- 27 Mar, 2018 16 commits
-
-
Parav Pandit authored
IB core maintains the GID cache entries for the GID table. This cache table has to be maintained regardless of HCA's support of GID table. For IB and iWarp ports, cache is created by querying the HCA. For RoCE cache is created based on netdev events. Therefore just refer to the RoCE port property of the {device, port} to decide whether to build cache by querying HCA or from netdev events. There is no need to check if HCA support GID table or not. ib_cache_update() referred to RoCE attribute before validating port. Though in all current callers port is valid, it is incorrect to query RoCE port property before validating the port. Therefore, rdma_protocol_roce() check is done after rdma_is_port_valid() verifies that port is valid. Fixes: 115b68aa ("IB/ocrdma: Removed GID add/del null routines") Reviewed-by:Daniel Jurgens <danielj@mellanox.com> Signed-off-by:
Parav Pandit <parav@mellanox.com> Signed-off-by:
Leon Romanovsky <leonro@mellanox.com> Signed-off-by:
Jason Gunthorpe <jgg@mellanox.com>
-
Parav Pandit authored
Even though API is only used by IPoIB driver, its incorrect to refer RoCE GID table property to search for GID. Look for only IB link layer to search for the GID. Fixes: dbb12562 ("IB/{core, ipoib}: Simplify ib_find_gid to search only for IB link layer") Signed-off-by:
-
Parav Pandit authored
ib_find_gid_by_filter() searches GID with filter only for RoCE link layer regardless of HCA's support for GID table. Therefore, right way to lookup is compare RoCE port property and not the GID table property. Fixes: 99b27e3b ("IB/cache: Add ib_find_gid_by_filter cache API") Signed-off-by:
-
Parav Pandit authored
Due to following reasons, GID table event is generated regardless of GID table property. 1. GID table cache is maintained at ib core layer regardless of link layer. 2. GID change event has no relation with IB link layer. 3. GID change event also doesn't depend on whether HCA supports GID table or not. Fixes: f3906bd3 ("IB/core: Refactor GID cache's ib_dispatch_event") Signed-off-by:
-
Parav Pandit authored
Due to below reasons, it is better to not support alternate path receive messages for RoCE in near term. 1. Alternate path for RoCE is not supported at rdmacm layer. 2. It is not supported in uverbs/core layer for RoCE. 3. Alternate path for IPv6 for link local address cannot resolve route determinstically without a valid incoming interface id whose usecase make sense only with dual port mode. 4. init_av_from_path while processing LAP messages for IB and RoCE can lead to adding duplicate entry of AV into the port list, leads to list corruption. 5. rdma-core userspace a well known userspace implementation has removed support of libucm which use ucm.ko module, which is the only module that can trigger alternate path related messages. 6. ucm kernel module is requested to be removed from the IB core in patch [1]. [1] https://patchwork.kernel.org/patch/10268503/Signed-off-by:
-
Mark Bloch authored
Currently access to hardware stats buffer isn't protected, this can result in multiple writes and reads at the same time to the same memory location. This can lead to providing an incorrect value to the user. Add a mutex to protect against it. Fixes: b40f4757 ("IB/core: Make device counter infrastructure dynamic") Signed-off-by:
Mark Bloch <markb@mellanox.com> Signed-off-by:
-
Majd Dibbiny authored
In some firmware configuration, UMR usage from Virtual Functions is restricted. This information is published to the driver using new capability bits. Avoid using UMRs in these cases and use the Firmware slow-path flow to create mkeys and populate them with Virtual to Physical address translation. Older drivers that do not have this patch, will end up using memory keys that aren't populated with Virtual to Physical address translation that is done part of the UMR work. Reviewed-by:
Majd Dibbiny <majd@mellanox.com> Signed-off-by:
Laurence Oberman <loberman@redhat.com> Signed-off-by:
-
Majd Dibbiny authored
When working with RC QPs, the FW sets the ECN capable bits for all the RoCE v2 packets. On the other hand, for UD QPs, the driver needs to set the the ECN capable bits in the Address Handler since the HW generates each packet according to the Address Handler and not the QP context. If ECN is not enabled in NIC or switch, these bits are ignored. Fixes: 2811ba51 ("IB/mlx5: Add RoCE fields to Address Vector") Reviewed-by:
