Flow steering is supposed to return EOPNOTSUPP error
for unsupported fields and not ENOTSUPP error.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Set flow tag in flow table entry, when IB_FLOW_SPEC_ACTION_TAG
is part of the flow specifications.
Flow tag doesn't support multicast flows, so it's passing to
hardware only when used.
Signed-off-by: Moses Reuben <mosesr@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The struct ib_uverbs_flow_spec_action_tag associates a tag_id with the
flow defined by any number of other flow_spec entries which can reference
L2, L3, and L4 packet contents.
Use of ib_uverbs_flow_spec_action_tag allows the consumer to identify
the set of rules which where matched by
the packet by examining the tag_id in the CQE.
Signed-off-by: Moses Reuben <mosesr@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This specification identifies flow with a specific tag-id.
This tag-id will be reported in the CQE.
Signed-off-by: Moses Reuben <mosesr@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Flow rules with same match criteria and value should be mapped to
the same flow table entry regardless the flow tag identifier.
Flow tag is part of flow table entry context and not of the
destination, therefore we should return error when we try to add
destination to flow table entry with different flow tag.
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Commit 9603b61de1 ("mlx5: Move pci device handling from mlx5_ib
to mlx5_core") moved prof_sel module parameter from mlx5_ib to mlx5_core
and marked it as deprecated in 2014.
Three years after deprecation, it is time to remove the deprecated
module parameter.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Jack Morgenstein <jackm@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
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.
Fixes: 2811ba51b0 ("IB/mlx5: Add RoCE fields to Address Vector")
Cc: Achiad Shochat <achiad@mellanox.com>
Signed-off-by: Majd Dibbiny <majd@mellanox.com>
Reviewed-by: Moni Shoua <monis@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
According to the device specification, we need to check that the
has_smi bit is set in vport context before allowing send SMP
MADs from VF.
Fixes: e126ba97db ('mlx5: Add driver for Mellanox Connect-IB adapters')
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Reviewed-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Check the has_smi bit in vport context and class version of MADs
before allowing MADs processing to take place.
MAD_IFC SMI commands can be executed only if smi bit is set.
Fixes: e126ba97db ('mlx5: Add driver for Mellanox Connect-IB adapters')
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Parvi Kaustubhi <parvik@mellanox.com>
Reviewed-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Make sure that the Q counters are supported by the FW before trying
to allocate/deallocte them, this will avoid driver load failure when
they aren't supported by the FW.
Fixes: 0837e86a7a ('IB/mlx5: Add per port counters')
Cc: <stable@vger.kernel.org> # v4.7+
Signed-off-by: Kamal Heib <kamalh@mellanox.com>
Reviewed-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In case of unsupported singature, we returned positive
value, while the better approach is to return -EINVAL.
In addition, in this change, the error print is enriched
to provide an actual supplied signature type.
Fixes: e6631814fb ("IB/mlx5: Support IB_WR_REG_SIG_MR")
Cc: Sagi Grimberg <sagi@grimberg.me>
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
When we initialize buffer to create SRQ in kernel,
the number of pages was less than actually used in
following mlx5_fill_page_array().
Fixes: e126ba97db ("mlx5: Add driver for Mellanox Connect-IB adapters")
Cc: <stable@vger.kernel.org> # v3.10+
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The following series of patches optimizes the usage of the UAR area which is
contained within the BAR 0-1. Previous versions of the firmware and the driver
assumed each system page contains a single UAR. This patch set will query the
firmware for a new capability that if published, means that the firmware can
support UARs of fixed 4K regardless of system page size. In the case of
powerpc, where page size equals 64KB, this means we can utilize 16 UARs per
system page. Since user space processes by default consume eight UARs per
context this means that with this change a process will need a single system
page to fulfill that requirement and in fact make use of more UARs which is
better in terms of performance.
In addition to optimizing user-space processes, we introduce an allocator
that can be used by kernel consumers to allocate blue flame registers
(which are areas within a UAR that are used to write doorbells). This provides
further optimization on using the UAR area since the Ethernet driver makes
use of a single blue flame register per system page and now it will use two
blue flame registers per 4K.
The series also makes changes to naming conventions and now the terms used in
the driver code match the terms used in the PRM (programmers reference manual).
Thus, what used to be called UUAR (micro UAR) is now called BFREG (blue flame
register).
In order to support compatibility between different versions of
library/driver/firmware, the library has now means to notify the kernel driver
that it supports the new scheme and the kernel can notify the library if it
supports this extension. So mixed versions of libraries can run concurrently
without any issues.
