For setups having a large amount of PCI devices, it makes sense to limit the
number of MSIX vectors per PCI device, in order to avoid running out of IRQ
vectors.
MFC after: 1 week
Sponsored by: Mellanox Technologies
The scatter list is formed by the chunks of MCLBYTES each, and larger
than default packets are returned to the stack as the mbuf chain.
Submitted by: kib@
MFC after: 1 week
Sponsored by: Mellanox Technologies
This deduplicates the code a bit, and also implicitly adds missing
callout_stop() to in[6]_lltable_delete_entry() functions.
PR: 209682, 225927
Submitted by: hselasky (previous version)
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D4605
To access the data, set sysctl dev.mce.N.conf.debug_stats to 1.
This enables the sysctl node dev.mce.N.hw_ctx_debug. Its content is
the mapping of each channel' number to used receive queue and associated
completion queue, set of the transmit queues numbers and corresponding
completion queues.
Trimmed example output:
channel 30 rq 188 cq 1085
channel 30 tc 0 sq 187 cq 1084
channel 31 rq 191 cq 1087
channel 31 tc 0 sq 190 cq 1086
MFC after: 1 week
Sponsored by: Mellanox Technologies
and csqrtl().
When one component is huge and the other is tiny, scaling down the tiny
component gave spurious underflow.
When both components are denormal, not scaling them up gave inaccuracies
of 34+ ulps on not very carefully selected args. Fixing this reduces the
maximum error to 1.6 ulps on the same set of args (mosly not denormal ones).
The scaling used multiplication of a complex variable by 2, but clang messes
this on amd64 up by losing the sign of -0.0. Calculate the components
separately, as is well known to be needed for operations on more exceptional
values.
Device detach and setting error state may deadlock over the interface mutex
like this:
a) Detach code in mlx5en waits until error state is set while the interface
mutex is locked.
b) The set error handler needs to lock the interface mutex before it can
set the error state.
The solution is to use atomics to set the error state.
MFC after: 1 week
Sponsored by: Mellanox Technologies
When resetting mlx5core instances it can happen that the order of attach and
detach for mlx5ib instances is changed. Take the unit number for mlx5_%d from
the parent PCI device, similarly to what is done in mlx5en(4), so that there
is a direct relationship between mce<N> and mlx5_<N>.
MFC after: 1 week
Sponsored by: Mellanox Technologies
The DSCP feature is controlled using a set of sysctl(8) fields under
the qos sysctl directory entry for mlx5en(4).
For Routable RoCE QPs, the DSCP should be set in the QP's address path.
The DSCP's value is derived from the traffic class.
Linux commit:
ed88451e1f2d400fd6a743d0a481631cf9f97550
MFC after: 1 week
Sponsored by: Mellanox Technologies
When creating address handle from multicast GID, set MAC according to
the appropriate formula instead of searching for it in the GID table:
- For IPv4 multicast GID use ip_eth_mc_map().
- For IPv6 multicast GID use ipv6_eth_mc_map().
Linux commit:
9636a56fa864464896bf7d1272c701f2b9a57737
MFC after: 1 week
Sponsored by: Mellanox Technologies
The return status of ib_init_ah_from_mcmember() is ignored by
cma_ib_mc_handler(). Honor it and return error event if ah attribute
initialization failed.
Linux commit:
6d337179f28cc50ddd7e224f677b4cda70b275fc
MFC after: 1 week
Sponsored by: Mellanox Technologies
ah_attr contains the port number to which cm_id is bound. However, while
searching for GID table for matching GID entry, the port number is
ignored.
This could cause the wrong GID to be used when the ah_attr is converted to
an AH.
Linux commit:
563c4ba3bd2b8b0b21c65669ec2226b1cfa1138b
MFC after: 1 week
Sponsored by: Mellanox Technologies
The current implementation assumes a static mapping between
the TOS bits and the priority code point, PCP bits.
MFC after: 1 week
Sponsored by: Mellanox Technologies
When a loopback address is detected use the network interface which
has the loopback flag set to trigger loopback logic in address resolve.
MFC after: 1 week
Sponsored by: Mellanox Technologies
The ib_uverbs_create_ah() ind ib_uverbs_modify_qp() calls receive
the port number from user input as part of its attributes and assumes
it is valid. Down on the stack, that parameter is used to access kernel
data structures. If the value is invalid, the kernel accesses memory
it should not. To prevent this, verify the port number before using it.
Linux commit:
5ecce4c9b17bed4dc9cb58bfb10447307569b77b
a62ab66b13a0f9bcb17b7b761f6670941ed5cd62
5a7a88f1b488e4ee49eb3d5b82612d4d9ffdf2c3
MFC after: 1 week
Sponsored by: Mellanox Technologies
RoCEv1 does not use the IPv6 stack to resolve the link local DGID since it
uses GID address. It forms the DMAC directly from the DGID.
