It is currently fairly difficult to follow what clk_rate_request are
issued, and how they have been modified once done.
Indeed, there's multiple paths that can be taken, some functions are
recursive and will just forward the request to its parent, etc.
Adding a lot of debug prints is just not very convenient, so let's add
trace events for the clock requests, one before they are submitted and
one after they are returned.
That way we can simply toggle the tracing on without modifying the
kernel code and without affecting performances or the kernel logs too
much.
Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
Link: https://lore.kernel.org/r/20221018-clk-rate-request-tracing-v2-2-5170b363c413@cerno.tech
Signed-off-by: Stephen Boyd <sboyd@kernel.org>
If a cookie expires from the LRU and the LRU_DISCARD flag is set, but
the state machine has not run yet, it's possible another thread can call
fscache_use_cookie and begin to use it.
When the cookie_worker finally runs, it will see the LRU_DISCARD flag
set, transition the cookie->state to LRU_DISCARDING, which will then
withdraw the cookie. Once the cookie is withdrawn the object is removed
the below oops will occur because the object associated with the cookie
is now NULL.
Fix the oops by clearing the LRU_DISCARD bit if another thread uses the
cookie before the cookie_worker runs.
BUG: kernel NULL pointer dereference, address: 0000000000000008
...
CPU: 31 PID: 44773 Comm: kworker/u130:1 Tainted: G E 6.0.0-5.dneg.x86_64 #1
Hardware name: Google Compute Engine/Google Compute Engine, BIOS Google 08/26/2022
Workqueue: events_unbound netfs_rreq_write_to_cache_work [netfs]
RIP: 0010:cachefiles_prepare_write+0x28/0x90 [cachefiles]
...
Call Trace:
netfs_rreq_write_to_cache_work+0x11c/0x320 [netfs]
process_one_work+0x217/0x3e0
worker_thread+0x4a/0x3b0
kthread+0xd6/0x100
Fixes: 12bb21a29c ("fscache: Implement cookie user counting and resource pinning")
Reported-by: Daire Byrne <daire.byrne@gmail.com>
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Tested-by: Daire Byrne <daire@dneg.com>
Link: https://lore.kernel.org/r/20221117115023.1350181-1-dwysocha@redhat.com/ # v1
Link: https://lore.kernel.org/r/20221117142915.1366990-1-dwysocha@redhat.com/ # v2
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add prepare_ondemand_read() callback dedicated for the on-demand read
scenario, so that callers from this scenario can be decoupled from
netfs_io_subrequest.
The original cachefiles_prepare_read() is now refactored to a generic
routine accepting a parameter list instead of netfs_io_subrequest.
There's no logic change, except that the debug id of subrequest and
request is removed from trace_cachefiles_prep_read().
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
Acked-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20221124034212.81892-2-jefflexu@linux.alibaba.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
The extent_io_tree::private_data was meant to be a preparatory work for
the metadata inode rework but that never materialized. Now it's used
only for an inode so it's better to change the appropriate type and
rename it.
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Add tracepoint events for recording the CXL uncorrectable and correctable
errors. For uncorrectable errors, there is additional data of 512B from
the header log register (CXL spec rev3 8.2.4.16.7). The trace event will
intake a dynamic array that will dump the entire Header Log data. If
multiple errors are set in the status register, then the
'first error' field (CXL spec rev3 v8.2.4.16.6) is read from the Error
Capabilities and Control Register in order to determine the error.
This implementation does not include CXL IDE Error details.
Cc: Steven Rostedt <rostedt@goodmis.org>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Link: https://lore.kernel.org/r/166974413388.1608150.5875712482260436188.stgit@djiang5-desk3.ch.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Since commit ac33e91e2d ("blk-iocost: implement vtime loss
compensation") rename original vtime_rate to vtime_base_rate
and current vtime_rate is original vtime_rate with compensation.
