| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
Merge pull request #14372 from bwbarrett/cleanup/spdx Add SPDX headers to Open MPI | 11 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
btl/tcp: lock the goodbye path, and say what it leaves queued A MCA_BTL_TCP_HDR_TYPE_FIN says the peer has let go of the socket it arrives on, either because its endpoint is being destructed or because it dropped a duplicate dial in the simultaneous-connect duel. Whatever we still have queued for that peer survives the close -- only the FAILED path of mca_btl_tcp_endpoint_close() flushes the queue -- but nothing re-drives it before the next send to the same peer, and there may not be one. Those fragments are then neither sent nor completed, and until now that was silent. At the FIN rather than in endpoint_close(), whose other callers keep the endpoint and let the queue ride across to the next socket: the duel closes the dial it drops with fragments legitimately waiting on it, and mca_btl_tcp_endpoint_connected() puts them on the adopted socket a few lines later. Only verbose. Completing them with an error would be the honest answer for a destructed peer, and wrong for the duel, where the endpoint is a single send away from working again. Do all of it under endpoint_send_lock. Every other path in mca_btl_tcp_frag_recv() that closes an endpoint holds that lock across the state change and the close; this one never has. It gets away with it only because the two statements run back to back: the state goes to MCA_BTL_TCP_CLOSED and endpoint_close() reads it a moment later, so the MCA_BTL_TCP_FAILED branch -- the one that walks the send queue, completes every fragment with OPAL_ERR_UNREACH and reports the peer upstream -- is not entered. Nothing stops another thread from moving the endpoint to MCA_BTL_TCP_FAILED in between, and that branch would then run against a concurrent sender with nothing held. The queue head read for the verbose line wants the same lock in any case, being a send-side field read from the receive path. The caller holds endpoint_recv_lock already, and the eof and error paths above reach for the send lock while holding it, so this is the order the file has established. While here, say btl_endpoint rather than frag->endpoint, as every other line in the function does. Signed-off-by: George Bosilca <gbosilca@nvidia.com> | 3 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 | |
Add SPDX Identifiers to OMPI Tree Signed-off-by: Brian Barrett <bbarrett@amazon.com> | 13 天前 |