- 27 Jan, 2020 1 commit
-
-
Håkon Bugge authored
Using CX-3 virtual functions, either from a bare-metal machine or pass-through from a VM, MAD packets are proxied through the PF driver. Since the VF drivers have separate name spaces for MAD Transaction Ids (TIDs), the PF driver has to re-map the TIDs and keep the book keeping in a cache. Following the RDMA Connection Manager (CM) protocol, it is clear when an entry has to evicted from the cache. When a DREP is sent from mlx4_ib_multiplex_cm_handler(), id_map_find_del() is called. Similar when a REJ is received by the mlx4_ib_demux_cm_handler(), id_map_find_del() is called. This function wipes out the TID in use from the IDR or XArray and removes the id_map_entry from the table. In short, it does everything except the topping of the cake, which is to remove the entry from the list and free it. In other words, for the REJ case enumerated above, one id_map_entry will be leaked. For the other case above, a DREQ has been received first. The reception of the DREQ will trigger queuing of a delayed work to delete the id_map_entry, for the case where the VM doesn't send back a DREP. In the normal case, the VM _will_ send back a DREP, and id_map_find_del() will be called. But this scenario introduces a secondary leak. First, when the DREQ is received, a delayed work is queued. The VM will then return a DREP, which will call id_map_find_del(). As stated above, this will free the TID used from the XArray or IDR. Now, there is window where that particular TID can be re-allocated, lets say by an outgoing REQ. This TID will later be wiped out by the delayed work, when the function id_map_ent_timeout() is called. But the id_map_entry allocated by the outgoing REQ will not be de-allocated, and we have a leak. Both leaks are fixed by removing the id_map_find_del() function and only using schedule_delayed(). Of course, a check in schedule_delayed() to see if the work already has been queued, has been added. Another benefit of always using the delayed version for deleting entries, is that we do get a TimeWait effect; a TID no longer in use, will occupy the XArray or IDR for CM_CLEANUP_CACHE_TIMEOUT time, without any ability of being re-used for that time period. Fixes: 3cf69cc8 ("IB/mlx4: Add CM paravirtualization") Link: https://lore.kernel.org/r/20200123155521.1212288-1-haakon.bugge@oracle.comSigned-off-by:
Håkon Bugge <haakon.bugge@oracle.com> Signed-off-by:
Manjunath Patil <manjunath.b.patil@oracle.com> Reviewed-by:
Rama Nichanamatlu <rama.nichanamatlu@oracle.com> Reviewed-by:
Jack Morgenstein <jackm@dev.mellanox.co.il> Signed-off-by:
Jason Gunthorpe <jgg@mellanox.com>
-
- 25 Jan, 2020 11 commits
-
-
Dillon Brock authored
Correcting a minor spelling mistake in the comments. Link: https://lore.kernel.org/r/20200118162542.15188-1-dab9861@gmail.comSigned-off-by:
Dillon Brock <dab9861@gmail.com> Acked-by:
Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by:
-
Nathan Chancellor authored
Clang warns: drivers/infiniband/hw/hfi1/msix.c:136:22: warning: overlapping comparisons always evaluate to false [-Wtautological-overlap-compare] if (type < IRQ_SDMA && type >= IRQ_OTHER) ~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~ 1 warning generated. It is impossible for something to be less than 0 (IRQ_SDMA) and greater than or equal to 3 (IRQ_OTHER) at the same time. A logical OR should have been used to keep the same logic as before. Link: https://lore.kernel.org/r/20200116222658.5285-1-natechancellor@gmail.com Link: https://github.com/ClangBuiltLinux/linux/issues/841 Fixes: 13d2a838 ("IB/hfi1: Decouple IRQ name from type") Signed-off-by:Nathan Chancellor <natechancellor@gmail.com> Reviewed-by:
Nick Desaulniers <ndesaulniers@google.com> Acked-by:
-
Jason Gunthorpe authored
All accesses now use the new IBA acessor scheme, so delete the structs entirely and generate the structures from the schema file. Link: https://lore.kernel.org/r/20200116170037.30109-8-jgg@ziepe.caTested-by:
Leon Romanovsky <leonro@mellanox.com> Reviewed-by:
-
Jason Gunthorpe authored
Use a Coccinelle spatch to replace CM structure members used as structures, arrays, or pointers with IBA_GET/SET versions. Applied with $ spatch --sp-file edits.sp --in-place drivers/infiniband/core/cm.c The spatch file was generated using the template pattern: @@ expression src; expression len; {struct} *msg; @@ - memcpy(msg->{old_name}, src, len) + IBA_SET_MEM({new_name}, msg, src, len) @@ {struct} *msg; identifier x; @@ - msg->{old_name}.x + IBA_GET_MEM_PTR({new_name}, msg)->x @@ {struct} *msg; @@ - &msg->{old_name} + IBA_GET_MEM_PTR({new_name}, msg) For GIDs: @@ {struct} *msg; @@ - msg->{old_name} + *IBA_GET_MEM_PTR({new_name}, msg) For non-GIDs: @@ {struct} *msg; @@ - msg->{old_name} + IBA_GET_MEM_PTR({new_name}, msg) Iterated for every remaining IBA_CHECK_OFF()/IBA_CHECK_GET() pairing. Touched up with clang-format after. Link: https://lore.kernel.org/r/20200116170037.30109-7-jgg@ziepe.caTested-by: -
Jason Gunthorpe authored
Use a Coccinelle spatch script to replace use of simple CM structure members with IBA_GET/SET versions. Applied with $ spatch --sp-file edits.sp --in-place drivers/infiniband/core/cm.c The spatch file was generated using the template pattern: @@ expression val; {struct} *msg; @@ - msg->{old_name} = val + IBA_SET({new_name}, msg, be{bits}_to_cpu(val)) @@ {struct} *msg; @@ - msg->{old_name} + cpu_to_be{bits}(IBA_GET({new_name}, msg)) Iterated for every IBA_CHECK_OFF that isn't a CM_FIELD_MLOC. And the below iterated over all byte sizes to remove doubled byte swaps: @@ expression val; @@ -be{bits}_to_cpu(cpu_to_be{bits}(val)) +val (and __be_to_cpu and ntoh varients) Touched up with clang-format after. Link: https://lore.kernel.org/r/20200116170037.30109-6-jgg@ziepe.caTested-by: -
Jason Gunthorpe authored
Use a Coccinelle spatch script to replace CM helper functions that return/accept BE values with IBA_GET/SET versions. Applied with $ spatch --sp-file edits.sp --in-place drivers/infiniband/core/cm.c The spatch file was generated using the template pattern: @@ expression val; {struct} *msg; @@ - {old_setter}(msg, val) + IBA_SET({new_name}, msg, be{bits}_to_cpu(val)) @@ {struct} *msg; @@ - {old_getter}(msg) + cpu_to_be{bits}(IBA_GET({new_name}, msg)) Iterated for every IBA_CHECK_GET_BE()/IBA_CHECK_SET_BE() pairing. And the below iterated over all byte sizes to remove doubled byte swaps: @@ expression val; @@ -be{bits}_to_cpu(cpu_to_be{bits}(val)) +val (and __be_to_cpu and ntoh varients) Touched up with clang-format after. Link: https://lore.kernel.org/r/20200116170037.30109-5-jgg@ziepe.caTested-by: -
Jason Gunthorpe authored
Use a Coccinelle spatch to replace CM helper functions with IBA_GET/SET versions. Applied with $ spatch --sp-file edits.sp --in-place drivers/infiniband/core/cm.c The spatch file was generated using the template pattern: @@ expression val; {struct} *msg; @@ - {old_setter} + IBA_SET({new_name}, msg, val) @@ {struct} *msg; @@ - {old_getter} + IBA_GET({new_name}, msg) Iterated for every IBA_CHECK_GET()/IBA_CHECK_GET() pairing. Touched up with clang-format after. Link: https://lore.kernel.org/r/20200116170037.30109-4-jgg@ziepe.caTested-by: -
Leon Romanovsky authored
There is no separation between RDMA-CM wire format as it is declared in IBTA and kernel logic which implements needed support. Such situation causes to many mistakes in conversion between big-endian (wire format) and CPU format used by kernel. It also mixes RDMA core code with combination of uXX and beXX variables. The idea that all accesses to IBA definitions will go through special GET/SET macros to ensure that no conversion mistakes are made. The shifting and masking required to read the value is automatically deduced using the field offset description from the tables in the IBA specification. This starts with the CM MADs described in IBTA release 1.3 volume 1. To confirm