- 09 May, 2018 5 commits
-
-
Doug Ledford authored
Several items of conflict have arisen between the RDMA stack's for-rc branch and upcoming for-next work: 9fd4350b ("IB/rxe: avoid double kfree_skb") directly conflicts with 2e473507 ("IB/rxe: optimize the function duplicate_request") Patches already submitted by Intel for the hfi1 driver will fail to apply cleanly without this merge Other people on the mailing list have notified that their upcoming patches also fail to apply cleanly without this merge Signed-off-by:
Doug Ledford <dledford@redhat.com>
-
Idan Burstein authored
As most kernel RDMA ULPs, (e.g. NVMe over Fabrics in its default "register_always=Y" mode) registers and invalidates user buffer upon each IO. Today the mlx5 driver is posting the registration work request using scatter/gather entry for the MTT/KLM list. The fetch of the MTT/KLM list becomes the bottleneck in number of IO operation could be done by NVMe over Fabrics host driver on a single adapter as shown below. This patch is adding the support for inline registration work request upon MTT/KLM list of size <=64B. The result for NVMe over Fabrics is increase of > x3.5 for small IOs as shown below, I expect other ULPs (e.g iSER, SRP, NFS over RDMA) performance to be enhanced as well. The following results were taken against a single NVMe-oF (RoCE link layer) subsystem with a single namespace backed by null_blk using fio benchmark (with rw=randread, numjobs=48, iodepth={16,64}, ioengine=libaio direct=1): ConnectX-5 (pci Width x16) --------------------------- Block Size s/g reg_wr inline reg_wr ++++++++++ +++++++++++++++ ++++++++++++++++ 512B 1302.8K/34.82% 4951.9K/99.02% 1KB 1284.3K/33.86% 4232.7K/98.09% 2KB 1238.6K/34.1% 2797.5K/80.04% 4KB 1169.3K/32.46% 1941.3K/61.35% 8KB 1013.4K/30.08% 1236.6K/39.47% 16KB 695.7K/20.19% 696.9K/20.59% 32KB 350.3K/9.64% 350.6K/10.3% 64KB 175.86K/5.27% 175.9K/5.28% ConnectX-4 (pci Width x8) --------------------------- Block Size s/g reg_wr inline reg_wr ++++++++++ +++++++++++++++ ++++++++++++++++ 512B 1285.8K/42.66% 4242.7K/98.18% 1KB 1254.1K/41.74% 3569.2K/96.00% 2KB 1185.9K/39.83% 2173.9K/75.58% 4KB 1069.4K/36.46% 1343.3K/47.47% 8KB 755.1K/27.77% 748.7K/29.14% Tested-by:Nitzan Carmi <nitzanc@mellanox.com> Signed-off-by:
Idan Burstein <idanb@mellanox.com> Signed-off-by:
Max Gurtovoy <maxg@mellanox.com> Signed-off-by:
Leon Romanovsky <leonro@mellanox.com> Signed-off-by:
-
Leon Romanovsky authored
The zgid is already provided by IB/core, so there is no need in locally defined variable, let's drop it and reuse common one. Signed-off-by:
-
Parav Pandit authored
_gid_table_setup_one() only performs GID table cache memory allocation, marks entries as invalid (free) and marks the reserved entries. At this point GID table is empty and no entries are added. On dual port device if _gid_table_setup_one() fails to allocate the gid table for 2nd port, there is no need to perform cleanup_gid_table_port() to delete GID entries, as GID table is empty. Therefore make use of existing gid_table_release_one() routine which frees the GID table memory and avoid code duplication. Reviewed-by:
Daniel Jurgens <danielj@mellanox.com> Signed-off-by:
Parav Pandit <parav@mellanox.com> Signed-off-by:
Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by:
-
Parav Pandit authored
gid_table_reserve_default() always returns zero. Make it return void and simplify error checking. rdma_port is already calculated, use that while calling gid_table_reserve_default() instead of recalculating it. Reviewed-by:
-
- 03 May, 2018 12 commits
-
-
Steve Wise authored
Provide a cxgb4-specific function to fill in qp state details. This allows dumping important c4iw_qp state useful for debugging. Included in the dump are the t4_sq, t4_rq structs, plus a dump of the t4_swsqe and t4swrqe descriptors for the first and last pending entries. Signed-off-by:
Steve Wise <swise@opengridcomputing.com> Signed-off-by:
-
Steve Wise authored
These help rdma drivers to fill out the driver entries. Signed-off-by:
-
Steve Wise authored
Each driver can register a "fill entry" function with the restrack core. This function will be called when filling out a resource, allowing the driver to add driver-specific details. The details consist of a nltable of nested attributes, that are in the form of <key, [print-type], value> tuples. Both key and value attributes are mandatory. The key nlattr must be a string, and the value nlattr can be one of the driver attributes that are generic, but typed, allowing the attributes to be validated. Currently the driver nlattr types include string, s32, u32, s64, and u64. The print-type nlattr allows a driver to specify an alternative display format for user tools displaying the attribute. For example, a u32 attribute will default to "%u", but a print-type attribute can be included for it to be displayed in hex. This allows the user tool to print the number in the format desired by the driver driver. More attrs can be defined as they become needed by drivers. Signed-off-by:
-
Steve Wise authored
Add a specific RDMA_NLDEV_ATTR_PAD attribute to be used for 64b attribute padding. To preserve the ABI, make this attribute equal to RDMA_NLDEV_ATTR_UNSPEC, which has a value of 0, because that has been used up until now as the pad attribute. Change all the previous use of 0 as the pad with this new enum. Signed-off-by:
-
Parav Pandit authored
When commit [1] was added, SGID was queried to derive the SMAC address. Then, later on during a refactor [2], SMAC was no longer needed. However, the now useless GID query remained. Then during additional code changes later on, the GID query was being done in such a way that it caused iWARP queries to start breaking. Remove the useless GID query and resolve the iWARP breakage at the same time. This is discussed in [3]. [1] commit dd5f03be ("IB/core: Ethernet L2 attributes in verbs/cm structures") [2] commit 5c266b23 ("IB/cm: Remove the usage of smac and vid of qp_attr and cm_av") [3] https://www.spinics.net/lists/linux-rdma/msg63951.htmlSuggested-by:
Shiraz Saleem <shiraz.saleem@intel.com> Signed-off-by:
-
Jack Morgenstein authored
When the kernel was compiled using the UBSAN option, we saw the following stack trace: [ 1184.827917] UBSAN: Undefined behaviour in drivers/infiniband/hw/mlx4/mr.c:349:27 [ 1184.828114] signed integer overflow: [ 1184.828247] -2147483648 - 1 cannot be represented in type 'int' The problem was caused by calling round_up in procedure mlx4_ib_umem_calc_optimal_mtt_size (on line 349, as noted in the stack trace) with the second parameter (1 << block_shift) (which is an int). The second parameter should have been (1ULL << block_shift) (which is an unsigned long long). (1 << block_shift) is treated by the compiler as an int (because 1 is an integer). Now, local variable block_shift is initialized to 31. If block_shift is 31, 1 << block_shift is 1 << 31 = 0x80000000=-214748368. This is the most negative int value. Inside the round_up macro, there is a cast applied to ((1 << 31) - 1). However, this cast is applied AFTER ((1 << 31) - 1) is calculated. Since (1 << 31) is treated as an int, we get the negative overflow identified by UBSAN in the process of calculating ((1 << 31) - 1). The fix is to change (1 << block_shift) to (1ULL << block_shift) on line 349. Fixes: 9901abf5 ("IB/mlx4: Use optimal numbers of MTT entries") Signed-off-by:
Jack Morgenstein <jackm@dev.mellanox.co.il> Signed-off-by:
-
Sebastian Sanchez authored