Thanks,
Eli and Matan
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJYc9kSAAoJEEg/ir3gV/o+a0EH/jEGiopH7CHc4T4nXT1I4kQa
TicrkMNV3Sr9MBWwn8TLOyx+Fi1dex4cumrJI/BNVjC6h/nS6JHbslYoZxTkX9lT
L0vRsHJBVr/PODqimIGNnlJFBPhNJSGiHG4JHlJHlpvcGNahitN3gXmUjcRNju+V
ExnvgwWzAXM0qg1qWf5A/3HmqbtYES1rJXQUsimtc2QAif/SIayBD4fEA8x5zNBA
i0p8xcDrzUqmeblkpnsJA3w40s1rsuqvJnvLPDpbpKENtHfw1UFZ2987P7LvOrIv
NF/mZBkStC0gOZX6dLEAdoZXL1gTsJX19hTkUMfYH4BHqHARa2/oCS3wcCf1Giw=
=C+cp
-----END PGP SIGNATURE-----
Merge tag 'mlx5-4kuar-for-4.11' of git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux
Saeed Mahameed says:
====================
mlx5 4K UAR
The following series of patches optimizes the usage of the UAR area which is
contained within the BAR 0-1. Previous versions of the firmware and the driver
assumed each system page contains a single UAR. This patch set will query the
firmware for a new capability that if published, means that the firmware can
support UARs of fixed 4K regardless of system page size. In the case of
powerpc, where page size equals 64KB, this means we can utilize 16 UARs per
system page. Since user space processes by default consume eight UARs per
context this means that with this change a process will need a single system
page to fulfill that requirement and in fact make use of more UARs which is
better in terms of performance.
In addition to optimizing user-space processes, we introduce an allocator
that can be used by kernel consumers to allocate blue flame registers
(which are areas within a UAR that are used to write doorbells). This provides
further optimization on using the UAR area since the Ethernet driver makes
use of a single blue flame register per system page and now it will use two
blue flame registers per 4K.
The series also makes changes to naming conventions and now the terms used in
the driver code match the terms used in the PRM (programmers reference manual).
Thus, what used to be called UUAR (micro UAR) is now called BFREG (blue flame
register).
In order to support compatibility between different versions of
library/driver/firmware, the library has now means to notify the kernel driver
that it supports the new scheme and the kernel can notify the library if it
supports this extension. So mixed versions of libraries can run concurrently
without any issues.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_get_info() has to lock the socket, so lets lock it
for an extended critical section, so that various fields
have consistent values.
This solves an annoying issue that some applications
reported when multiple counters are updated during one
particular rx/rx event, and TCP_INFO was called from
another cpu.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Alexei Starovoitov says:
====================
bpf: verifier improvements
A number of bpf verifier improvements from Gianluca.
See individual patches for details.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
since ARG_PTR_TO_STACK is no longer just pointer to stack
rename it to ARG_PTR_TO_MEM and adjust comment.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, helpers that read and write from/to the stack can do so using
a pair of arguments of type ARG_PTR_TO_STACK and ARG_CONST_STACK_SIZE.
ARG_CONST_STACK_SIZE accepts a constant register of type CONST_IMM, so
that the verifier can safely check the memory access. However, requiring
the argument to be a constant can be limiting in some circumstances.
Since the current logic keeps track of the minimum and maximum value of
a register throughout the simulated execution, ARG_CONST_STACK_SIZE can
be changed to also accept an UNKNOWN_VALUE register in case its
boundaries have been set and the range doesn't cause invalid memory
accesses.
One common situation when this is useful:
int len;
char buf[BUFSIZE]; /* BUFSIZE is 128 */
if (some_condition)
len = 42;
else
len = 84;
some_helper(..., buf, len & (BUFSIZE - 1));
The compiler can often decide to assign the constant values 42 or 48
into a variable on the stack, instead of keeping it in a register. When
the variable is then read back from stack into the register in order to
be passed to the helper, the verifier will not be able to recognize the
register as constant (the verifier is not currently tracking all
constant writes into memory), and the program won't be valid.
However, by allowing the helper to accept an UNKNOWN_VALUE register,
this program will work because the bitwise AND operation will set the
range of possible values for the UNKNOWN_VALUE register to [0, BUFSIZE),
so the verifier can guarantee the helper call will be safe (assuming the
argument is of type ARG_CONST_STACK_SIZE_OR_ZERO, otherwise one more
check against 0 would be needed). Custom ranges can be set not only with
ALU operations, but also by explicitly comparing the UNKNOWN_VALUE
register with constants.
Another very common example happens when intercepting system call
arguments and accessing user-provided data of variable size using
bpf_probe_read(). One can load at runtime the user-provided length in an
UNKNOWN_VALUE register, and then read that exact amount of data up to a
compile-time determined limit in order to fit into the proper local
storage allocated on the stack, without having to guess a suboptimal
access size at compile time.
