- 07 Aug, 2017 1 commit
-
-
Doug Ledford authored
Merge tag 'rdma-rc-2017-07-26' of git://git.kernel.org/pub/scm/linux/kernel/git/leon/linux-rdma into leon-ipoib IPoIB fixes for 4.13 The patchset provides various fixes for IPoIB. It is combination of fixes to various issues discovered during verification along with static checkers cleanup patches. Most of the patches are from pre-git era and hence lack of Fixes lines. There is one exception in this IPoIB group - addition of patch revert: Revert "IB/core: Allow QP state transition from reset to error", but it followed by proper fix to the annoying print, so I thought it is appropriate to include it. Signed-off-by:
Doug Ledford <dledford@redhat.com>
-
- 04 Aug, 2017 5 commits
-
-
Dan Carpenter authored
The hns_roce_v1_create_lp_qp() returns NULL on error, not error pointers. Fixes: bfcc681b ("IB/hns: Fix the bug when free mr") Signed-off-by:
Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by:
-
Leon Romanovsky authored
The extended address vector is the highest bit in be32 variable, but it was compared with the lowest. This patch fixes the endianness of that check and removes already declared define. Fixes: 17d2f88f ("IB/mlx5: Add ODP atomics support") Reviewed-by:
Artemy Kovalyov <artemyko@mellanox.com> Signed-off-by:
Leon Romanovsky <leonro@mellanox.com> Signed-off-by:
-
Yishai Hadas authored
Uverbs device should be cleaned up only when there is no potential usage of. As part of ib_uverbs_remove_one which might be triggered upon reset flow the device reference count is decreased as expected and leave the final cleanup to the FDs that were opened. Current code increases reference count upon opening a new command FD and decreases it upon closing the file. The event FD is opened internally and rely on the command FD by taking on it a reference count. In case that the command FD was closed and just later the event FD we may ensure that the device resources as of srcu are still alive as they are still in use. Fixing the above by moving the reference count decreasing to the place where the command FD is really freed instead of doing that when it was just closed. fixes: 036b1063 ("IB/uverbs: Enable device removal when there are active user space applications") Signed-off-by:
Yishai Hadas <yishaih@mellanox.com> Reviewed-by:
Matan Barak <matanb@mellanox.com> Reviewed-by:
Jason Gunthorpe <jgunthorpe@obsidianresearch.com> Tested-by:
Leon Romanovsky <leon@kernel.org> Signed-off-by:
-
Leon Romanovsky authored
initialize to zero the response structure to prevent the leakage of "resp.reserved" field. drivers/infiniband/core/uverbs_cmd.c:1178 ib_uverbs_resize_cq() warn: check that 'resp.reserved' doesn't leak information Fixes: 33b9b3ee ("IB: Add userspace support for resizing CQs") Signed-off-by:
Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by:
-
Parav Pandit authored
Currently while resolving IP address to MAC address single delayed work is used for resolving multiple such resolve requests. This singled work is essentially performs two tasks. (a) any retry needed to resolve and (b) it executes the callback function for all completed requests While work is executing callbacks, any new work scheduled on for this workqueue is lost because workqueue has completed looking at all pending requests and now looking at callbacks, but work is still under execution. Any further retry to look at pending requests in process_req() after executing callbacks would lead to similar race condition (may be reduce the probably further but doesn't eliminate it). Retrying to enqueue work that from queue_req() context is not something rest of the kernel modules have followed. Therefore fix in this patch utilizes kernel facility to enqueue multiple work items to a workqueue. This ensures that no such requests gets lost in synchronization. Request list is still maintained so that rdma_cancel_addr() can unlink the request and get the completion with error sooner. Neighbour update event handling continues