From patchwork Fri Jul 6 14:18:07 2018 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Qi Zhang X-Patchwork-Id: 42516 Return-Path: X-Original-To: patchwork@dpdk.org Delivered-To: patchwork@dpdk.org Received: from [92.243.14.124] (localhost [127.0.0.1]) by dpdk.org (Postfix) with ESMTP id B337E1BE90; Fri, 6 Jul 2018 16:18:20 +0200 (CEST) Received: from mga07.intel.com (mga07.intel.com [134.134.136.100]) by dpdk.org (Postfix) with ESMTP id BB69C1BE8B for ; Fri, 6 Jul 2018 16:18:17 +0200 (CEST) X-Amp-Result: SKIPPED(no attachment in message) X-Amp-File-Uploaded: False Received: from fmsmga006.fm.intel.com ([10.253.24.20]) by orsmga105.jf.intel.com with ESMTP/TLS/DHE-RSA-AES256-GCM-SHA384; 06 Jul 2018 07:18:15 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.51,316,1526367600"; d="scan'208";a="243511280" Received: from dpdk51.sh.intel.com ([10.67.110.190]) by fmsmga006.fm.intel.com with ESMTP; 06 Jul 2018 07:17:49 -0700 From: Qi Zhang To: thomas@monjalon.net, anatoly.burakov@intel.com Cc: konstantin.ananyev@intel.com, dev@dpdk.org, bruce.richardson@intel.com, ferruh.yigit@intel.com, benjamin.h.shelton@intel.com, narender.vangati@intel.com, Qi Zhang Date: Fri, 6 Jul 2018 22:18:07 +0800 Message-Id: <20180706141826.204044-1-qi.z.zhang@intel.com> X-Mailer: git-send-email 2.13.6 In-Reply-To: <20180607123849.14439-1-qi.z.zhang@intel.com> References: <20180607123849.14439-1-qi.z.zhang@intel.com> Subject: [dpdk-dev] [PATCH v9 00/19] enable hotplug on multi-process X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" v9: - Move hotplug IPC from rte_eth_dev_attach/rte_eth_dev_detach to eal_dev_hotplug_add and eal_dev_hotplug_remove, now all kinds of devices will be synced in multi-process. - Fix couple issue when a device is bound to vfio. 1) The device can't be detached clearly in a secondary process, which also cause it can't be attached again, due to the error that /dev/vfio/ is still busy.(see Patch 3/19 and 4/19) 2) repeat detach/attach device will cause "cannot find TAILQ entry for PCI device" due to incorrect PCI address compare. (see patch 2/19). - Removed device lock. - Removed private device support. - Fix commit log grammar issue v8: - update rte_eal_version.map due to new API added. - minor reword on release note. - minor fix on commit log and code style. NOTE: Some issues which is not related with this patchset is expected when play with hotplug_mp sample as belows. - Attach a PCI device twice may cause device can't be detached below fix is required: https://patches.dpdk.org/patch/42030/ - ixgbe device can't detached, below fix is required https://patches.dpdk.org/patch/42031/ v7: - update rte_ethdev_version.map for new APIs. - improve code readability in __handle_secondary_request by use goto. - add comments to explain why need to call rte_eal_alarm_set. - add error log when process_mp_init_callbacks failed. - reword release notes base on Anatoly's suggestion. - add back previous "Acked-by" and "Reviewed-by" in commit log. NOTE: current patchset depends on below IPC fix, or it may not be able to attach a shared vdev. https://patches.dpdk.org/patch/41647/ v6: - remove bus->scan_one, since ABI break is not necessary. - remove patch for failsafe PMD since it will not support secondary. - fix wrong implemenation on ixgbe. - add rte_eth_dev_release_port_private into rte_eth_dev_pci_generic_remove for secondary process, so we don't need to patch on PMD if PMD use the default remove function. - add release notes update. - agreed to use strdup(peer) as workaround for repling a sync request in seperate thread. v5: - since we will keep mp thread separate from interrupt thread, it is not necessary to use temporary thread, we use rte_eal_alarm_set. - remove the change in rte_eth_dev_release_port, since there is a better way to prevent rte_eth_dev_release_port be called after rte_eth_dev_release_port_private. - fix the issue that lock does not take effect on secondary due to previous re-work - fix the issue when the first attached device is a private device from secondary. (patch 8/24) - work around for reply a sync request in separate thread, this is still an open and in discussion as below. https://mails.dpdk.org/archives/dev/2018-June/105359.html v4: - since mp thread will be merged to interrupt thread, the fix on v3 for sync IPC deadlock will not work. the new version enable the machanism to invoke a mp action callback in a temporary thread to avoid the IPC deadlock, with this, secondary to primary request impelemtation also be simplified, since we can use sync request directly in a separate thread. v3: - enable mp init callback register to help non-eal module to initialize mp channel during rte_eal_init - fix when attach share device from secondary. 