Also, in case the helpers accepting the UNKNOWN_VALUE register operate
in raw mode, disable the raw mode so that the program is required to
initialize all memory, since there is no guarantee the helper will fill
it completely, leaving possibilities for data leak (just relevant when
the memory used by the helper is the stack, not when using a pointer to
map element value or packet). In other words, ARG_PTR_TO_RAW_STACK will
be treated as ARG_PTR_TO_STACK.
Signed-off-by: Gianluca Borello <g.borello@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
commit 484611357c ("bpf: allow access into map value arrays")
introduces the ability to do pointer math inside a map element value via
the PTR_TO_MAP_VALUE_ADJ register type.
The current support doesn't handle the case where a PTR_TO_MAP_VALUE_ADJ
is spilled into the stack, limiting several use cases, especially when
generating bpf code from a compiler.
Handle this case by explicitly enabling the register type
PTR_TO_MAP_VALUE_ADJ to be spilled. Also, make sure that min_value and
max_value are reset just for BPF_LDX operations that don't result in a
restore of a spilled register from stack.
Signed-off-by: Gianluca Borello <g.borello@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Enable helpers to directly access a map element value by passing a
register type PTR_TO_MAP_VALUE (or PTR_TO_MAP_VALUE_ADJ) to helper
arguments ARG_PTR_TO_STACK or ARG_PTR_TO_RAW_STACK.
This enables several use cases. For example, a typical tracing program
might want to capture pathnames passed to sys_open() with:
struct trace_data {
char pathname[PATHLEN];
};
SEC("kprobe/sys_open")
void bpf_sys_open(struct pt_regs *ctx)
{
struct trace_data data;
bpf_probe_read(data.pathname, sizeof(data.pathname), ctx->di);
/* consume data.pathname, for example via
* bpf_trace_printk() or bpf_perf_event_output()
*/
}
Such a program could easily hit the stack limit in case PATHLEN needs to
be large or more local variables need to exist, both of which are quite
common scenarios. Allowing direct helper access to map element values,
one could do:
struct bpf_map_def SEC("maps") scratch_map = {
.type = BPF_MAP_TYPE_PERCPU_ARRAY,
.key_size = sizeof(u32),
.value_size = sizeof(struct trace_data),
.max_entries = 1,
};
SEC("kprobe/sys_open")
int bpf_sys_open(struct pt_regs *ctx)
{
int id = 0;
struct trace_data *p = bpf_map_lookup_elem(&scratch_map, &id);
if (!p)
return;
bpf_probe_read(p->pathname, sizeof(p->pathname), ctx->di);
/* consume p->pathname, for example via
* bpf_trace_printk() or bpf_perf_event_output()
*/
}
And wouldn't risk exhausting the stack.
Code changes are loosely modeled after commit 6841de8b0d ("bpf: allow
helpers access the packet directly"). Unlike with PTR_TO_PACKET, these
changes just work with ARG_PTR_TO_STACK and ARG_PTR_TO_RAW_STACK (not
ARG_PTR_TO_MAP_KEY, ARG_PTR_TO_MAP_VALUE, ...): adding those would be
trivial, but since there is not currently a use case for that, it's
reasonable to limit the set of changes.
Also, add new tests to make sure accesses to map element values from
helpers never go out of boundary, even when adjusted.
Signed-off-by: Gianluca Borello <g.borello@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the logic to check memory accesses to a PTR_TO_MAP_VALUE_ADJ from
check_mem_access() to a separate helper check_map_access_adj(). This
enables to use those checks in other parts of the verifier as well,
where boundaries on PTR_TO_MAP_VALUE_ADJ might need to be checked, for
example when checking helper function arguments. The same thing is
already happening for other types such as PTR_TO_PACKET and its
check_packet_access() helper.
The code has been copied verbatim, with the only difference of removing
the "off += reg->max_value" statement and moving the sum into the call
statement to check_map_access(), as that was only needed due to the
earlier common check_map_access() call.
Signed-off-by: Gianluca Borello <g.borello@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ursula Braun says:
====================
net/smc: Shared Memory Communications - RDMA
here is now V4 of the SMC-R patches having processed your feedback from end
of November. The most important change is the replacement of sysfs by a
generic netlink solution in patch 04. And I tried to get rid of the __packed
attributes. There are still a few usages left due to SMC-R protocol defined
structures.
V4 changes:
The order of patches 03 and 04 for pnet table management and SMC IB-client
establishing has been exchanged, since pnet table management is now built on
top of smc_ib_devices.
Patch 01: Use EXPORT_SYMBOL_GPL().
Patch 02: Define "use_fallback" as bool.
Get rid of useless smc_sock fields clearing in smc_sock_alloc(),
since sk_alloc() clears out the memory.
Patch 03: Postpone smc_ib_remember_port_attr() call till ib_device is
mentioned in the pnet table.