to be handled in same way as before. Additionally process_req() work entry cancels any pending work for a request that gets completed while processing those requests. Originally ib_addr was ST workqueue, but it became MT work queue with patch of [1]. This patch again makes it similar to ST so that neighbour update events handler work item doesn't race with other work items. In one such below trace, (though on 4.5 based kernel) it can be seen that process_req() never executed the callback, which is likely for an event that was schedule by queue_req() when previous callback was getting executed by workqueue. [<ffffffff816b0dde>] schedule+0x3e/0x90 [<ffffffff816b3c45>] schedule_timeout+0x1b5/0x210 [<ffffffff81618c37>] ? ip_route_output_flow+0x27/0x70 [<ffffffffa027f9c9>] ? addr_resolve+0x149/0x1b0 [ib_addr] [<ffffffff816b228f>] wait_for_completion+0x10f/0x170 [<ffffffff810b6140>] ? try_to_wake_up+0x210/0x210 [<ffffffffa027f220>] ? rdma_copy_addr+0xa0/0xa0 [ib_addr] [<ffffffffa0280120>] rdma_addr_find_l2_eth_by_grh+0x1d0/0x278 [ib_addr] [<ffffffff81321297>] ? sub_alloc+0x77/0x1c0 [<ffffffffa02943b7>] ib_init_ah_from_wc+0x3a7/0x5a0 [ib_core] [<ffffffffa0457aba>] cm_req_handler+0xea/0x580 [ib_cm] [<ffffffff81015982>] ? __switch_to+0x212/0x5e0 [<ffffffffa04582fd>] cm_work_handler+0x6d/0x150 [ib_cm] [<ffffffff810a14c1>] process_one_work+0x151/0x4b0 [<ffffffff810a1940>] worker_thread+0x120/0x480 [<ffffffff816b074b>] ? __schedule+0x30b/0x890 [<ffffffff810a1820>] ? process_one_work+0x4b0/0x4b0 [<ffffffff810a1820>] ? process_one_work+0x4b0/0x4b0 [<ffffffff810a6b1e>] kthread+0xce/0xf0 [<ffffffff810a6a50>] ? kthread_freezable_should_stop+0x70/0x70 [<ffffffff816b53a2>] ret_from_fork+0x42/0x70 [<ffffffff810a6a50>] ? kthread_freezable_should_stop+0x70/0x70 INFO: task kworker/u144:1:156520 blocked for more than 120 seconds. "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. kworker/u144:1 D ffff883ffe1d7600 0 156520 2 0x00000080 Workqueue: ib_addr process_req [ib_addr] ffff883f446fbbd8 0000000000000046 ffff881f95280000 ffff881ff24de200 ffff883f66120000 ffff883f446f8008 ffff881f95280000 ffff883f6f9208c4 ffff883f6f9208c8 00000000ffffffff ffff883f446fbbf8 ffffffff816b0dde [1] http://lkml.iu.edu/hypermail/linux/kernel/1608.1/05834.htmlSigned-off-by:
Parav Pandit <parav@mellanox.com> Reviewed-by:
Mark Bloch <markb@mellanox.com> Signed-off-by:
-
- 23 Jul, 2017 11 commits
-
-
Erez Shitrit authored
Modify QP can fail and it can be acceptable, like when moving from RST to ERR state, all the rest are not acceptable and a message to the log should be printed. The current code prints on all failures and many messages like: "Failed to modify QP to ERROR state" appear, even when supported by the state machine of the QP object. Signed-off-by:
Erez Shitrit <erezsh@mellanox.com> Signed-off-by:
-
Leon Romanovsky authored
The commit ebc9ca43 ("IB/core: Allow QP state transition from reset to error") allowed transition from Reset to Error state for the QPs. This behavior doesn't follow the IBTA specification 1.3, which in 10.3.1 QUEUE PAIR AND EE CONTEXT STATES section. The quote from the spec: "An error can be forced from any state, except Reset, with the Modify QP/EE Verb." Signed-off-by:
-
Leon Romanovsky authored
There is no need to assign "p" pointer twice. This patch fixes the following smatch warning: drivers/infiniband/ulp/ipoib/ipoib_cm.c:517 ipoib_cm_rx_handler() warn: missing break? reassigning 'p->id' Fixes: 839fcaba ("IPoIB: Connected mode experimental support") Signed-off-by:
-
Leon Romanovsky authored
Refactor error paths in ipoib_add_port() function. The code flow ensures that the function terminates on every error flow and it makes redundant all "else" cases. The functions are called during the flow are returning "result < 0", in case of error, so there is no need to check it explicitly. Fixes: 58e9cc90 ("IB/IPoIB: Fix bad error flow in ipoib_add_port()") Signed-off-by:
-
Feras Daoud authored
Add SRIOV VF support to get traffic statistics. Signed-off-by:
Feras Daoud <ferasda@mellanox.com> Signed-off-by:
-
Alex Vesker authored
Update the multicast counter when multicast packets are received and provide this information through ethtool support. Signed-off-by:
Alex Vesker <valex@mellanox.com> Signed-off-by:
-
Feras Daoud authored
Set IPOIB_NEIGH_TBL_FLUSH bit after initializing the neighbor flushed completion, otherwise the garbage collector may signal a completion while it is not initialized yet. Fixes: b63b70d8 ("IPoIB: Use a private hash table for path lookup in xmit path") Signed-off-by:
-
Alex Vesker authored
Don't allow negative values to max_nonsrq_conn_qp. There is no functional impact on a negative value but it is logicically incorrect. Fixes: 68e995a2 ("IPoIB/cm: Add connected mode support for devices without SRQs") Signed-off-by:
-
Erez Shitrit authored
While cleaning neighs and there is a send-only mcast neigh, the driver should wait to finish its join process before trying to remove it. Without this patch, we will see messages like: "ipoib_mcast_leave on an in-flight join" and unexpected results in the join_complete. Signed-off-by:
-
Erez Shitrit authored
The work mcast_task can re-queue itself, so instead of doing cancel && flush_workqueue, that still can leave a queued task on the air, use cancel_delayed_work_sync. Also, no need to use lock over the cancel, the original lock was due to bit assignment setting (IPOIB_MCAST_RUN) that is not in use anymore. Signed-off-by:
-
Feras Daoud authored
A potential race between light_event and interface restart may attach multicast group to an already attached QP. Scenario: light_event flow goes through ipoib_mcast_dev_flush function, if a context switch occurs before calling ipoib_mcast_remove_list, then we may face a situation where the broadcast of the priv is null and the corresponding QP is not detached yet. If an "interface restart" runs during the previous context switch, the following scenario occurs: When the device goes up, ipoib_ib_dev_up function will be called, it will send a new registration request to the broadcast group and then attach the group to the QP that was not detached before. IPOIB_FLUSH_LIGHT INTERFACE RESTART __ipoib_ib_dev_flush | | | | | | | ipoib_mcast_dev_flush | Move mcast list and broadcast to remove_list | | | | | Context Switch--> | | ipoib_ib_dev_down | | | | | ipoib_ib_dev_up | | | | | ipoib_mcast_join_task | allocate new broadcast | | | | | Attach QP to multicast group | | | | | <--Context Switch ipoib_mcast_leave Detach QP from multicast group Signed-off-by:
-
- 20 Jul, 2017 23 commits
-
-
Ismail, Mustafa authored
Initialize the port_num for iWARP in rdma_init_qp_attr. Fixes: 5ecce4c9("Check port number supplied by user verbs cmds") Cc: <stable@vger.kernel.org> # v2.6.14+ Reviewed-by:
Steve Wise <swise@opengridcomputing.com> Signed-off-by:
Mustafa Ismail <mustafa.ismail@intel.com> Tested-by:
Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by:
-
Ismail, Mustafa authored
The port number is only valid if IB_QP_PORT is set in the mask. So only check port number if it is valid to prevent modify_qp from failing due to an invalid port number. Fixes: 5ecce4c9("Check port number supplied by user verbs cmds") Cc: <stable@vger.kernel.org> # v2.6.14+ Reviewed-by:
-
Kalderon, Michal authored
Once in_dev_get is called to receive in_device pointer, the in_device reference counter is increased, but if there are no ipv4 addresses configured on the net-device the ifa_list will be null, resulting in a flow that doesn't call in_dev_put to decrease the ref_cnt. This was exposed when running RoCE over ipv6 without any ipv4 addresses configured Fixes: commit 8e3867310c90 ("IB/cma: Fix a race condition in iboe_addr_get_sgid()") Signed-off-by:Michal Kalderon <Michal.Kalderon@cavium.com> Signed-off-by:
Ariel Elior <Ariel.Elior@cavium.com> Signed-off-by:
-
Sagi Grimberg authored
We might get some bogus error completions in case the target will remotely invalidate the rkey and the HCA will need to retransmit from this buffer. Signed-off-by:
