command parser > error paths."); > - Address Darrell's comments: > - The changes in dp_packet_resize__() were trying to alleviate the call > to OVS_NOT_REACHED() for DPDK packets, by trying to reuse the available > tailroom space when no more headroom space is available, and vice-versa. > However, this was breaking the API for the dp_packet_resize__() > function (only called in dp_packet_prealloc_tailroom() and > dp_packet_prealloc_headroom()), which doesn't seem to suit the purposes > for DPDK packets. > Instead, and because this is isolate funtionality, revert to the > previous state where dp_packet_resize__() is not supported for DPDK > packets. Hence, then patch 08/14 has been dropped. > - Additionally, fix the tests that were relying on the removed > functionality. > > v5: - Rebase on master 030958a0cc ("conntrack: Fix conn_update_state_alg use > after free."); > - Address Eelco's comments: > - Remove dpdk_mp_sweep() call in netdev_dpdk_mempool_configure(), a > leftover from rebase. Only call should be in dpdk_mp_get(); > - Remove NEWS line added by mistake during rebase (about adding > experimental vhost zero copy support). > - Address Ian's comments: > - Drop patch 01 from previous series entirely; > - Patch (now) 01/14 adds a new call to dpdk_buf_size() inside > dpdk_mp_create() to get the correct "mbuf_size" to be used; > - Patch (now) 11/14 modifies dpdk_mp_create() to check if multi-segment > mbufs is enabled, in which case it calculates the new "mbuf_size" to be > used; > - In free_dpdk_buf() and dpdk_buf_alloc(), don't lock and unlock > conditionally. > - Add "per-port-memory=true" to test "Multi-segment mbufs Tx" as the current > DPDK set up in system-dpdk-testsuite can't handle higher MTU sizes using > the shared mempool model (runs out of memory); > - Add new examples for when multi-segment mbufs are enabled in > topics/dpdk/memory.rst, and a reference to topics/dpdk/jumbo-frames.rst > (now patch 11/14). > > v4: - Rebase on master b22fb00 ("ovn-nbctl: Clarify error messages in qos-add > command."): > - A new patch (now 01/15) has been introduced to differentiate between > MTU and mbuf size when creating the mempool. This is because as part of > the new support for both per port and shared mempools, mempools were > being reused based on the mbuf size, and for multi-segment mbufs the > mbuf size can end up being the same for all mbufs; > - A couple of other patches (now 02/15 and 12/15) have been modified as > part of the rebase, but only to adapt to the code changes to "Support > both shared and per port mempools.", no functionality should have been > changed. > > v3: > - Address Eelco's comment: > - Fix the ovs_assert() introduced in v2 in __packet_set_data(), which > wasn't correctly asserting that the passed 'v' was smaller than the > first mbuf's buf_len. > > v2: > - Rebase on master e7cd8cf ("db-ctl-base: Don't die in cmd_destroy() on > error."); > - Address Eelco's comments: > - In free_dpdk_buf(), use mbuf's struct address in dp_packet instead of > casting; > - Remove unneeded variable in dp_packet_set_size(), pointing to the > first mbuf in the chain; > - Assert in dp_packet_set_size() to enforce that "pkt_len == v" is > always true for DPBUF_DPDK packets; > - Assert in __packet_set_data() to enforce that data_off never goes > beyond the first mbuf in the chain. > > v1: > - v8 should have been sent as v1 really, as that's usually the approach > followed in OvS. That clearly didn't happen, so restarting the series > now. This also helps making it clear it is no longer an RFC series; > - Rebase on master e461481 ("ofp-meter: Fix ofp_print_meter_flags() > output."); > - Address Eelco's comments: > - Change dp_packet_size() so that for `DPBUF_DPDK` packets their > `pkt_len` and `data_len` can't be set to values bigger than the > available space. Also fix assigment to `data_len` which was > incorrectly being set to just`pkt_len`; > - Improve `nonpmd_mp_mutex` comment with a better explanation as to > why the mutex is needed; > - Fix dp_packet_clear() to not call rte_pktmbuf_reset() for non > `DPBUF_DPDK` packets; > - Dropped `if` clause in dp_packet_l4_size(), keep just the `else`; > - Change dp_packet_clone_with_headroom() to use rte_pktmbuf_read() for > copying `DPBUF_DPDK` packets' data. Also, change it to return > appropriate and meaningful errors, instead of just "0" or "1"; > - Change dpdk_prep_tx_buf() name's to dpdk_clone_dp_packet_to_mbuf(), > and reuse dp_packet_mbuf_write() instead of manual copy; > - Add note vswitchd/vswitch.xml to make it clear the enabling of > dpdk-multi-seg-mbufs requires a restart; > - Change dpdk_mp_create() to increase # mbufs used under the multi-segment > mbufs approach; > - Increase test coverage by adding "end-to-end" tests that verify that > "dpdk-multi-seg-mbufs" is disabled by default and that a packet is > successfully sent out; > - Some helper funcs such as dp_packet_tail() and dp_packet_end() were moved > back to be common between `DPBUF_DPDK` and non `DPBUF_DPDK` packets, to > minimise changes; > - Add some extra notes to "Performance notes" in jumbo-frames.rst doc, > after further testing; > - Structure changes: > - Drop patch 07/13 which is now unneeded; > - Two more patches added for extra test coverage. This is what accounts > for the increase in size (+1 patch) in the series. > > v8 (non-RFC): > - Rebase on master 88125d6 ("rhel: remove ovs-sim man page from > temporary directory (also for RHEL)"); > - Address Ciara's and Ilya's comments: > - Drop the dp_packet_mbuf_tail() function and use only the > already existing dp_packet_tail(); > - Fix bug in dpdk_do_tx_copy() where the error return from > dpdk_prep_tx_buf() was being wrongly checked; > - Use dpdk_buf_alloc() and free_dpdk_buf() instead of > rte_pktmbuf_alloc() and rte_pktmbuf_free(); > - Fix some other code style and duplication issues pointed out. > - Refactored dp_packet_shift(), dp_packet_resize__() and > dp_packet_put*() functions to work within the bounds of existing > mbufs only; > - Fix dp_packet_clear() which wasn't correctly clearing / freeing > other mbufs in the chain for chains with more than a single mbuf; > - dp_packet layer functions (such as dp_packet_l3()) now check if > the header is within the first mbuf, when using mbufs; > - Move patch 08/13 to before patch 04/13, since dp_packet_set_size() > was refactored to use free_dpdk_buf(); > - Fix wrong rte_memcpy() when performing dp_packet_clone() which was > leading to memory corruption; > - Modified the added tests to account for some of the above changes; > - Run performance tests, compiling results and adding them to the > cover letter; > - Add a multi-seg mbufs explanation to the jumbo-frames.rst doc, > together with a "Performance notes" sub-section reflecting the > findings mentioned above in the cover letter. > > v7: - Rebase on master 5e720da ("erspan: fix invalid erspan version."); > - Address Ilya comments; > - Fix non-DPDK build; > - Serialise the access of non pmds to allocation and free of mbufs by > using a newly introduced mutex. > - Add a new set of tests that integrates with the recently added DPDK > testsuite. These focus on allocating dp_packets, with a single or > multiple mbufs, from an instantiated mempool and performing several > operations on those, verifying if the data at the end matches what's > expected; > - Fix bugs found by the new tests: > - dp_packet_shift() wasn't taking into account shift lefts; > - dp_packet_resize__() was misusing and miscalculating the tailrooms > and headrooms, ending up calculating the wrong number of segments > that needed allocation; > - An mbuf's end was being miscalculated both in dp_packet_tail, > dp_packet_mbuf_tail() and dp_packet_end(); > - dp_packet_set_size() was not updating the number of chained segments > 'nb_segs'; > - Add support for multi-seg mbufs in dp_packet_clear(). > > v6: - Rebase on master 7c0cb29 ("conntrack-tcp: Handle tcp session > reuse."); > - Further improve dp_packet_put_uninit() and dp_packet_shift() to > support multi-seg mbufs; > - Add support for multi-seg mbufs in dp_packet_l4_size() and > improve other helper funcs, such as dp_packet_tail() and dp_ > packet_tailroom(). > - Add support for multi-seg mbufs in dp_packet_put(), dp_packet_ > put_zeros(), as well as dp_packet_resize__() - allocating new > mbufs and linking them together; > Restructured patchset: > - Squash patch 5 into patch 6, since they were both related to > copying data while handling multi-seg mbufs; > - Split patch 4 into two separate patches - one that introduces the > changes in helper functions to deal with multi-seg mbufs and > two others for the shift() and put_uninit() functionality; > - Move patch 4 to before patch 3, so that ihelper functions come > before functionality improvement that rely on those helpers. > > v5: - Rebased on master e5e22dc ("datapath-windows: Prevent ct-counters > from getting redundantly incremented"); > - Sugesh's comments have been addressed: > - Changed dp_packet_set_data() and dp_packet_set_size() logic to > make them independent of each other; > - Dropped patch 3 now that dp_packet_set_data() and dp_packet_set_ > size() are independent; > - dp_packet_clone_with_headroom() now has split functions for > handling DPDK sourced packets and non-DPDK packets; > - Modified various functions in dp-packet.h to account for multi-seg > mbufs - dp_packet_put_uninit(), dp_packet_tail(), dp_packet_tail() > and dp_packet_at(); > - Added support for shifting