Patch 04: Replace sysfs-usage by a generic netlink approach for pnet table
configuration.
Change layout of pnet table entries to reference net_device and
ib_device instead of dealing with names of net_devices and
ib_devices.
Patch 05: Adapt "use_fallback" usages to new type bool.
Get rid of useless smc_sock fields clearing in smc_sock_alloc()
Avoid __packed where possible.
Check if clc responses are not too big.
Patch 09: Postpone smc_setup_per_ibdev till the first connection with this
ib_device is really created.
Patch 11: Get rid of __packed usage.
V3 changes:
Patch 05: Remove unneeded DEFINE_WAIT
Patch 06: Improve synchronization of link group creation
Patch 07: Rename peer_rmbe_len into peer_rmbe_size to be more consistent
Patch 09: Avoid calls of ib_get_memory_region with IB_ACCESS_LOCAL_WRITE,
use new default local_dma_lkey from protection domain as lkey
instead.
Remove no longer needed function smc_ib_dereg_memory_region().
Patch 14: Switch to state ACTIVE only if still in state INIT.
Return 0 for recvmsg invoked in a socket closing state.
Allow getname call in state APPCLOSEWAIT1
Do not trigger destruction of a socket-in-error queued in accept
queue.
During cleanup of accept queue, make sure sockets are destructed,
and sockets in fallback mode are handled appropriately.
When freeing sndbufs/rmbs, remove them from their list and free
the entry.
Use add_wait_queue() and remove_wait_queue() in close wait
functions.
If actively closing a socket in state for PEERFINCLOSEWAIT, keep
this state.
If passively closing a socket while bytes are to be received, move
to state APPCLOSEWAIT1.
If actively aborting a socket, skip sending the close_abort flag,
since RDMA communication is no longer possible.
When terminating a link group, do not schedule link group freeing a
2nd time, since already done when unregistering the last remaining
connection.
Patch 15: Introduce smc_diag module for monitoring SMC protocol sockets.
This replaces the old patch 0015 dealing with procfs.
V2 changes:
Patch 0002: Add SMC versions for family key strings in net/core/sock.c.
Patch 0006: initialize rb_tree.
Patch 0007: Get rid of unneeded use of xchg() in smc_sndbuf_unuse() and
smc_rmb_unuse().
Patch 0008: Correct error checking logic for ib_function calls.
Define struct smc_link field wr_tx_id as atomic_long_t.
Use "do_div" instead of "%" to be architecture-independent.
Patch 0009: Correct error checking logic for ib_function calls.
Patch 0011: Remove xchg() calls in cursor handling. Use atomic64_t for cursor
overlays on 64-bit architectures. If not available, use plain u64
and add locking for cursor reading and writing.
Implement smc_curs_add() without modulo operator "%".
Patch 0012: Remove xchg() calls in cursor handling.
Implement smc_tx_rdma_writes() without module operator "%".
Patch 0013: Remove xchg() calls in cursor handling.
Patch 0014: Return type bool in smc_wr_tx_has_pending().
Remove unneeded semicolon in smc_close_shutdown_write().
Call smc_close_active() in non-fallback case only.
Get rid of duplicate schedule of sock_put_work().
Take nested sock_lock in smc_listen_work().
Start close stream_wait in case of prepared sends only.
Patch 0015: Remove unneeded socket ref_count in smc_proc_seq_show().
Take lock before list_empty check in smc_proc_sock_list_del().
These patches are the initial part of the implementation of the
"Shared Memory Communications-RDMA" (SMC-R) protocol as defined in
RFC7609 [1]. While SMC-R does not aim to replace TCP,
it taps a wealth of existing data center TCP socket applications
to become more efficient without the need for rewriting them.
SMC-R uses RDMA over Converged Ethernet (RoCE) to save CPU consumption.
For instance, when running 10 parallel connections with uperf, we measured
a decrease of 60% in CPU consumption with SMC-R compared to TCP/IP
(with throughput and latency comparable;
measured on x86_64 with the same RoCE card and port).
SMC-R does not require an RDMA communication manager (RDMA CM).
SMC-R inherits TCP qualities such as reliable connections, host-based
firewall packet filtering (on connection establishment) and unmodified
application of communication encryption such as TLS (transport layer
security) or SSL (secure sockets layer). Since original TCP is used to
establish SMC-R connections, load balancers and packet inspection based
on TCP/IP connection establishment continue to work for SMC-R.