When IRQ affinity is set and the interrupt type is unknown, a cpu mask allocated within the function is never freed. Fix this memory leak by allocating memory within the scope where it is used. Reviewed-by:
Mike Marciniszyn <mike.marciniszyn@intel.com> Reviewed-by:
Michael J. Ruhl <michael.j.ruhl@intel.com> Signed-off-by:
Sebastian Sanchez <sebastian.sanchez@intel.com> Signed-off-by:
-
Sebastian Sanchez authored
When allocating device data, if there's an allocation failure, the already allocated memory won't be freed such as per-cpu counters. Fix memory leaks in exception path by creating a common reentrant clean up function hfi1_clean_devdata() to be used at driver unload time and device data allocation failure. To accomplish this, free_platform_config() and clean_up_i2c() are changed to be reentrant to remove dependencies when they are called in different order. This helps avoid NULL pointer dereferences introduced by this patch if those two functions weren't reentrant. In addition, set dd->int_counter, dd->rcv_limit, dd->send_schedule and dd->tx_opstats to NULL after they're freed in hfi1_clean_devdata(), so that hfi1_clean_devdata() is fully reentrant. Reviewed-by:
-
Sebastian Sanchez authored
When an invalid num_vls is used as a module parameter, the code execution follows an exception path where the macro dd_dev_err() expects dd->pcidev->dev not to be NULL in hfi1_init_dd(). This causes a NULL pointer dereference. Fix hfi1_init_dd() by initializing dd->pcidev and dd->pcidev->dev earlier in the code. If a dd exists, then dd->pcidev and dd->pcidev->dev always exists. BUG: unable to handle kernel NULL pointer dereference at 00000000000000f0 IP: __dev_printk+0x15/0x90 Workqueue: events work_for_cpu_fn RIP: 0010:__dev_printk+0x15/0x90 Call Trace: dev_err+0x6c/0x90 ? hfi1_init_pportdata+0x38d/0x3f0 [hfi1] hfi1_init_dd+0xdd/0x2530 [hfi1] ? pci_conf1_read+0xb2/0xf0 ? pci_read_config_word.part.9+0x64/0x80 ? pci_conf1_write+0xb0/0xf0 ? pcie_capability_clear_and_set_word+0x57/0x80 init_one+0x141/0x490 [hfi1] local_pci_probe+0x3f/0xa0 work_for_cpu_fn+0x10/0x20 process_one_work+0x152/0x350 worker_thread+0x1cf/0x3e0 kthread+0xf5/0x130 ? max_active_store+0x80/0x80 ? kthread_bind+0x10/0x10 ? do_syscall_64+0x6e/0x1a0 ? SyS_exit_group+0x10/0x10 ret_from_fork+0x35/0x40 Cc: <stable@vger.kernel.org> # 4.9.x Reviewed-by:
-
Mike Marciniszyn authored
AHG may be armed to use the stored header, which by design is limited to edits in the PSN/A 32 bit word (bth2). When the code is trying to send a BECN, the use of the stored header will lose the BECN bit. Fix by avoiding AHG when getting ready to send a BECN. This is accomplished by always claiming the packet is not a middle packet which is an AHG precursor. BECNs are not a normal case and this should not hurt AHG optimizations. Cc: <stable@vger.kernel.org> # 4.14.x Reviewed-by:
-
Michael J. Ruhl authored
The module parameter num_user_context is defined as 'int' and defaults to -1. The module_param_named() says that it is uint. Correct module_param_named() type information and update the modinfo text to reflect the default value. Reviewed-by:
-
Mike Marciniszyn authored
The code for handling a marked UD packet unconditionally returns the dlid in the header of the FECN marked packet. This is not correct for multicast packets where the DLID is in the multicast range. The subsequent attempt to send the CNP with the multicast lid will cause the chip to halt the ack send context because the source lid doesn't match the chip programming. The send context will be halted and flush any other pending packets in the pio ring causing the CNP to not be sent. A part of investigating the fix, it was determined that the 16B work broke the FECN routine badly with inconsistent use of 16 bit and 32 bits types for lids and pkeys. Since the port's source lid was correctly 32 bits the type mixmatches need to be dealt with at the same time as fixing the CNP header issue. Fix these issues by: - Using the ports lid for as the SLID for responding to FECN marked UD packets - Insure pkey is always 16 bit in this and subordinate routines - Insure lids are 32 bits in this and subordinate routines Cc: <stable@vger.kernel.org> # 4.14.x Fixes: 88733e3b ("IB/hfi1: Add 16B UD support") Reviewed-by:
Don Hiatt <don.hiatt@intel.com> Reviewed-by:
-
- 01 May, 2018 3 commits
-
-
Colin Ian King authored
Trivial fix to spelling mistake in DP_ERR error message Signed-off-by:
Colin Ian King <colin.king@canonical.com> Signed-off-by:
-
YueHaibing authored
Use the recently introduced helper to replace the pattern of skb_put_zero/__skb_put() && memset(). Signed-off-by:
YueHaibing <yuehaibing@huawei.com> Reviewed-by:
-
Parav Pandit authored
Make security.c depends on CONFIG_SECURITY_INFINIBAND. Reviewed-by:
-
- 30 Apr, 2018 1 commit
-
-
Håkon Bugge authored
Currently, the kernel protects access to the agent ID allocator on a per port basis using a spinlock, so it is impossible for two apps/threads on the same port to get the same TID, but it is entirely possible for two threads on different ports to end up with the same TID. As this can be confusing (regardless of it being legal according to the IB Spec 1.3, C13-18.1.1, in section 13.4.6.4 - TransactionID usage), and as the rdma-core user space API for /dev/umad devices implies unique TIDs even across ports, make the TID an atomic type so that no two allocations, regardless of port number, will be the same. Signed-off-by:
Håkon Bugge <haakon.bugge@oracle.com> Reviewed-by:
Ira Weiny <ira.weiny@intel.com> Reviewed-by:
Zhu Yanjun <yanjun.zhu@oracle.com> Signed-off-by:
-
- 27 Apr, 2018 19 commits
-
-
Bharat Potnuri authored
When a CQ is shared by multiple QPs, c4iw_flush_hw_cq() needs to acquire corresponding QP lock before moving the CQEs into its corresponding SW queue and accessing the SQ contents for completing a WR. Ignore CQEs if corresponding QP is already flushed. Cc: stable@vger.kernel.org Signed-off-by:
Potnuri Bharat Teja <bharat@chelsio.com> Reviewed-by:
-
Ariel Levkovich authored
This patch fixes a crash that happens due to access to an uninitialized DM pointer within the MR object. The change makes sure the DM pointer in the MR object is set to NULL during a non-DM MR creation to prevent a false indication that this MR is related to a DM in the dereg flow. Fixes: be934cca ("IB/uverbs: Add device memory registration ioctl support") Reported-by:
Lijun Ou <oulijun@huawei.com> Signed-off-by:
Ariel Levkovich <lariel@mellanox.com> Signed-off-by:
-
Ariel Levkovich authored
This patch adds a check in the ib_uverbs_rereg_mr flow to make sure there's no attempt to rereg a device memory MR to regular MR. In such case the command will fail with -EINVAL status. fixes: be934cca ("IB/uverbs: Add device memory registration ioctl support") Signed-off-by:
-
Leon Romanovsky authored
Despite being advertised to user space application, the RSS inner header flag was filtered by checks at the beginning of QP creation routine. Cc: <stable@vger.kernel.org> # 4.15 Fixes: 4d02ebd9 ("IB/mlx4: Fix RSS hash fields restrictions") Fixes: 07d84f7b ("IB/mlx4: Add support to RSS hash for inner headers") Signed-off-by:
-
oulijun authored
This patch fixes two spelling errors. Signed-off-by:
-
oulijun authored
When generated bad work reqeust, it needs to report to user. This patch mainly fixes it. Signed-off-by:
-
oulijun authored
When posting a work reqeust, it need to update the owner bit of send wqe. This patch mainly fix the bug when posting multiply work request. Signed-off-by:
-
oulijun authored
This patch update the order of cleaning hem table for trrl_table and irrl_table as well as mtt_cqe_table and mtt_table. Signed-off-by:
-
oulijun authored
Only when the IB_QP_PATH_DEST_QPN flag of attr_mask is set is it valid to assign the dqpn field of qp context Signed-off-by:
-
oulijun authored
This patch deletes some unnecessary attr_mask if condition in hip08 according to the IB protocol. Signed-off-by:
-
oulijun authored
Only when the IB_QP_PATH_MTU flag of attr_mask is set it is valid to assign the mtu field of qp context when qp type is not GSI and UD. Signed-off-by:
-
oulijun authored
According to RoCE protocol, it is possible to transition from error to error state for modifying qp in hip08. This patch fix it. Signed-off-by:
-
oulijun authored
RDMA read operation is not supported inline data. If user cofigures issue a RDMA read and use inline data, it will happen a hardware error. Signed-off-by:
-
oulijun authored
During init hem table, type should be used instead of table->type which is finally initializaed with type. Signed-off-by:
Yixian Liu <liuyixian@huawei.com> Signed-off-by:
-
Zhu Yanjun authored
When skb is sent, it will pass the following functions in soft roce. rxe_send [rdma_rxe] ip_local_out __ip_local_out ip_output ip_finish_output ip_finish_output2 dev_queue_xmit __dev_queue_xmit dev_hard_start_xmit In the above functions, if error occurs in the above functions or iptables rules drop skb after ip_local_out, kfree_skb will be called. So it is not necessary to call kfree_skb in soft roce module again. Or else crash will occur. The steps to reproduce: server client --------- --------- |1.1.1.1|<----rxe-channel--->|1.1.1.2| --------- --------- On server: rping -s -a 1.1.1.1 -v -C 10000 -S 512 On client: rping -c -a 1.1.1.1 -v -C 10000 -S 512 The kernel configs CONFIG_DEBUG_KMEMLEAK and CONFIG_DEBUG_OBJECTS are enabled on both server and client. When rping runs, run the following command in server: iptables -I OUTPUT -p udp --dport 4791 -j DROP Without this patch, crash will occur. CC: Srinivas Eeda <srinivas.eeda@oracle.com> CC: Junxiao Bi <junxiao.bi@oracle.com> Signed-off-by:Yuval Shaia <yuval.shaia@oracle.com> Signed-off-by:
-
Jianchao Wang authored
w/o RXE_START_MASK, the last_psn of IB_OPCODE_RC_SEND_ONLY_INV will not be updated in update_wqe_psn, and the corresponding wqe will not be acked in rxe_completer due to its last_psn is zero. Finally, the other wqe will also not be able to be acked, because the wqe of IB_OPCODE_RC_SEND_ONLY_INV with last_psn 0 is still there. This causes large amount of io timeout when nvmeof is over rxe. Add RXE_START_MASK for IB_OPCODE_RC_SEND_ONLY_INV to fix this. Signed-off-by:
Jianchao Wang <jianchao.w.wang@oracle.com> Reviewed-by:
-
Colin Ian King authored
In the cases where iwpm_hash_bucket is NULL and where function get_mapinfo_hash_bucket returns NULL then the map_info is never added to hash_bucket_head and hence there is a leak of map_info. Fix this by nullifying hash_bucket_head and if that is null we know that that map_info was not added to hash_bucket_head and hence map_info should be free'd. Detected by CoverityScan, CID#1222481 ("Resource Leak") Fixes: 30dc5e63 ("RDMA/core: Add support for iWARP Port Mapper user space service") Signed-off-by: -
Luc Van Oostenryck authored
The method ndo_start_xmit() is defined as returning an 'netdev_tx_t', which is a typedef for an enum type, but the implementation in this driver returns an 'int'. Fix this by returning 'netdev_tx_t' in this driver too. Signed-off-by:
Luc Van Oostenryck <luc.vanoostenryck@gmail.com> Reviewed-by:
-
Luc Van Oostenryck authored
The method ndo_start_xmit() is defined as returning an 'netdev_tx_t', which is a typedef for an enum type, but the implementation in this driver returns an 'int'. Fix this by returning 'netdev_tx_t' in this driver too. Signed-off-by:
-
