{"payload":{"feedbackUrl":"https://github.com/orgs/community/discussions/53140","repo":{"id":68788712,"defaultBranch":"master","name":"rdma-core","ownerLogin":"yishaih","currentUserCanPush":false,"isFork":true,"isEmpty":false,"createdAt":"2016-09-21T06:50:55.000Z","ownerAvatar":"https://avatars.githubusercontent.com/u/4038053?v=4","public":true,"private":false,"isOrgOwned":false},"refInfo":{"name":"","listCacheKey":"v0:1626766025.4446878","currentOid":""},"activityList":{"items":[{"before":"b1ec2acc561afd92a20b4eff9d4806b650c21ebe","after":"c4a5ac8bba611206e062c0955fb605bfc0f48b0f","ref":"refs/heads/rdmacm_misc","pushedAt":"2024-05-05T08:41:22.000Z","pushType":"push","commitsCount":4428,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"librdmacm: Fix an overflow bug in qsort comparison function\n\nThe comparison function dev_cmp() doesn't work with 64b pointers in some\ncases, as it casts the pointer to int. For example it's not able to sort\nthis list:\n  {0xfffe0c2f0b00, 0xaaac741b4a90, 0xaaac741b4d70}\n\nFixes: e5d371cb0af0 (\"librdmacm: Globally store and sort IB device list\")\nSigned-off-by: Mark Zhang <markzhang@nvidia.com>\nReviewed-by: Leon Romanovsky <leonro@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"librdmacm: Fix an overflow bug in qsort comparison function"}},{"before":"e692a549428d5f8253955acc0597b5d3160d9d5c","after":"cebd491f520fbed90b19b01aad280bc43338ebaa","ref":"refs/heads/mlx5_dr","pushedAt":"2024-03-26T12:07:40.000Z","pushType":"push","commitsCount":64,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, Using sq ts format when RoCE is disabled\n\nWhen creating a QP in SW steering, one of the attributes is TS format\n(timestamp). In some devices, we have a limitation that all QPs should\nhave the same ts_format.\n\nThe qp_ts_format cap resides under the RoCE caps table, and the cap is\nnot queried when RoCE is disabled.\nConsider the case when the system is configured to run with REAL_TIME_TS\nformat (0x2), some QPs are created, and we disable RoCE.\nWhen trying to construct a new QP, we can't use the qp_ts_format, that\nis queried from the RoCE caps table.\nThat leads to passing 0x0 (FREE_RUNNING_TS) as the value of the\nqp_ts_format, which is different than the ts_format of the previous QPs\nREAL_TIME_TS format (0x2) and the creation will fail.\n\nTo resolve this, we will read the sq_ts_format, which also reflect the\nsupported ts format for the QP when RoCE is disabled, and override it\nwith qp_ts_format from the RoCE caps table if RoCE is enabled.\n\nFixes: ed504ab47384 (\"mlx5: DR, Set QP timestamp mode to default\")\nSigned-off-by: Itamar Gozlan <igozlan@nvidia.com>\nSigned-off-by: Alex Vesker <valex@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, Using sq ts format when RoCE is disabled"}},{"before":"50122fbe13850ad86ee732b182a686e62077033b","after":"e692a549428d5f8253955acc0597b5d3160d9d5c","ref":"refs/heads/mlx5_dr","pushedAt":"2024-01-11T07:55:34.000Z","pushType":"push","commitsCount":150,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, Fix ASO CT action applying in cross domain\n\nFix ASO CT action applying in cross domain, instead of checking if the\nGVMI's of the domains are different, check if the domains are different,\nsince we can have both domains on the same GVMI.\nThis fix will allow two different domains on the same GVMI to share the\nsame ASO CT action.\n\nFixes: 730329f392f3 (\"mlx5: DR, Support ASO CT cross domain\")\nSigned-off-by: Hamdan Igbaria <hamdani@nvidia.com>\nReviewed-by: Erez Shitrit <erezsh@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, Fix ASO CT action applying in cross domain"}},{"before":"4192b0ef5c5cee57219a0615766a9db981c2321a","after":"5afb79f9d86fd4ed7ff789515f021df2b056d3cf","ref":"refs/heads/mlx5_misc","pushedAt":"2023-12-31T14:35:28.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, Can't go to uplink vport on RX rule\n\nGo-To-Vport action on RX is not allowed when the vport is uplink.  