This ensures that future programmers do not have to remember to re-size
the bitmaps due to adding new values. Although this is unlikely for this
driver, it may happen and it's best to prevent it from ever being an
issue.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Replace bitwise operators and #defines with a BITMAP and enumeration
values. This is similar to how we handle the "state" values as well.
This has two distinct advantages over the old method. First, we ensure
correctness of operations which are currently problematic due to race
conditions. Suppose that two kernel threads are running, such as the
watchdog and an ethtool ioctl, and both modify flags. We'll say that the
watchdog is CPU A, and the ethtool ioctl is CPU B.
CPU A sets FLAG_1, which can be seen as
CPU A read FLAGS
CPU A write FLAGS | FLAG_1
CPU B sets FLAG_2, which can be seen as
CPU B read FLAGS
CPU A write FLAGS | FLAG_2
However, "|=" and "&=" operators are not actually atomic. So this could
be ordered like the following:
CPU A read FLAGS -> variable
CPU B read FLAGS -> variable
CPU A write FLAGS (variable | FLAG_1)
CPU B write FLAGS (variable | FLAG_2)
Notice how the 2nd write from CPU B could actually undo the write from
CPU A because it isn't guaranteed that the |= operation is atomic.
In practice the race windows for most flag writes is incredibly narrow
so it is not easy to isolate issues. However, the more flags we have,
the more likely they will cause problems. Additionally, if such
a problem were to arise, it would be incredibly difficult to track down.
Second, there is an additional advantage beyond code correctness. We can
now automatically size the BITMAP if more flags were added, so that we
do not need to remember that flags is u32 and thus if we added too many
flags we would over-run the variable. This is not a likely occurrence
for fm10k driver, but this patch can serve as an example for other
drivers which have many more flags.
This particular change does have a bit of trouble converting some of the
idioms previously used with the #defines for flags. Specifically, when
converting FM10K_FLAG_RSS_FIELD_IPV[46]_UDP flags. This whole operation
was actually quite problematic, because we actually stored flags
separately. This could more easily show the problem of the above
re-ordering issue.
This is really difficult to test whether atomics make a difference in
practical scenarios, but you can ensure that basic functionality remains
the same. This patch has a lot of code coverage, but most of it is
relatively simple.
While we are modifying these files, update their copyright year.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This was accidentally removed when we defeatured the full 1588 Clock
support. We need to report the Rx descriptor timestamp value so that
applications built on top of the IES API can function properly.
Additionally, remove the FM10K_FLAG_RX_TS_ENABLED, as it is not used now
that 1588 functionality has been removed.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
On packet RX, we perform a dma sync for cpu before passing the
packet up. Here we limit that sync to the actual length of the
incoming packet, rather than always syncing the entire buffer.
Signed-off-by: Scott Peterson <scott.d.peterson@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Trivial change here to cleanup a checkpatch.pl warning that got
introduced when changing to alloc_workqueue.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
In preparation for adding Geneve Rx offload support, refactor the
current VXLAN offload flow to be a bit more generic so that it will be
easier to add the new Geneve code. The fm10k hardware supports one VXLAN
and one Geneve tunnel, so we will eventually treat the VXLAN and Geneve
tunnels identically. To this end, factor out the code that handles the
current list so that we can use the generic flow for both tunnels in the
next patch.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
When fm10k_poll fully cleans rings it returns 0. This is incorrect as it
messes up the budget accounting in the core NAPI code. Fix this by
returning actual work done, capped at budget - 1 since the core doesn't
expect a return of the full budget when the driver modifies the NAPI
status.
Cc: Paolo Abeni <pabeni@redhat.com>
Cc: Venkatesh Srinivas <venkateshs@google.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
While technically not needed, as all our uses of ACCESS_ONCE are scalar
types, we already use READ_ONCE in a few places, and for code
readability we can swap all the uses of the older ACCESS_ONCE into
READ_ONCE.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
A previous patch added support to check for hardware Tx pending in the
fm10k_down routine. This support was intended to ensure that we
accurately check what the hardware state is. However, checking for Tx
hangs in this manor during the hotpath results in a large performance
hit. Avoid this by making the hotpath check use the SW counters instead.
Fixes: a0f53cf49cb0 ("fm10k: use actual hardware registers when checking for pending Tx", 2016-06-08)
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The pci_enable_msix_range() function returns a positive value of the
number of allocated vectors if it succeeds. On failure it returns
a negative error code. Return this code properly so that the error
message printed by the driver will show the actual error code instead of
being masked by -ENOMEM.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
It turns out that sometimes during a reset the Tx queues will be
temporarily stuck longer than .stop_hw() expects. Work around this issue
by attempting to .stop_hw() first. If it tails, wait a number of
attempts until the Tx queues appear to be drained. After this, attempt
stop_hw() again. This ensures that we avoid waiting if we don't need to,
such as during the first initialization of a VF, and give the proper
amount of time necessary to recover from most situations. It is possible
that the hardware is actually stuck. For PFs, this is usually fixed by
a datapath reset. Unfortunately the VF cannot request a similar reset
for itself.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
alloc_workqueue replaces deprecated create_workqueue().