Linux commit:
56d0a7d9a0f045ee27a001762deac28c7d28e2e4
MFC after: 1 week
Sponsored by: Mellanox Technologies
This patch fixes the kernel crash that occurs during ib_dealloc_device()
called due to provider driver fails with an error after
ib_alloc_device() and before it can register using ib_register_device().
This crashed seen in tha lab as below which can occur with any IB device
which fails to perform its device initialization before invoking
ib_register_device().
This patch avoids touching cache and port immutable structures if device
is not yet initialized.
It also releases related memory when cache and port immutable data
structure initialization fails during register_device() state.
Linux commit:
4be3a4fa51f432ef045546d16f25c68a1ab525b9
MFC after: 1 week
Sponsored by: Mellanox Technologies
Garbage supplied by user will cause to UCMA module provide zero
memory size for memcpy(), because it wasn't checked, it will
produce unpredictable results in rdma_resolve_addr().
There are several places in the ucma ABI where userspace can pass in a
sockaddr but set the address family to AF_IB. When that happens,
rdma_addr_size() will return a size bigger than sizeof struct sockaddr_in6,
and the ucma kernel code might end up copying past the end of a buffer
not sized for a struct sockaddr_ib.
Fix this by introducing new variants
int rdma_addr_size_in6(struct sockaddr_in6 *addr);
int rdma_addr_size_kss(struct __kernel_sockaddr_storage *addr);
that are type-safe for the types used in the ucma ABI and return 0 if the
size computed is bigger than the size of the type passed in. We can use
these new variants to check what size userspace has passed in before
copying any addresses.
Linux commit:
2975d5de6428ff6d9317e9948f0968f7d42e5d74
09abfe7b5b2f442a85f4c4d59ecf582ad76088d7
84652aefb347297aa08e91e283adf7b18f77c2d5
MFC after: 1 week
Sponsored by: Mellanox Technologies
This was done by auditing all callers of ucma_get_ctx and switching the
ones that unconditionally touch ->device to ucma_get_ctx_dev. This covers
a little less than half of the call sites.
The 11 remaining call sites to ucma_get_ctx() were manually audited.
Linux commit:
4b658d1bbc16605330694bb3ef2570c465ef383d
8b77586bd8fe600d97f922c79f7222c46f37c118
MFC after: 1 week
Sponsored by: Mellanox Technologies
Attempt to modify XRC_TGT QP type from the user space (ibv_xsrq_pingpong
invocation) will trigger the following kernel panic. It is caused by the
fact that such QPs missed uobject initialization.
Linux commit:
f45765872e7aae7b81feb3044aaf9886b21885ef
MFC after: 1 week
Sponsored by: Mellanox Technologies
As part of ib_uverbs_remove_one which might be triggered upon
reset flow, we trigger IB_EVENT_DEVICE_FATAL event to userspace
application.
If device was removed after uverbs fd was opened but before
ib_uverbs_get_context was called, the event file will be accessed
before it was allocated, result in NULL pointer dereference:
Linux commit:
870201f95fcbd19538aef630393fe9d583eff82e
MFC after: 1 week
Sponsored by: Mellanox Technologies
The attempt to join multicast group without ensuring that CMA device
exists will lead to the following crash reported by syzkaller.
Linux commit:
7688f2c3bbf55e52388e37ac5d63ca471a7712e1
MFC after: 1 week
Sponsored by: Mellanox Technologies
Prior to access UCMA commands, the context should be initialized
and connected to CM_ID with ucma_create_id(). In case user skips
this step, he can provide non-valid ctx without CM_ID and cause
to multiple NULL dereferences.
Also there are situations where the create_id can be raced with
other user access, ensure that the context is only shared to
other threads once it is fully initialized to avoid the races.
Linux commit:
e8980d67d6017c8eee8f9c35f782c4bd68e004c9
MFC after: 1 week
Sponsored by: Mellanox Technologies
When receiving a PCP change all GID entries are reloaded.
This ensures the relevant GID entries use prio tagging,
by setting VLAN present and VLAN ID to zero.
The priority for prio tagged traffic is set using the regular
rdma_set_service_type() function.
Fake the real network device to have a VLAN ID of zero
when prio tagging is enabled. This is logic is hidden inside
the rdma_vlan_dev_vlan_id() function which must always be used
to retrieve the VLAN ID throughout all of ibcore and the
infiniband network drivers.
The VLAN presence information then propagates through all
of ibcore and so incoming connections will have the VLAN
bit set. The incoming VLAN ID is then checked against the
return value of rdma_vlan_dev_vlan_id().
MFC after: 1 week
Sponsored by: Mellanox Technologies
cma_iboe_set_mgid() is updated to reflect the RoCEv2 GID check.