If\nthis happens on RX domain, the rule insertion fails, as the rule is\nillegal.\nIf this happens on RX side of the FDB domain, then the TX part of the\nrule is legal, and the RX part will be replaced with drop.\n\nFixes: 828351a7da47 (\"mlx5: Expose steering action functionality\")\nSigned-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>\nReviewed-by: Erez Shitrit <erezsh@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, Can't go to uplink vport on RX rule"}},{"before":"5b61a5af4acb87c55f8a2e7af41f42adfd8587bb","after":"4192b0ef5c5cee57219a0615766a9db981c2321a","ref":"refs/heads/mlx5_misc","pushedAt":"2023-12-28T13:35:20.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Use correct value of max WQE size for DC\n\nThe max WQE size for DCI might be different than the max SQ WQE of an RC\nQP. So, in case we have the DC cap over DEVX, simply use it instead of\nthe QP one.\n\nThis change enables an application getting an early error in the control\npath rather than delaying the error to the data path, in case its input\nfor the DCI QP creation is not valid.\n\nSigned-off-by: Or Har-Toov <ohartoov@nvidia.com>\nReviewed-by: Michael Guralnik <michaelgur@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Use correct value of max WQE size for DC"}},{"before":"c63abc09fe0f0d2c94d01e6bffa493d71fa86a0c","after":"5b61a5af4acb87c55f8a2e7af41f42adfd8587bb","ref":"refs/heads/mlx5_misc","pushedAt":"2023-12-28T12:13:15.000Z","pushType":"push","commitsCount":19,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Use correct value of max WQE size for DC\n\nMax WQE size is the max descriptor size in bytes for the SQ of DC type.\nSQ WQEs must not be bigger than this value. We check this constrain and\nfail it on control path if it is not happening.\nSo far we used RC value for DC but we have different RC and DC\nlimitations, so checking the RC cap on DC caused a failure on data path\ninstead of control path.\n\nFixes: b9967a9d722a (\"mlx5: Create DC transport QPs\")\nSigned-off-by: Or Har-Toov <ohartoov@nvidia.com>\nReviewed-by: Michael Guralnik <michaelgur@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Use correct value of max WQE size for DC"}},{"before":"93d83000546af7493979a98caf829a43c3a94f6c","after":"c63abc09fe0f0d2c94d01e6bffa493d71fa86a0c","ref":"refs/heads/mlx5_misc","pushedAt":"2023-12-14T13:48:54.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Provide reg C0 value for matching egress traffic\n\nWhen provided the value can be used to match egress traffic from the\nlocal vport in the FDB.\n\nSigned-off-by: Mark Bloch <mbloch@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Provide reg C0 value for matching egress traffic"}},{"before":"27257eb5cb6f8ff5e14d974f4a7a82f56cba829a","after":"93d83000546af7493979a98caf829a43c3a94f6c","ref":"refs/heads/mlx5_misc","pushedAt":"2023-11-28T15:34:30.000Z","pushType":"push","commitsCount":27,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Provide reg C0 value for matching egress traffic\n\nWhen provided the value can be used to match egress traffic from the\nlocal vport in the FDB.\n\nSigned-off-by: Mark Bloch <mbloch@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Provide reg C0 value for matching egress traffic"}},{"before":"4827e3388baf96f2bf630246567ab40467b7352c","after":"27257eb5cb6f8ff5e14d974f4a7a82f56cba829a","ref":"refs/heads/mlx5_misc","pushedAt":"2023-11-05T12:11:06.000Z","pushType":"push","commitsCount":75,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Fix the rdma_tracepoint() code\n\nAs part of adding rdma_tracepoint() to _mlx5_post_send() the total\nlength of the SGs is calculated unconditionally even if the trace point\nmechanism is compiled out.\n\nThe above extra code in the data path must be prevented.\n\nAs such, the trace point was changed to include some exiting data as of\nnum_sge.\n\nIn addition, drop the dummy trace point function and have an empty\ndefinition for rdma_tracepoint as was done for rxe and the efa drivers.