2006-11-09 Roland Dreier <rdreier@cisco.com> * Release version 1.0.3. 2006-10-17 Roland Dreier <rdreier@cisco.com> * src/cq.c, src/qp.c, src/srq.c: Convert existing uses of mb() to rmb() or wmb() as appropriate. In fact all memory barriers were really just wmb(), except for the barrier between reading a CQE's ownership bit and contents, which should be rmb(). * src/mthca.h: Add compatibility defines of rmb()/wmb() so that libmthca continues to build against old libibverbs releases. 2006-10-03 Roland Dreier <rdreier@cisco.com> * src/cq.c (mthca_poll_one): Annotate so that Valgrind knows contents of CQ entry are all valid after they are written by HCA. (set_cqe_hw): Annotate so that CQ entries owned by hardware are not defined. * src/mthca.h: Add wrapper for VALGRIND_MAKE_MEM_DEFINED so that it can be used in .c files without worrying about whether Valgrind is installed or enabled. * configure.in: Add support for Valgrind annotation (enabled with --with-valgrind option to configure). * src/verbs.c (mthca_create_qp): Set reserved fields to 0 to avoid future problems and also to make Valgrind a little quieter. 2006-09-04 Roland Dreier <rdreier@cisco.com> * src/verbs.c (mthca_destroy_qp): Avoid potential AB-BA deadlock when destroying QPs by always taking CQ locks in a consistent order (lowest CQN first). The old code always took the send_cq lock first, which is prone to deadlock if the send_cq of one QP is the recv_cq of another QP destroyed at the same time. This bug was pointed out by Dotan Barak and Jack Morgenstein. 2006-08-23 Roland Dreier <rdreier@cisco.com> * src/verbs.c (mthca_resize_cq): Add a test for IBV_CMD_RESIZE_CQ_HAS_RESP_PARAMS to make libmthca work with newer libibverbs libraries that add two parameters to ibv_cmd_resize_cq(). 2006-08-09 Michael S. Tsirkin <mst@mellanox.co.il> * src/qp.c (mthca_tavor_post_send, mthca_arbel_post_send): Fence bit must be set in both doorbell and WQE. 2006-08-03 Jack Morgenstein <jackm@mellanox.co.il> * src/mthca.h: Include <stddef.h> to get definition of offsetof(). 2006-08-03 Michael S. Tsirkin <mst@mellanox.co.il> * src/verbs.c (mthca_create_srq): Limit SRQ max_wr to avoid integer overflow. 2006-07-26 Roland Dreier <rdreier@cisco.com> * src/mthca.h, src/ah.c, src/cq.c, src/memfree.c, src/qp.c, src/srq.c, src/verbs.c: Convert internal allocations for AH pages (for non-memfree HCAs), CQ buffers, doorbell pages (for memfree HCAs), QP buffers and SRQ buffers to use the new buffer allocator. This makes libmthca fork()-clean when built against libibverbs 1.1. * src/buf.c (mthca_alloc_buf, mthca_free_buf): Add new functions to wrap up allocating page-aligned buffers. The new functions will call ibv_dontfork_range()/ibv_dofork_range() to do proper madvise()ing to handle fork(), if applicable. * configure.in: Check for ibv_dontfork_range() and ibv_dontfork_range(). 2006-07-04 Dotan Barak <dotanb@mellanox.co.il> * src/verbs.c (mthca_create_cq, mthca_resize_cq): Passing huge size values to create_cq/resize_cq causes a hang in align_cq_size(). Fix this by validating input, similiar to what we do for mthca_create_qp() and mthca_create_srq(). 2006-06-13 Roland Dreier <rdreier@cisco.com> * Release version 1.0.2. 2006-06-13 Michael S. Tsirkin <mst@mellanox.co.il> * src/cq.c (mthca_poll_one): Add workaround for MemFree FW bug that causes wrong WQE addr to be reported. 2006-05-24 Roland Dreier <rdreier@cisco.com> * src/mthca.c: If <sysfs/libsysfs.h> is detected, include it explicitly. This lets things build when sysfs headers are installed on the build system even when building against libibverbs 1.1 (which does not include sysfs headers implicitly). * src/ah.c, src/cq.c, src/memfree.c, src/mthca.c, src/qp.c, src/srq.c: Add include of <string.h>, since it may no long be implicitly included from libsysfs headers when building with libibverbs 1.1. 2006-05-24 Michael S. Tsirkin <mst@mellanox.co.il> * src/srq.c (mthca_tavor_post_srq_recv): Fix posting of lists of receives that have exactly a multiple of 256 entries (same as QP bug fixed below). 2006-05-22 Roland Dreier <rdreier@cisco.com> * configure.in, src/mthca.c (openib_driver_init): Check for the presence of <sysfs/libsysfs.h>, and if it is not installed, don't export the old openib_driver_init() entry point. 2006-05-18 Michael S. Tsirkin <mst@mellanox.co.il> * src/qp.c (mthca_tavor_post_recv): Fix posting of lists of receives that have exactly a multiple of 256 entries. 