-
Matan Barak authored
The ioctl() UAPIs are meant to be used by both user-space and kernel ioctl() handlers. Mostly, these UAPI structs tend to consist of simple types, but sometimes user-space pointers may be passed between user-space and kernel. We would like to avoid dereferencing a user-space pointer in the kernel, thus - we always define RDMA_UAPI_PTR as a __aligned_u64 type. Fixes: 1f7ff9d5 ('IB/uverbs: Move to new headers and make naming consistent') Signed-off-by:
Matan Barak <matanb@mellanox.com> Signed-off-by:
-
Jason Gunthorpe authored
The design of the uAPI had intended all structs to share the same layout on 32 and 64 bit compiles. Unfortunately over the years some errors have crept in. This series fixes all the incompatabilities. It goes along with a userspace rdma-core series that causes the providers to use these structs directly and then does various self-checks on the command formation. Those checks were combined with output from pahole on 32 and 64 bit compiles to confirm that the structure layouts are the same. This series does not make implicit padding explicit, as long as the implicit padding is the same on 32 and 64 bit compiles. Finally, the issue is put to rest by using __aligned_u64 in the uapi headers, if new code copies that type, and is checked in userspace, it is unlikely we will see problems in future. There are two patches that break the ABI for a 32 bit kernel, one for rxe and one for mlx4. Both patches have notes, but the overall feeling from Doug and I is that providing compat is just too difficult and not necessary since there is no real user of a 32 bit userspace and 32 bit kernel for various good reasons. The 32 bit userspace / 64 bit kernel case however does seem to have some real users and does need to work as designed. * 32compat: RDMA: Change all uapi headers to use __aligned_u64 instead of __u64 RDMA/rxe: Fix uABI structure layouts for 32/64 compat RDMA/mlx4: Fix uABI structure layouts for 32/64 compat RDMA/qedr: Fix uABI structure layouts for 32/64 compat RDMA/ucma: Fix uABI structure layouts for 32/64 compat RDMA: Remove minor pahole differences between 32/64
-
Jason Gunthorpe authored
The new auditing standard for the subsystem will be to only use __aligned_64 in uapi headers to try and prevent 32/64 compat bugs from existing in the future. Changing all existing usage will help ensure new developers copy the right idea. The before/after of this patch was tested using pahole on 32 and 64 bit compiles to confirm it has no change in the structure layout, so this patch is a NOP. Signed-off-by: -
Jason Gunthorpe authored
With 32 bit compilation several of the fields become misaligned here. Fixing this is an ABI break for 32 bit rxe and it is in well used portions of the rxe ABI. To handle this we bump the ABI version, as expected. However the user space driver doesn't handle it properly today, so all existing user space continues to work. Updated userspace will start to require the necessary kernel version. We don't expect there to be any 32 bit users of rxe. Most likely cases, such as ARM 32 already generally don't work because rxe does not handle the CPU cache properly on its shared with userspace pages. Signed-off-by: -
Jason Gunthorpe authored
rss_caps in struct mlx4_uverbs_ex_query_device_resp is misaligned on 32 bit compared to 64 bit, add explicit padding. The rss caps were introduced recently and are very rarely used in user space, mainly for DPDK. We don't expect there to be a real 32 bit user, so this change is done without compat considerations. Fixes: 09d208b2 ("IB/mlx4: Add report for RSS capabilities by vendor channel") Signed-off-by:
-
Jason Gunthorpe authored
struct qedr_alloc_ucontext_resp is a different length in 32 and 64 bit compiles due to implicit compiler padding. The structs alloc_pd_uresp, create_cq_uresp and create_qp_uresp are not padded by the compiler, but in user space the compiler pads them due to the way the core and driver structs are concatenated. Make this padding explicit and consistent for future sanity. The kernel driver can already handle the user buffer being smaller than required and copies correctly, so no compat or ABI break happens from introducing the explicit padding. Acked-by:
Michal Kalderon <michal.kalderon@cavium.com> Signed-off-by:
-
Jason Gunthorpe authored
The rdma_ucm_event_resp is a different length on 32 and 64 bit compiles. The kernel requires it to be the expected length or longer so 32 bit builds running on a 64 bit kernel will not work. Retain full compat by having all kernels accept a struct with or without the trailing reserved field. Signed-off-by: -
Jason Gunthorpe authored
To help automatic detection we want pahole to report the same struct layouts for 32 and 64 bit compiles. These cases are all implicit padding added at the end of embedded structs as part of a union. The added reserved fields have no impact on the ABI. Signed-off-by:
-
- 23 Mar, 2018 3 commits
-
-
Jason Gunthorpe authored
The udata's for alloc_pd cannot contain u64s due to alignment constraints. Switch the two never-used u64's to arrays of u32 to reduce the required struct alignment to 4 bytes. These reserved fields are totally unnecessary, never written and never read. Signed-off-by: -
Steve Wise authored
This is useful to rdma ULPs. Signed-off-by:
Steve Wise <swise@opengridcomputing.com> Signed-off-by:
-
Parav Pandit authored
Currently CM request for RoCE follows following flow. rdma_create_id() rdma_resolve_addr() rdma_resolve_route() For RC QPs: rdma_connect() ->cma_connect_ib() ->ib_send_cm_req() ->cm_init_av_by_path() ->ib_init_ah_attr_from_path() For UD QPs: rdma_connect() ->cma_resolve_ib_udp() ->ib_send_cm_sidr_req() ->cm_init_av_by_path() ->ib_init_ah_attr_from_path() In both the flows, route is already resolved before sending CM requests. Therefore, code is refactored to avoid resolving route second time in ib_cm layer. ib_init_ah_attr_from_path() is extended to resolve route when it is not yet resolved for RoCE link layer. This is achieved by caller setting route_resolved field in path record whenever it has route already resolved. Signed-off-by:
-
- 22 Mar, 2018 4 commits
-
-
Parav Pandit authored
ib_query_gid() in commit [1] refers to RoCE GID table capability of the HCA using rdma_cap_roce_gid_table(). ib_core maintains the GID table cache regardless of the HCA provider drivers capability to maintain RoCE GID table. Therefore, whether to return a GID table entry from the software cache or from HCA should be done based on whether the port is RoCE or not. [1] commit 03db3a2d ("IB/core: Add RoCE GID table management") Reviewed-by:
-
Leon Romanovsky authored
The restrack clean routine had simple, but powerful WARN_ON check to see if all resources are cleared prior to releasing device. The WARN_ON check performed very well, but lack of information which device caused to resource leak, the object type and origin made debug to be fun and challenging at the same time. The fact that all dumps were the same because restrack_clean() is called in dealloc() didn't help either. So let's fix spelling error and convert WARN_ON to be more debug friendly. The dmesg cut below gives example of how the output will look output for the case fixed in patch [1] [ 438.421372] restrack: ------------[ cut here ]------------ [ 438.423448] restrack: BUG: RESTRACK detected leak of resources on mlx5_2 [ 438.425600] restrack: Kernel PD object allocated by mlx5_ib is not freed [ 438.427753] restrack: Kernel CQ object allocated by mlx5_ib is not freed [ 438.429660] restrack: ------------[ cut here ]------------ [1] https://patchwork.kernel.org/patch/10298695/ Cc: Michal Kalderon <Michal.Kalderon@cavium.com> Cc: Chuck Lever <chuck.lever@oracle.com> Reviewed-by:
-
Yixian Liu authored
Upon detecting both kernel and user space support record doorbell, the kernel needs to enable this capability in hardware by db_en, and it should take place before cq context configuration in hns_roce_cq_alloc. Currently, db_en is configured after cq alloc and db_map_user has similar problem. Reported-by:
Xiping Zhang <zhangxiping3@huawei.com> Fixes: 9b44703d ("RDMA/hns: Support cq record doorbell for the user space") Signed-off-by:
Yixian Liu <liuyixian@huawei.com> Signed-off-by:
Lijun Ou <oulijun@huawei.com> Signed-off-by:
-
Jason Gunthorpe authored
Open coding a loose value is not acceptable for describing the uABI in RDMA. Provide the missing struct. Acked-by:
-
- 21 Mar, 2018 1 commit
-
-
Sinan Kaya authored