A dedicated workqueue has been used since the workitem (viz
fm10k_service_task, which manages and runs other subtasks) is involved in
normal device operation and requires forward progress under memory
pressure.
create_workqueue has been replaced with alloc_workqueue with max_active
as 0 since there is no need for throttling the number of active work
items.
Since network devices may be used in memory reclaim path,
WQ_MEM_RECLAIM has been set to guarantee forward progress.
flush_workqueue is unnecessary since destroy_workqueue() itself calls
drain_workqueue() which flushes repeatedly till the workqueue
becomes empty. Hence the call to flush_workqueue() has been dropped.
Signed-off-by: Bhaktipriya Shridhar <bhaktipriya96@gmail.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
While reviewing the i40e driver changes to support page based receive I
realized that I had overlooked the fact that the fm10k hardware required a
512 byte alignment for Rx buffers. This patch is meant to address that by
changing the alignment for Rx buffers to 512 bytes instead of allowing it
to be L1 cache aligned.
Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Check for and handle IPv6 extended headers so that Tx checksum offload
can be done. Also use skb_checksum_help for unexpected cases. This was
originally discovered in ixgbe.
Reported-by: Mark Rustad <mark.d.rustad@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Update every header file and other locations to consistently use
Intel(R) instead of just Intel. Also update copyright year of files
which we modified.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
When writing a new default redirection table, we needed to populate
a new RSS table using ethtool_rxfh_indir_default. We populated this
table into a region of memory allocated using kcalloc, but never checked
this for NULL. Fix this by moving the default table generation into
fm10k_write_reta. If this function is passed a table, use it. Otherwise,
generate the default table using ethtool_rxfh_indir_default, 4 at at
time.
Fixes: 0ea7fae440 ("fm10k: use ethtool_rxfh_indir_default for default redirection table")
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The 1588 support within fm10k does not work correctly with the current
version of the switch management software, and likely never worked
correctly to begin with. Remove support for PTP/1588. Update copyright
year for all these files while we're touching them.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Use DRV_SUMMARY, similar to DRV_VERSION so that we don't have to
duplicate the driver summary in multiple places.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch enables bulk free in Tx cleanup for fm10k and cleans up the
boolean logic in the polling routines for fm10k in the hopes of avoiding
any mix-ups similar to what occurred with i40e and i40evf.
Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The fm10k driver used its own code for generating a default indirection
table on device load, which was not the same as the default generated by
ethtool when indir_size of 0 is passed to SRXFH. Take advantage of
ethtool_rxfh_indir_default() and simplify code to write the redirection
table to reduce some code duplication.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
In fm10k_set_num_queues, we previously assigned the base template. This
would always be overwritten by either fm10k_set_qos_queues or
fm10k_set_rss_queues. In either case, we don't need the base values, so
we can just remove them.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Use BIT() macro instead.
Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The semantic patch that makes this change is available
in scripts/coccinelle/misc/compare_const_fl.cocci.
More information about semantic patching is available at
http://coccinelle.lip6.fr/
Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Pull networking updates from David Miller:
"Highlights:
1) Support more Realtek wireless chips, from Jes Sorenson.
2) New BPF types for per-cpu hash and arrap maps, from Alexei
Starovoitov.
3) Make several TCP sysctls per-namespace, from Nikolay Borisov.
4) Allow the use of SO_REUSEPORT in order to do per-thread processing
of incoming TCP/UDP connections. The muxing can be done using a
BPF program which hashes the incoming packet. From Craig Gallek.
5) Add a multiplexer for TCP streams, to provide a messaged based
interface. BPF programs can be used to determine the message
boundaries. From Tom Herbert.
6) Add 802.1AE MACSEC support, from Sabrina Dubroca.
7) Avoid factorial complexity when taking down an inetdev interface
with lots of configured addresses. We were doing things like
traversing the entire address less for each address removed, and
flushing the entire netfilter conntrack table for every address as
well.
8) Add and use SKB bulk free infrastructure, from Jesper Brouer.
9) Allow offloading u32 classifiers to hardware, and implement for
ixgbe, from John Fastabend.
10) Allow configuring IRQ coalescing parameters on a per-queue basis,
from Kan Liang.