packet data in multi-seg mbufs, using > dp_packet_shift(); > - Fixed some minor inconsistencies. > > Note that some of the changes in v5 have been contributed by Mark > Kavanagh as well. > > v4: - restructure patchset > - account for 128B ARM cacheline when sizing mbufs > > Mark Kavanagh (4): > netdev-dpdk: fix mbuf sizing > dp-packet: Init specific mbuf fields. > netdev-dpdk: copy large packet to multi-seg. mbufs > netdev-dpdk: support multi-segment jumbo frames > > Michael Qiu (1): > dp-packet: copy data from multi-seg. DPDK mbuf > > Tiago Lam (9): > dp-packet: Fix allocated size on DPDK init. > netdev-dpdk: Serialise non-pmds mbufs' alloc/free. > dp-packet: Fix data_len handling multi-seg mbufs. > dp-packet: Handle multi-seg mbufs in helper funcs. > dp-packet: Handle multi-seg mubfs in shift() func. > dp-packet: Add support for data "linearization". > dpdk-tests: Add unit-tests for multi-seg mbufs. > dpdk-tests: Accept other configs in OVS_DPDK_START > dpdk-tests: End-to-end tests for multi-seg mbufs. > > Documentation/topics/dpdk/jumbo-frames.rst | 67 ++++ > Documentation/topics/dpdk/memory.rst | 36 ++ > NEWS | 1 + > lib/bfd.c | 3 +- > lib/conntrack.c | 17 +- > lib/dp-packet.c | 192 ++++++++- > lib/dp-packet.h | 339 ++++++++++++++-- > lib/dpdk.c | 8 + > lib/dpif-netlink.c | 2 +- > lib/dpif.c | 2 +- > lib/netdev-bsd.c | 2 +- > lib/netdev-dpdk.c | 242 +++++++++-- > lib/netdev-dpdk.h | 2 + > lib/netdev-dummy.c | 5 +- > lib/netdev-linux.c | 5 +- > lib/netdev-native-tnl.c | 10 +- > lib/odp-execute.c | 2 +- > lib/ofp-print.c | 2 +- > lib/ovs-lldp.c | 3 +- > lib/packets.c | 3 +- > ofproto/ofproto-dpif-sflow.c | 2 +- > ofproto/ofproto-dpif-upcall.c | 2 +- > ofproto/ofproto-dpif-xlate.c | 12 +- > tests/automake.mk | 10 +- > tests/dpdk-packet-mbufs.at | 7 + > tests/system-dpdk-macros.at | 6 +- > tests/system-dpdk-testsuite.at | 1 + > tests/system-dpdk.at | 65 +++ > tests/test-dpdk-mbufs.c | 619 +++++++++++++++++++++++++++++ > vswitchd/vswitch.xml | 22 + > 30 files changed, 1571 insertions(+), 118 deletions(-) > create mode 100644 tests/dpdk-packet-mbufs.at > create mode 100644 tests/test-dpdk-mbufs.c >
Hi, Tiago. I applied that whole patch-set and looked through the code. Didn't finish review yet, but have a few concerns about current implementation: 1. dp_packet_l3/l4/others returns the pointer and checks only that offset is less than the size of the first segment. But this check doesn't guarantee that even one byte of this header located in the first segment. Users will commonly read/write using that pointer and crash/read wrong memory regions. 2. mbuf modifications with buf_addr, data_off, etc changes looks very dangerous. At least, in your current implementation dp_packet_get_allocated() will return wrong value for reallocated this way packet, because you're not updating nb_segs. This is really hard to track all the mbuf fields that should be stored and every DPDK upgrade may introduce very complex issues. Maybe it's better to just clone the packet in a usual way? 3. dp_packet_linear_ofs(..) must be called before any call of dp_packet_l3/l4/others ? Otherwise you will get a wrong pointer. This is a bad concept from the API point of view. Also, some users in your current implementation uses both dp_packet_linear_ofs(pkt, pkt->l3_ofs) and dp_packet_l3(pkt) at the same time for unknown reason. See handle_ftp_ctl() for example. 4. handle_ftp_ctl() copies the whole packet even if function will return after the first simple check. 5. You missed copying of 'rss_hash_valid' in dp_packet_clone_with_headroom(). In general, there are a lot of code duplications which is not a good thing in so complicated code. 6. No need to have netdev_dpdk_is_multi_segment_mbufs_enabled() in a common header. 7. dp_packet_reset_packet() is broken. dp_packet_set_size() should not change the real packet, i.e. free the segments. In your implementation dp_packet_set_size() crops the packet from the tail and after that dp_packet_set_data() moves the head forward and reduces packet size again. As a result dp_packet_reset_packet() eats the requested number of bytes from both sides of the packet. 8. dp_packet_l3/l4/others returns NULL in case of headers placement not in a first segment. But callers never checks the result. They expect that these functions always return valid pointer for the packets that contains these headers. In reality, you can not be sure that all the headers are in the first segment. For example in case of few levels of tunnelling. Best regards, Ilya Maximets. On 30.08.2018 10:36, Lam, Tiago wrote: > Hi Ilya, > > Since you've shown some concerns on the latest revision, does this > address you concerns? More specifically, does patch 09/14 address the > points you made on v7 of the series, where dp_packet_data() was being > used to traverse the whole data, even if segmented? > > Thanks, > Tiago. > > On 24/08/2018 17:24, Tiago Lam wrote: >> Overview >> ======== >> This patchset introduces support for multi-segment mbufs to OvS-DPDK. >> Multi-segment mbufs are typically used when the size of an mbuf is >> insufficient to contain the entirety of a packet's data. Instead, the >> data is split across numerous mbufs, each carrying a portion, or >> 'segment', of the packet data. Mbufs are chained via their 'next' >> attribute (an mbuf pointer). >> >> The main motivation behind the support for multi-segment mbufs is to >> later introduce TSO (use case i. below) / GRO in OvS-DPDK, which is >> planned to be introduced after this series. >> >> Use Cases >> ========= >> i. Handling oversized (guest-originated) frames, which are marked >> for hardware accelration/offload (TSO, for example). >> >> Packets which originate from a non-DPDK source may be marked for >> offload; as such, they may be larger than the permitted ingress >> interface's MTU, and may be stored in an oversized dp-packet. In >> order to transmit such packets over a DPDK port, their contents >> must be copied to a DPDK mbuf (via dpdk_do_tx_copy). However, in >> its current implementation, that function only copies data into >> a single mbuf; if the space available in the mbuf is exhausted, >> but not all packet data has been copied, then it is lost. >> Similarly, when cloning a DPDK mbuf, it must be considered >> whether that mbuf contains multiple segments. Both issues are >> resolved within this patchset. >> >> ii. Handling jumbo frames. >> >> While OvS already supports jumbo frames, it does so by increasing >> mbuf size, such that the entirety of a jumbo frame may be handled >> in a single mbuf. This is certainly the preferred, and most >> performant approach (and remains the default). >> >> Enabling multi-segment mbufs >> ============================ >> Multi-segment and single-segment mbufs are mutually exclusive, and the >> user must decide on which approach to adopt on init. The introduction >> of a new OVSDB field, 'dpdk-multi-seg-mbufs', facilitates this. >> >> This is a global boolean value, which determines how jumbo frames are >> represented across all DPDK ports. In the absence of a user-supplied >> value, 'dpdk-multi-seg-mbufs' defaults to false, i.e. multi-segment >> mbufs must be explicitly enabled / single-segment mbufs remain the >> default. >> >> Setting the field is identical to setting existing DPDK-specific OVSDB >> fields: >> >> ovs-vsctl set Open_vSwitch . other_config:dpdk-init=true >> ovs-vsctl set Open_vSwitch . other_config:dpdk-lcore-mask=0x10 >> ovs-vsctl set Open_vSwitch . other_config:dpdk-socket-mem=4096,0 >> ==> ovs-vsctl set Open_vSwitch . other_config:dpdk-multi-seg-mbufs=true >> >> Performance notes (based on v8, 1st non-RFC) >> ================= >> In order to test for regressions in performance, tests were run on top >> of master 88125d6 and v8 of this patchset, both with the multi-segment >> mbufs option enabled and disabled. >> >> VSperf was used to run the phy2phy_cont and pvp_cont tests with varying >> packet sizes of 64B, 1500B and 7000B, on a 10Gbps interface. >> >> Test | Size | Master | Multi-seg disabled | Multi-seg enabled >> ------------------------------------------------------------- >> p2p | 64 | ~22.7 | ~22.65 | ~18.3 >> p2p | 1500 | ~1.6 | ~1.6 | ~1.6 >> p2p | 7000 | ~0.36 | ~0.36 | ~0.36 >> pvp | 64 | ~6.7 | ~6.7 | ~6.3 >> pvp | 1500 | ~1.6 | ~1.6 | ~1.6 >> pvp | 7000 | ~0.36 | ~0.36 | ~0.36 >> >> Packet size is in bytes, while all packet rates are reported in mpps >> (aggregated). >> >> No noticeable regression has been observed (certainly everything is >> within the ± 5% margin of existing performance), aside from the 64B >> packet size case when multi-segment mbuf is enabled. This is >> expected, however, because of how Tx vectoriszed functions are >> incompatible with multi-segment mbufs on some PMDs. The PMD under >> use during these tests was the i40e (on a Intel X710 NIC), which >> indeed doesn't support vectorized Tx functions with multi-segment >> mbufs. >> >> --- >> v9: - Rebase on master e4e2009 ("tunnel, tests: Sort flow output in ERSPAN >> v1/v2 metadata"); >> - Simplify patch 09/14. The functions introduced in packets.c were dropped >> so the code in netdev-native-tnl.c remains largely the same. These can be >> introduced at a later time, if needed (maybe when csum across segmented >> data is introduced); >> >> v8: - Rebase on master 1dd218a ("ovsdb-idl: Fix recently introduced Python 3 >> tests."); >> - Address Ian's comment: >> - Fix sparse warnings on patch 07/14 and 12/14 by allocating memory >> dynamically. >> - Address Ilya's