Skip to content

L
linux
Switch branch/tag
  1. 01 Feb, 2024 3 commits
    • Shyam Prasad N's avatar
      cifs: failure to add channel on iface should bump up weight · 6aac002b
      Shyam Prasad N authored 
      After the interface selection policy change to do a weighted
round robin, each iface maintains a weight_fulfilled. When the
weight_fulfilled reaches the total weight for the iface, we know
that the weights can be reset and ifaces can be allocated from
scratch again.

During channel allocation failures on a particular channel,
weight_fulfilled is not incremented. If a few interfaces are
inactive, we could end up in a situation where the active
interfaces are all allocated for the total_weight, and inactive
ones are all that remain. This can cause a situation where
no more channels can be allocated further.

This change fixes it by increasing weight_fulfilled, even when
channel allocation failure happens. This could mean that if
there are temporary failures in channel allocation, the iface
weights may not strictly be adhered to. But that's still okay.

Fixes: a6d8fb54 ("cifs: distribute channels across interfaces based on speed")
Signed-off-by: default avatarShyam Prasad N <sprasad@microsoft.com>
Signed-off-by: default avatarSteve French <stfrench@microsoft.com>
      6aac002b
    • Shyam Prasad N's avatar
      cifs: do not search for channel if server is terminating · 88675b22
      Shyam Prasad N authored 
      In order to scale down the channels, the following sequence
of operations happen:
1. server struct is marked for terminate
2. the channel is deallocated in the ses->chans array
3. at a later point the cifsd thread actually terminates the server

Between 2 and 3, there can be calls to find the channel for
a server struct. When that happens, there can be an ugly warning
that's logged. But this is expected.

So this change does two things:
1. in cifs_ses_get_chan_index, if server->terminate is set, return
2. always make sure server->terminate is set with chan_lock held
Signed-off-by: default avatarShyam Prasad N <sprasad@microsoft.com>
Signed-off-by: default avatarSteve French <stfrench@microsoft.com>
      88675b22
    • Shyam Prasad N's avatar
      cifs: avoid redundant calls to disable multichannel · e77e15fa
      Shyam Prasad N authored 
      When the server reports query network interface info call
as unsupported following a tree connect, it means that
multichannel is unsupported, even if the server capabilities
report otherwise.

When this happens, cifs_chan_skip_or_disable is called to
disable multichannel on the client. However, we only need
to call this when multichannel is currently setup.

Fixes: f591062b ("cifs: handle servers that still advertise multichannel after disabling")
Signed-off-by: default avatarShyam Prasad N <sprasad@microsoft.com>
Signed-off-by: default avatarSteve French <stfrench@microsoft.com>
      e77e15fa
  3. 31 Jan, 2024 1 commit
    • Shyam Prasad N's avatar
      cifs: make sure that channel scaling is done only once · ee36a3b3
      Shyam Prasad N authored 
      Following a successful cifs_tree_connect, we have the code
to scale up/down the number of channels in the session.
However, it is not protected by a lock today.

As a result, this code can be executed by several processes
that select the same channel. The core functions handle this
well, as they pick chan_lock. However, we've seen cases where
smb2_reconnect throws some warnings.

To fix that, this change introduces a flags bitmap inside the
cifs_ses structure. A new flag type is used to ensure that
only one process enters this section at any time.
Signed-off-by: default avatarShyam Prasad N <sprasad@microsoft.com>
Signed-off-by: default avatarSteve French <stfrench@microsoft.com>
      ee36a3b3
  5. 29 Jan, 2024 1 commit
  7. 28 Jan, 2024 7 commits
  9. 27 Jan, 2024 9 commits
  11. 26 Jan, 2024 19 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7