2006-05-08 Jack Morgenstein <jackm@mellanox.co.il> * src/mthca.c: Add include files needed for open() if HAVE_IBV_READ_SYSFS_FILE is not defined (so libmthca includes a private local definition of ibv_read_sysfs_file()). 2006-04-11 Roland Dreier <rdreier@cisco.com> * src/mthca.c (ibv_driver_init, openib_driver_init): Add new forward-compatible driver entry point. Make old entry point a simple wrapper for the new one. 2006-03-14 Roland Dreier <rdreier@cisco.com> * Release version 1.0.1. * Makefile.am (EXTRA_DIST): Remove debian/ directory from tarballs, since Debian policy is that upstream tarballs should not include it. 2006-03-13 Roland Dreier <rdreier@cisco.com> * Release version 1.0. 2006-02-27 Dotan Barak <dotanb@mellanox.co.il> * src/qp.c (mthca_tavor_post_send, mthca_arbel_post_send): Add support for IBV_SEND_FENCE flag. 2006-02-16 Roland Dreier <rdreier@cisco.com> * src/memfree.c (mthca_alloc_db): Introduce a temporary variable to pass to posix_memalign() to avoid "warning: dereferencing type-punned pointer will break strict-aliasing rules." * Release version 1.0-rc7. 2006-02-15 Roland Dreier <rdreier@cisco.com> * src/verbs.c (mthca_create_qp): Update to add new response and response size parameters for libibverbs ibv_cmd_create_qp(). 2006-02-14 Roland Dreier <rdreier@cisco.com> * Release version 1.0-rc6. 2006-02-13 Dotan Barak <dotanb@mellanox.co.il> * src/verbs.c (mthca_query_qp, mthca_query_srq): Add query QP and query SRQ verbs. 2006-01-31 Roland Dreier <rdreier@cisco.com> * src/mthca.h: Remove useless "extern" from function declarations. 2006-01-30 Michael S. Tsirkin <mst@mellanox.co.il> * src/qp.c (mthca_tavor_post_recv, mthca_arbel_post_recv): Pass recv_cq to wq_overflow() so we lock the correct CQ. Noticed by Yossi Leybovich. 2006-01-26 Roland Dreier <rdreier@cisco.com> * src/mthca.h, src/verbs.c, src/cq.c, src/mthca.c: Add implementation of resize CQ operation. * src/mthca-abi.h: Add mthca-specific resize CQ ABI. 2006-01-22 Roland Dreier <rdreier@cisco.com> * Release version 1.0-rc5. 2006-01-11 Jack Morgenstein <jackm@mellanox.co.il> * src/verbs.c (mthca_free_pd): Free pointer to correct structure (we get lucky now, but don't rely on this). * src/mthca.c (mthca_free_context): Free context's PD so we don't leak it. 2006-01-06 Michael S. Tsirkin <mst@mellanox.co.il> * src/verbs.c (mthca_destroy_qp): Jack Morgenstein has discovered the following race condition in libmthca: Thread A destroys QP A at the kernel side by calling ibv_cmd_destroy_qp, but its time-slice is over before removing it from the user-space qp_table removal. Thread B allocates QP B, receiving a QP number that matches the just-destroyed QP A in the low 16 bits. Thread B will now over-write the slot in qp_table which was used for QP A. Thread A wakes up and clears qp_table slot, in effect removing QP B from qp_table. As a solution, remove the QP from qp_table before calling ibv_cmd_destroy_qp. This also makes sense since operations are performed in the reverse order in create_qp. * src/cq.c (handle_error_cqe): Fill in vendor_err field for completions with error. 2006-01-05 Jack Morgenstein <jackm@mellanox.co.il> * src/verbs.c (mthca_destroy_qp, mthca_destroy_srq): Free QP/SRQ object to avoid memory leak. 2005-12-15 Jack Morgenstein <jackm@mellanox.co.il> * src/cq.c (mthca_cq_clean): When cleaning up a CQ, we should free an SRQ WQE if and only if the CQE is a receive. 2005-12-15 Michael S. Tsirkin <mst@mellanox.co.il> * src/qp.c (mthca_store_qp): Don't increment qp_table ref count if allocation fails. 2005-11-29 Michael S. Tsirkin <mst@mellanox.co.il> * src/qp.c (mthca_arbel_post_send): Add handling for posting long send lists for mem-free HCAs. * src/qp.c (mthca_tavor_post_recv): Fix posting long receive lists: nreq is set to zero early on, so we need to use MTHCA_TAVOR_MAX_WQES_PER_RECV_DB as the increment to rq.head. 2005-11-28 Roland Dreier <roland@cisco.com> * src/qp.c (mthca_init_qp_indices): Set qp->sq.last and qp->rq.last so that QP is fully reset when the indices are reinited on transition to RESET state. (mthca_tavor_post_send, mthca_arbel_post_send): Don't create an inline send segment when a work request is posted that has the inline flag set but no gather entries included. 