comments: >> - netdev_linux_tap_batch_send() and udp_extract_tnl_md() now linearize >> the data before hand, beforing write()'ing or performing the checksums >> in the data; >> - Some other cases have been found and adapted; The new patch 09/14 >> introduced in the series is where the "linearization" logic is >> introduced and, as a consequence, some users of the dp_packet API, >> which were assuming the data is held contiguously in memory, are >> changed to use the new APIs. >> - Add support for multi-segment mbufs to dp_packet_equal() (patch 06/14); >> - Fix a bug in patch 08/14 where the call to dp_packet_copy_mbuf_flags() in >> dp_packet_clone_with_headroom() was setting incorrectly the nb_segs field >> on the destination mbuf; >> - Add unit-tests for dp_packet_equal() and (new) dp_packet_linear_data() to >> patch 12/14; >> - Add a comment to jumbo-frames.rst under topics/dpdk/ to warn how >> multi-segments mbufs may affect performance when using large packets >> across DPDK and non-DPDK ports. >> >> v7: - Rebase on master 024810c ("Prepare for post-2.10.0 (2.10.90)."); >> - Add Ben's proposed fix for automake's warning; >> - Add a note to cover letter to explain this is preperatory work for TSO / >> GRO. >> >> v6: - Rebase on master d1b235d ("tests: Add test for ovn-nbctl's command parser >> error paths."); >> - Address Darrell's comments: >> - The changes in dp_packet_resize__() were trying to alleviate the call >> to OVS_NOT_REACHED() for DPDK packets, by trying to reuse the available >> tailroom space when no more headroom space is available, and vice-versa. >> However, this was breaking the API for the dp_packet_resize__() >> function (only called in dp_packet_prealloc_tailroom() and >> dp_packet_prealloc_headroom()), which doesn't seem to suit the purposes >> for DPDK packets. >> Instead, and because this is isolate funtionality, revert to the >> previous state where dp_packet_resize__() is not supported for DPDK >> packets. Hence, then patch 08/14 has been dropped. >> - Additionally, fix the tests that were relying on the removed >> functionality. >> >> v5: - Rebase on master 030958a0cc ("conntrack: Fix conn_update_state_alg use >> after free."); >> - Address Eelco's comments: >> - Remove dpdk_mp_sweep() call in netdev_dpdk_mempool_configure(), a >> leftover from rebase. Only call should be in dpdk_mp_get(); >> - Remove NEWS line added by mistake during rebase (about adding >> experimental vhost zero copy support). >> - Address Ian's comments: >> - Drop patch 01 from previous series entirely; >> - Patch (now) 01/14 adds a new call to dpdk_buf_size() inside >> dpdk_mp_create() to get the correct "mbuf_size" to be used; >> - Patch (now) 11/14 modifies dpdk_mp_create() to check if multi-segment >> mbufs is enabled, in which case it calculates the new "mbuf_size" to be >> used; >> - In free_dpdk_buf() and dpdk_buf_alloc(), don't lock and unlock >> conditionally. >> - Add "per-port-memory=true" to test "Multi-segment mbufs Tx" as the current >> DPDK set up in system-dpdk-testsuite can't handle higher MTU sizes using >> the shared mempool model (runs out of memory); >> - Add new examples for when multi-segment mbufs are enabled in >> topics/dpdk/memory.rst, and a reference to topics/dpdk/jumbo-frames.rst >> (now patch 11/14). >> >> v4: - Rebase on master b22fb00 ("ovn-nbctl: Clarify error messages in qos-add >> command."): >> - A new patch (now 01/15) has been introduced to differentiate between >> MTU and mbuf size when creating the mempool. This is because as part of >> the new support for both per port and shared mempools, mempools were >> being reused based on the mbuf size, and for multi-segment mbufs the >> mbuf size can end up being the same for all mbufs; >> - A couple of other patches (now 02/15 and 12/15) have been modified as >> part of the rebase, but only to adapt to the code changes to "Support >> both shared and per port mempools.", no functionality should have been >> changed. >> >> v3: >> - Address Eelco's comment: >> - Fix the ovs_assert() introduced in v2 in __packet_set_data(), which >> wasn't correctly asserting that the passed 'v' was smaller than the >> first mbuf's buf_len. >> >> v2: >> - Rebase on master e7cd8cf ("db-ctl-base: Don't die in cmd_destroy() on >> error."); >> - Address Eelco's comments: >> - In free_dpdk_buf(), use mbuf's struct address in dp_packet instead of >> casting; >> - Remove unneeded variable in dp_packet_set_size(), pointing to the >> first mbuf in the chain; >> - Assert in dp_packet_set_size() to enforce that "pkt_len == v" is >> always true for DPBUF_DPDK packets; >> - Assert in __packet_set_data() to enforce that data_off never goes >> beyond the first mbuf in the chain. >> >> v1: >> - v8 should have been sent as v1 really, as that's usually the approach >> followed in OvS. That clearly didn't happen, so restarting the series >> now. This also helps making it clear it is no longer an RFC series; >> - Rebase on master e461481 ("ofp-meter: Fix ofp_print_meter_flags() >> output."); >> - Address Eelco's comments: >> - Change dp_packet_size() so that for `DPBUF_DPDK` packets their >> `pkt_len` and `data_len` can't be set to values bigger than the >> available space. Also fix assigment to `data_len` which was >> incorrectly being set to just`pkt_len`; >> - Improve `nonpmd_mp_mutex` comment with a better explanation as to >> why the mutex is needed; >> - Fix dp_packet_clear() to not call rte_pktmbuf_reset() for non >> `DPBUF_DPDK` packets; >> - Dropped `if` clause in dp_packet_l4_size(), keep just the `else`; >> - Change dp_packet_clone_with_headroom() to use rte_pktmbuf_read() for >> copying `DPBUF_DPDK` packets' data. Also, change it to return >> appropriate and meaningful errors, instead of just "0" or "1"; >> - Change dpdk_prep_tx_buf() name's to dpdk_clone_dp_packet_to_mbuf(), >> and reuse dp_packet_mbuf_write() instead of manual copy; >> - Add note vswitchd/vswitch.xml to make it clear the enabling of >> dpdk-multi-seg-mbufs requires a restart; >> - Change dpdk_mp_create() to increase # mbufs used under the multi-segment >> mbufs approach; >> - Increase test coverage by adding "end-to-end" tests that verify that >> "dpdk-multi-seg-mbufs" is disabled by default and that a packet is >> successfully sent out; >> - Some helper funcs such as dp_packet_tail() and dp_packet_end() were moved >> back to be common between `DPBUF_DPDK` and non `DPBUF_DPDK` packets, to >> minimise changes; >> - Add some extra notes to "Performance notes" in jumbo-frames.rst doc, >> after further testing; >> - Structure changes: >> - Drop patch 07/13 which is now unneeded; >> - Two more patches added for extra test coverage. This is what accounts >> for the increase in size (+1 patch) in the series. >> >> v8 (non-RFC): >> - Rebase on master 88125d6 ("rhel: remove ovs-sim man page from >> temporary directory (also for RHEL)"); >> - Address Ciara's and Ilya's comments: >> - Drop the dp_packet_mbuf_tail() function and use only the >> already existing dp_packet_tail(); >> - Fix bug in dpdk_do_tx_copy() where the error return from >> dpdk_prep_tx_buf() was being wrongly checked; >> - Use dpdk_buf_alloc() and free_dpdk_buf() instead of >> rte_pktmbuf_alloc() and rte_pktmbuf_free(); >> - Fix some other code style and duplication issues pointed out. >> - Refactored dp_packet_shift(), dp_packet_resize__() and >> dp_packet_put*() functions to work within the bounds of existing >> mbufs only; >> - Fix dp_packet_clear() which wasn't correctly clearing / freeing >> other mbufs in the chain for chains with more than a single mbuf; >> - dp_packet layer functions (such as dp_packet_l3()) now check if >> the header is within the first mbuf, when using mbufs; >> - Move patch 08/13 to before patch 04/13, since dp_packet_set_size() >> was refactored to use free_dpdk_buf(); >> - Fix wrong rte_memcpy() when performing dp_packet_clone() which was >> leading to memory corruption; >> - Modified the added tests to account for some of the above changes; >> - Run performance tests, compiling results and adding them to the >> cover letter; >> - Add a multi-seg mbufs explanation to the jumbo-frames.rst doc, >> together with a "Performance notes" sub-section reflecting the >> findings mentioned above in the cover letter. >> >> v7: - Rebase on master 5e720da ("erspan: fix invalid erspan version."); >> - Address Ilya comments; >> - Fix non-DPDK build; >> - Serialise the access of non pmds to allocation and free of mbufs by >> using a newly introduced mutex. >> - Add a new set of tests that integrates with the recently added DPDK >> testsuite. These focus on allocating dp_packets, with a single or >> multiple mbufs, from an instantiated mempool and performing several >> operations on those, verifying if the data at the end matches what's >> expected; >> - Fix bugs found by the new tests: >> - dp_packet_shift() wasn't taking into account shift lefts; >> - dp_packet_resize__() was misusing and miscalculating the tailrooms >> and headrooms, ending up calculating the wrong number of segments >> that needed allocation; >> - An mbuf's end was being miscalculated both in dp_packet_tail, >> dp_packet_mbuf_tail() and dp_packet_end(); >> - dp_packet_set_size() was not updating the number of chained segments >> 'nb_segs'; >> - Add support for multi-seg mbufs in dp_packet_clear(). >> >> v6: - Rebase on master 7c0cb29 ("conntrack-tcp: Handle tcp session >> reuse."); >> - Further improve dp_packet_put_uninit() and dp_packet_shift() to >> support