11) Extend ethtool so that larger link mode masks can be supported.
From David Decotigny.
12) Introduce devlink, which can be used to configure port link types
(ethernet vs Infiniband, etc.), port splitting, and switch device
level attributes as a whole. From Jiri Pirko.
13) Hardware offload support for flower classifiers, from Amir Vadai.
14) Add "Local Checksum Offload". Basically, for a tunneled packet
the checksum of the outer header is 'constant' (because with the
checksum field filled into the inner protocol header, the payload
of the outer frame checksums to 'zero'), and we can take advantage
of that in various ways. From Edward Cree"
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1548 commits)
bonding: fix bond_get_stats()
net: bcmgenet: fix dma api length mismatch
net/mlx4_core: Fix backward compatibility on VFs
phy: mdio-thunder: Fix some Kconfig typos
lan78xx: add ndo_get_stats64
lan78xx: handle statistics counter rollover
RDS: TCP: Remove unused constant
RDS: TCP: Add sysctl tunables for sndbuf/rcvbuf on rds-tcp socket
net: smc911x: convert pxa dma to dmaengine
team: remove duplicate set of flag IFF_MULTICAST
bonding: remove duplicate set of flag IFF_MULTICAST
net: fix a comment typo
ethernet: micrel: fix some error codes
ip_tunnels, bpf: define IP_TUNNEL_OPTS_MAX and use it
bpf, dst: add and use dst_tclassid helper
bpf: make skb->tc_classid also readable
net: mvneta: bm: clarify dependencies
cls_bpf: reset class and reuse major in da
ldmvsw: Checkpatch sunvnet.c and sunvnet_common.c
ldmvsw: Add ldmvsw.c driver code
...
The success of CMA allocation largely depends on the success of
migration and key factor of it is page reference count. Until now, page
reference is manipulated by direct calling atomic functions so we cannot
follow up who and where manipulate it. Then, it is hard to find actual
reason of CMA allocation failure. CMA allocation should be guaranteed
to succeed so finding offending place is really important.
In this patch, call sites where page reference is manipulated are
converted to introduced wrapper function. This is preparation step to
add tracepoint to each page reference manipulation function. With this
facility, we can easily find reason of CMA allocation failure. There is
no functional change in this patch.
In addition, this patch also converts reference read sites. It will
help a second step that renames page._count to something else and
prevents later attempt to direct access to it (Suggested by Andrew).
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Also print an error message incase we do have to reconfigure as this
should no longer happen anymore due to ethtool changes. If it somehow
does occur, user should be made aware of it.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Cleans up checkpatch GLOBAL_INITIALIZERS error
Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
If the q_vector allocation fails we should free the resources associated
with the MSI-X vector table.
Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Reviewed-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
We haven't bumped the driver version in a while despite many fixes being
pulled in from the out-of-tree Sourceforge driver. Update the version to
match.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The existing adaptive ITR algorithm is overly restrictive. It throttles
incorrectly for various traffic rates, and does not produce good
performance. The algorithm now allows for more interrupts per second,
and does some calculation to help improve for smaller packet loads. In
addition, take into account the new itr_scale from the hardware which
indicates how much to scale due to PCIe link speed.
Reported-by: Matthew Vick <matthew.vick@intel.com>
Reported-by: Alex Duyck <alexander.duyck@gmail.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Define a macro for identifying when the itr value is dynamic or
adaptive. The concept was taken from i40e. This helps make clear what
the check is, and reduces the line length to something more reasonable
in a few places.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch corrects an issue in which the polling routine would increase
the budget for Rx to at least 1 per queue if multiple queues were present.
This would result in Rx packets being processed when the budget was 0 which
is meant to indicate that no Rx can be handled.
Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
As per Eric Dumazet's previous patches:
(see commit (24d2e4a507) - tg3: use napi_complete_done())
Quoting verbatim:
Using napi_complete_done() instead of napi_complete() allows
us to use /sys/class/net/ethX/gro_flush_timeout
GRO layer can aggregate more packets if the flush is delayed a bit,
without having to set too big coalescing parameters that impact
latencies.
</end quote>
Tested
configuration: low latency via ethtool -C ethx adaptive-rx off
rx-usecs 10 adaptive-tx off tx-usecs 15
workload: streaming rx using netperf TCP_MAERTS
igb:
MIGRATED TCP MAERTS TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.0.1 () port 0 AF_INET : demo
...
Interim result: 941.48 10^6bits/s over 1.000 seconds ending at 1440193171.589
Alignment Offset Bytes Bytes Recvs Bytes Sends
Local Remote Local Remote Xfered Per Per
Recv Send Recv Send Recv (avg) Send (avg)
8 8 0 0 1176930056 1475.36 797726 16384.00 71905
MIGRATED TCP MAERTS TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.0.1 () port 0 AF_INET : demo
...