Linux commit:
5c181bda77f409d89ad513528eccac5f3a416474
MFC after: 1 week
Sponsored by: Mellanox Technologies
RoCEv2 Annex states that for RoCEv2 over IPv4, the corresponding
IPv4 address is encoded into the GID according to the following rule:
GID= :ffff:<IPv4 address>
Remove the 0xff0e prefix for RoCEv2 packets with IPv4 and leave it
zeroed and change rdma_is_multicast_addr() to consider the new logic.
Linux commit:
be1d325a335840a86c133a56c6a911c368bac0fd
1c3aea2bc8f0b2e5b57375ead40457ff75a3a2ec
MFC after: 1 week
Sponsored by: Mellanox Technologies
The Infiniband spec defines "A multicast address is defined by a
MGID and a MLID" (section 10.5).
Add check to verify that the MLID value is in the correct address
range.
RoCE Annex (A16.9.10/11) declares that during attach (detach) QP to a
multicast group, if the QP is associated with a RoCE port, the
multicast group MLID is unused and is ignored.
During attach or detach multicast, when the QP is associated with a
port, it is enough to check the port's link layer and validate the
LID only if it is Infiniband. Otherwise, avoid validating the
multicast LID.
Linux commit:
8561eae60ff9417a50fa1fb2b83ae950dc5c1e21
5236333592244557a19694a51337df6ac018f0a7
MFC after: 1 week
Sponsored by: Mellanox Technologies
Implement a more generic solution for detecting loopback.
The problem was that the default netdevice was resolved
for loopback also when VLAN was used. Use real network
device instead of loopback device for bound device
interface.
How to test:
ucmatose -b 127.0.0.1 -p 20090
ucmatose -s 5.6.5.1 -p 20090
Note that RDMA treats the IPv4 and IPv6 loopback
addresses like any address.
MFC after: 1 week
Sponsored by: Mellanox Technologies
A list of MGID/MLID pairs is built when doing a multicast attach. When
the multicast detach is called, the list is searched, and regardless of
the search outcome, the driver detach is called.
If an MGID/MLID pair is not on the list, driver detach should not be
called, and an error should be returned. Calling the driver without
removing an MGID/MLID pair from the list can leave the core and driver
out of sync.
Linux commit:
20c7840a77ddcb2ed2fbd66e8197db2868495751
MFC after: 1 week
Sponsored by: Mellanox Technologies
When two handlers used the same object in the old schema, we blocked
the process in the kernel. The new schema just returns -EBUSY. This
could lead to different behaviour in applications between the old
schema and the new schema. In most cases, using such handlers
concurrently could lead to crashing the process. For example, if
thread A destroys a QP and thread B modifies it, we could have the
destruction happens before the modification. In this case, we are
accessing freed memory which could lead to crashing the process.
This is true for most cases. However, attaching and detaching
a multicast address from QP concurrently is safe. Therefore, we
preserve the original behaviour by adding a lock there.
Linux commit:
f48b726920d96dcd1860df06143bdea7d6d7dcc3
MFC after: 1 week
Sponsored by: Mellanox Technologies
When resolving an IP address in ibcore, only update the source address
upon normal completion. The ibcore address resolve function does not
care about the scope ID value of the IPv6 link-local addresses and expects
this information has already been extracted into the bound_dev_if field.
Because the same IPv6 link-local address can exist on multiple interfaces
the ibcore address resolver gets confused and returns ENETUNREACH.
Instead of updating both source address and bound_dev_if just keep the
address set to any address until resolving completes. For the sake of code
symmetry a similar change has been applied to the IPv4 address resolve path.
MFC after: 1 week
Sponsored by: Mellanox Technologies
When setting a large address resolve timeout it was observed that the
address resolving would succeed at the timeout and not when the address
was available. Make sure the address resolving requests are processed no
slower than one time every second.
While at it use "int" for jiffies instead of "unsigned long" to match
FreeBSD ticks.
MFC after: 1 week
Sponsored by: Mellanox Technologies
independent of the precision in most cases. This is mainly to simplify
checking for errors. r176266 did this for e_pow[f].c using a less
refined expression that often didn't work. r176276 fixes an error in
the log message for r176266. The main refinement is to always expand
to long double precision. See old log messages (especially these 2)
and the comment on the macro for more general details.
Specific details:
- using nan_mix() consistently for the new and old pow*() functions was
the only thing needed to make my consistency test for powl() vs pow()
pass on amd64.
- catrig[fl].c already had all the refinements, but open-coded.
- e_atan2[fl].c, e_fmod[fl].c and s_remquo[fl] only had primitive NaN
mixing.
- e_hypot[fl].c already had a different refined version of r176266. Refine
this further. nan_mix() is not directly usable here since we want to
clear the sign bit.
- e_remainder[f].c already had an earlier version of r176266.
- s_ccosh[f].c,/s_csinh[f].c already had a version equivalent to r176266.
Refine this further. nan_mix() is not directly usable here since the
expression has to handle some non-NaN cases.
- s_csqrt.[fl]: the mixing was special and mostly wrong. Partially fix the
special version.
- s_ctanh[f].c already had a version of r176266.