multi-seg mbufs; >> - Add support for multi-seg mbufs in dp_packet_l4_size() and >> improve other helper funcs, such as dp_packet_tail() and dp_ >> packet_tailroom(). >> - Add support for multi-seg mbufs in dp_packet_put(), dp_packet_ >> put_zeros(), as well as dp_packet_resize__() - allocating new >> mbufs and linking them together; >> Restructured patchset: >> - Squash patch 5 into patch 6, since they were both related to >> copying data while handling multi-seg mbufs; >> - Split patch 4 into two separate patches - one that introduces the >> changes in helper functions to deal with multi-seg mbufs and >> two others for the shift() and put_uninit() functionality; >> - Move patch 4 to before patch 3, so that ihelper functions come >> before functionality improvement that rely on those helpers. >> >> v5: - Rebased on master e5e22dc ("datapath-windows: Prevent ct-counters >> from getting redundantly incremented"); >> - Sugesh's comments have been addressed: >> - Changed dp_packet_set_data() and dp_packet_set_size() logic to >> make them independent of each other; >> - Dropped patch 3 now that dp_packet_set_data() and dp_packet_set_ >> size() are independent; >> - dp_packet_clone_with_headroom() now has split functions for >> handling DPDK sourced packets and non-DPDK packets; >> - Modified various functions in dp-packet.h to account for multi-seg >> mbufs - dp_packet_put_uninit(), dp_packet_tail(), dp_packet_tail() >> and dp_packet_at(); >> - Added support for shifting packet data in multi-seg mbufs, using >> dp_packet_shift(); >> - Fixed some minor inconsistencies. >> >> Note that some of the changes in v5 have been contributed by Mark >> Kavanagh as well. >> >> v4: - restructure patchset >> - account for 128B ARM cacheline when sizing mbufs >> >> Mark Kavanagh (4): >> netdev-dpdk: fix mbuf sizing >> dp-packet: Init specific mbuf fields. >> netdev-dpdk: copy large packet to multi-seg. mbufs >> netdev-dpdk: support multi-segment jumbo frames >> >> Michael Qiu (1): >> dp-packet: copy data from multi-seg. DPDK mbuf >> >> Tiago Lam (9): >> dp-packet: Fix allocated size on DPDK init. >> netdev-dpdk: Serialise non-pmds mbufs' alloc/free. >> dp-packet: Fix data_len handling multi-seg mbufs. >> dp-packet: Handle multi-seg mbufs in helper funcs. >> dp-packet: Handle multi-seg mubfs in shift() func. >> dp-packet: Add support for data "linearization". >> dpdk-tests: Add unit-tests for multi-seg mbufs. >> dpdk-tests: Accept other configs in OVS_DPDK_START >> dpdk-tests: End-to-end tests for multi-seg mbufs. >> >> Documentation/topics/dpdk/jumbo-frames.rst | 67 ++++ >> Documentation/topics/dpdk/memory.rst | 36 ++ >> NEWS | 1 + >> lib/bfd.c | 3 +- >> lib/conntrack.c | 17 +- >> lib/dp-packet.c | 192 ++++++++- >> lib/dp-packet.h | 339 ++++++++++++++-- >> lib/dpdk.c | 8 + >> lib/dpif-netlink.c | 2 +- >> lib/dpif.c | 2 +- >> lib/netdev-bsd.c | 2 +- >> lib/netdev-dpdk.c | 242 +++++++++-- >> lib/netdev-dpdk.h | 2 + >> lib/netdev-dummy.c | 5 +- >> lib/netdev-linux.c | 5 +- >> lib/netdev-native-tnl.c | 10 +- >> lib/odp-execute.c | 2 +- >> lib/ofp-print.c | 2 +- >> lib/ovs-lldp.c | 3 +- >> lib/packets.c | 3 +- >> ofproto/ofproto-dpif-sflow.c | 2 +- >> ofproto/ofproto-dpif-upcall.c | 2 +- >> ofproto/ofproto-dpif-xlate.c | 12 +- >> tests/automake.mk | 10 +- >> tests/dpdk-packet-mbufs.at | 7 + >> tests/system-dpdk-macros.at | 6 +- >> tests/system-dpdk-testsuite.at | 1 + >> tests/system-dpdk.at | 65 +++ >> tests/test-dpdk-mbufs.c | 619 +++++++++++++++++++++++++++++ >> vswitchd/vswitch.xml | 22 + >> 30 files changed, 1571 insertions(+), 118 deletions(-) >> create mode 100644 tests/dpdk-packet-mbufs.at >> create mode 100644 tests/test-dpdk-mbufs.c >> > >
Hi Ilya, Thanks for the comments, I have some responses in-line. I'll cook up the fixes for the next iteration, but would be interested in getting on your views, mainly on 1., 2. and 3.. On 30/08/2018 12:06, Ilya Maximets wrote: > Hi, Tiago. I applied that whole patch-set and looked through the > code. Didn't finish review yet, but have a few concerns about current > implementation: > > 1. dp_packet_l3/l4/others returns the pointer and checks only that > offset is less than the size of the first segment. But this check > doesn't guarantee that even one byte of this header located in > the first segment. Users will commonly read/write using that > pointer and crash/read wrong memory regions. For packets with small / few headers, there won't be any problems. But for packets where there's the need to access bigger / more headers, that's why dp_packet_linear_ofs() has been introduced. It copies the data so it can be accessed contiguously. More generally, the idea of the new linear APIs is - when a caller knows it will need all the data, such as a calculating a checksum or calling a write() it will need to use the linear APIs, instead of the the normal APIs. > > 2. mbuf modifications with buf_addr, data_off, etc changes > looks very dangerous. At least, in your current implementation > dp_packet_get_allocated() will return wrong value for reallocated > this way packet, because you're not updating nb_segs. > This is really hard to track all the mbuf fields that should be stored > and every DPDK upgrade may introduce very complex issues. > Maybe it's better to just clone the packet in a usual way? > > 3. dp_packet_linear_ofs(..) must be called before any call of > dp_packet_l3/l4/others ? Otherwise you will get a wrong pointer. > This is a bad concept from the API point of view. > Also, some users in your current implementation uses both > dp_packet_linear_ofs(pkt, pkt->l3_ofs) and dp_packet_l3(pkt) > at the same time for unknown reason. See handle_ftp_ctl() for > example. I see your point; The idea here was avoiding changing the callers of the dp_packet API as much as possible. If we make dp_packet_linear_data() call dp_packet_clone() then the caller will have to to update its pointer to the packet. In some cases this pointer is even `CONST *`. Hence why I opted for this approach. > > 4. handle_ftp_ctl() copies the whole packet even if function will > return after the first simple check.> > 5. You missed copying of 'rss_hash_valid' in dp_packet_clone_with_headroom(). > In general, there are a lot of code duplications which is > not a good thing in so complicated code. I don't understand this one; `rss_hash_valid` is only supposed to be used when OvS has been compiled without DPDK. With regards to the duplication, a previous review pointed out that there were two many compilation flags around dp_packet_clone_with_headroom(), hence why it has been broken out into two different implementations (which leads to some duplication). > > 6. No need to have netdev_dpdk_is_multi_segment_mbufs_enabled() in > a common header. > > 7. dp_packet_reset_packet() is broken. > dp_packet_set_size() should not change the real packet, i.e. free > the segments. In your implementation dp_packet_set_size() crops > the packet from the tail and after that dp_packet_set_data() > moves the head forward and reduces packet size again. As a result > dp_packet_reset_packet() eats the requested number of bytes from > both sides of the packet. You're right, I'll fix 6 and 7. However, dp_packet_size() was implemented like this (frees the tail segments) so that the tail of the packet is always in the last segment mbuf. This is to avoid returning a pointer to data that's segment in dp_packet_tail() or dp_packet_end(). > > 8. dp_packet_l3/l4/others returns NULL in case of headers placement > not in a first segment. But callers never checks the result. > They expect that these functions always return valid pointer > for the packets that contains these headers. In reality, > you can not be sure that all the headers are in the first segment. > For example in case of few levels of tunnelling. It is true that some places are not checking for NULL, but I believe I've covered most cases by using dp_packet_linear_ofs(). Are you talking of any case in specific? But given this is hidden behind the "dpdk-multi-seg-mbufs" flags I'd prefer to treat it on a case by case scenario other than "polluting" the code with NULL checks where none is required (most of the cases, I would say, won't require it). Thanks, Tiago.