1) dead lock due to sync IPC be invoked in rte_malloc in primary process when handle secondary request to attach device, the solution is primary process to issue share device attach/detach in interrupt thread. 2) return port_id not correct. - check nb_sent and nb_received in sync IPC. - fix memory leak duirng error handling at attach_on_secondary. - improve clean_lock_callback to only lock/unlock spinlock once - improve error code return in check-reply during async IPC. - remove rte_ prefix of internal function in ethdev_mp.c - sample code improvement. 1) rename sample to "hotplug_mp", and move to example/multi-process. 2) cleanup header include. 3) call rte_eal_cleanup before exit. v2: - rename rte_ethdev_mp.* to ethdev_mp.* - rename rte_ethdev_lock.* to ethdev_lock.* - move internal funciton to ethdev_private.h - separate rte_eth_dev_[un]lock into rte_eth_dev_[un]lock and rte_eth_dev_[un]lock_with_callback - lock callbacks will be removed automatically after device is detached. - add experimental tag for all new APIs. - fix coding style issue. - fix wrong lisence header in sample code. - fix spelling - fix meson.build. - improve comments. Background: =========== Currently secondary process will only sync ethdev from primary process at init stage, but it will not be aware if device is attached/detached on primary process at runtime. While there is the requirement from application that take primary-secondary process model. The primary process work as a resource management process, it will create/destroy virtual device at runtime, while the secondary process deal with the network stuff with these devices. Solution: ========= So the orignial intention is to fix this gap, but beyond that the patch set provide a more comprehesive solution to handle different hotplug cases in multi-process situation, it cover below scenario: 1. Attach a device from the primary 2. Detach a device from the primary 3. Attach a device from a secondary 4. Detach a device from a secondary In primary-secondary process model, we assume ethernet devices are shared by default. that means attach or detach a device on any process will broadcast to all other processes through mp channel then device information will be synchronized on all processes. Any failure during attaching or detaching process will cause inconsistent status between processes, so proper rollback action should be considered. Scenario for Case 1, 2: attach device from primary a) primary attach the new device if failed goto h). b) primary send attach sync request to all secondary. c) secondary receive request and attach device and send reply. d) primary check the reply if all success go to i). e) primary send attach rollback sync request to all secondary. f) secondary receive the request and detach device and send reply. g) primary receive the reply and detach device as rollback action. h) attach fail i) attach success detach device from primary a) primary perform pre-detach check, if device is locked, goto i). b) primary send pre-detach sync request to all secondary. c) secondary perform pre-detach check and send reply. d) primary check the reply if any fail goto i). e) primary send detach sync request to all secondary f) secondary detach the device and send reply (assume no fail) g) primary detach the device. h) detach success i) detach failed Scenario for case 3, 4: attach device from secondary: a) seconary send asycn request to primary and wait on a condition which will be released by matched response from primary. b) primary receive the request and attach the new device if failed goto i). c) primary forward attach request to all secondary as async request (because this in mp thread context, use sync request will deadlock, same reason for all following async request.) d) secondary receive request and attach device and send reply. e) primary check the reply if all success go to j). f) primary send attach rollback async request to all secondary. g) secondary receive the request and detach device and send reply. h) primary receive the reply and detach device as rollback action. i) send fail response to secondary, goto k). j) send success response to secondary. k) secondary process receive response and return. detach device from secondary: a) secondary send async request to primary and wait on a condition which will be released by matched response from primary. b) primary receive the request and perform pre-detach check, if device is