that the new macros behave the same as the old accessors a self-test is included in this patch. Each macro replacing a straightforward struct field compile-time tests that the new field has the same offsetof() and width as the old field. For the fields with accessor functions a runtime test, the 'all ones' value is placed in a dummy message and read back in several ways to confirm that both approaches give identical results. Later patches in this series delete the self test. This creates a tested table of new field name, old field name(s) and some meta information like BE coding for the functions which will be used in the next patches. Link: https://lore.kernel.org/r/20200116170037.30109-3-jgg@ziepe.ca Link: https://lore.kernel.org/r/20191212093830.316934-5-leon@kernel.orgSigned-off-by:
-
Jason Gunthorpe authored
Access the two fields through wrappers, like all other fields, to make it clearer what is happening. Link: https://lore.kernel.org/r/20200116170037.30109-2-jgg@ziepe.caTested-by:
-
Danit Goldberg authored
A user can change the operational GUID (a.k.a affective GUID) through link/infiniband. Therefore it is preferred to return the currently set GUID if it exists instead of the operational. This way the PF can query which VF GUID will be set in the next bind. In order to align with MAC address, zero is returned if administrative GUID is not set. For example, before setting administrative GUID: $ ip link show ib0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 4092 qdisc mq state UP mode DEFAULT group default qlen 256 link/infiniband 00:00:00:08:fe:80:00:00:00:00:00:00:52:54:00:c0:fe:12:34:55 brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff vf 0 link/infiniband 00:00:00:08:fe:80:00:00:00:00:00:00:52:54:00:c0:fe:12:34:55 brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff, spoof checking off, NODE_GUID 00:00:00:00:00:00:00:00, PORT_GUID 00:00:00:00:00:00:00:00, link-state auto, trust off, query_rss off Then: $ ip link set ib0 vf 0 node_guid 11:00:af:21:cb:05:11:00 $ ip link set ib0 vf 0 port_guid 22:11:af:21:cb:05:11:00 After setting administrative GUID: $ ip link show ib0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 4092 qdisc mq state UP mode DEFAULT group default qlen 256 link/infiniband 00:00:00:08:fe:80:00:00:00:00:00:00:52:54:00:c0:fe:12:34:55 brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff vf 0 link/infiniband 00:00:00:08:fe:80:00:00:00:00:00:00:52:54:00:c0:fe:12:34:55 brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff, spoof checking off, NODE_GUID 11:00:af:21:cb:05:11:00, PORT_GUID 22:11:af:21:cb:05:11:00, link-state auto, trust off, query_rss off Fixes: 9c0015ef ("IB/mlx5: Implement callbacks for getting VFs GUID attributes") Link: https://lore.kernel.org/r/20200116120048.12744-1-leon@kernel.orgSigned-off-by:
Danit Goldberg <danitg@mellanox.com> Signed-off-by:
-
Jason Gunthorpe authored
The set of entry->driver_removed is missing locking, protect it with xa_lock() which is held by the only reader. Otherwise readers may continue to see driver_removed = false after rdma_user_mmap_entry_remove() returns and may continue to try and establish new mmaps. Fixes: 3411f9f0 ("RDMA/core: Create mmap database and cookie helper functions") Link: https://lore.kernel.org/r/20200115202041.GA17199@ziepe.caReviewed-by:
Gal Pressman <galpress@amazon.com> Acked-by:
Michal Kalderon <michal.kalderon@marvell.com> Signed-off-by:
-
- 21 Jan, 2020 1 commit
-
-
Jason Gunthorpe authored
