The ACK timeout determines how long we take to notice a failed
Reliable Connection. Reducing it from the arbitrary value of 19 down
to 14 reduces the individual ACK timeout from around 2.1s to 67ms;
this in turn reduces the time to tear down and re-establish a broken
SRP session from around 30s to around 1s.
[hermon] Randomise the high-order bits of queue pair numbers
The Infiniband Communication Manager will refuse to establish a
connection if it believes the connection is already established.
There is no immediately obvious way to ask it to tear down the
existing connection and replace it; to issue a DREP we would need to
know the local and remote communication IDs used for the previous
connection setup.
We can work around this by randomising the high-order bits of the
queue pair number; these have no significance to the hardware, but are
sufficient to convince the IB CM that this is a different connection.
[infiniband] Change IB_{QPN,QKEY,QPT} names from {SMA,GMA} to {SMI,GSI}
The IBA specification refers to management "interfaces" and "agents".
The interface is the component that connects to the queue pair and
sends and receives MADs; the agent is the component that constructs
the reply to the MAD.
Rename the IB_{QPN,QKEY,QPT} constants as a first step towards making
this separation in gPXE.
[hermon] Allow software GMA to receive packets destined for QP1
The Linux IB Communication Manager will always send MADs to QP1,
rather than back to the originating QP. On Hermon, QP1 is by default
handled by the embedded firmware. We can change this, but the cost is
that we have to handle both QP0 and QP1 (i.e. we have to provide SMA
as well as GMA service in software), and we have to use MLX queues
rather than standard UD queues (i.e. we have to construct the UD
datagrams by hand).
There doesn't seem to be any viable way around this situation, ugly
though it is.
[infiniband] Add infrastructure for RC queue pairs
Queue pairs are now assumed to be created in the INIT state, with a
call to ib_modify_qp() required to bring the queue pair to the RTS
state.
ib_modify_qp() no longer takes a modification list; callers should
modify the relevant queue pair parameters (e.g. qkey) directly and
then call ib_modify_qp() to synchronise the changes to the hardware.
The packet sequence number is now a property of the queue pair, rather
than of the device.
Each queue pair may have an associated address vector. For RC queue
pairs, this is the address vector that will be programmed in to the
hardware as the remote address. For UD queue pairs, it will be used
as the default address vector if none is supplied to ib_post_send().
[hermon] Fix permissions broken in 3a799e9 ("Add PCI ID for ConnectX QDR card")
The patch file supplied for commit 3a799e9 ("[hermon] Add PCI ID for
ConnectX QDR card") accidentally marked drivers/infiniband/hermon.c as
being executable.
[i386] Change [u]int32_t to [unsigned] int, rather than [unsigned] long
This brings us in to line with Linux definitions, and also simplifies
adding x86_64 support since both platforms have 2-byte shorts, 4-byte
ints and 8-byte long longs.
[infiniband] Respect hop pointer when building directed route SMP return path
The return path in directed route SMPs lists the egress ports in order
from SM to node, rather than from node to SM.
To write to the correct offset within the return path, we need to
parse the hop pointer. This is held within the class-specific data
portion of the MAD header, which was previously unused by us and
defined to be a uint16_t. Define this field to be a union type; this
requires some rearrangement of ib_mad.h and corresponding changes to
ipoib.c.
[linda] Add support for QLogic 7220-based Infiniband HCAs
These cards very nearly support our current IB Verbs model. There is
one minor difference: multicast packets will always be delivered by
the hardware to QP0, so the driver has to redirect them to the
appropriate QP. This means that QP owners may see receive completions
for buffers that they never posted. Nothing in our current codebase
will break because of this.
[infiniband] Add raw packet parser and constructor
This can be used with cards that require the driver to construct and
parse packet headers manually. Headers are optionally handled
out-of-line from the packet payload, since some such cards will split
received headers into a separate ring buffer.
[infiniband] Split subnet management agent client out into ib_smc.c
Not all Infiniband cards have embedded subnet management agents.
Split out the code that communicates with such an embedded SMA into a
separate ib_smc.c file, and have drivers call ib_smc_update()
explicitly when they suspect that the answers given by the embedded
SMA may have changed.
[infiniband] Pass address vector in receive completions
Receive completion handlers now get passed an address vector
containing the information extracted from the packet headers
(including the GRH, if present), and only the payload remains in the
I/O buffer.
This breaks the symmetry between transmit and receive completions, so
remove the ib_completer_t type and use an ib_completion_queue_operations
structure instead.
Rename the "destination QPN" and "destination LID" fields in struct
ib_address_vector to reflect its new dual usage.
Since the ib_completion structure now contains only an IB status code,
("syndrome") replace it with a generic gPXE integer status code.
[infiniband] Flush uncompleted work queue entries at QP teardown
Avoid leaking I/O buffers in ib_destroy_qp() by completing any
outstanding work queue entries with a generic error code. This
requires the completion handlers to be available to ib_destroy_qp(),
which is done by making them static configuration parameters of the CQ
(set by ib_create_cq()) rather than being provided on each call to
ib_poll_cq().
This mimics the functionality of netdev_{tx,rx}_flush(). The netdev
flush functions would previously have been catching any I/O buffers
leaked by the IPoIB data queue (though not by the IPoIB metadata
queue).
[hermon] Allocate only as much memory as is needed for ICM and ICM AUX
Use individual page mappings rather than a single whole-region
mapping, to avoid the waste of memory that occurs due to the
constraint that each mapped block must be aligned on its own size.
[arbel] Allocate sufficient space for firmware buffer
We were accidentally allocating only half the required amount of
memory (given the alignment method) for the firmware buffer, leading
to conflicts between the firmware buffer and gPXE code/data segments.
[hermon] Allocate sufficient space for firmware buffer
We were accidentally allocating only half the required amount of
memory (given the alignment method) for the firmware buffer, leading
to conflicts between the firmware buffer and gPXE code/data segments.
[Infiniband] Add preliminary multiple port support for Hermon cards
Infiniband devices no longer block waiting for link-up in
register_ibdev().
Hermon driver needs to create an event queue and poll for link-up events.
Infiniband core needs to reread MAD parameters when link state changes.
IPoIB needs to cope with Infiniband link parameters being only partially
available at probe and open time.
[Infiniband] Add preliminary support for multi-port devices.
Arbel and Hermon cards both have multiple ports. Add the
infrastructure required to register each port as a separate IB
device. Don't yet register more than one port, since registration
will currently fail unless a valid link is detected.
Use ib_*_{set,get}_{drv,owner}data wrappers to access driver- and
owner-private data on Infiniband structures.
Pull out common code for handling management datagrams from arbel.c
and hermon.c into infiniband.c.
Add port number to struct ib_device.
Add open(), close() and mad() methods to struct ib_device_operations.
Michael Brown