Hi Ilya, On 30/08/2018 12:06, Ilya Maximets wrote: > Hi, Tiago. I applied that whole patch-set and looked through the > code. Didn't finish review yet, but have a few concerns about current > implementation: > > 1. dp_packet_l3/l4/others returns the pointer and checks only that > offset is less than the size of the first segment. But this check > doesn't guarantee that even one byte of this header located in > the first segment. Users will commonly read/write using that > pointer and crash/read wrong memory regions. > > 2. mbuf modifications with buf_addr, data_off, etc changes > looks very dangerous. At least, in your current implementation > dp_packet_get_allocated() will return wrong value for reallocated > this way packet, because you're not updating nb_segs. > This is really hard to track all the mbuf fields that should be stored > and every DPDK upgrade may introduce very complex issues. > Maybe it's better to just clone the packet in a usual way? > > 3. dp_packet_linear_ofs(..) must be called before any call of > dp_packet_l3/l4/others ? Otherwise you will get a wrong pointer. > This is a bad concept from the API point of view. > Also, some users in your current implementation uses both > dp_packet_linear_ofs(pkt, pkt->l3_ofs) and dp_packet_l3(pkt) > at the same time for unknown reason. See handle_ftp_ctl() for > example. I've fixed the other issues, but would like to clarify the above two points (2. and 3.) before sending v10. With regards to the way dp_packet_linear_data() is making the conversion between DPDK and non-DPDK packets, it seems to me that the options we have here are, either construct another packet entirely and store the copied data in the new packet, or modify the current packet internally to hold the new data. I chose the latter to abstract the changes from the caller, so the caller doesn't have to update its pointer to the packet. I tend to agree the current way it is being done (saving `buf_addr` and such) is prone to error. However, I've noticed that in DPDK 18.05 a new function has been introduced, rte_pktmbuf_attach_extbuff() [1], which, theoretically (I haven't played with it just yet), will enable us to attach an external buffer into an mbuf. This would allow us to remove the process currently in place once we integrate with DPDK 18.11, by just attaching the malloc()'ed buffer and detaching before returning the mbuf to the pool. Then, the purpose is not to use dp_packet_linear_ofs() before or after dp_packet_l3/l4/others, but instead. That is, if one plans to traverse the data, dp_packet_linear_ofs() is the function to use. Otherwise dp_packet_l3() is the one. The caller just needs to be aware that by calling the linear APIs the address of where the data is will change. Does that make sense? What do you think? Tiago. [1] http://doc.dpdk.org/api/rte__mbuf_8h.html#a67b6d31fd6f79b55dca6565e18431668 > > 4. handle_ftp_ctl() copies the whole packet even if function will > return after the first simple check. > > 5. You missed copying of 'rss_hash_valid' in dp_packet_clone_with_headroom(). > In general, there are a lot of code duplications which is > not a good thing in so complicated code. > > 6. No need to have netdev_dpdk_is_multi_segment_mbufs_enabled() in > a common header. > > 7. dp_packet_reset_packet() is broken. > dp_packet_set_size() should not change the real packet, i.e. free > the segments. In your implementation dp_packet_set_size() crops > the packet from the tail and after that dp_packet_set_data() > moves the head forward and reduces packet size again. As a result > dp_packet_reset_packet() eats the requested number of bytes from > both sides of the packet. > > 8. dp_packet_l3/l4/others returns NULL in case of headers placement > not in a first segment. But callers never checks the result. > They expect that these functions always return valid pointer > for the packets that contains these headers. In reality, > you can not be sure that all the headers are in the first segment. > For example in case of few levels of tunnelling. > > Best regards, Ilya Maximets. > > On 30.08.2018 10:36, Lam, Tiago wrote: >> Hi Ilya, >> >> Since you've shown some concerns on the latest revision, does this >> address you concerns? More specifically, does patch 09/14 address the >> points you made on v7 of the series, where dp_packet_data() was being >> used to traverse the whole data, even if segmented? >> >> Thanks, >> Tiago. >> >> On 24/08/2018 17:24, Tiago Lam wrote: >>> Overview >>> ======== >>> This patchset introduces support for multi-segment mbufs to OvS-DPDK. >>> Multi-segment mbufs are typically used when the size of an mbuf is >>> insufficient to contain the entirety of a packet's data. Instead, the >>> data is split across numerous mbufs, each carrying a portion, or >>> 'segment', of the packet data. Mbufs are chained via their 'next' >>> attribute (an mbuf pointer). >>> >>> The main motivation behind the support for multi-segment mbufs is to >>> later introduce TSO (use case i. below) / GRO in OvS-DPDK, which is >>> planned to be introduced after this series. >>> >>> Use Cases >>> ========= >>> i. Handling oversized (guest-originated) frames, which are marked >>> for hardware accelration/offload (TSO, for example). >>> >>> Packets which originate from a non-DPDK source may be marked for >>> offload; as such, they may be larger than the permitted ingress >>> interface's MTU, and may be stored in an oversized dp-packet. In >>> order to transmit such packets over a DPDK port, their contents >>> must be copied to a DPDK mbuf (via dpdk_do_tx_copy). However, in >>> its current implementation, that function only copies data into >>> a single mbuf; if the space available in the mbuf is exhausted, >>> but not all packet data has been copied, then it is lost. >>> Similarly, when cloning a DPDK mbuf, it must be considered >>> whether that mbuf contains multiple segments. Both issues are >>> resolved within this patchset. >>> >>> ii. Handling jumbo frames. >>> >>> While OvS already supports jumbo frames, it does so by increasing >>> mbuf size, such that the entirety of a jumbo frame may be handled >>> in a single mbuf. This is certainly the preferred, and most >>> performant approach (and remains the default). >>> >>> Enabling multi-segment mbufs >>> ============================ >>> Multi-segment and single-segment mbufs are mutually exclusive, and the >>> user must decide on which approach to adopt on init. The introduction >>> of a new OVSDB field, 'dpdk-multi-seg-mbufs', facilitates this. >>> >>> This is a global boolean value, which determines how jumbo frames are >>> represented across all DPDK ports. In the absence of a user-supplied >>> value, 'dpdk-multi-seg-mbufs' defaults to false, i.e. multi-segment >>> mbufs must be explicitly enabled / single-segment mbufs remain the >>> default. >>> >>> Setting the field is identical to setting existing DPDK-specific OVSDB >>> fields: >>> >>> ovs-vsctl set Open_vSwitch . other_config:dpdk-init=true >>> ovs-vsctl set Open_vSwitch . other_config:dpdk-lcore-mask=0x10 >>> ovs-vsctl set Open_vSwitch . other_config:dpdk-socket-mem=4096,0 >>> ==> ovs-vsctl set Open_vSwitch . other_config:dpdk-multi-seg-mbufs=true >>> >>> Performance notes (based on v8, 1st non-RFC) >>> ================= >>> In order to test for regressions in performance, tests were run on top >>> of master 88125d6 and v8 of this patchset, both with the multi-segment >>> mbufs option enabled and disabled. >>> >>> VSperf was used to run the phy2phy_cont and pvp_cont tests with varying >>> packet sizes of 64B, 1500B and 7000B, on a 10Gbps interface. >>> >>> Test | Size | Master | Multi-seg disabled | Multi-seg enabled >>> ------------------------------------------------------------- >>> p2p | 64 | ~22.7 | ~22.65 | ~18.3 >>> p2p | 1500 | ~1.6 | ~1.6 | ~1.6 >>> p2p | 7000 | ~0.36 | ~0.36 | ~0.36 >>> pvp | 64 | ~6.7 | ~6.7 | ~6.3 >>> pvp | 1500 | ~1.6 | ~1.6 | ~1.6 >>> pvp | 7000 | ~0.36 | ~0.36 | ~0.36 >>> >>> Packet size is in bytes, while all packet rates are reported in mpps >>> (aggregated). >>> >>> No noticeable regression has been observed (certainly everything is >>> within the ± 5% margin of existing performance), aside from the 64B >>> packet size case when multi-segment mbuf is enabled. This is >>> expected, however, because of how Tx vectoriszed functions are >>> incompatible with multi-segment mbufs on some PMDs. The PMD under >>> use during these tests was the i40e (on a Intel X710 NIC), which >>> indeed doesn't support vectorized Tx functions with multi-segment >>> mbufs. >>> >>> --- >>> v9: - Rebase on master e4e2009 ("tunnel, tests: Sort flow output in ERSPAN >>> v1/v2 metadata"); >>> - Simplify patch 09/14. The functions introduced in packets.c were dropped >>> so the code in netdev-native-tnl.c remains largely the same. These can be >>> introduced at a later time, if needed (maybe when csum across segmented >>> data is introduced); >>> >>> v8: - Rebase on master 1dd218a ("ovsdb-idl: Fix recently introduced Python 3 >>> tests."); >>> - Address Ian's comment: >>> - Fix sparse warnings on patch 07/14 and 12/14 by allocating memory >>> dynamically. >>> - Address Ilya's comments: >>> - netdev_linux_tap_batch_send() and udp_extract_tnl_md() now linearize >>> the data before hand, beforing write()'ing or performing the checksums >>> in the data; >>> - Some other cases have been found and adapted; The new patch 09/14 >>> introduced in the series is where the "linearization" logic is >>> introduced and, as a consequence, some users of the dp_packet API, >>> which were assuming the data is held contiguously in memory, are >>> changed to use the new APIs. >>> - Add support for multi-segment mbufs to dp_packet_equal() (patch 06/14); >>> - Fix a bug in patch 08/14 where the call to dp_packet_copy_mbuf_flags() in >>> dp_packet_clone_with_headroom() was setting incorrectly the nb_segs field >>> on the destination mbuf; >>> - Add unit-tests for dp_packet_equal() and (new) dp_packet_linear_data() to >>> patch 12/14; >>> - Add a comment to jumbo-frames.rst under topics/dpdk/ to warn how >>> multi-segments mbufs may affect performance when using large packets >>> across DPDK and non-DPDK ports. >>> >>> v7: - Rebase on master 024810c ("Prepare for post-2.10.0 (2.10.90)."); >>> - Add Ben's proposed fix for automake's warning; >>> - Add a note to cover letter to explain this is preperatory work for TSO / >>> GRO. >>> >>> v6: - Rebase on master d1b235d ("tests: Add test for ovn-nbctl's command parser >>> error paths."); >>> - Address Darrell's comments: >>> - The changes in dp_packet_resize__() were trying to alleviate the call >>> to OVS_NOT_REACHED() for DPDK packets, by trying to reuse the available >>> tailroom space when no more headroom space is available, and