From https://git.kernel.org/pub/scm/linux/kernel/git/leon/linux-rdma Leon Romanovsky says: ==================== Use ODP MRs for kernel ULPs The following series extends MR creation routines to allow creation of user MRs through kernel ULPs as a proxy. The immediate use case is to allow RDS to work over FS-DAX, which requires ODP (on-demand-paging) MRs to be created and such MRs were not possible to create prior this series. The first part of this patchset extends RDMA to have special verb ib_reg_user_mr(). The common use case that uses this function is a userspace application that allocates memory for HCA access but the responsibility to register the memory at the HCA is on an kernel ULP. This ULP acts as an agent for the userspace application. The second part provides advise MR functionality for ULPs. This is integral part of ODP flows and used to trigger pagefaults in advance to prepare memory before running working set. The third part is actual user of those in-kernel APIs. ==================== * tag 'rds-odp-for-5.5': net/rds: Use prefetch for On-Demand-Paging MR net/rds: Handle ODP mr registration/unregistration net/rds: Detect need of On-Demand-Paging memory registration RDMA/mlx5: Fix handling of IOVA != user_va in ODP paths IB/mlx5: Mask out unsupported ODP capabilities for kernel QPs RDMA/mlx5: Don't fake udata for kernel path IB/mlx5: Add ODP WQE handlers for kernel QPs IB/core: Add interface to advise_mr for kernel users IB/core: Introduce ib_reg_user_mr IB: Allow calls to ib_umem_get from kernel ULPs Signed-off-by:
-
- 18 Jan, 2020 2 commits
-
-
Hans Westgaard Ry authored
Try prefetching pages when using On-Demand-Paging MR using ib_advise_mr. Signed-off-by:
Hans Westgaard Ry <hans.westgaard.ry@oracle.com> Acked-by:
Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by:
-
Hans Westgaard Ry authored
On-Demand-Paging MRs are registered using ib_reg_user_mr and unregistered with ib_dereg_mr. Signed-off-by:
-
- 16 Jan, 2020 22 commits
-
-
Jack Morgenstein authored
In procedure mlx4_ib_add_gid(), if the driver is unable to update the FW gid table, there is a memory leak in the driver's copy of the gid table: the gid entry's context buffer is not freed. If such an error occurs, free the entry's context buffer, and mark the entry as available (by setting its context pointer to NULL). Fixes: e26be1bf ("IB/mlx4: Implement ib_device callbacks") Link: https://lore.kernel.org/r/20200115085050.73746-1-leon@kernel.orgSigned-off-by:
Parav Pandit <parav@mellanox.com> Signed-off-by:
-
Avihai Horon authored
Introduce the following RoCE accelerator counters: * roce_adp_retrans - number of adaptive retransmission for RoCE traffic. * roce_adp_retrans_to - number of times RoCE traffic reached time out due to adaptive retransmission. * roce_slow_restart - number of times RoCE slow restart was used. * roce_slow_restart_cnps - number of times RoCE slow restart generate CNP packets. * roce_slow_restart_trans - number of times RoCE slow restart change state to slow restart. Link: https://lore.kernel.org/r/20200115145459.83280-3-leon@kernel.orgSigned-off-by:
Avihai Horon <avihaih@mellanox.com> Reviewed-by:
Maor Gottlieb <maorg@mellanox.com> Signed-off-by:
-
Michael Guralnik authored
Enable relaxed ordering in the mkey context when requested. As relaxed ordering is not currently supported in UMR, disable UMR usage for relaxed ordering MRs. Link: https://lore.kernel.org/r/1578506740-22188-11-git-send-email-yishaih@mellanox.comSigned-off-by:
Michael Guralnik <michaelgur@mellanox.com> Signed-off-by:
Yishai Hadas <yishaih@mellanox.com> Signed-off-by:
-
Michael Guralnik authored
Add the core support field to METHOD_GET_CONTEXT, this field should represent capabilities that are not device-specific. Return support for optional access flags for memory regions. User-space will use this capability to mask the optional access flags for unsupporting kernels. Link: https://lore.kernel.org/r/1578506740-22188-10-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Michael Guralnik authored
Add a new relaxed ordering access flag for memory regions. Using memory regions with relaxed ordeing set can enhance performance. This access flag is handled in a best-effort manner, drivers should ignore if they don't support setting relaxed ordering. Link: https://lore.kernel.org/r/1578506740-22188-9-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Michael Guralnik authored