The current rate showed in tracepoint is mixed with vtime_rate
and vtime_base_rate:
1) In function ioc_adjust_base_vrate, the first trace_iocost_ioc_vrate_adj
shows vtime_rate, the second trace_iocost_ioc_vrate_adj shows
vtime_base_rate.
2) In function iocg_activate shows vtime_rate by calling
TRACE_IOCG_PATH(iocg_activate...
3) In function ioc_check_iocgs shows vtime_rate by calling
TRACE_IOCG_PATH(iocg_idle...
Trace vtime_base_rate instead of vtime_rate as:
1) Before commit ac33e91e2d ("blk-iocost: implement vtime loss
compensation"), the traced rate is without compensation, so still
show rate without compensation.
2) The vtime_base_rate is more stable while vtime_rate heavily depends on
excess budeget on current period which may change abruptly in next period.
Signed-off-by: Kemeng Shi <shikemeng@huawei.com>
Acked-by: Tejun Heo <tj@kernel.org>
Link: https://lore.kernel.org/r/20221018121932.10792-4-shikemeng@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For ACKs generated inside the I/O thread, transmit the ACK at the point of
generation. Where the ACK is generated outside of the I/O thread, it's
offloaded to the I/O thread to transmit it.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Add a tracepoint to log when a cwnd reset occurs due to lack of
transmission on a call.
Add stat counters to count transmission underflows (ie. when we have tx
window space, but sendmsg doesn't manage to keep up), cwnd resets and
transmission failures.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Move the functions from the call->processor and local->processor work items
into the domain of the I/O thread.
The call event processor, now called from the I/O thread, then takes over
the job of cranking the call state machine, processing incoming packets and
transmitting DATA, ACK and ABORT packets. In a future patch,
rxrpc_send_ACK() will transmit the ACK on the spot rather than queuing it
for later transmission.
The call event processor becomes purely received-skb driven. It only
transmits things in response to events. We use "pokes" to queue a dummy
skb to make it do things like start/resume transmitting data. Timer expiry
also results in pokes.
The connection event processor, becomes similar, though crypto events, such
as dealing with CHALLENGE and RESPONSE packets is offloaded to a work item
to avoid doing crypto in the I/O thread.
The local event processor is removed and VERSION response packets are
generated directly from the packet parser. Similarly, ABORTs generated in
response to protocol errors will be transmitted immediately rather than
being pushed onto a queue for later transmission.
Changes:
========
ver #2)
- Fix a couple of introduced lock context imbalances.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
A received skbuff needs a ref when it gets put on a call data queue or conn
packet queue, and rxrpc_input_packet() and co. jump through a lot of hoops
to avoid double-dropping the skbuff ref so that we can avoid getting a ref
when we queue the packet.
Change this so that the skbuff ref is unconditionally dropped by the caller
of rxrpc_input_packet(). An additional ref is then taken on the packet if
it is pushed onto a queue.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Move DATA transmission into the call processor work item. In a future
patch, this will be called from the I/O thread rather than being itsown
work item.
This will allow DATA transmission to be driven directly by incoming ACKs,
pokes and timers as those are processed.
The Tx queue is also split: The queue of packets prepared by sendmsg is now
places in call->tx_sendmsg and the packet dispatcher decants the packets
into call->tx_buffer as space becomes available in the transmission
window. This allows sendmsg to run ahead of the available space to try and
prevent an underflow in transmission.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Copy client call parameters into rxrpc_call earlier so that that can be
used to convey them to the connection code - which can then be offloaded to
the I/O thread.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Provide a means by which an event notification can be sent to a call such
that the I/O thread can process it rather than it being done in a separate
workqueue. This will allow a lot of locking to be removed.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Currently, rxrpc gives the connection's work item a ref on the connection
when it queues it - and this is called from the timer expiration function.
The problem comes when queue_work() fails (ie. the work item is already
queued): the timer routine must put the ref - but this may cause the
cleanup code to run.
This has the unfortunate effect that the cleanup code may then be run in
softirq context - which means that any spinlocks it might need to touch
have to be guarded to disable softirqs (ie. they need a "_bh" suffix).
