Skip to content

L
linux
Switch branch/tag
  1. 06 Apr, 2017 6 commits
    • Jacob Keller's avatar
      fm10k: allow service task to reschedule itself · b4fd8ffc
      Jacob Keller authored 
      If some code path executes fm10k_service_event_schedule(), it is
guaranteed that we only queue the service task once, since we use
__FM10K_SERVICE_SCHED flag. Unfortunately this has a side effect that if
a service request occurs while we are currently running the watchdog, it
is possible that we will fail to notice the request and ignore it until
the next time the request occurs.

This can cause problems with pf/vf mailbox communication and other
service event tasks. To avoid this, introduce a FM10K_SERVICE_REQUEST
bit. When we successfully schedule (and set the _SCHED bit) the service
task, we will clear this bit. However, if we are unable to currently
schedule the service event, we just set the new SERVICE_REQUEST bit.

Finally, after the service event completes, we will re-schedule if the
request bit has been set.

This should ensure that we do not miss any service event schedules,
since we will re-schedule it once the currently running task finishes.
This means that for each request, we will always schedule the service
task to run at least once in full after the request came in.

This will avoid timing issues that can occur with the service event
scheduling. We do pay a cost in re-running many tasks, but all the
service event tasks use either flags to avoid duplicate work, or are
tolerant of being run multiple times.
Signed-off-by: default avatarJacob Keller <jacob.e.keller@intel.com>
Tested-by: default avatarKrishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: default avatarJeff Kirsher <jeffrey.t.kirsher@intel.com>
      b4fd8ffc
    • Jacob Keller's avatar
      fm10k: future-proof state bitmaps using DECLARE_BITMAP · 46929557
      Jacob Keller authored 
      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: default avatarJacob Keller <jacob.e.keller@intel.com>
Tested-by: default avatarKrishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: default avatarJeff Kirsher <jeffrey.t.kirsher@intel.com>
      46929557
    • Jacob Keller's avatar
      fm10k: use a BITMAP for flags to avoid race conditions · 3ee7b3a3
      Jacob Keller authored 
      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: default avatarJacob Keller <jacob.e.keller@intel.com>
Tested-by: default avatarKrishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: default avatarJeff Kirsher <jeffrey.t.kirsher@intel.com>
      3ee7b3a3
    • Phil Turnbull's avatar
      fm10k: correctly check if interface is removed · 540fca35
      Phil Turnbull authored 
      FM10K_REMOVED expects a hardware address, not a 'struct fm10k_hw'.

Fixes: 5cb8db4a ("fm10k: Add support for VF")
Signed-off-by: default avatarPhil Turnbull <phil.turnbull@oracle.com>
Tested-by: default avatarKrishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: default avatarJeff Kirsher <jeffrey.t.kirsher@intel.com>
      540fca35
    • Jarod Wilson's avatar
      bonding: attempt to better support longer hw addresses · faeeb317
      Jarod Wilson authored 
      People are using bonding over Infiniband IPoIB connections, and who knows
what else. Infiniband has a hardware address length of 20 octets
(INFINIBAND_ALEN), and the network core defines a MAX_ADDR_LEN of 32.
Various places in the bonding code are currently hard-wired to 6 octets
(ETH_ALEN), such as the 3ad code, which I've left untouched here. Besides,
only alb is currently possible on Infiniband links right now anyway, due
to commit 1533e773, so the alb code is where most of the changes are.

One major component of this change is the addition of a bond_hw_addr_copy
function that takes a length argument, instead of using ether_addr_copy
everywhere that hardware addresses need to be copied about. The other
major component of this change is converting the bonding code from using
struct sockaddr for address storage to struct sockaddr_storage, as the
former has an address storage space of only 14, while the latter is 128
minus a few, which is necessary to support bonding over device with up to
MAX_ADDR_LEN octet hardware addresses. Additionally, this probably fixes
up some memory corruption issues with the current code, where it's
possible to write an infiniband hardware address into a sockaddr declared
on the stack.

Lightly tested on a dual mlx4 IPoIB setup, which properly shows a 20-octet
hardware address now:

$ cat /proc/net/bonding/bond0
Ethernet Channel Bonding Driver: v3.7.1 (April 27, 2011)

Bonding Mode: fault-tolerance (active-backup) (fail_over_mac active)
Primary Slave: mlx4_ib0 (primary_reselect always)
Currently Active Slave: mlx4_ib0
MII Status: up
MII Polling Interval (ms): 100
Up Delay (ms): 100
Down Delay (ms): 100

Slave Interface: mlx4_ib0
MII Status: up
Speed: Unknown
Duplex: Unknown
Link Failure Count: 0
Permanent HW addr:
80:00:02:08:fe:80:00:00:00:00:00:00:e4:1d:2d:03:00:1d:67:01
Slave queue ID: 0

Slave Interface: mlx4_ib1
MII Status: up
Speed: Unknown
Duplex: Unknown
Link Failure Count: 0
Permanent HW addr:
80:00:02:09:fe:80:00:00:00:00:00:01:e4:1d:2d:03:00:1d:67:02
Slave queue ID: 0

Also tested with a standard 1Gbps NIC bonding setup (with a mix of
e1000 and e1000e cards), running LNST's bonding tests.

CC: Jay Vosburgh <j.vosburgh@gmail.com>
CC: Veaceslav Falico <vfalico@gmail.com>
CC: Andy Gospodarek <andy@greyhouse.net>
CC: netdev@vger.kernel.org
Signed-off-by: default avatarJarod Wilson <jarod@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      faeeb317
    • Edward Cree's avatar
      sfc: don't insert mc_list on low-latency firmware if it's too long · 148cbab6
      Edward Cree authored 
      If the mc_list is longer than 256 addresses, we enter mc_promisc mode.
If we're in mc_promisc mode and the firmware doesn't support cascaded
 multicast, normally we also insert our mc_list, to prevent stealing by
 another VI.  However, if the mc_list was too long, this isn't really
 helpful - the MC groups that didn't fit in the list can still get
 stolen, and having only some of them stealable will probably cause
 more confusing behaviour than having them all stealable.  Since
 inserting 256 multicast filters takes a long time and can lead to MCDI
 state machine timeouts, just skip the mc_list insert in this overflow
 condition.
Signed-off-by: default avatarEdward Cree <ecree@solarflare.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      148cbab6
  3. 05 Apr, 2017 34 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7