\n\nFixes: ec4b65b3541a (\"mlx5: Add tracepoint for post send\")\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Fix the rdma_tracepoint() code"}},{"before":"62de4fdfa9444ca0088930467deccaa3b66c5129","after":"91b4ed8dd8460b9c547bde47c4ae6c57f73b61af","ref":"refs/heads/verbs_misc","pushedAt":"2023-09-27T10:24:08.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"verbs: Add support for XDR speed in ibv_query_port\n\nPreviously we had 8 bit field to represent the link speed in query_port,\nso speeds that were represented by the 9th bit couldn't be reported.\n\nAdd new u32 field to be able to report speeds that are in the 9th bit or\nbeyond. The old active speed field still reports the speed correctly, in\ncase the speed is lower than XDR, reports NDR otherwise.\n\nIn addition, speed_str() of ibv_devinfo was updated to consider the new\nXDR rate, and read the rate from active_speed_ex when valid.\n\nThe ibv_query_port() man page was updated with the expected usage of\nthose two fields.\n\nSigned-off-by: Or Har-Toov <ohartoov@nvidia.com>\nReviewed-by: Patrisious Haddad <phaddad@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"verbs: Add support for XDR speed in ibv_query_port"}},{"before":"c45290abe1cb4565c547806c1b8cc91955bbdbe2","after":"62de4fdfa9444ca0088930467deccaa3b66c5129","ref":"refs/heads/verbs_misc","pushedAt":"2023-09-20T15:21:47.000Z","pushType":"push","commitsCount":643,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"verbs: Add support for XDR speed in ibv_query_port\n\nPreviously we had 8 bit field to represent the link speed in query_port,\nso speeds that were represented by the 9th bit couldn't be reported.\n\nAdd new u32 field to be able to report speeds that are in the 9th bit or\nbeyond. The old active speed field still reports the speed correctly, in\ncase the speed is lower than XDR, reports NDR otherwise.\n\nIn addition, speed_str() of ibv_devinfo was updated to consider the new\nXDR rate, and read the rate from active_speed_ex when valid.\n\nThe ibv_query_port() man page was updated with the expected usage of\nthose two fields.\n\nSigned-off-by: Or Har-Toov <ohartoov@nvidia.com>\nReviewed-by: Patrisious Haddad <phaddad@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"verbs: Add support for XDR speed in ibv_query_port"}},{"before":"89d59d81012bc1b48c95eb95906c13fdedab9e13","after":"4827e3388baf96f2bf630246567ab40467b7352c","ref":"refs/heads/mlx5_misc","pushedAt":"2023-08-24T08:49:46.000Z","pushType":"push","commitsCount":77,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Enhance SRQ performance\n\nEnhance SRQ performance by revising the method of advancing to the next\nWQE. Unlike the previous approach, which relied on the first freed WQE\nas the next post receive target, the new algorithm locates the nearest\navailable WQE (or contiguous chunk of WQEs if available) to the tail\nindex.\n\nA WQE is considered to be available when it's under SW ownership and is\nnot part of the SRQ queues (main/wait queue), in other words, it's not\nincluded in the linked lists extending from the queues' head to tail.\n\nIn the context of ODP, where a page-faulted WQE requires Kernel\nintervention to resolve and handle its addresses, a wait queue is used\nas a cooldown mechanism. This mechanism places such WQEs in a wait queue\n(or at the end of the main queue when a wait queue is unavailable),\ndelaying potential post-receives until they are properly handled.\n\nConsequently, in the presence of a page-faulted WQE, the SRQ LL is\npopulated with all available WQEs, regardless of their contiguity,\nreducing the risk of inadvertently overwriting a page-faulted WQE with a\nfuture post-SRQ receive operation.