(1) Don't give a ref to the work item.
(2) Simplify handling of service connections by adding a separate active
count so that the refcount isn't also used for this.
(3) Connection destruction for both client and service connections can
then be cleaned up by putting rxrpc_put_connection() out of line and
making a tidy progression through the destruction code (offloaded to a
workqueue if put from softirq or processor function context). The RCU
part of the cleanup then only deals with the freeing at the end.
(4) Make rxrpc_queue_conn() return immediately if it sees the active count
is -1 rather then queuing the connection.
(5) Make sure that the cleanup routine waits for the work item to
complete.
(6) Stash the rxrpc_net pointer in the conn struct so that the rcu free
routine can use it, even if the local endpoint has been freed.
Unfortunately, neither the timer nor the work item can simply get around
the problem by just using refcount_inc_not_zero() as the waits would still
have to be done, and there would still be the possibility of having to put
the ref in the expiration function.
Note the connection work item is mostly going to go away with the main
event work being transferred to the I/O thread, so the wait in (6) will
become obsolete.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Currently, rxrpc gives the call timer a ref on the call when it starts it
and this is passed along to the workqueue by the timer expiration function.
The problem comes when queue_work() fails (ie. the work item is already
queued): the timer routine must put the ref - but this may cause the
cleanup code to run.
This has the unfortunate effect that the cleanup code may then be run in
softirq context - which means that any spinlocks it might need to touch
have to be guarded to disable softirqs (ie. they need a "_bh" suffix).
Fix this by:
(1) Don't give a ref to the timer.
(2) Making the expiration function not do anything if the refcount is 0.
Note that this is more of an optimisation.
(3) Make sure that the cleanup routine waits for timer to complete.
However, this has a consequence that timer cannot give a ref to the work
item. Therefore the following fixes are also necessary:
(4) Don't give a ref to the work item.
(5) Make the work item return asap if it sees the ref count is 0.
(6) Make sure that the cleanup routine waits for the work item to
complete.
Unfortunately, neither the timer nor the work item can simply get around
the problem by just using refcount_inc_not_zero() as the waits would still
have to be done, and there would still be the possibility of having to put
the ref in the expiration function.
Note the call work item is going to go away with the work being transferred
to the I/O thread, so the wait in (6) will become obsolete.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Currently mm_struct maintains rss_stats which are updated on page fault
and the unmapping codepaths. For page fault codepath the updates are
cached per thread with the batch of TASK_RSS_EVENTS_THRESH which is 64.
The reason for caching is performance for multithreaded applications
otherwise the rss_stats updates may become hotspot for such applications.
However this optimization comes with the cost of error margin in the rss
stats. The rss_stats for applications with large number of threads can be
very skewed. At worst the error margin is (nr_threads * 64) and we have a
lot of applications with 100s of threads, so the error margin can be very
high. Internally we had to reduce TASK_RSS_EVENTS_THRESH to 32.
Recently we started seeing the unbounded errors for rss_stats for specific
applications which use TCP rx0cp. It seems like vm_insert_pages()
codepath does not sync rss_stats at all.
This patch converts the rss_stats into percpu_counter to convert the error
margin from (nr_threads * 64) to approximately (nr_cpus ^ 2). However
this conversion enable us to get the accurate stats for situations where
accuracy is more important than the cpu cost.
This patch does not make such tradeoffs - we can just use
percpu_counter_add_local() for the updates and percpu_counter_sum() (or
percpu_counter_sync() + percpu_counter_read) for the readers. At the
moment the readers are either procfs interface, oom_killer and memory
reclaim which I think are not performance critical and should be ok with
slow read. However I think we can make that change in a separate patch.
Link: https://lkml.kernel.org/r/20221024052841.3291983-1-shakeelb@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
As with PG_arch_2, this flag is only allowed on 64-bit architectures due
to the shortage of bits available. It will be used by the arm64 MTE code
in subsequent patches.