Sagi Grimberg <sagi@grimberg.me> Signed-off-by:
-
Vijay Immanuel authored
If we modified the qp to ERROR state, and drained the recieve queue, post_recv must trigger the responder task to complete the drain work request. Cc: Bart Van Assche <bart.vanassche@sandisk.com> Signed-off-by:
Vijay Immanuel <vijayi@attalasystems.com> Signed-off-by:
-
Amrani, Ram authored
Wrap ib_copy_to_udata with a function that ensures that the data being copied over to user space isn't longer than the allowed. Fixes: cecbcddf ("qedr: Add support for QP verbs") Fixes: a7efd777 ("qedr: Add support for PD,PKEY and CQ verbs") Fixes: ac1b36e5 ("qedr: Add support for user context verbs") Signed-off-by:
Ram Amrani <Ram.Amrani@cavium.com> Signed-off-by:
-
Ganesh Goudar authored
Only use the read sge lkey/addr and the remote rkey/addr if the length of the read is not zero. Otherwise the read response might be treated as the RTR read response and not delivered to the application. Or worse Terminator hardware will fail a 0B read if the STAG is 0 even if the read length is 0. Signed-off-by:
Ganesh Goudar <ganeshgr@chelsio.com> Signed-off-by:
-
Håkon Bugge authored
CM REQs cannot be successfully retried, because a new pv_cm_id is created for each request, without checking if one already exists. By checking if an id exists before creating one, the bug is fixed. This bug can be provoked by running an RDMA CM user-land application, but inserting a five seconds delay before the rdma_accept() call on the passive side. This delay is larger than the default CMA timeout, and triggers a retry from the active side. The retried REQ will use another pv_cm_id (the cm_id on the wire). This confuses the CM protocol and two REJs are sent from the passive side. Here is an excerpt from ibdump running without the patch: 3.285092 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello) 7.382711 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello) 7.382861 LID: 4 -> LID: 4 InfiniBand 290 CM: ConnectReject 7.387644 LID: 4 -> LID: 4 InfiniBand 290 CM: ConnectReject and here is the same with bug fix applied: 3.251010 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello) 7.349387 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello) 8.258443 LID: 4 -> LID: 4 SDP 290 CM: ConnectReply(SDP Hello) 8.259890 LID: 4 -> LID: 4 InfiniBand 290 CM: ReadyToUse Suggested-by:
Venkat Venkatsubra <venkat.x.venkatsubra@oracle.com> Signed-off-by:
Håkon Bugge <haakon.bugge@oracle.com> Reported-by:
Wei Lin Guay <wei.lin.guay@oracle.com> Tested-by:
Yuval Shaia <yuval.shaia@oracle.com> Acked-by:
Jack Morgenstein <jackm@dev.mellanox.co.il> Signed-off-by:
-
Kaike Wan authored
Current computation of qp->timeout_jiffies in rvt_modify_qp() will cause overflow due to the fact that the input to the function usecs_to_jiffies is only 32-bit ( unsigned int). Overflow will occur when attr->timeout is equal to or greater than 30. The consequence is unnecessarily excessive retry and thus degradation of the system performance. This patch fixes the problem by limiting the input to 5-bit and calling usecs_to_jiffies() before multiplying the scaling factor. Reviewed-by:
Kaike Wan <kaike.wan@intel.com> Signed-off-by:
-
Matan Barak authored
Delete unused variables to prevent sparse warnings. Fixes: db1b5ddd ("IB/core: Rename uverbs event file structure") Fixes: fd3c7904 ("IB/core: Change idr objects to use the new schema") Signed-off-by:
-
Selvin Xavier authored
Local ack delay exposed by the driver is 0 which means infinite QP timeout. Reporting the default value to 16 (approx 260ms) Signed-off-by:
Selvin Xavier <selvin.xavier@broadcom.com> Signed-off-by:
-
Selvin Xavier authored
While invoking the req_notify_cq hook, ULPs can request whether the CQs have any CQEs pending. If CQEs are pending, drivers can indicate it by returning 1 for req_notify_cq. The stack will poll CQ again till CQ is empty. This patch peeks the CQ for any valid entries and return accordingly. Signed-off-by:
-
Devesh Sharma authored
Fix the incorrect reporting of number of polled entries by taking into account the max CQ depth in the driver. Signed-off-by:
Devesh Sharma <devesh.sharma@broadcom.com> Signed-off-by:
-
Devesh Sharma authored
Driver shall check if the host system bios has enabled Atomic operations capability in PCI Device Control 2 register of the pci-device. Expose the ATOMIC_HCA flag only if the Atomic operations capability is set. Signed-off-by:
-
Somnath Kotur authored
Starting FW version 20.6.47, firmware is keeping separate statistics for L2 and RDMA. However, driver needs to specify RDMA or not when allocating stat_ctx. Signed-off-by:
-
Eddie Wai authored
There's a couple of bugs in the support of max_rd_atomic and max_dest_rd_atomic. In the modify_qp, if the requested max_rd_atomic, which is the ORRQ size, is greater than what the chip can support, then we have to cap the request to chip max as we can't have the HW overflow the ORRQ. Capping the max_rd_atomic support internally is okay to do as the remaining read/atomic WRs will still be sitting in the SQ. However, for the max_dest_rd_atomic, the driver has to error out as this dictates the IRRQ size and we can't control what the remote side sends. Signed-off-by:
Eddie Wai <eddie.wai@broadcom.com> Signed-off-by:
-
Selvin Xavier authored
- Report supported value for max_mr_size to IB stack in query_device. Also, check and log if MR size requested by application in reg_user_mr() is greater than value currently supported by driver. - Report only 4K page size support for now - Fix Max_QP value returned by ibv_devinfo -vv. In case of PF, FW reserves 129 QPs for creating QP1s of VFs and PF. So the max_qp value reported by FW for PF doesn'tt include the QP1. Fixing this issue by adding 1 with the value reported by FW. Signed-off-by:
-
Selvin Xavier authored
This fix is added only to avoid system crash in some a specific scenario. When bnxt_re driver is loaded and if user tries to change interface mac address, delete GID fails because QP1 is still associated with existing MAC (default GID). If the above command fails GID tables are not modified in the h/w or driver, but the GID context memory is freed. Now, if the user changes the mac back to the original value, another add_gid comes to the driver where the driver reports that the GID is already present in its table and tries to access the context which was already freed. So, in this case, in order to avoid NULL pointer de-reference, this patch removes the context memory free if delete_gid fails and the same context memory is re-used in new add_gid. Memory cleanup will be taken care during driver unload, while deleting the GID table. Signed-off-by:
Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Signed-off-by:
-
Somnath Kotur authored
Posting WQE size of 2 results in a WQE_FORMAT_ERROR thrown by the HW as it requires host to supply WQE Size with room for atleast one SGE so that the resulting WQE size be atleast 3. Signed-off-by:
Somnath Kotur <somnath.kotur@broadcom.com> Signed-off-by:
-
Devesh Sharma authored
The driver must free the DPI during the dealloc_ucontext instead of freeing it during dealloc_pd. However, the DPI allocation scheme remains unchanged. Signed-off-by:
-
Dan Carpenter authored
"umem" is a valid pointer. We intended to print "*umem" or even just "err" instead. Signed-off-by:
-
Dan Carpenter authored
If either of these allocations fail then we return ERR_PTR(0). That's equivalent to NULL and results in a NULL pointer dereference in the caller. Fixes: fe2caefc ("RDMA/ocrdma: Add driver for Emulex OneConnect IBoE RDMA adapter") Signed-off-by:
-
Dan Carpenter authored
We should preserve the original "status" error code instead of resetting it to zero. Returning ERR_PTR(0) is the same as NULL and results in a NULL dereference in the callers. I added a printk() on error instead. Fixes: 45e86b33 ("RDMA/ocrdma: Cache recv DB until QP moved to RTR") Signed-off-by:
-