Interim result: 941.49 10^6bits/s over 0.997 seconds ending at 1440193142.763
Alignment Offset Bytes Bytes Recvs Bytes Sends
Local Remote Local Remote Xfered Per Per
Recv Send Recv Send Recv (avg) Send (avg)
8 8 0 0 1175182320 50476.00 23282 16384.00 71816
i40e:
Hard to test because the traffic is incoming so fast (24Gb/s) that GRO
always receives 87kB, even at the highest interrupt rate.
Other drivers were only compile tested.
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Check for actual value NETREG_UNINITIALIZED in case it ever changes from
the current value of zero.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Add a private ethtool flag to enable display of these statistics, which
are generally less useful. However, sometimes it can be useful for
debugging purposes. The most useful portion is the ability to see what
the PF thinks the VF mailboxes look like.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch ensures that VLAN traffic on the default VID will go to the
corresponding VLAN device if it exists. To do this, mask the rx_ring VID
if we have an active VLAN on that VID.
For this to work correctly, we need to update fm10k_process_skb_fields
to correctly mask off the VLAN_PRIO_MASK bits and compare them
separately, otherwise we incorrectly compare the priority bits with the
cleared flag. This also happens to fix a related bug where having
priority bits set causes us to incorrectly classify traffic.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change pulls out the optimization that assumed that all fragments
would be limited to page size. That hasn't been the case for some time now
and to assume this is incorrect as the TCP allocator can provide up to a
32K page fragment.
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Acked-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Commit c48a11c7ad ("netvm: propagate page->pfmemalloc to skb") added
checks for page->pfmemalloc to __skb_fill_page_desc():
if (page->pfmemalloc && !page->mapping)
skb->pfmemalloc = true;
It assumes page->mapping == NULL implies that page->pfmemalloc can be
trusted. However, __delete_from_page_cache() can set set page->mapping
to NULL and leave page->index value alone. Due to being in union, a
non-zero page->index will be interpreted as true page->pfmemalloc.
So the assumption is invalid if the networking code can see such a page.
And it seems it can. We have encountered this with a NFS over loopback
setup when such a page is attached to a new skbuf. There is no copying
going on in this case so the page confuses __skb_fill_page_desc which
interprets the index as pfmemalloc flag and the network stack drops
packets that have been allocated using the reserves unless they are to
be queued on sockets handling the swapping which is the case here and
that leads to hangs when the nfs client waits for a response from the
server which has been dropped and thus never arrive.
The struct page is already heavily packed so rather than finding another
hole to put it in, let's do a trick instead. We can reuse the index
again but define it to an impossible value (-1UL). This is the page
index so it should never see the value that large. Replace all direct
users of page->pfmemalloc by page_is_pfmemalloc which will hide this
nastiness from unspoiled eyes.
The information will get lost if somebody wants to use page->index
obviously but that was the case before and the original code expected
that the information should be persisted somewhere else if that is
really needed (e.g. what SLAB and SLUB do).
[akpm@linux-foundation.org: fix blooper in slub]
Fixes: c48a11c7ad ("netvm: propagate page->pfmemalloc to skb")
Signed-off-by: Michal Hocko <mhocko@suse.com>
Debugged-by: Vlastimil Babka <vbabka@suse.com>
Debugged-by: Jiri Bohac <jbohac@suse.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Miller <davem@davemloft.net>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: <stable@vger.kernel.org> [3.6+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This change folds the fm10k_pull_tail call into fm10k_add_rx_frag. The
advantage to doing this is that the fragment doesn't have to be modified
after it is added to the skb.
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The netpoll path will call napi->poll with a budget of 0 in order to clean
the Tx rings only. This change updates the fm10k driver so that it will
correctly support that instead of cleaning 1 Rx frame if a budget of 0 is
received.
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With the recent driver changes, bump the version.
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Since we run the watchdog periodically, which might take a while and
potentially monopolize the system default workqueue, create our own
separate work queue. This also helps reduce and stabilize latency
between scheduling the work in our interrupt and actually performing
the work. Still use a timer for the regular scheduled interval but
queue the work onto its own work queue.
It seemed overkill to create a single workqueue per interface, so we
just spawn a single work queue for all interfaces upon driver load. For
this reason, use a multi-threaded workqueue with one thread per
processor, rather than single threaded queue.
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Acked-by: Matthew Vick <matthew.vick@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Since we already print this message when a reset is requested via the
RESET_REQUESTED flag, we do not need to print it before setting the
flag.
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Acked-by: Matthew Vick <matthew.vick@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