\n\nHere are some results from perftest traffic with max-wr size of 256 and\nmessage size of 4KB (1KB as the max MTU on UD) on a ConnectX-6 Dx device\n(results in Gb/sec):\n+------+---------+-----------+----------+----------------+\n| #QPs | QP Type |  W/O SRQ  | Prev SRQ |    New SRQ     |\n+------+---------+-----------+----------+----------------+\n|  10  |    UD   |   55.88   |   15.20  |  55.71 (+267%) |\n|  10  |    RC   |   92.58   |   60.67  |  79.85 (+31%)  |\n|   5  |    RC   |   92.57   |   58.08  |  86.24 (+48%)  |\n|   3  |    RC   |   92.57   |   59.33  |  91.94 (+55%)  |\n+------+---------+-----------+----------+----------------+\n\nSigned-off-by: Edward Srouji <edwards@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Enhance SRQ performance"}},{"before":"5e221817dc1e27f448f4c0c9181eb5ff2581c182","after":"50122fbe13850ad86ee732b182a686e62077033b","ref":"refs/heads/mlx5_dr","pushedAt":"2023-07-19T15:58:53.000Z","pushType":"push","commitsCount":82,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, Optimize args sending on QP while in multi QP\n\nIn multi QP delay the write of the args to the rule creating\nphase, on the same QP of the rule using the action, this way\nwe will save writing of the args on all the QPs when creating\nit while the multi QP feature is enabled.\n\nSigned-off-by: Hamdan Igbaria <hamdani@nvidia.com>\nSigned-off-by: Alex Vesker <valex@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, Optimize args sending on QP while in multi QP"}},{"before":"60541727f76bf7e8c044e30856dbfeb17bae66a1","after":"3ec6fb5c89676e823fef9b874ebc2184196eed42","ref":"refs/heads/mlx5_vfio","pushedAt":"2023-06-26T09:11:04.644Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Refactor how mlx5_vfio_reg_mr calculate the iova page size\n\nCurrently for some virtual addresses mlx5_vfio_reg_mr returns very large\niova page size, which may not be supported. For example vaddr\n0x7f7fff9a8000 and length 0x3c32000 gets iova page size (1ULL << 40),\nwhile the default IOVA range is (1ULL << 39). In this case \"ENOSPC\"\nwould be returned.\n\nThis patch refactors the way how to calculate iova page size, to make\nsure it won't waste more than 1G of IOVA address space.\n\nSigned-off-by: Mark Zhang <markzhang@nvidia.com>\nReviewed-by: Jason Gunthorpe <jgg@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Refactor how mlx5_vfio_reg_mr calculate the iova page size"}},{"before":"6b5a79a597cec4a6b7691225c56c3c5cf30cbc9a","after":"60541727f76bf7e8c044e30856dbfeb17bae66a1","ref":"refs/heads/mlx5_vfio","pushedAt":"2023-06-26T08:57:04.819Z","pushType":"push","commitsCount":847,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Refactor how mlx5_vfio_reg_mr calculate the iova page size\n\nCurrently for some virtual addresses mlx5_vfio_reg_mr returns very large\niova page size, which may not be supported. For example vaddr\n0x7f7fff9a8000 and length 0x3c32000 gets iova page size (1ULL << 40),\nwhile the default IOVA range is (1ULL << 39). In this case \"ENOSPC\"\nwould be returned.\n\nThis patch refactors the way how to calculate iova page size, to make\nsure it won't waste more then 1G of IOVA address space.\n\nSigned-off-by: Mark Zhang <markzhang@nvidia.com>\nReviewed-by: Jason Gunthorpe <jgg@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Refactor how mlx5_vfio_reg_mr calculate the iova page size"}},{"before":"a09e57981e97242f988bd7bea9fa9c7677ba7a59","after":"89d59d81012bc1b48c95eb95906c13fdedab9e13","ref":"refs/heads/mlx5_misc","pushedAt":"2023-06-25T15:23:11.423Z","pushType":"push","commitsCount":82,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"tests: Test atomic traffic on DC\n\nWhile testing atomic traffic, max_dc_read_atom value is checked by\nposting maximum number of WR possible.\n\nSigned-off-by: Or Har-Toov <ohartoov@nvidia.com>\nSigned-off-by: Edward Srouji <edwards@nvidia.com>","shortMessageHtmlLink":"tests: Test atomic traffic on DC"}},{"before":"79a8cc2511dba94ae09aa33850dfbb3175a4c481","after":"5e221817dc1e27f448f4c0c9181eb5ff2581c182","ref":"refs/heads/mlx5_dr","pushedAt":"2023-05-24T06:53:56.095Z","pushType":"push","commitsCount":46,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, New ECPF checking\n\nThere's a new dedicated capability field indicating port number of\nE-Switch manager.