lkml.org 
[lkml]   [2020]   [Nov]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RFC net-next 1/2] ethtool: add support for controling the type of adaptive coalescing
From
Date


On 2020/11/20 6:02, Michal Kubecek wrote:
> On Thu, Nov 19, 2020 at 04:56:42PM +0800, tanhuazhong wrote:
>> On 2020/11/19 12:15, Andrew Lunn wrote:
>>>> diff --git a/include/uapi/linux/ethtool.h b/include/uapi/linux/ethtool.h
>>>> index 9ca87bc..afd8de2 100644
>>>> --- a/include/uapi/linux/ethtool.h
>>>> +++ b/include/uapi/linux/ethtool.h
>>>> @@ -433,6 +433,7 @@ struct ethtool_modinfo {
>>>> * a TX interrupt, when the packet rate is above @pkt_rate_high.
>>>> * @rate_sample_interval: How often to do adaptive coalescing packet rate
>>>> * sampling, measured in seconds. Must not be zero.
>>>> + * @use_dim: Use DIM for IRQ coalescing, if adaptive coalescing is enabled.
>>>> *
>>>> * Each pair of (usecs, max_frames) fields specifies that interrupts
>>>> * should be coalesced until
>>>> @@ -483,6 +484,7 @@ struct ethtool_coalesce {
>>>> __u32 tx_coalesce_usecs_high;
>>>> __u32 tx_max_coalesced_frames_high;
>>>> __u32 rate_sample_interval;
>>>> + __u32 use_dim;
>>>> };
>>>
>>> You cannot do this.
>>>
>>> static noinline_for_stack int ethtool_set_coalesce(struct net_device *dev,
>>> void __user *useraddr)
>>> {
>>> struct ethtool_coalesce coalesce;
>>> int ret;
>>>
>>> if (!dev->ethtool_ops->set_coalesce)
>>> return -EOPNOTSUPP;
>>>
>>> if (copy_from_user(&coalesce, useraddr, sizeof(coalesce)))
>>> return -EFAULT;
>>>
>>> An old ethtool binary is not going to set this extra last byte to
>>> anything meaningful. You cannot tell if you have an old or new user
>>> space, so you have no idea if it put anything into use_dim, or if it
>>> is random junk.
>
> Even worse, as there is no indication of data length, ETHTOOL_GCOALESCE
> ioctl request from old ethtool on new kernel would result in kernel
> writing past the end of userspace buffer.
>
>>> You have to leave the IOCTL interface unchanged, and limit this new
>>> feature to the netlink API.
>>>
>>
>> Hi, Andrew.
>> thanks for pointing out this problem, i will fix it.
>> without callling set_coalesce/set_coalesce of ethtool_ops, do you have any
>> suggestion for writing/reading this new attribute to/from the driver? add a
>> new field in net_device or a new callback function in ethtool_ops seems not
>> good.
>
> We could use a similar approach as struct ethtool_link_ksettings, e.g.
>
> struct kernel_ethtool_coalesce {
> struct ethtool_coalesce base;
> /* new members which are not part of UAPI */
> }
>
> get_coalesce() and set_coalesce() would get pointer to struct
> kernel_ethtool_coalesce and ioctl code would be modified to only touch
> the base (legacy?) part.
> > While already changing the ops arguments, we could also add extack
> pointer, either as a separate argument or as struct member (I slightly
> prefer the former).
>


Hi, Michal.

If changing the ops arguments, each driver who implement
set_coalesce/get_coalesce of ethtool_ops need to be updated. Is it
acceptable adding two new ops to get/set ext_coalesce info (like
ecc31c60240b ("ethtool: Add link extended state") does)? Maybe i can
send V2 in this way, and then could you help to see which one is more
suitable?


> BtW, please don't forget to update the message descriptions in
> Documentation/networking/ethtool-netlink.rst
>
> Michal
>
> .
>

\
 
 \ /
  Last update: 2020-11-20 04:01    [W:0.060 / U:0.896 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site