Signed-off-by: Peter Collingbourne <pcc@google.com>
Cc: Will Deacon <will@kernel.org>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Steven Price <steven.price@arm.com>
[catalin.marinas@arm.com: added flag preserving in __split_huge_page_tail()]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20221104011041.290951-5-pcc@google.com
Commit 4beba9486a ("mm: Add PG_arch_2 page flag") introduced a new
page flag for all 64-bit architectures. However, even if an architecture
is 64-bit, it may still have limited spare bits in the 'flags' member of
'struct page'. This may happen if an architecture enables SPARSEMEM
without SPARSEMEM_VMEMMAP as is the case with the newly added loongarch.
This architecture port needs 19 more bits for the sparsemem section
information and, while it is currently fine with PG_arch_2, adding any
more PG_arch_* flags will trigger build-time warnings.
Add a new CONFIG_ARCH_USES_PG_ARCH_X option which can be selected by
architectures that need more PG_arch_* flags beyond PG_arch_1. Select it
on arm64.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
[pcc@google.com: fix build with CONFIG_ARM64_MTE disabled]
Signed-off-by: Peter Collingbourne <pcc@google.com>
Reported-by: kernel test robot <lkp@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Steven Price <steven.price@arm.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20221104011041.290951-2-pcc@google.com
The trace events have a __bitmask field that can be used for anything
that requires bitmasks. Although currently it is only used for CPU
masks, it could be used in the future for any type of bitmasks.
There is some user space tooling that wants to know if a field is a CPU
mask and not just some random unsigned long bitmask. Introduce
"__cpumask()" helper functions that work the same as the current
__bitmask() helpers but displays in the format file:
field:__data_loc cpumask_t *[] mask; offset:36; size:4; signed:0;
Instead of:
field:__data_loc unsigned long[] mask; offset:32; size:4; signed:0;
The main difference is the type. Instead of "unsigned long" it is
"cpumask_t *". Note, this type field needs to be a real type in the
__dynamic_array() logic that both __cpumask and__bitmask use, but the
comparison field requires it to be a scalar type whereas cpumask_t is a
structure (non-scalar). But everything works when making it a pointer.
Valentin added changes to remove the need of passing in "nr_bits" and the
__cpumask will always use nr_cpumask_bits as its size.
Link: https://lkml.kernel.org/r/20221014080456.1d32b989@rorschach.local.home
Requested-by: Valentin Schneider <vschneid@redhat.com>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Signed-off-by: Valentin Schneider <vschneid@redhat.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Due to compiler optimizations like inlining, there are cases where
MMIO traces using _THIS_IP_ for caller information might not be
sufficient to provide accurate debug traces.
1) With optimizations (Seen with GCC):
In this case, _THIS_IP_ works fine and prints the caller information
since it will be inlined into the caller and we get the debug traces
on who made the MMIO access, for ex:
rwmmio_read: qcom_smmu_tlb_sync+0xe0/0x1b0 width=32 addr=0xffff8000087447f4
rwmmio_post_read: qcom_smmu_tlb_sync+0xe0/0x1b0 width=32 val=0x0 addr=0xffff8000087447f4
2) Without optimizations (Seen with Clang):
_THIS_IP_ will not be sufficient in this case as it will print only
the MMIO accessors itself which is of not much use since it is not
inlined as below for example:
rwmmio_read: readl+0x4/0x80 width=32 addr=0xffff8000087447f4
rwmmio_post_read: readl+0x48/0x80 width=32 val=0x4 addr=0xffff8000087447f4
So in order to handle this second case as well irrespective of the compiler
optimizations, add _RET_IP_ to MMIO trace to make it provide more accurate
debug information in all these scenarios.