On the other hand, using SMC-R implies:
- either involving a preload library when invoking the unchanged TCP-application
or slightly modifying the source by simply changing the socket family in
the socket() call
- accepting extra overhead and latency in connection establishment due to
SMC Connection Layer Control (CLC) handshake
- explicit coupling of RoCE ports with Ethernet ports
- not routable as currently built on RoCE V1
- bypassing of packet-based networking features
- filtering (netfilter)
- sniffing (libpcap, packet sockets, (E)BPF)
- traffic control (scheduling, shaping)
- bypassing of IP-header based socket options
- bypassing of memory buffer (pressure) management
- unusable together with IPsec
Overview of the SMC-R Protocol described in informational RFC 7609
SMC-R is an open protocol that provides RDMA capabilities over RoCE
transparently for applications exploiting TCP sockets.
A new socket protocol family PF_SMC is introduced.
There are no changes required to applications using the sockets API for TCP
stream sockets other than the specification of the new socket family AF_SMC.
Unmodified applications can be used by means of a dynamic preload shared
library which rewrites the socket API call
socket(AF_INET, SOCK_STREAM, IPPROTO_TCP) into
socket(AF_SMC, SOCK_STREAM, IPPROTO_TCP).
SMC-R re-uses the address family AF_INET for all addressing purposes around
struct sockaddr.
SMC-R system architecture layers:
+=============================================================================+
| | unmodified TCP application |
| native SMC application +--------------------------------------+
| | dynamic preload shared library |
+=============================================================================+
| SMC socket |
+-----------------------------------------------------------------------------+
| | TCP socket (for connection establishment and fallback) |
| IB verbs +--------------------------------------------------------+
| | IP |
+--------------------+--------------------------------------------------------+
| RoCE device driver | some network device driver |
+=============================================================================+
Terms:
A link group is determined by an ordered peer pair of TCP client and TCP server
(IP addresses and subnet). Reversed client server roles cause an own link group.
A link is a logical point-to-point connection based on an
infiniband reliable connected queue pair (RC-QP) between two RoCE ports
(MACs and GIDs) of a peer pair.
A link group can have 1..8 links for failover and load balancing.
This initial Linux implementation always has 1 link per link group.
Each link group on a peer can have 1..255 remote memory buffers (RMBs).
If more RMBs are needed, a peer can open another link group
(this initial Linux implementation) or fall back to TCP.
Each RMB has its own particular size and its own (R)DMA mapping and credentials
(rtoken consisting of rkey and RDMA "virtual address").
This initial Linux implementation uses physically contiguous memory for RMBs
but we are working towards scattered memory because of memory fragmentation.
Each RMB has 1..255 RMB elements (RMBEs) of equal size
to provide multiplexing of connections within an RMB.
An RMBE is the RDMA Write destination organized as wrapping ring buffer
for data transmit of a particular connection in one direction
(duplex by means of mirror symmetry as with TCP).
This initial Linux implementation always has 1 RMBE per RMB
and thus an individual RMB for each connection.
SMC-R connection establishment with subsequent data transfer:
CLIENT SERVER
TCP three-way handshake:
regular TCP SYN
-------------------------------------------------------->
regular TCP SYN ACK
<--------------------------------------------------------
regular TCP ACK
-------------------------------------------------------->
SMC Connection Layer Control (CLC) handshake
exchanges RDMA credentials between peers:
via above TCP connection: SMC CLC Proposal
-------------------------------------------------------->
via above TCP connection: SMC CLC Accept
<--------------------------------------------------------
via above TCP connection: SMC CLC Confirm
-------------------------------------------------------->
SMC Link Layer Control (LLC) (only once per link, i.e. 1st conn. of link group):
RoCE RC-QP: SMC LLC Confirm Link
<========================================================
RoCE RC-QP: SMC LLC Confirm Link response
========================================================>
SMC data transmission (incl. SMC Connection Data Control (CDC) message):
RoCE RC-QP: RDMA Write
========================================================>
RoCE RC-QP: SMC CDC message (flow control)
========================================================>
...
RoCE RC-QP: RDMA Write
<========================================================
RoCE RC-QP: SMC CDC message (flow control)
<========================================================
...
Data flow within an established connection:
+----------------------------------------------------------------------------
| SENDER
| sendmsg()
| |
| | produces into sndbuf [sender's process context]
| v
| +--------+
| | sndbuf | [ring buffer]
| +--------+
| |
| | consumes from sndbuf and produces into receiver's RMBE [any context]
| | by sending RDMA Write followed by SMC CDC message over RoCE RC-QP
| |
+----|-----------------------------------------------------------------------
|
+----|-----------------------------------------------------------------------
| v RECEIVER
| +------+
| | RMBE | [ring buffer, can have size different from sender's sndbuf]
| | | [RMBE represents rcvbuf, no further de-coupling as on sender side]
| +------+
| |
| | consumes from RMBE [receiver's process context]
| v
| recvmsg()
+----------------------------------------------------------------------------
Flow control ("cursor" updates) by means of SMC CDC messages:
SENDER RECEIVER
sends updates via CDC-------------+ sends updates via CDC
on consuming from sndbuf | on consuming from RMBE
and producing into RMBE | by means of recvmsg()
| |
| |
+-----------------------------------|------------+
| |
+--v-------------------------+ +--v-----------------------+
| receiver's consumer cursor | | sender's producer cursor----+
+----------------|-----------+ +--------------------------+ |
| |
| receiver's RMBE |
| +--------------------------+ |
| | | |
+--------------------------------+ | |
| | | |
| v | |
| +------------| |
|-------------+////////////| |
|//RDMA data written by////| |
|////sender that is////////| |
|/available to be consumed/| |
|///////// +---------------| |
|----------+^ | |
| | | |
| +-----------------+
| |
+--------------------------+
Sending updates of the producer cursor is immediate for low latency;
something like Nagle's algorithm (absence of TCP_NODELAY) is optional and
currently not part of this initial Linux implementation.