Code includes wmb() followed by writel(). writel() already has a barrier on some architectures like arm64. This ends up CPU observing two barriers back to back before executing the register write. Since code already has an explicit barrier call, changing writel() to writel_relaxed(). Signed-off-by:
Sinan Kaya <okaya@codeaurora.org> Reviewed-by:
-
- 19 Mar, 2018 16 commits
-
-
Matan Barak authored
Enable the ioctl() uAPI for IB by default if the standard write() uAPI (INFINIBAND_USER_ACCESS) is enabled. Verbs that are also available under the old write() uAPI are put inside a new INFINIBAND_EXP_LEGACY_VERBS_NEW_UAPI Kconfig. Reviewed-by:
Yishai Hadas <yishaih@mellanox.com> Signed-off-by:
-
Matan Barak authored
Previously, adding driver specific attributes required drivers to declare all the hierarchy - object tree, object, methods and the attributes themselves. A common use case is adding a few attributes to an existing common method. In order to simplify the driver's code, we add some macros to do all these declarations automatically. Signed-off-by:
-
Matan Barak authored
Currently, all objects are declared in uverbs_std_types. This could lead to a huge file once we implement all objects, methods and handlers. Moving each object to its own file to keep the files smaller and more readable. uverbs_std_types.c will only contain the parsing tree definition and objects without any methods. Signed-off-by:
-
Matan Barak authored
The ioctl() based uverbs is based on merging feature trees. This teaches the generic parser how to parse methods according to the provider's support. In order to support merging with the common objects, exporting the common-object-tree to the provider drivers. Reviewed-by:
-
Matan Barak authored
Previously, we've used UVERBS_ATTR_SPEC_F_MIN_SZ for extending existing attributes. The behavior of this flag was the kernel accepts anything bigger than the minimum size it specified. This is unsafe, since in order to safely extend an attribute, we need to make sure unknown size is zeroed. Replacing UVERBS_ATTR_SPEC_F_MIN_SZ with UVERBS_ATTR_SPEC_F_MIN_SZ_OR_ZERO, which essentially checks that the unknown size is zero. In addition, attributes are now decorated with UVERBS_ATTR_TYPE and UVERBS_ATTR_STRUCT, so we can provide the minimum and known length. Users of this flag needs to use copy_from_or_zero functions/macros. Reviewed-by:
-
Matan Barak authored
Downstream patches extend uverbs_attr_spec with new fields. In order to save space, we move the type and flags fields to the various attribute flavors contained in the union. Reviewed-by:
-
Matan Barak authored
Extending uverbs_ioctl header with driver_id and another reserved field. driver_id should be used in order to identify the driver. Since every driver could have its own parsing tree, this is necessary for strace support. Downstream patches take off the EXPERIMENTAL flag from the ioctl() IB support and thus we add some reserved fields for future usage. Reviewed-by:
-
Matan Barak authored
Use macros to make names consistent in ioctl() uAPI: The ioctl() uAPI works with object-method hierarchy. The method part also states which handler should be executed when this method is called from user-space. Therefore, we need to tie method, method's id, method's handler and the object owning this method together. Previously, this was done through explicit developer chosen names. This makes grepping the code harder. Changing the method's name, method's handler and object's name to be automatically generated based on the ids. The headers are split in a way so they be included and used by user-space. One header strictly contains structures that are used directly by user-space applications, where another header is used for internal library (i.e. libibverbs) to form the ioctl() commands. Other header simply contains the required general command structure. Reviewed-by:
-
Bart Van Assche authored