vice-versa. >>> However, this was breaking the API for the dp_packet_resize__() >>> function (only called in dp_packet_prealloc_tailroom() and >>> dp_packet_prealloc_headroom()), which doesn't seem to suit the purposes >>> for DPDK packets. >>> Instead, and because this is isolate funtionality, revert to the >>> previous state where dp_packet_resize__() is not supported for DPDK >>> packets. Hence, then patch 08/14 has been dropped. >>> - Additionally, fix the tests that were relying on the removed >>> functionality. >>> >>> v5: - Rebase on master 030958a0cc ("conntrack: Fix conn_update_state_alg use >>> after free."); >>> - Address Eelco's comments: >>> - Remove dpdk_mp_sweep() call in netdev_dpdk_mempool_configure(), a >>> leftover from rebase. Only call should be in dpdk_mp_get(); >>> - Remove NEWS line added by mistake during rebase (about adding >>> experimental vhost zero copy support). >>> - Address Ian's comments: >>> - Drop patch 01 from previous series entirely; >>> - Patch (now) 01/14 adds a new call to dpdk_buf_size() inside >>> dpdk_mp_create() to get the correct "mbuf_size" to be used; >>> - Patch (now) 11/14 modifies dpdk_mp_create() to check if multi-segment >>> mbufs is enabled, in which case it calculates the new "mbuf_size" to be >>> used; >>> - In free_dpdk_buf() and dpdk_buf_alloc(), don't lock and unlock >>> conditionally. >>> - Add "per-port-memory=true" to test "Multi-segment mbufs Tx" as the current >>> DPDK set up in system-dpdk-testsuite can't handle higher MTU sizes using >>> the shared mempool model (runs out of memory); >>> - Add new examples for when multi-segment mbufs are enabled in >>> topics/dpdk/memory.rst, and a reference to topics/dpdk/jumbo-frames.rst >>> (now patch 11/14). >>> >>> v4: - Rebase on master b22fb00 ("ovn-nbctl: Clarify error messages in qos-add >>> command."): >>> - A new patch (now 01/15) has been introduced to differentiate between >>> MTU and mbuf size when creating the mempool. This is because as part of >>> the new support for both per port and shared mempools, mempools were >>> being reused based on the mbuf size, and for multi-segment mbufs the >>> mbuf size can end up being the same for all mbufs; >>> - A couple of other patches (now 02/15 and 12/15) have been modified as >>> part of the rebase, but only to adapt to the code changes to "Support >>> both shared and per port mempools.", no functionality should have been >>> changed. >>> >>> v3: >>> - Address Eelco's comment: >>> - Fix the ovs_assert() introduced in v2 in __packet_set_data(), which >>> wasn't correctly asserting that the passed 'v' was smaller than the >>> first mbuf's buf_len. >>> >>> v2: >>> - Rebase on master e7cd8cf ("db-ctl-base: Don't die in cmd_destroy() on >>> error."); >>> - Address Eelco's comments: >>> - In free_dpdk_buf(), use mbuf's struct address in dp_packet instead of >>> casting; >>> - Remove unneeded variable in dp_packet_set_size(), pointing to the >>> first mbuf in the chain; >>> - Assert in dp_packet_set_size() to enforce that "pkt_len == v" is >>> always true for DPBUF_DPDK packets; >>> - Assert in __packet_set_data() to enforce that data_off never goes >>> beyond the first mbuf in the chain. >>> >>> v1: >>> - v8 should have been sent as v1 really, as that's usually the approach >>> followed in OvS. That clearly didn't happen, so restarting the series >>> now. This also helps making it clear it is no longer an RFC series; >>> - Rebase on master e461481 ("ofp-meter: Fix ofp_print_meter_flags() >>> output."); >>> - Address Eelco's comments: >>> - Change dp_packet_size() so that for `DPBUF_DPDK` packets their >>> `pkt_len` and `data_len` can't be set to values bigger than the >>> available space. Also fix assigment to `data_len` which was >>> incorrectly being set to just`pkt_len`; >>> - Improve `nonpmd_mp_mutex` comment with a better explanation as to >>> why the mutex is needed; >>> - Fix dp_packet_clear() to not call rte_pktmbuf_reset() for non >>> `DPBUF_DPDK` packets; >>> - Dropped `if` clause in dp_packet_l4_size(), keep just the `else`; >>> - Change dp_packet_clone_with_headroom() to use rte_pktmbuf_read() for >>> copying `DPBUF_DPDK` packets' data. Also, change it to return >>> appropriate and meaningful errors, instead of just "0" or "1"; >>> - Change dpdk_prep_tx_buf() name's to dpdk_clone_dp_packet_to_mbuf(), >>> and reuse dp_packet_mbuf_write() instead of manual copy; >>> - Add note vswitchd/vswitch.xml to make it clear the enabling of >>> dpdk-multi-seg-mbufs requires a restart; >>> - Change dpdk_mp_create() to increase # mbufs used under the multi-segment >>> mbufs approach; >>> - Increase test coverage by adding "end-to-end" tests that verify that >>> "dpdk-multi-seg-mbufs" is disabled by default and that a packet is >>> successfully sent out; >>> - Some helper funcs such as dp_packet_tail() and dp_packet_end() were moved >>> back to be common between `DPBUF_DPDK` and non `DPBUF_DPDK` packets, to >>> minimise changes; >>> - Add some extra notes to "Performance notes" in jumbo-frames.rst doc, >>> after further testing; >>> - Structure changes: >>> - Drop patch 07/13 which is now unneeded; >>> - Two more patches added for extra test coverage. This is what accounts >>> for the increase in size (+1 patch) in the series. >>> >>> v8 (non-RFC): >>> - Rebase on master 88125d6 ("rhel: remove ovs-sim man page from >>> temporary directory (also for RHEL)"); >>> - Address Ciara's and Ilya's comments: >>> - Drop the dp_packet_mbuf_tail() function and use only the >>> already existing dp_packet_tail(); >>> - Fix bug in dpdk_do_tx_copy() where the error return from >>> dpdk_prep_tx_buf() was being wrongly checked; >>> - Use dpdk_buf_alloc() and free_dpdk_buf() instead of >>> rte_pktmbuf_alloc() and rte_pktmbuf_free(); >>> - Fix some other code style and duplication issues pointed out. >>> - Refactored dp_packet_shift(), dp_packet_resize__() and >>> dp_packet_put*() functions to work within the bounds of existing >>> mbufs only; >>> - Fix dp_packet_clear() which wasn't correctly clearing / freeing >>> other mbufs in the chain for chains with more than a single mbuf; >>> - dp_packet layer functions (such as dp_packet_l3()) now check if >>> the header is within the first mbuf, when using mbufs; >>> - Move patch 08/13 to before patch 04/13, since dp_packet_set_size() >>> was refactored to use free_dpdk_buf(); >>> - Fix wrong rte_memcpy() when performing dp_packet_clone() which was >>> leading to memory corruption; >>> - Modified the added tests to account for some of the above changes; >>> - Run performance tests, compiling results and adding them to the >>> cover letter; >>> - Add a multi-seg mbufs explanation to the jumbo-frames.rst doc, >>> together with a "Performance notes" sub-section reflecting the >>> findings mentioned above in the cover letter. >>> >>> v7: - Rebase on master 5e720da ("erspan: fix invalid erspan version."); >>> - Address Ilya comments; >>> - Fix non-DPDK build; >>> - Serialise the access of non pmds to allocation and free of mbufs by >>> using a newly introduced mutex. >>> - Add a new set of tests that integrates with the recently added DPDK >>> testsuite. These focus on allocating dp_packets, with a single or >>> multiple mbufs, from an instantiated mempool and performing several >>> operations on those, verifying if the data at the end matches what's >>> expected; >>> - Fix bugs found by the new tests: >>> - dp_packet_shift() wasn't taking into account shift lefts; >>> - dp_packet_resize__() was misusing and miscalculating the tailrooms >>> and headrooms, ending up calculating the wrong number of segments >>> that needed allocation; >>> - An mbuf's end was being miscalculated both in dp_packet_tail, >>> dp_packet_mbuf_tail() and dp_packet_end(); >>> - dp_packet_set_size() was not updating the number of chained segments >>> 'nb_segs'; >>> - Add support for multi-seg mbufs in dp_packet_clear(). >>> >>> v6: - Rebase on master 7c0cb29 ("conntrack-tcp: Handle tcp session >>> reuse."); >>> - Further improve dp_packet_put_uninit() and dp_packet_shift() to >>> support multi-seg mbufs; >>> - Add support for multi-seg mbufs in dp_packet_l4_size() and >>> improve other helper funcs, such as dp_packet_tail() and dp_ >>> packet_tailroom(). >>> - Add support for multi-seg mbufs in dp_packet_put(), dp_packet_ >>> put_zeros(), as well as dp_packet_resize__() - allocating new >>> mbufs and linking them together; >>> Restructured patchset: >>> - Squash patch 5 into patch 6, since they were both related to >>> copying data while handling multi-seg mbufs; >>> - Split patch 4 into two separate patches - one that introduces the >>> changes in helper functions to deal with multi-seg mbufs and >>> two others for the shift() and put_uninit() functionality; >>> - Move patch 4 to before patch 3, so that ihelper functions come >>> before functionality improvement that rely on those helpers. >>> >>> v5: - Rebased on master e5e22dc ("datapath-windows: Prevent ct-counters >>> from getting redundantly incremented"); >>> - Sugesh's comments have been addressed: >>> - Changed dp_packet_set_data() and dp_packet_set_size() logic to >>> make them independent of each other; >>> - Dropped patch 3 now that dp_packet_set_data() and dp_packet_set_ >>> size() are independent; >>> - dp_packet_clone_with_headroom() now has split functions for >>> handling DPDK sourced packets and non-DPDK packets; >>> - Modified various functions in dp-packet.h to account for multi-seg >>> mbufs - dp_packet_put_uninit(), dp_packet_tail(), dp_packet_tail() >>> and dp_packet_at(); >>> - Added support for shifting packet data in multi-seg mbufs, using >>> dp_packet_shift(); >>> - Fixed some minor inconsistencies. >>> >>> Note that some of the changes in v5 have been contributed by Mark >>> Kavanagh as well. >>> >>> v4: - restructure patchset >>> - account for 128B ARM cacheline when sizing mbufs >>> >>> Mark Kavanagh (4): >>> netdev-dpdk: fix mbuf sizing >>> dp-packet: Init specific mbuf fields. >>> netdev-dpdk: copy large packet to multi-seg. mbufs >>> netdev-dpdk: support multi-segment jumbo frames >>> >>> Michael Qiu (1): >>> dp-packet: copy data from multi-seg. DPDK mbuf >>> >>> Tiago Lam (9): >>> dp-packet: Fix allocated size on DPDK init. >>> netdev-dpdk: Serialise non-pmds mbufs' alloc/free. >>> dp-packet: Fix data_len handling multi-seg mbufs. >>> dp-packet: Handle multi-seg mbufs in helper funcs. >>> dp-packet: Handle multi-seg mubfs in shift() func. >>> dp-packet: Add support for data "linearization". >>> dpdk-tests: Add unit-tests for multi-seg mbufs. >>> dpdk-tests: Accept other configs in OVS_DPDK_START >>> dpdk-tests: End-to-end tests for multi-seg mbufs. >>> >>> Documentation/topics/dpdk/jumbo-frames.rst | 67 ++++ >>> Documentation/topics/dpdk/memory.rst | 36 ++ >>> NEWS | 1 + >>> lib/bfd.c | 3 +- >>> lib/conntrack.c | 17 +- >>> lib/dp-packet.c | 192 ++++++++- >>> lib/dp-packet.h | 339 ++++++++++++++-- >>> lib/dpdk.c | 8 + >>> lib/dpif-netlink.c | 2 +- >>> lib/dpif.c | 2 +- >>> lib/netdev-bsd.c | 2 +- >>> lib/netdev-dpdk.c | 242 +++++++++-- >>> lib/netdev-dpdk.h | 2 + >>> lib/netdev-dummy.c | 5 +- >>> lib/netdev-linux.c | 5 +- >>> lib/netdev-native-tnl.c | 10 +- >>> lib/odp-execute.c | 2 +- >>> lib/ofp-print.c | 2 +- >>> lib/ovs-lldp.c | 3 +- >>> lib/packets.c | 3 +- >>> ofproto/ofproto-dpif-sflow.c | 2 +- >>> ofproto/ofproto-dpif-upcall.c | 2 +- >>> ofproto/ofproto-dpif-xlate.c | 12 +- >>> tests/automake.mk | 10 +- >>> tests/dpdk-packet-mbufs.at | 7 + >>> tests/system-dpdk-macros.at | 6 +- >>> tests/system-dpdk-testsuite.at | 1 + >>> tests/system-dpdk.at | 65 +++ >>> tests/test-dpdk-mbufs.c | 619 +++++++++++++++++++++++++++++ >>> vswitchd/vswitch.xml | 22 + >>> 30 files changed, 1571 insertions(+), 118 deletions(-) >>> create mode 100644 tests/dpdk-packet-mbufs.at >>> create mode 100644 tests/test-dpdk-mbufs.c >>> >> >>
Overview ======== This patchset introduces support for multi-segment mbufs to OvS-DPDK. Multi-segment mbufs are typically used when the size of an mbuf is insufficient to contain the entirety of a packet's data. Instead, the data is split across numerous mbufs, each carrying a portion, or 'segment', of the packet data. Mbufs are chained via their 'next' attribute (an mbuf pointer). The main motivation behind the support for multi-segment mbufs is to later introduce TSO (use case i. below) / GRO in OvS-DPDK, which is planned to be introduced after this series. Use Cases ========= i. Handling oversized (guest-originated) frames, which are marked for hardware accelration/offload (TSO, for example). Packets which originate from a non-DPDK source may be marked for offload; as such, they may be larger than the permitted ingress interface's MTU, and may be stored in an oversized dp-packet. In order to transmit such packets over a DPDK port, their contents must be copied to a DPDK mbuf (via dpdk_do_tx_copy). However, in its current implementation, that function only copies data into a single mbuf; if the space available in the mbuf is exhausted, but not all packet data has been copied, then it is lost. Similarly, when cloning a DPDK mbuf, it must be considered whether that mbuf contains multiple segments. Both issues are resolved within this patchset. ii. Handling jumbo frames. While OvS already supports jumbo frames, it does so by increasing mbuf size, such that the entirety of a jumbo frame may be handled in a single mbuf. This is certainly the preferred, and most performant approach (and remains the default). Enabling multi-segment mbufs ============================ Multi-segment and single-segment mbufs are mutually exclusive, and the user must decide on which approach to adopt on init. The introduction of a new OVSDB field, 'dpdk-multi-seg-mbufs', facilitates this. This is a global boolean value, which determines how jumbo frames are represented across all DPDK ports. In the absence of a user-supplied value, 'dpdk-multi-seg-mbufs' defaults to false, i.e. multi-segment mbufs must be explicitly enabled / single-segment mbufs remain the default. Setting the field is identical to setting existing DPDK-specific OVSDB fields: ovs-vsctl set Open_vSwitch . other_config:dpdk-init=true ovs-vsctl set Open_vSwitch . other_config:dpdk-lcore-mask=0x10 ovs-vsctl set Open_vSwitch . other_config:dpdk-socket-mem=4096,0 ==> ovs-vsctl set Open_vSwitch . other_config:dpdk-multi-seg-mbufs=true Performance notes (based on v8, 1st non-RFC) ================= In order to test for regressions in performance, tests were run on top of master 88125d6 and v8 of this patchset, both with the multi-segment mbufs option enabled and disabled. VSperf was used to run the phy2phy_cont and pvp_cont tests with varying packet sizes of 64B, 1500B and 7000B, on a 10Gbps interface. Test | Size | Master | Multi-seg disabled | Multi-seg enabled ------------------------------------------------------------- p2p | 64 | ~22.7 | ~22.65 | ~18.3 p2p | 1500 | ~1.6 | ~1.6 | ~1.6 p2p | 7000 | ~0.36 | ~0.36 | ~0.36 pvp | 64 | ~6.7 | ~6.7 | ~6.3 pvp | 1500 | ~1.6 | ~1.6 | ~1.6 pvp | 7000 | ~0.36 | ~0.36 | ~0.36 Packet size is in bytes, while all packet rates are reported in mpps (aggregated). No noticeable regression has been observed (certainly everything is within the ± 5% margin of existing performance), aside from the 64B packet size case when multi-segment mbuf is enabled. This is expected, however, because of how Tx vectoriszed functions are incompatible with multi-segment mbufs on some PMDs. The PMD under use during these tests was the i40e (on a Intel X710 NIC), which indeed doesn't support vectorized Tx functions with multi-segment mbufs. --- v9: - Rebase on master e4e2009 ("tunnel, tests: Sort flow output in ERSPAN v1/v2 metadata"); - Simplify patch 09/14. The functions introduced in packets.c were dropped so the code in netdev-native-tnl.c remains largely the same. These can be introduced at a later time, if needed (maybe when csum across segmented data is introduced); v8: - Rebase on master 1dd218a ("ovsdb-idl: Fix recently introduced Python 3 tests."); - Address Ian's comment: - Fix sparse warnings on patch 07/14 and 12/14 by allocating memory dynamically. - Address Ilya's comments: - netdev_linux_tap_batch_send() and udp_extract_tnl_md() now linearize the data before hand, beforing write()'ing or performing the checksums in the data; - Some other cases have been found and adapted; The new patch 09/14 introduced in the series is where the "linearization" logic is introduced and, as a consequence, some users of the dp_packet API, which were assuming the data is held contiguously in memory, are changed to use the new APIs. - Add support for multi-segment mbufs to dp_packet_equal() (patch 06/14); - Fix a bug in patch 08/14 where the call to dp_packet_copy_mbuf_flags() in dp_packet_clone_with_headroom() was setting incorrectly the nb_segs field on the destination mbuf; - Add unit-tests for dp_packet_equal() and (new) dp_packet_linear_data() to patch 12/14; - Add a comment to jumbo-frames.rst under topics/dpdk/ to warn how multi-segments mbufs may affect performance when using large packets across DPDK and non-DPDK ports. v7: - Rebase on master 024810c ("Prepare for post-2.10.0 (2.10.90)."); - Add Ben's proposed fix for automake's warning; - Add a note to cover letter to explain this is preperatory work for TSO / GRO. v6: - Rebase on master d1b235d ("tests: Add test for ovn-nbctl's command parser error paths."); - Address Darrell's comments: - The changes in dp_packet_resize__() were trying to alleviate the call to OVS_NOT_REACHED() for DPDK packets, by trying to reuse the available tailroom space when no more headroom space is available, and vice-versa. However, this was breaking the API for the dp_packet_resize__() function (only called in dp_packet_prealloc_tailroom() and dp_packet_prealloc_headroom()), which doesn't seem to suit the purposes for DPDK packets. Instead, and because this is isolate funtionality, revert to the previous state where dp_packet_resize__() is not supported for DPDK packets. Hence, then patch 08/14 has been dropped. - Additionally, fix the tests that were relying on the removed functionality. v5: - Rebase on master 030958a0cc ("conntrack: Fix conn_update_state_alg use after free."); - Address Eelco's comments: - Remove dpdk_mp_sweep() call in netdev_dpdk_mempool_configure(), a leftover from rebase. Only call should be in dpdk_mp_get(); - Remove NEWS line added by mistake during rebase (about adding experimental vhost zero copy support). - Address Ian's comments: - Drop patch 01 from previous series entirely; - Patch (now) 01/14 adds a new call to dpdk_buf_size() inside dpdk_mp_create() to get the correct "mbuf_size" to be used; - Patch (now) 11/14 modifies dpdk_mp_create() to check if multi-segment mbufs is enabled, in which case it calculates the new "mbuf_size" to be used; - In free_dpdk_buf() and dpdk_buf_alloc(), don't lock and unlock conditionally. - Add "per-port-memory=true" to test "Multi-segment mbufs Tx" as the current DPDK set up in