Sending updates of the consumer cursor is conditional to avoid the
silly window syndrome.
Normal connection termination:
Normal connection termination starts transitioning from socket state
ACTIVE via either "Active Close" or "Passive Close".
shutdown rdwr +-----------------+
or close, +-------------->| INIT / CLOSED |<-------------+
send PeerCon|nClosed +-----------------+ | PeerConnClosed
| | | received
| connection | established |
| V |
+----------------+ +-----------------+ +----------------+
|AppFinCloseWait | | ACTIVE | |PeerFinCloseWait|
+----------------+ +-----------------+ +----------------+
| | | |
| Active Close: | |Passive Close: |
| close or | |PeerConnClosed or |
| shutdown wr or| |PeerDoneWriting |
| shutdown rdwr | |received |
| V V |
PeerConnClo|sed +--------------+ +-------------+ | close or
received +--<----|PeerCloseWait1| |AppCloseWait1|--->----+ shutdown rdwr,
| +--------------+ +-------------+ | send
| PeerDoneWri|ting | shutdown wr, | PeerConnClosed
| received | send Pee|rDoneWriting |
| V V |
| +--------------+ +-------------+ |
+--<----|PeerCloseWait2| |AppCloseWait2|--->----+
+--------------+ +-------------+
In state CLOSED, the socket can be destructed only, once the application has
issued a close().
Abnormal connection termination:
+-----------------+
+-------------->| INIT / CLOSED |<-------------+
| +-----------------+ |
| |
| +-----------------------+ |
| | Any state | |
PeerConnAbo|rt | (before setting | | send
received | | PeerConnClosed | | PeerConnAbort
| | indicator in | |
| | peer's RMBE) | |
| +-----------------------+ |
| | | |
| Active Abort: | | Passive Abort: |
| problem, | | PeerConnAbort |
| send | | received, |
| PeerConnAbort,| | ECONNRESET |
| ECONNABORTED | | |
| V V |
| +--------------+ +--------------+ |
+-------|PeerAbortWait | | ProcessAbort |------+
+--------------+ +--------------+
Implementation notes beyond RFC 7609:
A PNET table in sysfs provides the mapping between network device names and
RoCE Infiniband device names for the transparent switch of data communication.
A PNET table can contain an arbitrary number of PNETIDs.
Each PNETID contains exactly one (Ethernet) network device name
and one or more RoCE Infiniband device names.
Each device name can only exist in at most one PNETID (no overlapping).
This initial Linux implementation allows at most one RoCE Infiniband device
name per PNETID.
After a new TCP connection is established, the network device name
used for egress traffic with the TCP connection's local source IP address
is used as key to lookup the unique PNETID, and the RoCE Infiniband device
of this PNETID is used to switch data communication from TCP to RDMA
during SMC CLC handshake.
Problem determination:
A protocol dissector is available with upstream wireshark for formatting
SMC-R related RoCE LAN traffic.
[https://code.wireshark.org/review/gitweb?p=wireshark.git;a=blob;f=epan/dissectors/packet-smcr.c]
We are working on enhancing the Linux implementation to cover:
- Improve default socket closing asynchronicity
- Address corner cases with many parallel connections
- Tracing
- Integrated load balancing and fail-over within a link group
- Splice and sendpage support
- IPv6 addressing support
- Keepalive, Cork
- Namespaces support
- Urgent data
- More socket options
- Diagnostics
- Statistics support
- SNMP support
References:
[1] SMC-R Informational RFC: http://www.rfc-editor.org/info/rfc7609
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Support for SMC socket monitoring via netlink sockets of protocol
NETLINK_SOCK_DIAG.
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
smc_shutdown() and smc_release() handling
delayed linkgroup cleanup for linkgroups without connections
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
move RMBE data into user space buffer and update managing cursors
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
copy data to kernel send buffer, and trigger RDMA write
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
send and receive CDC messages (via IB message send and CQE)
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
send and receive LLC messages CONFIRM_LINK (via IB message send and CQE)
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Prepare the link for RDMA transport:
Create a queue pair (QP) and move it into the state Ready-To-Receive (RTR).