As part of adding a range of optional access flags that drivers need to be able to accept, mask this range inside efa driver. This will prevent the driver from failing when an access flag from that range is passed. Link: https://lore.kernel.org/r/1578506740-22188-8-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Michael Guralnik authored
Define a range of access flags that are defined to be optional, both uverbs and drivers should enable getting them and use if they are applicable This will be used, for example, for the relaxed ordering access flag which unsupporting drivers can ignore. Link: https://lore.kernel.org/r/1578506740-22188-7-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Michael Guralnik authored
Verify that MR access flags that are passed from user are all supported ones, otherwise an error is returned. Fixes: 4fca0377 ("IB/uverbs: Move ib_access_flags and ib_read_counters_flags to uapi") Link: https://lore.kernel.org/r/1578506740-22188-6-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Jason Gunthorpe authored
Allow future extensions of the get context command through the uverbs ioctl kabi. Unlike the uverbs version this does not return an async_fd as well, that has to be done with another command. Link: https://lore.kernel.org/r/1578506740-22188-5-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Jason Gunthorpe authored
This lock only serializes ucontext creation. Instead of checking the ucontext_lock during destruction hold the existing hw_destroy_rwsem during creation, which is the standard pattern for object creation. The simplification of locking is needed for the next patch. Link: https://lore.kernel.org/r/1578506740-22188-4-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Jason Gunthorpe authored
Allow the async FD to be allocated separately from the context. This is necessary to introduce the ioctl to create a context, as an ioctl should only ever create a single uobject at a time. If multiple async FDs are created then the first one is used to deliver affiliated events from any ib_uevent_object, with all subsequent ones will receive only unaffiliated events. Link: https://lore.kernel.org/r/1578506740-22188-3-git-send-email-yishaih@mellanox.comSigned-off-by:
-
Jason Gunthorpe authored
From the mlx5-next branch at git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux Merged due to dependencies in the next patches. * branch 'mlx5-next': net/mlx5: Expose relaxed ordering bits net/mlx5: Add RoCE accelerator counters
-
Michael Guralnik authored
Expose relaxed ordering bits in HCA capability and mkey context structs. Signed-off-by:
-
Leon Romanovsky authored
Add RoCE accelerator definitions. Signed-off-by: -
Hans Westgaard Ry authored
Add code to check if memory intended for RDMA is FS-DAX-memory. RDS will fail with error code EOPNOTSUPP if FS-DAX-memory is detected. Signed-off-by:
-
Jason Gunthorpe authored
Till recently it was not possible for userspace to specify a different IOVA, but with the new ibv_reg_mr_iova() library call this can be done. To compute the user_va we must compute: user_va = (iova - iova_start) + user_va_start while being cautious of overflow and other math problems. The iova is not reliably stored in the mmkey when the MR is created. Only the cached creation path (the common one) set it, so it must also be set when creating uncached MRs. Fix the weird use of iova when computing the starting page index in the MR. In the normal case, when iova == umem.address: iova & (~(BIT(page_shift) - 1)) == ALIGN_DOWN(umem.address, odp->page_size) == ib_umem_start(odp) And when iova is different using it in math with a user_va is wrong. Finally, do not allow an implicit ODP to be created with a non-zero IOVA as we have no support for that. Fixes: 7bdf65d4 ("IB/mlx5: Handle page faults") Signed-off-by:
Moni Shoua <monis@mellanox.com> Signed-off-by:
-
Moni Shoua authored
