Re: [PATCH] can: use sock_efree instead of own destructor

On 10.03.2015 04:48, Florian Westphal wrote:

It is identical to the can destructor.

Signed-off-by: Florian Westphal <fw@strlen.de>

Acked-by: Oliver Hartkopp <socketcan@hartkopp.net>

Hello Florian,

the other callers use it in the same way so it's a good simplification.
Btw. the name of sock_efree() is a bit misleading - nothing is free'd here.

Won't it be better to rename sock_efree(skb) with sock_put_skb(skb) or 
something like that? sock_efree() has no comment why it's named like this.

Regards,
Oliver

ps. changed from linux ML to netdev and linux-can ML in CC

On Tue, 2015-03-10 at 07:14 +0100, Oliver Hartkopp wrote:

the other callers use it in the same way so it's a good simplification.
Btw. the name of sock_efree() is a bit misleading - nothing is free'd here.

Won't it be better to rename sock_efree(skb) with sock_put_skb(skb) or 
something like that? sock_efree() has no comment why it's named like this.

I would prefer name stays as is. It eases searches in changelogs to not
change function names unless really needed, for backports and code
maintenance.


# git log | grep sock_efree
    net: merge cases where sock_efree and sock_edemux are the same function
    Since sock_efree and sock_demux are essentially the same code for non-TCP
    In addition I have added a destructor named sock_efree which is meant to

sock_efree was added in commit 62bccb8cdb69051b95a55ab0c489e3cab261c8ef

I guess Alexander chose the name close to sock_edemux() and sock_rfree() ones.

This makes sense to me at least.

Yes - in connection with sock_rfree() for the read buffer destructur and
sock_wfree() for the write buffer it can make sense to name a function
sock_efree() as an unassigned destructor - which does not fiddle with rmem nor
wmen.

But both sock_efree() and sock_edemux() lack some comment - especially when it
makes sense to use them from non-INET contexts which Florian suggested.

Maybe Alexander can send a patch which adds a comment, as I don't know if I
would find the best words for it.

Regards,
Oliver

On 03/10/2015 01:22 PM, Eric Dumazet wrote:

On Tue, 2015-03-10 at 07:14 +0100, Oliver Hartkopp wrote:

quoted

the other callers use it in the same way so it's a good simplification.
Btw. the name of sock_efree() is a bit misleading - nothing is free'd here.

Won't it be better to rename sock_efree(skb) with sock_put_skb(skb) or 
something like that? sock_efree() has no comment why it's named like this.

I would prefer name stays as is. It eases searches in changelogs to not
change function names unless really needed, for backports and code
maintenance.


# git log | grep sock_efree
    net: merge cases where sock_efree and sock_edemux are the same function
    Since sock_efree and sock_demux are essentially the same code for non-TCP
    In addition I have added a destructor named sock_efree which is meant to

sock_efree was added in commit 62bccb8cdb69051b95a55ab0c489e3cab261c8ef

I guess Alexander chose the name close to sock_edemux() and sock_rfree() ones.

This makes sense to me at least.

On Tue, 2015-03-10 at 14:36 +0100, Oliver Hartkopp wrote:

Yes - in connection with sock_rfree() for the read buffer destructur and
sock_wfree() for the write buffer it can make sense to name a function
sock_efree() as an unassigned destructor - which does not fiddle with rmem nor
wmen.

But both sock_efree() and sock_edemux() lack some comment - especially when it
makes sense to use them from non-INET contexts which Florian suggested.

Maybe Alexander can send a patch which adds a comment, as I don't know if I
would find the best words for it.

Please do not top post on netdev.

If you cannot find best words for it, maybe a comment would not be
useful : Very often, best comments are added by people that had problems
to understand the code ;)

I do not trust comments, I prefer using "git grep" or tools like that to
check call sites.

In this particular case this becomes clear, while being concise.

# git grep -n sock_efree
include/net/sock.h:1526:void sock_efree(struct sk_buff *skb);
include/net/sock.h:1530:#define sock_edemux(skb) sock_efree(skb)
net/core/skbuff.c:3625: clone->destructor = sock_efree;
net/core/sock.c:1658:void sock_efree(struct sk_buff *skb)
net/core/sock.c:1662:EXPORT_SYMBOL(sock_efree);
net/ipv4/udp.c:1992:    skb->destructor = sock_efree;

On 03/10/2015 03:36 PM, Eric Dumazet wrote:

On Tue, 2015-03-10 at 14:36 +0100, Oliver Hartkopp wrote:

quoted

Maybe Alexander can send a patch which adds a comment, as I don't know if I
would find the best words for it.

Please do not top post on netdev.

? I did not.

If you cannot find best words for it, maybe a comment would not be
useful : Very often, best comments are added by people that had problems
to understand the code ;)

Yes that's why I asked the original author to do so - he hopefully understood
his code :-)

I do not trust comments, I prefer using "git grep" or tools like that to
check call sites.

In this particular case this becomes clear, while being concise.

# git grep -n sock_efree

A good hint! Will use this first in the future.

Thanks,
Oliver

On 03/10/2015 07:36 AM, Eric Dumazet wrote:

On Tue, 2015-03-10 at 14:36 +0100, Oliver Hartkopp wrote:

quoted

Yes - in connection with sock_rfree() for the read buffer destructur and
sock_wfree() for the write buffer it can make sense to name a function
sock_efree() as an unassigned destructor - which does not fiddle with rmem nor
wmen.

But both sock_efree() and sock_edemux() lack some comment - especially when it
makes sense to use them from non-INET contexts which Florian suggested.

Maybe Alexander can send a patch which adds a comment, as I don't know if I
would find the best words for it.

Please do not top post on netdev.

If you cannot find best words for it, maybe a comment would not be
useful : Very often, best comments are added by people that had problems
to understand the code ;)

I do not trust comments, I prefer using "git grep" or tools like that to
check call sites.

In this particular case this becomes clear, while being concise.

# git grep -n sock_efree
include/net/sock.h:1526:void sock_efree(struct sk_buff *skb);
include/net/sock.h:1530:#define sock_edemux(skb) sock_efree(skb)
net/core/skbuff.c:3625: clone->destructor = sock_efree;
net/core/sock.c:1658:void sock_efree(struct sk_buff *skb)
net/core/sock.c:1662:EXPORT_SYMBOL(sock_efree);
net/ipv4/udp.c:1992:    skb->destructor = sock_efree;

As I recall one of the reasons for leaving it as sock_efree was because 
sock_put just calls sk_free if it has decremented the reference count to 
0.  Also as Eric pointed out the name is actually pretty close to the 
existing destructors sock_edemux and sock_rfree.  The sock_efree name 
was based on the fact that the main consumer is usually either the UDP 
early receive path or the error handler/Timestamp path.

As far as the comments about it not freeing anything take a look at 
sock_rfree and tell me what it is actually freeing?  Last I knew it is 
simply dropping references to memory so that the socket can go ahead and 
either make use of them or close itself once all references have been 
freed.  That is the same thing this does, but it is holding references 
to a single byte of sk_wmem_alloc.

Patches are always welcome if you believe we need additional 
documentation, I had assumed this was pretty straight forward since the 
function was essentially just a wrapper for sock_put so I could use it 
as a destructor.

- Alex