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The base containers for RDMA transport are work requests and completion
queue entries processed through Infiniband verbs:
* allocate and initialize these areas
* map these areas to DMA
* implement the basic communication consisting of work request posting
and receival of completion queue events
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* allocate data RMB memory for sending and receiving
* size depends on the maximum socket send and receive buffers
* allocated RMBs are kept during life time of the owning link group
* map the allocated RMBs to DMA
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* create smc_connection for SMC-sockets
* determine suitable link group for a connection
* create a new link group if necessary
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* CLC (Connection Layer Control) handshake
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Connection creation with SMC-R starts through an internal
TCP-connection. The Ethernet interface for this TCP-connection is not
restricted to the Ethernet interface of a RoCE device. Any existing
Ethernet interface belonging to the same physical net can be used, as
long as there is a defined relation between the Ethernet interface and
some RoCE devices. This relation is defined with the help of an
identification string called "Physical Net ID" or short "pnet ID".
Information about defined pnet IDs and their related Ethernet
interfaces and RoCE devices is stored in the SMC-R pnet table.
A pnet table entry consists of the identifying pnet ID and the
associated network and IB device.
This patch adds pnet table configuration support using the
generic netlink message interface referring to network and IB device
by their names. Commands exist to add, delete, and display pnet table
entries, and to flush or display the entire pnet table.
There are cross-checks to verify whether the ethernet interfaces
or infiniband devices really exist in the system. If either device
is not available, the pnet ID entry is not created.
Loss of network devices and IB devices is also monitored;
a pnet ID entry is removed when an associated network or
IB device is removed.
Signed-off-by: Thomas Richter <tmricht@linux.vnet.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* enable smc module loading and unloading
* register new socket family
* basic smc socket creation and deletion
* use backing TCP socket to run CLC (Connection Layer Control)
handshake of SMC protocol
* Setup for infiniband traffic is implemented in follow-on patches.
For now fallback to TCP socket is always used.
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Reviewed-by: Utz Bacher <utz.bacher@de.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Direct call of tcp_set_keepalive() function from protocol-agnostic
sock_setsockopt() function in net/core/sock.c violates network
layering. And newly introduced protocol (SMC-R) will need its own
keepalive function. Therefore, add "keepalive" function pointer
to "struct proto", and call it from sock_setsockopt() via this pointer.
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Reviewed-by: Utz Bacher <utz.bacher@de.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Niklas Söderlund says:
====================
sh_eth: add wake-on-lan support via magic packet
This series adds support for Wake-on-Lan using Magic Packet for a few
models of the sh_eth driver. Patch 1/6 fix a naming error, patch 2/6
adds generic support to control and support WoL while patches 3/6 - 6/6
enable different models.
Based ontop of net-next master.
Changes since v2.
- Fix bookkeeping for "active_count" and "event_count" reported in
/sys/kernel/debug/wakeup_sources. Thanks Geert for noticing this.
- Add new patch 1/6 which corrects the name of ECMR_MPDE bit, suggested
by Sergei.
- s/sh7743/sh7734/ in patch 5/6. Thanks Geert for spotting this.
- Spelling improvements suggested by Sergei and Geert.
- Add Tested-by to 3/6 and 4/6.
Changes since v1.
- Split generic WoL functionality and device enablement to different
patches.
- Enable more devices then Gen2 after feedback from Geert and
datasheets.
- Do not set mdp->irq_enabled = false and remove specific MagicPacket
interrupt clearing, instead let sh_eth_error() clear the interrupt as
for other EMAC interrupts, thanks Sergei for the suggestion.
- Use the original return logic in sh_eth_resume().
- Moved sh_eth_private variable *clk to top of data structure to avoid
possible gaps due to alignment restrictions.
- Make wol_enabled in sh_eth_private part of the already existing
bitfield instead of a bool.
- Do not initiate mdp->wol_enabled to 0, the struct is kzalloc'ed so
it's already set to 0.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This is based on public datasheet for sh7763 which shows it has the
same behavior and registers for WoL as other versions of sh_eth.
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is based on public datasheet for sh7734 which shows it has the
same behavior and registers for WoL as other versions of sh_eth.
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
Geert Uytterhoeven reported WoL worked on his Armadillo board.
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Tested-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
Tested on Gen2 r8a7791/Koelsch.
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Tested-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add generic functionality to support Wake-on-LAN using MagicPacket which
are supported by at least a few versions of sh_eth. Only add
functionality for WoL, no specific sh_eth versions are marked to support
WoL yet.