system-dpdk-testsuite can't handle higher MTU sizes using the shared mempool model (runs out of memory); - Add new examples for when multi-segment mbufs are enabled in topics/dpdk/memory.rst, and a reference to topics/dpdk/jumbo-frames.rst (now patch 11/14). v4: - Rebase on master b22fb00 ("ovn-nbctl: Clarify error messages in qos-add command."): - A new patch (now 01/15) has been introduced to differentiate between MTU and mbuf size when creating the mempool. This is because as part of the new support for both per port and shared mempools, mempools were being reused based on the mbuf size, and for multi-segment mbufs the mbuf size can end up being the same for all mbufs; - A couple of other patches (now 02/15 and 12/15) have been modified as part of the rebase, but only to adapt to the code changes to "Support both shared and per port mempools.", no functionality should have been changed. v3: - Address Eelco's comment: - Fix the ovs_assert() introduced in v2 in __packet_set_data(), which wasn't correctly asserting that the passed 'v' was smaller than the first mbuf's buf_len. v2: - Rebase on master e7cd8cf ("db-ctl-base: Don't die in cmd_destroy() on error."); - Address Eelco's comments: - In free_dpdk_buf(), use mbuf's struct address in dp_packet instead of casting; - Remove unneeded variable in dp_packet_set_size(), pointing to the first mbuf in the chain; - Assert in dp_packet_set_size() to enforce that "pkt_len == v" is always true for DPBUF_DPDK packets; - Assert in __packet_set_data() to enforce that data_off never goes beyond the first mbuf in the chain. v1: - v8 should have been sent as v1 really, as that's usually the approach followed in OvS. That clearly didn't happen, so restarting the series now. This also helps making it clear it is no longer an RFC series; - Rebase on master e461481 ("ofp-meter: Fix ofp_print_meter_flags() output."); - Address Eelco's comments: - Change dp_packet_size() so that for `DPBUF_DPDK` packets their `pkt_len` and `data_len` can't be set to values bigger than the available space. Also fix assigment to `data_len` which was incorrectly being set to just`pkt_len`; - Improve `nonpmd_mp_mutex` comment with a better explanation as to why the mutex is needed; - Fix dp_packet_clear() to not call rte_pktmbuf_reset() for non `DPBUF_DPDK` packets; - Dropped `if` clause in dp_packet_l4_size(), keep just the `else`; - Change dp_packet_clone_with_headroom() to use rte_pktmbuf_read() for copying `DPBUF_DPDK` packets' data. Also, change it to return appropriate and meaningful errors, instead of just "0" or "1"; - Change dpdk_prep_tx_buf() name's to dpdk_clone_dp_packet_to_mbuf(), and reuse dp_packet_mbuf_write() instead of manual copy; - Add note vswitchd/vswitch.xml to make it clear the enabling of dpdk-multi-seg-mbufs requires a restart; - Change dpdk_mp_create() to increase # mbufs used under the multi-segment mbufs approach; - Increase test coverage by adding "end-to-end" tests that verify that "dpdk-multi-seg-mbufs" is disabled by default and that a packet is successfully sent out; - Some helper funcs such as dp_packet_tail() and dp_packet_end() were moved back to be common between `DPBUF_DPDK` and non `DPBUF_DPDK` packets, to minimise changes; - Add some extra notes to "Performance notes" in jumbo-frames.rst doc, after further testing; - Structure changes: - Drop patch 07/13 which is now unneeded; - Two more patches added for extra test coverage. This is what accounts for the increase in size (+1 patch) in the series. v8 (non-RFC): - Rebase on master 88125d6 ("rhel: remove ovs-sim man page from temporary directory (also for RHEL)"); - Address Ciara's and Ilya's comments: - Drop the dp_packet_mbuf_tail() function and use only the already existing dp_packet_tail(); - Fix bug in dpdk_do_tx_copy() where the error return from dpdk_prep_tx_buf() was being wrongly checked; - Use dpdk_buf_alloc() and free_dpdk_buf() instead of rte_pktmbuf_alloc() and rte_pktmbuf_free(); - Fix some other code style and duplication issues pointed out. - Refactored dp_packet_shift(), dp_packet_resize__() and dp_packet_put*() functions to work within the bounds of existing mbufs only; - Fix dp_packet_clear() which wasn't correctly clearing / freeing other mbufs in the chain for chains with more than a single mbuf; - dp_packet layer functions (such as dp_packet_l3()) now check if the header is within the first mbuf, when using mbufs; - Move patch 08/13 to before patch 04/13, since dp_packet_set_size() was refactored to use free_dpdk_buf(); - Fix wrong rte_memcpy() when performing dp_packet_clone() which was leading to memory corruption; - Modified the added tests to account for some of the above changes; - Run performance tests, compiling results and adding them to the cover letter; - Add a multi-seg mbufs explanation to the jumbo-frames.rst doc, together with a "Performance notes" sub-section reflecting the findings mentioned above in the cover letter. v7: - Rebase on master 5e720da ("erspan: fix invalid erspan version."); - Address Ilya comments; - Fix non-DPDK build; - Serialise the access of non pmds to allocation and free of mbufs by using a newly introduced mutex. - Add a new set of tests that integrates with the recently added DPDK testsuite. These focus on allocating dp_packets, with a single or multiple mbufs, from an instantiated mempool and performing several operations on those, verifying if the data at the end matches what's expected; - Fix bugs found by the new tests: - dp_packet_shift() wasn't taking into account shift lefts; - dp_packet_resize__() was misusing and miscalculating the tailrooms and headrooms, ending up calculating the wrong number of segments that needed allocation; - An mbuf's end was being miscalculated both in dp_packet_tail, dp_packet_mbuf_tail() and dp_packet_end(); - dp_packet_set_size() was not updating the number of chained segments 'nb_segs'; - Add support for multi-seg mbufs in dp_packet_clear(). v6: - Rebase on master 7c0cb29 ("conntrack-tcp: Handle tcp session reuse."); - Further improve dp_packet_put_uninit() and dp_packet_shift() to support multi-seg mbufs; - Add support for multi-seg mbufs in dp_packet_l4_size() and improve other helper funcs, such as dp_packet_tail() and dp_ packet_tailroom(). - Add support for multi-seg mbufs in dp_packet_put(), dp_packet_ put_zeros(), as well as dp_packet_resize__() - allocating new mbufs and linking them together; Restructured patchset: - Squash patch 5 into patch 6, since they were both related to copying data while handling multi-seg mbufs; - Split patch 4 into two separate patches - one that introduces the changes in helper functions to deal with multi-seg mbufs and two others for the shift() and put_uninit() functionality; - Move patch 4 to before patch 3, so that ihelper functions come before functionality improvement that rely on those helpers. v5: - Rebased on master e5e22dc ("datapath-windows: Prevent ct-counters from getting redundantly incremented"); - Sugesh's comments have been addressed: - Changed dp_packet_set_data() and dp_packet_set_size() logic to make them independent of each other; - Dropped patch 3 now that dp_packet_set_data() and dp_packet_set_ size() are independent; - dp_packet_clone_with_headroom() now has split functions for handling DPDK sourced packets and non-DPDK packets; - Modified various functions in dp-packet.h to account for multi-seg mbufs - dp_packet_put_uninit(), dp_packet_tail(), dp_packet_tail() and dp_packet_at(); - Added support for shifting packet data in multi-seg mbufs, using dp_packet_shift(); - Fixed some minor inconsistencies. Note that some of the changes in v5 have been contributed by Mark Kavanagh as well. v4: - restructure patchset - account for 128B ARM cacheline when sizing mbufs Mark Kavanagh (4): netdev-dpdk: fix mbuf sizing dp-packet: Init specific mbuf fields. netdev-dpdk: copy large packet to multi-seg. mbufs netdev-dpdk: support multi-segment jumbo frames Michael Qiu (1): dp-packet: copy data from multi-seg. DPDK mbuf Tiago Lam (9): dp-packet: Fix allocated size on DPDK init. netdev-dpdk: Serialise non-pmds mbufs' alloc/free. dp-packet: Fix data_len handling multi-seg mbufs. dp-packet: Handle multi-seg mbufs in helper funcs. dp-packet: Handle multi-seg mubfs in shift() func. dp-packet: Add support for data "linearization". dpdk-tests: Add unit-tests for multi-seg mbufs. dpdk-tests: Accept other configs in OVS_DPDK_START dpdk-tests: End-to-end tests for multi-seg mbufs. Documentation/topics/dpdk/jumbo-frames.rst | 67 ++++ Documentation/topics/dpdk/memory.rst | 36 ++ NEWS | 1 + lib/bfd.c | 3 +- lib/conntrack.c | 17 +- lib/dp-packet.c | 192 ++++++++- lib/dp-packet.h | 339 ++++++++++++++-- lib/dpdk.c | 8 + lib/dpif-netlink.c | 2 +- lib/dpif.c | 2 +- lib/netdev-bsd.c | 2 +- lib/netdev-dpdk.c | 242 +++++++++-- lib/netdev-dpdk.h | 2 + lib/netdev-dummy.c | 5 +- lib/netdev-linux.c | 5 +- lib/netdev-native-tnl.c | 10 +- lib/odp-execute.c | 2 +- lib/ofp-print.c | 2 +- lib/ovs-lldp.c | 3 +- lib/packets.c | 3 +- ofproto/ofproto-dpif-sflow.c | 2 +- ofproto/ofproto-dpif-upcall.c | 2 +- ofproto/ofproto-dpif-xlate.c | 12 +- tests/automake.mk | 10 +- tests/dpdk-packet-mbufs.at | 7 + tests/system-dpdk-macros.at | 6 +- tests/system-dpdk-testsuite.at | 1 + tests/system-dpdk.at | 65 +++ tests/test-dpdk-mbufs.c | 619 +++++++++++++++++++++++++++++ vswitchd/vswitch.xml | 22 + 30 files changed, 1571 insertions(+), 118 deletions(-) create mode 100644 tests/dpdk-packet-mbufs.at create mode 100644 tests/test-dpdk-mbufs.c