2005-11-09 Roland Dreier <roland@cisco.com> * src/srq.c (mthca_tavor_post_srq_recv), src/qp.c (mthca_tavor_post_recv): Fix bugs in long receive list handling; need to set nreq to 0 and not put 256 credits into the second doorbell word. * src/cq.c (mthca_cq_clean): Handle case where CQ indices wrap around by treating signed comparisons of prod_index and cq->cons_index carefully. 2005-11-09 Michael S. Tsirkin <mst@mellanox.co.il> * src/srq.c (mthca_tavor_post_srq_recv), src/qp.c (mthca_tavor_post_recv): Tavor requires that a doorbell be rung at least every 256 receives, so add code to ring doorbells in the middle of posting a huge list of receives. * src/qp.c (mthca_tavor_post_send, mthca_tavor_post_send): When posting atomic operations, could wqe size in "octowords" correctly. * src/ah.c (mthca_alloc_av): Don't free ah if page allocation fails. It will be freed where it's allocated, in the caller. 2005-11-08 Roland Dreier <roland@cisco.com> * src/qp.c, src/verbs.c, src/mthca.h: Delegate setting of QP capabilities (max_sge, max_inline_data, etc) to kernel. 2005-11-04 Roland Dreier <roland@cisco.com> * src/verbs.c (mthca_destroy_qp): Clean CQEs when we destroy a QP. (mthca_modify_qp): Clean CQEs when we move a QP to RESET state, and reset QP index pointers. * src/cq.c (mthca_cq_clean): Add function to clean out CQEs for QPs that are being destroyed or reset. 2005-10-30 Roland Dreier <roland@cisco.com> * src/srq.c (wqe_to_link): Change to use an offset of 12 (the imm field), because posting an SRQ WQE may actually change the ee_nds field and still cause free list corruption. A receive WQE will never have immediate data, so using imm is definitely safe. 2005-10-25 Roland Dreier <roland@cisco.com> * Release version 1.0-rc4. 2005-10-23 Roland Dreier <roland@cisco.com> * src/qp.c (mthca_return_cap, mthca_alloc_qp_buf), src/verbs.c (mthca_create_qp): Explicitly pass QP type to functions used while creating QP, since we can't rely on ibv_qp.qp_type to be set until after we return. This fixes breakage with UD QPs introduced in the last change below. 2005-10-19 Roland Dreier <roland@cisco.com> * src/mthca.h, src/verbs.c (mthca_create_qp), src/qp.c (mthca_tavor_post_send, mthca_arbel_post_send, mthca_alloc_qp_buf, mthca_return_cap): Eliminate struct mthca_qp.qpt field and use struct ibv_qp.qp_type instead (now that that field has been added in libibverbs). 2005-10-18 Roland Dreier <roland@cisco.com> * src/cq.c (handle_error_cqe, mthca_poll_one): Dump CQEs for local QP operation errors instead of all error statuses. 2005-10-06 Roland Dreier <roland@cisco.com> * src/srq.c (mthca_free_srq_wqe): Pass index instead of WQE address. The only caller already has the index handy, so there's no need to recalculate it here. * src/srq.c (mthca_tavor_post_srq_recv, mthca_arbel_post_srq_recv): Add an extra check so that we report the SRQ as full before using the one extra WQE we need internally. 2005-10-05 Roland Dreier <roland@cisco.com> * src/verbs.c (mthca_modify_srq): Fill in mthca_modify_srq(). 2005-09-29 Roland Dreier <roland@cisco.com> * src/verbs.c (mthca_query_device): Update to match new libibverbs API that requires device-specific libraries to format firmware version. 2005-09-25 Roland Dreier <roland@cisco.com> * src/cq.c, src/mthca.c, src/mthca.h, src/verbs.c: Update to match new libibverbs API introduced with completion channel implementation. 2005-09-13 Roland Dreier <roland@cisco.com> * src/qp.c (mthca_tavor_post_send, mthca_tavor_post_recv, mthca_arbel_post_send), src/srq.c (mthca_tavor_post_srq_recv): Apply Michael S. Tsirkin's patch to fix linking of WQEs on mem-free HCAs. While we're at it, simplify the Tavor WQE posting code as well -- there's no need for a conditional, just always lik the previous WQE. 2005-09-07 Roland Dreier <roland@cisco.com> * src/mthca.h: Get rid of ntohll() and htonll() now that libibverbs defines them in <infiniband/arch.h>. 2005-08-31 Roland Dreier <roland@cisco.com> * src/memfree.c (mthca_free_db): When we free a doorbell record, really mark it as free in the free bitmap. This we we don't eventually run out of doorbells if a consumer creates and frees a lot of objects. * src/memfree.c (mthca_alloc_db): Introduce MTHCA_FREE_MAP_SIZE so that we iterate over the correct number of entries in the mem-free doorbell record free maps. This fixes some off-by-a-factor-of-8 bugs that could lead to crashes. * src/verbs.c (mthca_create_cq): In the mem-free case, when creating a CQ fails to allocate an arm doorbell, make sure we free the set CI doorbell instead of the (non-existent) arm doorbell.