WoL is enabled in the suspend callback by setting MagicPacket detection
and disabling all interrupts expect MagicPacket. In the resume path the
driver needs to reset the hardware to rearm the WoL logic, this prevents
the driver from simply restoring the registers and to take advantage of
that sh_eth was not suspended to reduce resume time. To reset the
hardware the driver closes and reopens the device just like it would do
in a normal suspend/resume scenario without WoL enabled, but it both
closes and opens the device in the resume callback since the device
needs to be open for WoL to work.
One quirk needed for WoL is that the module clock needs to be prevented
from being switched off by Runtime PM. To keep the clock alive the
suspend callback need to call clk_enable() directly to increase the
usage count of the clock. Then when Runtime PM decreases the clock usage
count it won't reach 0 and be switched off.
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
This bit was wrongly named due to a typo, Sergei checked the SH7734/63
manuals and this bit should be named MPDE.
Suggested-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jesper Dangaard Brouer says:
====================
net: optimize ICMP-reply code path
This patchset is optimizing the ICMP-reply code path, for ICMP packets
that gets rate limited. A remote party can easily trigger this code
path by sending packets to port number with no listening service.
Generally the patchset moves the sysctl_icmp_msgs_per_sec ratelimit
checking to earlier in the code path and removes an allocation.
Use-case: The specific case I experienced this being a bottleneck is,
sending UDP packets to a port with no listener, which obviously result
in kernel replying with ICMP Destination Unreachable (type:3), Port
Unreachable (code:3), which cause the bottleneck.
After Eric and Paolo optimized the UDP socket code, the kernels PPS
processing capabilities is lower for no-listen ports, than normal UDP
sockets. This is bad for capacity planning when restarting a service.
UDP no-listen benchmark 8xCPUs using pktgen_sample04_many_flows.sh:
Baseline: 6.6 Mpps
Patch: 14.7 Mpps
Driver mlx5 at 50Gbit/s.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
It is possible to avoid the atomic operation in icmp{v6,}_xmit_lock,
by checking the sysctl_icmp_msgs_per_sec ratelimit before these calls,
as pointed out by Eric Dumazet, but the BH disabled state must be correct.
The icmp_global_allow() call states it must be called with BH
disabled. This protection was given by the calls icmp_xmit_lock and
icmpv6_xmit_lock. Thus, split out local_bh_disable/enable from these
functions and maintain it explicitly at callers.
Suggested-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch split the global and per (inet)peer ICMP-reply limiter
code, and moves the global limit check to earlier in the packet
processing path. Thus, avoid spending cycles on ICMP replies that
gets limited/suppressed anyhow.
The global ICMP rate limiter icmp_global_allow() is a good solution,
it just happens too late in the process. The kernel goes through the
full route lookup (return path) for the ICMP message, before taking
the rate limit decision of not sending the ICMP reply.
Details: The kernels global rate limiter for ICMP messages got added
in commit 4cdf507d54 ("icmp: add a global rate limitation"). It is
a token bucket limiter with a global lock. It brilliantly avoids
locking congestion by only updating when 20ms (HZ/50) were elapsed. It
can then avoids taking lock when credit is exhausted (when under
pressure) and time constraint for refill is not yet meet.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commit 9a99d4a50c ("icmp: avoid allocating large struct
on stack"), because struct icmp_bxm no really a large struct, and
allocating and free of this small 112 bytes hurts performance.
Fixes: 9a99d4a50c ("icmp: avoid allocating large struct on stack")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
-----BEGIN PGP SIGNATURE-----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=H/FD
-----END PGP SIGNATURE-----
Merge tag 'rxrpc-rewrite-20170109' of git://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs
David Howells says:
====================
afs: Refcount afs_call struct
These patches provide some tracepoints for AFS and fix a potential leak by
adding refcounting to the afs_call struct.
The patches are:
(1) Add some tracepoints for logging incoming calls and monitoring
notifications from AF_RXRPC and data reception.
(2) Get rid of afs_wait_mode as it didn't turn out to be as useful as
initially expected. It can be brought back later if needed. This
clears some stuff out that I don't then need to fix up in (4).
(3) Allow listen(..., 0) to be used to disable listening. This makes
shutting down the AFS cache manager server in the kernel much easier
and the accounting simpler as we can then be sure that (a) all
preallocated afs_call structs are relesed and (b) no new incoming
calls are going to be started.
For the moment, listening cannot be reenabled.
(4) Add refcounting to the afs_call struct to fix a potential multiple
release detected by static checking and add a tracepoint to follow the
lifecycle of afs_call objects.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Florian Fainelli says:
====================
net: dsa: Make dsa_switch_ops const
This patch series allows us to annotate dsa_switch_ops with a const
qualifier.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that we have properly encapsulated and made drivers utilize exported
functions, we can switch dsa_switch_ops to be a annotated with const.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