The ODP handler for WQEs in RQ or SRQ is not implented for kernel QPs. Therefore don't report support in these if query comes from a kernel user. Signed-off-by:
-
Leon Romanovsky authored
Kernel paths must not set udata and provide NULL pointer, instead of faking zeroed udata struct. Signed-off-by: -
Moni Shoua authored
One of the steps in ODP page fault handler for WQEs is to read a WQE from a QP send queue or receive queue buffer at a specific index. Since the implementation of this buffer is different between kernel and user QP the implementation of the handler needs to be aware of that and handle it in a different way. ODP for kernel MRs is currently supported only for RDMA_READ and RDMA_WRITE operations so change the handler to - read a WQE from a kernel QP send queue - fail if access to receive queue or shared receive queue is required for a kernel QP Signed-off-by:
-
Moni Shoua authored
Allow ULPs to call advise_mr, so they can control ODP regions in the same way as user space applications. Signed-off-by:
-
Moni Shoua authored
Add ib_reg_user_mr() for kernel ULPs to register user MRs. The common use case that uses this function is a userspace application that allocates memory for HCA access but the responsibility to register the memory at the HCA is on an kernel ULP. This ULP that acts as an agent for the userspace application. This function is intended to be used without a user context so vendor drivers need to be aware of calling reg_user_mr() device operation with udata equal to NULL. Among all drivers, i40iw is the only driver which relies on presence of udata, so check udata existence for that driver. Signed-off-by:
Guy Levi <guyle@mellanox.com> Signed-off-by:
-
Moni Shoua authored
So far the assumption was that ib_umem_get() and ib_umem_odp_get() are called from flows that start in UVERBS and therefore has a user context. This assumption restricts flows that are initiated by ULPs and need the service that ib_umem_get() provides. This patch changes ib_umem_get() and ib_umem_odp_get() to get IB device directly by relying on the fact that both UVERBS and ULPs sets that field correctly. Reviewed-by:
-
- 15 Jan, 2020 3 commits
-
-
Sergey Gorenko authored
Some SRP targets that do not support specification SRP-2, put the garbage to the reserved bits of the SRP login response. The problem was not detected for a long time because the SRP initiator ignored those bits. But now one of them is used as SRP_LOGIN_RSP_IMMED_SUPP. And it causes a critical error on the target when the initiator sends immediate data. The ib_srp module has a use_imm_date parameter to enable or disable immediate data manually. But it does not help in the above case, because use_imm_date is ignored at handling the SRP login response. The problem is definitely caused by a bug on the target side, but the initiator's behavior also does not look correct. The initiator should not use immediate data if use_imm_date is disabled by a user. This commit adds an additional checking of use_imm_date at the handling of SRP login response to avoid unexpected use of immediate data. Fixes: 882981f4 ("RDMA/srp: Add support for immediate data") Link: https://lore.kernel.org/r/20200115133055.30232-1-sergeygo@mellanox.comSigned-off-by:
Sergey Gorenko <sergeygo@mellanox.com> Reviewed-by:
Bart Van Assche <bvanassche@acm.org> Signed-off-by:
-
Rao Shoaib authored
The SGE buffer size and max_inline data should be derived from the size of the WQE. Each value individually sets the WQE size, so compute the actual sizes based on the actual WQE size and configure the QP with the maximums. Also fix the missing return of the actual maximum capability to the caller. Link: https://lore.kernel.org/r/1578962480-17814-3-git-send-email-rao.shoaib@oracle.comSigned-off-by:
Rao Shoaib <rao.shoaib@oracle.com> Reviewed-by:
-
Rao Shoaib authored
Introduce maximum WQE size to impose limits on max SGE's and inline data Link: https://lore.kernel.org/r/1578962480-17814-2-git-send-email-rao.shoaib@oracle.comSigned-off-by:
-