locked, goto j). c) primary send pre-detach async request to all secondary. d) secondary perform pre-detach check and send reply. e) primary check the reply if any fail goto j). f) primary send detach async request to all secondary g) secondary detach the device and send reply h) primary detach the device. i) send success response to secondary, goto k). j) send fail response to secondary. k) secondary process receive response and return. APIs chenages: ============== scope of rte_eal_hotplug_add and rte_eal_hotplug_remove is extended. In primary-secondary process model, rte_eal_hotplug_add will guarantee that device be attached on all processes, while rte_eal_hotplug_remove will guarantee device be detached on all processes. PMD Impact: =========== Currently device removing is not handled well in secondary process on most pmd drivers, rte_eth_dev_relase_port will be invoked and will mess up primary process since it reset all shared data. So we introduced new API rte_eth_dev_release_port_local which only reset ethdev's state to unsued but not touch shared data so other process will not be impacted. Since not all device driver is target to support primary-secondary process model, so the patch set only fix this on all Intel devices and vdev, it can be refereneced by other driver when equevalent fix is required Example: ======== The patchset also contains a example to demonstrate device hotplug in multi-process model, below are detail instructions. /* start sample code as primary then secondary */ ./hotplug_mp --proc-type=auto Command Line Example: >help >list /* attach a af_packet vdev */ >attach net_af_packet,iface=eth0 /* detach port 0 */ >detach 0 Qi Zhang (19): ethdev: add function to release port in local process bus/pci: fix PCI address compare bus/pci: enable vfio unmap resource for secondary vfio: remove uneccessary IPC for group fd clear eal: enable hotplug on multi-process eal: support attach or detach share device from secondary net/i40e: enable port detach on secondary process net/ixgbe: enable port detach on secondary process net/af_packet: enable port detach on secondary process net/bonding: enable port detach on secondary process net/kni: enable port detach on secondary process net/null: enable port detach on secondary process net/octeontx: enable port detach on secondary process net/pcap: enable port detach on secondary process net/softnic: enable port detach on secondary process net/tap: enable port detach on secondary process net/vhost: enable port detach on secondary process examples/multi_process: add hotplug sample doc: update release notes for multi process hotplug doc/guides/rel_notes/release_18_08.rst | 11 + drivers/bus/pci/linux/pci_vfio.c | 88 ++++++- drivers/net/af_packet/rte_eth_af_packet.c | 11 + drivers/net/bonding/rte_eth_bond_pmd.c | 11 + drivers/net/i40e/i40e_ethdev.c | 2 + drivers/net/ixgbe/ixgbe_ethdev.c | 3 + drivers/net/kni/rte_eth_kni.c | 11 + drivers/net/null/rte_eth_null.c | 16 +- drivers/net/octeontx/octeontx_ethdev.c | 16 ++ drivers/net/pcap/rte_eth_pcap.c | 15 +- drivers/net/softnic/rte_eth_softnic.c | 19 +- drivers/net/tap/rte_eth_tap.c | 17 +- drivers/net/vhost/rte_eth_vhost.c | 11 + examples/multi_process/Makefile | 1 + examples/multi_process/hotplug_mp/Makefile | 23 ++ examples/multi_process/hotplug_mp/commands.c | 197 ++++++++++++++ examples/multi_process/hotplug_mp/commands.h | 10 + examples/multi_process/hotplug_mp/main.c | 41 +++ lib/librte_eal/bsdapp/eal/Makefile | 1 + lib/librte_eal/common/eal_common_dev.c | 168 +++++++++++- lib/librte_eal/common/eal_private.h | 37 +++ lib/librte_eal/common/hotplug_mp.c | 346 +++++++++++++++++++++++++ lib/librte_eal/common/hotplug_mp.h | 44 ++++ lib/librte_eal/common/include/rte_bus.h | 3 + lib/librte_eal/common/include/rte_dev.h | 9 + lib/librte_eal/common/meson.build | 1 + lib/librte_eal/linuxapp/eal/Makefile | 1 + lib/librte_eal/linuxapp/eal/eal.c | 6 + lib/librte_eal/linuxapp/eal/eal_vfio.c | 45 +--- lib/librte_eal/linuxapp/eal/eal_vfio.h | 1 - lib/librte_eal/linuxapp/eal/eal_vfio_mp_sync.c | 8 - lib/librte_ethdev/rte_ethdev.c | 12 + lib/librte_ethdev/rte_ethdev_driver.h | 16 +- lib/librte_ethdev/rte_ethdev_pci.h | 8 + 34 files changed, 1147 insertions(+), 62 deletions(-) create mode 100644 examples/multi_process/hotplug_mp/Makefile create mode 100644 examples/multi_process/hotplug_mp/commands.c create mode 100644 examples/multi_process/hotplug_mp/commands.h create mode 100644 examples/multi_process/hotplug_mp/main.c create mode 100644 lib/librte_eal/common/hotplug_mp.c create mode 100644 lib/librte_eal/common/hotplug_mp.h