Before:
rwmmio_read: readl+0x4/0x80 width=32 addr=0xffff8000087447f4
rwmmio_post_read: readl+0x48/0x80 width=32 val=0x4 addr=0xffff8000087447f4
After:
rwmmio_read: qcom_smmu_tlb_sync+0xe0/0x1b0 -> readl+0x4/0x80 width=32 addr=0xffff8000087447f4
rwmmio_post_read: qcom_smmu_tlb_sync+0xe0/0x1b0 -> readl+0x4/0x80 width=32 val=0x0 addr=0xffff8000087447f4
Fixes: 210031971c ("asm-generic/io: Add logging support for MMIO accessors")
Signed-off-by: Sai Prakash Ranjan <quic_saipraka@quicinc.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
In DLM when we send a dlm message it is easy to add the lock resource
name, but additional lookup is required when to trace the receive
message side. The idea here is to move the lookup work to the user by
using a lookup to find the right send message with recv message. As note
DLM can't drop any message which is guaranteed by a special session
layer.
For doing the lookup a 3 tupel is required as an unique identification
which is dst nodeid, src nodeid and sequence number. This patch adds the
destination nodeid to the dlm message trace points. The source nodeid is
given by the h_nodeid field inside the header.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch renames seq to h_seq as it is named in the dlm header
structure.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
Long Li says:
====================
Introduce Microsoft Azure Network Adapter (MANA) RDMA driver [netdev prep]
The first 11 patches which modify the MANA Ethernet driver to support
RDMA driver.
* 'mana-shared-6.2' of https://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
net: mana: Define data structures for protection domain and memory registration
net: mana: Define data structures for allocating doorbell page from GDMA
net: mana: Define and process GDMA response code GDMA_STATUS_MORE_ENTRIES
net: mana: Define max values for SGL entries
net: mana: Move header files to a common location
net: mana: Record port number in netdev
net: mana: Export Work Queue functions for use by RDMA driver
net: mana: Set the DMA device max segment size
net: mana: Handle vport sharing between devices
net: mana: Record the physical address for doorbell page region
net: mana: Add support for auxiliary device
====================
Link: https://lore.kernel.org/all/1667502990-2559-1-git-send-email-longli@linuxonhyperv.com/
Signed-off-by: Leon Romanovsky <leon@kernel.org>
This patch adds tracepoints for send and recv cases of dlm messages and
dlm rcom messages. In case of send and dlm message we add the dlm rsb
resource name this dlm messages belongs to. This has the advantage to
follow dlm messages on a per lock basis. In case of recv message the
resource name can be extracted by follow the send message sequence
number.
The dlm message DLM_MSG_PURGE doesn't belong to a lock request and will
not set the resource name in a dlm_message trace. The same for all rcom
messages.
There is additional handling required for this debugging functionality
which is tried to be small as possible. Also the midcomms layer gets
aware of lock resource names, for now this is required to make a
connection between sequence number and lock resource names. It is for
debugging purpose only.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
rxrpc has a problem in its congestion management in that it saves the
congestion window size (cwnd) from one call to another, but if this is 0 at
the time is saved, then the next call may not actually manage to ever
transmit anything.
To this end:
(1) Don't save cwnd between calls, but rather reset back down to the
initial cwnd and re-enter slow-start if data transmission is idle for
more than an RTT.
(2) Preserve ssthresh instead, as that is a handy estimate of pipe
capacity. Knowing roughly when to stop slow start and enter
congestion avoidance can reduce the tendency to overshoot and drop
larger amounts of packets when probing.
In future, cwind growth also needs to be constrained when the window isn't
being filled due to being application limited.
Reported-by: Simon Wilkinson <sxw@auristor.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Improve the tracking of which packets need to be transmitted by saving the
last ACK packet that we receive that has a populated soft-ACK table rather
than marking packets. Then we can step through the soft-ACK table and look
at the packets we've transmitted beyond that to determine which packets we
might want to retransmit.
We also look at the highest serial number that has been acked to try and
guess which packets we've transmitted the peer is likely to have seen. If
necessary, we send a ping to retrieve that number.
One downside that might be a problem is that we can't then compare the
previous acked/unacked state so easily in rxrpc_input_soft_acks() - which
is a potential problem for the slow-start algorithm.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Change the way the Tx queueing works to make the following ends easier to
achieve:
(1) The filling of packets, the encryption of packets and the transmission
of packets can be handled in parallel by separate threads, rather than
rxrpc_sendmsg() allocating, filling, encrypting and transmitting each
packet before moving onto the next one.