\nThis patch queries that value and checks if current E-Switch manager is\nECPF or not. If the value is invalid, fallback to old way for ECPF\nchecking.\n\nSigned-off-by: Shun Hao <shunh@nvidia.com>\nSigned-off-by: Hamdan Igbaria <hamdani@nvidia.com>\nSigned-off-by: Alex Vesker <valex@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, New ECPF checking"}},{"before":"f3ae9a97ec028ceb6c5f862ef2974fcd8e29c63a","after":"79a8cc2511dba94ae09aa33850dfbb3175a4c481","ref":"refs/heads/mlx5_dr","pushedAt":"2023-04-13T14:06:20.000Z","pushType":"push","commitsCount":25,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, Remove ICM chunk fields that can be calculated separately\n\nReduce the chunk struct size by removing fields that can be calculated\nusing the existing fields.\n\nSigned-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>\nReviewed-by: Erez Shitrit <erezsh@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, Remove ICM chunk fields that can be calculated separately"}},{"before":"db9049c97ad0f93b9cab8c399a8c3b5cb21e1088","after":"a09e57981e97242f988bd7bea9fa9c7677ba7a59","ref":"refs/heads/mlx5_misc","pushedAt":"2023-04-04T10:28:09.000Z","pushType":"push","commitsCount":51,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: Report IBV_ATOMIC_GLOB when supported\n\nTurn on IBV_ATOMIC_GLOB atomic cap bit in case that PCI atomic\noperations of compare-and-swap and fetch-and-add of 8 bytes are\nsupported by the device.\n\nSigned-off-by: Maor Gottlieb <maorg@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: Report IBV_ATOMIC_GLOB when supported"}},{"before":"904b1fcdee9090659f3c3fbfe4c5af16263b8fbc","after":"f3ae9a97ec028ceb6c5f862ef2974fcd8e29c63a","ref":"refs/heads/mlx5_dr","pushedAt":"2023-03-20T08:03:17.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, Check fl_rc_qp_when_roce_disabled independently from ROCE\n\nSWS uses RC QP for writing STEs to ICM, the writes are done in LB and FL\n(force-loopback) QP is preferred for performance. FL is available when\nRoCE is enabled or disabled based on RoCE caps.\nThis patch also uses the HCA cap for an additional capability for RC QP\nFL support when RoCE capability is disabled.  Disabling ROCE may reduce\nmemory for virtio, but FL QP is still required for internal\nimplementation. The solution is to allow reading a copy of the bit\noutside the domain of ROCE caps.\n\nSigned-off-by: igozlan <igozlan@nvidia.com>\nSigned-off-by: Alex Vesker <valex@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, Check fl_rc_qp_when_roce_disabled independently from ROCE"}},{"before":"c6f26fa667fc0edcad3d982904c49348664ca6e0","after":"904b1fcdee9090659f3c3fbfe4c5af16263b8fbc","ref":"refs/heads/mlx5_dr","pushedAt":"2023-03-19T16:25:43.000Z","pushType":"push","commitsCount":78,"pusher":{"login":"yishaih","name":"Yishai Hadas","path":"/yishaih","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/4038053?s=80&v=4"},"commit":{"message":"mlx5: DR, Check fl_rc_qp_when_roce_disabled independently from ROCE\n\nSWS uses RC QP for writing STEs to ICM, the writes are done in LB and FL\n(force-loopback) QP is preferred for performance. FL is available when\nRoCE is enabled or disabled based on RoCE caps.\nThis patch also uses the HCA cap for an additional capability for RC QP\nFL support when RoCE capability is disabled.  Disabling ROCE may reduce\nmemory for virtio, but FL QP is still required for internal\nimplementation. The solution is to allow reading a copy of the bit\noutside the domain of ROCE caps.\n\nSigned-off-by: igozlan <igozlan@nvidia.com>\nSigned-off-by: Alex Vesker <valex@nvidia.com>\nSigned-off-by: Yishai Hadas <yishaih@nvidia.com>","shortMessageHtmlLink":"mlx5: DR, Check fl_rc_qp_when_roce_disabled independently from ROCE"}}],"hasNextPage":false,"hasPreviousPage":false,"activityType":"all","actor":null,"timePeriod":"all","sort":"DESC","perPage":30,"cursor":"djE6ks8AAAAEQcHp7AA","startCursor":null,"endCursor":null}},"title":"Activity · yishaih/rdma-core"}