According to the SRP standard the INITIATOR and TARGET PORT IDENTIFIER fields from the login request specify the I_T nexus. Whether or not an SRP target closes an existing connection for an I_T nexus when a login request is received depends on the value of the MULTICHANNEL field in the login request. The SRP initiator derives the value of the INITIATOR and TARGET PORT IDENTIFIER fields from the .id_ext, .ioc_guid, .initiator_ext .sgid members of the srp_target_port structure. This means that the .rdma_cm.dst check must be removed from srp_conn_unique(). This patch avoids that for target ports that have multiple addresses, e.g. an IPv4 and an IPv6 address, and if a connection is established to both target port addresses, that the initiator logs in alternatingly every 10 seconds to the other target port address. An SRP target must namely terminate all but one connections for a given I_T nexus if the MULTICHANNEL field has not been set in the login request. Fixes: 19f31343 ("IB/srp: Add RDMA/CM support") Signed-off-by:
Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by:
-
Bodong Wang authored
Enable RAW QP to be able to configure burst control by modify_qp. By using burst control with rate limiting, user can achieve best performance and accuracy. The burst control information is passed by user through udata. This patch also reports burst control capability for mlx5 related hardwares, burst control is only marked as supported when both packet_pacing_burst_bound and packet_pacing_typical_size are supported. Signed-off-by:
Bodong Wang <bodong@mellanox.com> Reviewed-by:
-
Bodong Wang authored
Add two new parameters: max_burst_sz and typical_pkt_size (both in bytes) to rate limit configurations. max_burst_sz: The device will schedule bursts of packets for an SQ connected to this rate, smaller than or equal to this value. Value 0x0 indicates packet bursts will be limited to the device defaults. This field should be used if bursts of packets must be strictly kept under a certain value. typical_pkt_size: When the rate limit is intended for a stream of similar packets, stating the typical packet size can improve the accuracy of the rate limiter. The expected packet size will be the same for all SQs associated with the same rate limit index. Ethernet driver is updated according to this change, but these two parameters will be kept as 0 due to lacking of proper way to get the configurations from user space which requires to change ndo_set_tx_maxrate interface. Signed-off-by:
-
Yixian Liu authored
The data in resp will be copied from kernel to userspace, thus it needs to be initialized to zeros to avoid copying uninited stack memory. Reported-by:
Dan Carpenter <dan.carpenter@oracle.com> Fixes: e088a685 ("RDMA/hns: Support rq record doorbell for the user space") Signed-off-by:
-
Parav Pandit authored
ib_peek_cq() verb doesn't seem be implemented in current code. There is some past reference to it at [1] about it being unimplemented. Lot of user documentation created out of kdoc refers to this unimplemented API. Therefore, remove unimplemented API. [1] http://lists.openfabrics.org/pipermail/ofw/2008-May/002465.htmlSigned-off-by:
-
Parav Pandit authored
Use rdma_is_port_valid() which performs port validity check instead of open coding the same check. Signed-off-by:
Yuval Shaia <yuval.shaia@oracle.com> Signed-off-by:
-
Jason Gunthorpe authored
What is going on here is a bit subtle, in the kernel there is no problem because the struct is copied using copy_from_user, so it can safely have an 8 byte alignment, however in userspace it must be constructed by concatenation with the ib_uverbs_alloc_pd_resp struct. This is due to the required memory layout to execute the command. Since ibv_uverbs_alloc_pd_resp is only 4 bytes long, this causes misalignment, and the user space will experience an unexpected padding. Currently it works around this via pointer maths. Make everything more robust by having the compiler reduce the alignment of the struct to 4. The userspace has assertions to ensure this works properly in all situations. Signed-off-by: -
Honggang Li authored
Before commit f1b65df5 ("IB/mlx5: Add support for active_width and active_speed in RoCE"), the mlx5_ib driver set the default active_width and active_speed to IB_WIDTH_4X and IB_SPEED_QDR. When the RoCE port is down, the RoCE port does not negotiate the active width with the remote side, causing the active width to be zero. When running userspace ibstat to view the port status, ibstat will panic as it reads an invalid width from sys file. This patch restores the original behavior. Fixes: f1b65df5 ("IB/mlx5: Add support for active_width and active_speed in RoCE"). Signed-off-by:
Honggang Li <honli@redhat.com> Reviewed-by:
Hal Rosenstock <hal@mellanox.com> Reviewed-by:
Noa Osherovich <noaos@mellanox.com> Signed-off-by:
-