(2) Get rid of the fixed-size ring which sets a hard limit on the number
of packets that can be retained in the ring. This allows the number
of packets to increase without having to allocate a very large ring or
having variable-sized rings.
[Note: the downside of this is that it's then less efficient to locate
a packet for retransmission as we then have to step through a list and
examine each buffer in the list.]
(3) Allow the filler/encrypter to run ahead of the transmission window.
(4) Make it easier to do zero copy UDP from the packet buffers.
(5) Make it easier to do zero copy from userspace to the packet buffers -
and thence to UDP (only if for unauthenticated connections).
To that end, the following changes are made:
(1) Use the new rxrpc_txbuf struct instead of sk_buff for keeping packets
to be transmitted in. This allows them to be placed on multiple
queues simultaneously. An sk_buff isn't really necessary as it's
never passed on to lower-level networking code.
(2) Keep the transmissable packets in a linked list on the call struct
rather than in a ring. As a consequence, the annotation buffer isn't
used either; rather a flag is set on the packet to indicate ackedness.
(3) Use the RXRPC_CALL_TX_LAST flag to indicate that the last packet to be
transmitted has been queued. Add RXRPC_CALL_TX_ALL_ACKED to indicate
that all packets up to and including the last got hard acked.
(4) Wire headers are now stored in the txbuf rather than being concocted
on the stack and they're stored immediately before the data, thereby
allowing zerocopy of a single span.
(5) Don't bother with instant-resend on transmission failure; rather,
leave it for a timer or an ACK packet to trigger.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Get rid of the Rx ring and replace it with a pair of queues instead. One
queue gets the packets that are in-sequence and are ready for processing by
recvmsg(); the other queue gets the out-of-sequence packets for addition to
the first queue as the holes get filled.
The annotation ring is removed and replaced with a SACK table. The SACK
table has the bits set that correspond exactly to the sequence number of
the packet being acked. The SACK ring is copied when an ACK packet is
being assembled and rotated so that the first ACK is in byte 0.
Flow control handling is altered so that packets that are moved to the
in-sequence queue are hard-ACK'd even before they're consumed - and then
the Rx window size in the ACK packet (rsize) is shrunk down to compensate
(even going to 0 if the window is full).
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Split up received jumbo packets into separate skbuffs by cloning the
original skbuff for each subpacket and setting the offset and length of the
data in that subpacket in the skbuff's private data. The subpackets are
then placed on the recvmsg queue separately. The security class then gets
to revise the offset and length to remove its metadata.
If we fail to clone a packet, we just drop it and let the peer resend it.
The original packet gets used for the final subpacket.
This should make it easier to handle parallel decryption of the subpackets.
It also simplifies the handling of lost or misordered packets in the
queuing/buffering loop as the possibility of overlapping jumbo packets no
longer needs to be considered.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Split the rxrpc_recvmsg tracepoint so that the tracepoints that are about
data packet processing (and which have extra pieces of information) are
separate from the tracepoint that shows the general flow of recvmsg().
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Clean up the rxrpc_propose_ACK() function. If deferred PING ACK proposal
is split out, it's only really needed for deferred DELAY ACKs. All other
ACKs, bar terminal IDLE ACK are sent immediately. The deferred IDLE ACK
submission can be handled by conversion of a DELAY ACK into an IDLE ACK if
there's nothing to be SACK'd.
Also, because there's a delay between an ACK being generated and being
transmitted, it's possible that other ACKs of the same type will be
generated during that interval. Apart from the ACK time and the serial
number responded to, most of the ACK body, including window and SACK
parameters, are not filled out till the point of transmission - so we can
avoid generating a new ACK if there's one pending that will cover the SACK
data we need to convey.
Therefore, don't propose a new DELAY or IDLE ACK for a call if there's one
already pending.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
